# crosstabs.com — full reference for LLMs Free SPSS alternative in your browser: open SPSS .sav files with labels and missing values, run crosstabs with significance tests, and export editable reports. Files stay on your device; no install and no Crosstabs account. AI clients can also use the local MCP server. Definitions, assumptions, formulas, and workflow limits below come from the same versioned evidence graph used by the public evidence page and JSON resource. Evidence graph version: 2026-10-28 Evidence HTML: https://www.crosstabs.com/evidence Evidence JSON: https://www.crosstabs.com/evidence.json Try it: https://www.crosstabs.com/workspace Local MCP setup and data boundaries: https://www.crosstabs.com/mcp. Install with uv tool install --python 3.12 crosstabs==1.5.3 (uv brings its own Python; or pipx install crosstabs==1.5.3 with Python 3.10+, or run uvx crosstabs). crosstabs-headless also needs Node.js 22+. Add crosstabs-headless and crosstabs as stdio MCP commands using their full paths (uv tool dir --bin prints the folder). The project server reads only the folders listed in CROSSTABS_MCP_ALLOWED_ROOTS (set it in that server's env in the client configuration) plus a private exchange folder; get_runtime_status reports them. Files stay on your device. Local PyPI package and Codex plugin documentation: https://www.crosstabs.com/mcp ## Free SPSS alternative: procedure coverage Crosstabs is a free SPSS alternative that runs in the browser: it opens SPSS .sav, .zsav and .por files with value labels and missing values, and files are not uploaded. Coverage as of 2026-10-26: All 59 SPSS command families covered · 115 Statistics procedures plus crosstabs · 117 procedure guides. No known gaps remain in the command families and options it tracks. Details: https://www.crosstabs.com/spss-alternative (Simplified Chinese: https://www.crosstabs.com/zh-cn/spss-alternative). - GET FILE (.sav) [available]: Opens compressed and uncompressed .sav files in the browser with variable labels, value labels and numeric user-missing codes and ranges. https://www.crosstabs.com/open-sav-file-without-spss - GET FILE / IMPORT (.zsav, .por) [available]: Reads .zsav and .por files with labels and missing-value definitions. https://www.crosstabs.com/help/import-formats - GET DATA / GET STATA / GET SAS [available]: Reads CSV/TSV, Excel (.xlsx, .xlsm, .xls, .xlsb), OpenDocument .ods/.fods, JSON, Stata .dta, SAS .sas7bdat/.xpt, R .rds/.RData, Parquet, Arrow/Feather and SQLite database files. https://www.crosstabs.com/help/import-formats - SAVE OUTFILE (.sav) [available]: Data › Export data writes the session's data, transformations and recodes included, as a byte-code compressed or uncompressed .sav file with variable and value labels, measurement levels, user-missing codes, an optional case weight and multiple-dichotomy sets, in the browser. https://www.crosstabs.com/help/export-formats - CROSSTABS /CELLS [available]: Counts, row, column and total percentages, expected counts and adjusted standardized residuals. https://www.crosstabs.com/crosstab-maker - CROSSTABS /STATISTICS=CHISQ [available]: Pearson and likelihood-ratio chi-square, continuity correction, linear-by-linear association, and Fisher's exact test (one- and two-sided for 2×2, exact Fisher–Freeman–Halton for larger tables). https://www.crosstabs.com/chi-square-calculator - CROSSTABS /STATISTICS (measures) [available]: Phi, Cramér's V, contingency coefficient, lambda, Goodman–Kruskal tau, uncertainty coefficient, gamma, Kendall's tau-b and tau-c, Somers' d and kappa, each with asymptotic standard error and significance in the workspace All statistics panel; quick calculators for single measures. https://www.crosstabs.com/cramers-v-calculator - CROSSTABS /STATISTICS (full set) [available]: Workspace Table mode, All statistics panel: all CROSSTABS statistics: Pearson, Yates, likelihood ratio, Fisher (one- and two-sided, and exact R×C), linear-by-linear; phi, V, C, lambda, Goodman–Kruskal tau and uncertainty coefficient; gamma, tau-b, tau-c and Somers d with ASE and p; Pearson r, Spearman and eta; kappa, McNemar and McNemar–Bowker; risk estimates; Cochran–Mantel–Haenszel, Breslow–Day and Tarone with a layer variable; and raw, standardized and adjusted residuals. https://www.crosstabs.com/spss/crosstabs-statistics - CTABLES (nesting, stacking, nets) [available]: Tab books with nested or stacked rows and banners, nets and subtotals, saved and re-run in one place. https://www.crosstabs.com/survey-tab-book - CTABLES /COMPARETEST (proportions) [available]: Column-proportion z-tests as significance letters, with confidence tiers, a minimum base and multiple-comparison correction. https://www.crosstabs.com/crosstab-significance-letters - CTABLES summary statistics [available]: Adds scale summary rows under any table or saved tab-book table — mean, median, mode, standard deviation, variance, standard error, minimum, maximum, range, sum, row % of sum, valid and total N and percentiles per banner column — with an SPSS-style column means test (pooled or pairwise variance, Bonferroni, Holm or Benjamini–Hochberg) shown as letters and kept in every export. https://www.crosstabs.com/spss/ctables-summary-statistics - MULT RESPONSE [available]: Multi-select questions as sets, with percentages of respondents or of responses kept distinct. https://www.crosstabs.com/multi-select-crosstab - FREQUENCIES [available]: Workspace Statistics mode: frequency tables with user-missing rows, valid and cumulative percent, five percentile methods, grouped percentiles, moments with standard errors, bar charts and histograms. https://www.crosstabs.com/spss/frequencies - DESCRIPTIVES [available]: Workspace Statistics mode: N, minimum, maximum, mean with standard error, SD, variance, skewness and kurtosis with standard errors, and listwise valid N. https://www.crosstabs.com/spss/descriptive-statistics - EXAMINE [available]: Workspace Statistics mode: descriptives with confidence intervals, 5% trimmed mean, four M-estimators, extreme values, Shapiro–Wilk and Lilliefors tests, four Levene variants and boxplot outlier rules. https://www.crosstabs.com/spss/explore-examine - MEANS [available]: Workspace Statistics mode: all 19 SPSS cell statistics in up to three crossed layers, with the ANOVA table, eta and the test for linearity. https://www.crosstabs.com/spss/compare-means - T-TEST [available]: Workspace Statistics mode: one-sample, independent-samples (Levene, pooled and Welch rows) and paired-samples t tests with Cohen d, Hedges g and Glass delta and their confidence intervals. https://www.crosstabs.com/spss/independent-samples-t-test - ONEWAY (post hoc) [available]: Workspace Statistics mode: ANOVA with Levene, Welch and Brown–Forsythe tests, eta², epsilon² and omega² with confidence intervals, polynomial contrasts, Tukey, Bonferroni, Šidák, Scheffé, LSD, Games–Howell, Dunnett, Hochberg's GT2, Gabriel, Tamhane's T2 and Dunnett's T3 and C post hoc tests, and homogeneous subsets for Tukey, S-N-K, Tukey's-b, Duncan, R-E-G-W F and Q and the Waller–Duncan k-ratio t test. https://www.crosstabs.com/spss/one-way-anova - NPAR TESTS /CHISQUARE [available]: Goodness-of-fit chi-square against equal or custom expected proportions. https://www.crosstabs.com/chi-square-goodness-of-fit-calculator - NPAR TESTS /MCNEMAR [available]: McNemar test for a paired 2×2 table. https://www.crosstabs.com/mcnemar-test-calculator - NPAR TESTS (rank tests) [available]: Workspace Statistics mode: chi-square goodness of fit, binomial, runs, one- and two-sample Kolmogorov–Smirnov, Mann–Whitney U, Moses, Wald–Wolfowitz, Kruskal–Wallis with Dunn pairwise comparisons, median test, Jonckheere–Terpstra, Wilcoxon, sign, McNemar, Friedman, Kendall W and Cochran Q, with exact p-values where SPSS gives them. https://www.crosstabs.com/spss/mann-whitney-u-test - CORRELATIONS / NONPAR CORR [available]: Workspace Statistics mode: Pearson (with Fisher-z confidence intervals), Spearman and Kendall tau-b matrices with one- or two-tailed significance, flags, pairwise or listwise deletion, and partial correlations. https://www.crosstabs.com/spss/pearson-spearman-correlation - RELIABILITY [available]: Workspace Statistics mode: Cronbach and standardized alpha, split-half, Guttman λ1–λ6, parallel and strictly parallel models, McDonald omega, item-total statistics, ANOVA with Friedman or Cochran, Tukey nonadditivity, Hotelling T² and intraclass correlations. https://www.crosstabs.com/spss/cronbach-alpha - FACTOR [available]: Workspace Statistics mode: principal components, principal axis and maximum-likelihood extraction, varimax, quartimax, equamax, oblimin and promax rotation, KMO and Bartlett, anti-image matrices, communalities, variance explained and scree plot, plus regression, Bartlett and Anderson–Rubin factor scores saved as new variables (FACTOR /SAVE). https://www.crosstabs.com/spss/factor-analysis - REGRESSION [available]: Workspace Statistics mode: Enter, Stepwise, Forward, Backward and Remove blocks; model summary with R² change and Durbin–Watson, ANOVA, coefficients with confidence intervals, part and partial correlations, tolerance and VIF, collinearity, casewise and residual diagnostics. https://www.crosstabs.com/spss/linear-regression - LOGISTIC REGRESSION [available]: Workspace Statistics mode: binary logistic regression with Enter, Forward:LR and Backward:LR, omnibus tests, −2LL with Cox & Snell and Nagelkerke R², Hosmer–Lemeshow, classification table, indicator-coded categorical covariates and odds ratios with confidence intervals. https://www.crosstabs.com/spss/binary-logistic-regression - PROCESS macro models (mediation and moderation) [available]: Workspace Statistics mode, independent implementation not affiliated with PROCESS: models 4 (parallel mediation), 1 (moderation), 6 (serial, two mediators), 7 and 14 (moderated mediation) with OLS or HC3 paths, seeded percentile bootstrap intervals for indirect effects, completely standardized effects, contrasts, Sobel, conditional effects, Johnson–Neyman, simple slopes and the index of moderated mediation. Continuous outcomes only. https://www.crosstabs.com/spss/mediation-moderation-process - GLM / UNIANOVA [available]: Workspace Statistics mode: GLM Univariate (factorial ANOVA and ANCOVA, Type I–IV sums of squares, estimated marginal means, Levene, ten post hoc tests and Waller–Duncan homogeneous subsets), GLM Multivariate (MANOVA with Box M), GLM Repeated Measures (Mauchly, Greenhouse–Geisser and Huynh–Feldt) and VARCOMP. https://www.crosstabs.com/spss/glm-univariate - QUICK CLUSTER [available]: SPSS-exact QUICK CLUSTER in workspace Statistics mode (initial centres, iteration history, ANOVA, memberships), plus deterministic k-means segmentation with profiles in the Segment Builder. https://www.crosstabs.com/tools/segment-builder - WEIGHT BY [available]: Declared frequency weights in saved projects. A weight is never guessed from a column name. https://www.crosstabs.com/survey-weighting-raking - Raking (SPSSINC RAKE) [available]: Guided raking to targets you enter, with trimming, convergence and effective-base diagnostics. https://www.crosstabs.com/survey-weighting-raking - CSTABULATE / CSDESCRIPTIVES [available]: Taylor-linearized or replicate-weight designs with design effects and Rao–Scott or adjusted-Wald tests for saved tables; CSDESCRIPTIVES (means, sums, ratios, population sizes by subpopulation) and CSTABULATE with odds ratios and risks in workspace Statistics mode. https://www.crosstabs.com/survey-weighting-raking - CSGLM / CSLOGISTIC / CSORDINAL / CSCOXREG [available]: Workspace Statistics mode, from a CSPLAN-style design (strata, clusters for one or two stages, population sizes): CSGLM, binary and multinomial CSLOGISTIC, CSORDINAL with five links and CSCOXREG, with Wald tests on design degrees of freedom and design effects; also design-aware models in Advanced methods. Main effects only. https://www.crosstabs.com/spss/complex-samples-glm - POWER PROPORTIONS [available]: Sample size for a margin of error and for comparing two proportions. https://www.crosstabs.com/tools/sample-size-calculator - OUTPUT EXPORT [available]: Exports to CSV, XLSX and PDF, plus editable PowerPoint and Word reports. https://www.crosstabs.com/help/export-formats - CORRESPONDENCE / ANACOR [available]: Workspace Statistics mode: inertia and chi-square, row and column point tables with contributions, five normalizations and a labelled perceptual map you can export. https://www.crosstabs.com/spss/correspondence-analysis - MULTIPLE CORRESPONDENCE / HOMALS [available]: Workspace Statistics mode: model summary with Cronbach alpha, discrimination measures, category points, object scores and a joint plot. https://www.crosstabs.com/spss/multiple-correspondence-analysis - PROXSCAL / ALSCAL (MDS) [available]: Workspace Statistics mode: PROXSCAL with ratio, interval or ordinal transformations, the six stress and fit measures, Shepard diagram and classical (Torgerson) scaling. Also ALSCAL (S-stress alternating least squares on squared distances, ratio or ordinal). https://www.crosstabs.com/spss/multidimensional-scaling - CATPCA / CATREG / PRINCALS / OVERALS [available]: Workspace Statistics mode: nominal, ordinal and numeric scaling levels by alternating least squares; eigenvalues, loadings and quantifications for CATPCA; R², standardized coefficients and Pratt importance for CATREG; PRINCALS and OVERALS with multiple nominal, single nominal, ordinal and numeric levels, loss and fit per set and canonical correlations. Spline levels are not included. https://www.crosstabs.com/spss/categorical-pca - ORTHOPLAN [available]: Workspace Statistics mode: orthogonal plans with holdout cards, an orthogonality check and SPSS-style CSV cards. https://www.crosstabs.com/spss/orthogonal-design - CONJOINT [available]: Workspace Statistics mode: part-worth utilities with standard errors, importance values, Pearson R and Kendall tau (including holdouts), reversals and max-utility, BTL and logit simulations, per subject and pooled. https://www.crosstabs.com/spss/conjoint-analysis - TREE (CHAID, CRT) [available]: Workspace Statistics mode: CHAID and Exhaustive CHAID with Bonferroni-adjusted merging and splitting, and CRT with Gini or twoing, surrogates and cost-complexity pruning; interactive tree, gains, risk and classification tables, and cross-validation. Also QUEST (ANOVA F, chi-square and Levene split-variable selection with quadratic-discriminant split points). https://www.crosstabs.com/spss/chaid-decision-tree - TWOSTEP CLUSTER [available]: Workspace Statistics mode: CF-tree pre-clustering, log-likelihood or Euclidean distance, automatic cluster count by BIC or AIC, cluster profiles, silhouette and predictor importance. https://www.crosstabs.com/spss/twostep-cluster - CLUSTER [available]: Workspace Statistics mode: seven linkage methods, agglomeration schedule, memberships for a range of solutions, silhouette and a rescaled dendrogram. https://www.crosstabs.com/spss/hierarchical-cluster - DISCRIMINANT [available]: Workspace Statistics mode: direct or stepwise (Wilks) entry, Box M, eigenvalues and canonical correlations, standardized, structure and unstandardized coefficients, centroids, Fisher classification functions and original and leave-one-out classification. https://www.crosstabs.com/spss/discriminant-analysis - PROXIMITIES [available]: Workspace Statistics mode: interval, count and 27 binary measures, seven standardizations and transforms, between cases or between variables. https://www.crosstabs.com/spss/distance-matrix - NOMREG / PLUM / GENLIN / GENLOG [available]: Workspace Statistics mode: NOMREG (with forward, backward and stepwise effect selection), PLUM (five links, parallel-lines test), GENLIN with seven distributions and GEE, PROBIT with fiducial limits, CURVEFIT, NLR/CNLR, WLS, 2SLS, quantile regression, HILOGLINEAR and GENLOG. Also LOGLINEAR with custom designs, deviation contrasts and the analysis of dispersion, and GENLINMIXED by REPL-PQL with random intercepts and slopes. Main effects only for the regression-type designs. https://www.crosstabs.com/spss/generalized-linear-models - KM / COXREG / SURVIVAL [available]: Workspace Statistics mode: Kaplan–Meier with log-rank, Breslow and Tarone–Ware tests and survival plots, actuarial life tables with the Gehan test, and Cox regression with Breslow ties, omnibus tests, hazard ratios, forward and backward stepwise selection (conditional, LR or Wald) and a time-dependent covariate in T_. https://www.crosstabs.com/spss/kaplan-meier - ROC / WEIGHTED KAPPA / Meta Analysis [available]: Workspace Statistics mode: ROC curves with AUC, standard error and coordinates, DeLong comparison of two AUCs and precision–recall curves, weighted and Fleiss multirater kappa, PROPORTIONS with every test and interval method in the manual, ratio statistics, and meta-analysis of binary and continuous outcomes with seven τ² estimators, Knapp–Hartung, Egger's test, forest plots and meta-regression. https://www.crosstabs.com/spss/roc-curve - MVA / MULTIPLE IMPUTATION / BOOTSTRAP [available]: Workspace Statistics mode: MVA with EM estimates and Little's MCAR test, pattern analysis, multiple imputation (FCS or monotone, regression, PMM or logistic) with Rubin's rules pooling of descriptives and linear regression, RMV previews, and seeded bootstrap percentile and BCa intervals. https://www.crosstabs.com/spss/missing-value-analysis - BAYES procedures [available]: Workspace Statistics mode: Bayes factors and posterior distributions with credible intervals for one-sample (normal, binomial, Poisson), paired and independent samples, Pearson correlation, one-way ANOVA, linear regression, contingency tables and repeated measures. https://www.crosstabs.com/spss/bayesian-independent-samples - ARIMA / EXSMOOTH / TSMODEL [available]: Workspace Statistics mode, with a define-dates step that turns respondent rows into periods: Expert Modeler, exponential smoothing (12 models), seasonal ARIMA by exact maximum likelihood, seasonal decomposition, ACF/PACF, CCF, AREG, univariate SPECTRA, CREATE and FIT. ARIMA can detect all seven TSMODEL outlier types (additive, innovational, level shift, transient, seasonal additive, local trend and additive patch) and estimate them jointly with the model; transfer functions are not included. https://www.crosstabs.com/spss/arima-forecasting - MIXED / KNN / NAIVE BAYES [available]: Workspace Statistics mode: linear mixed models (random effects, repeated covariance structures, REML or ML, Type III tests with Satterthwaite df), nearest-neighbor analysis with cross-validated k and feature selection, and naive Bayes with forward selection. Neural networks are not included yet. https://www.crosstabs.com/spss/linear-mixed-models - CSSELECT [available]: Workspace Statistics mode: seeded simple random, systematic, PPS with replacement, PPS systematic and Sampford selection, stratified and clustered, with inclusion probabilities and weights, as a download. https://www.crosstabs.com/spss/complex-sample-selection - RANK / OPTIMAL BINNING / SAMPLE [available]: Workspace Statistics mode: eight rank types with four tie rules and normal scores, supervised MDLP or equal-frequency binning, and seeded case samples, all as previews and downloads that never change your data. https://www.crosstabs.com/spss/rank-cases - SPSS syntax (.sps) [available]: Reports › Syntax runs a .sps file or pasted syntax on the open data: COMPUTE, RECODE, IF, SELECT IF, FILTER BY, WEIGHT BY, VARIABLE LABELS, VALUE LABELS, MISSING VALUES and VARIABLE LEVEL, then FREQUENCIES, DESCRIPTIVES, CROSSTABS, rows-by-columns CTABLES, MEANS, T-TEST, ONEWAY, CORRELATIONS, REGRESSION, RELIABILITY and FACTOR. Every other command is listed with the reason it did not run. https://www.crosstabs.com/spss-crosstabs-syntax - ERROR BARS / PPLOT / SPCHART [available]: Workspace Statistics mode: error bars for CI, SE or SD by group, P-P and Q-Q plots against 12 distributions, and X-bar/R, X-bar/s, I-MR, p, np, c and u control charts with 11 rule checks. https://www.crosstabs.com/spss/control-charts - SIMULATION [available]: Workspace Statistics mode: Monte Carlo simulation of a target expression from inputs with chosen distributions, with summary, percentiles, histogram and sensitivity. https://www.crosstabs.com/spss/monte-carlo-simulation - MLP / RBF / PLS / LINEAR / SELECTPRED / AIM [available]: Workspace Statistics mode: multilayer perceptron and radial basis function networks with seeded partitions, classification tables, ROC, gains and predictor importance; PLS regression with VIP; automatic linear modeling with data preparation, stepwise or best-subsets selection; predictor screening and ranking; and attribute importance for clusters. Network weights will not match SPSS exactly because its random start is undocumented. Also DETECTANOMALY (TwoStep peer groups, anomaly index and reasons), bagging and boosting ensembles of linear models, and an Automated Data Preparation (ADP) preview with a download of the transformed fields. https://www.crosstabs.com/spss/multilayer-perceptron Sourced comparisons with other free and online SPSS alternatives: Crosstabs vs PSPP, JASP and jamovi https://www.crosstabs.com/crosstabs-vs-pspp-jasp-jamovi (中文 https://www.crosstabs.com/zh-cn/crosstabs-vs-pspp-jasp-jamovi); Crosstabs vs SPSSAU and SPSSPRO https://www.crosstabs.com/crosstabs-vs-spssau-spsspro (中文 https://www.crosstabs.com/zh-cn/crosstabs-vs-spssau-spsspro). SPSS is a trademark of IBM Corp. Crosstabs is not affiliated with or endorsed by IBM. ## SPSS procedure library 117 procedure pages, updated 2026-10-18, each in English and Simplified Chinese. Hub: https://www.crosstabs.com/spss (简体中文: https://www.crosstabs.com/zh-cn/spss). Every page's worked example is computed by the Crosstabs engine on seeded synthetic data at build time. ### Crosstabs / 交叉表 #### Crosstabs Statistics — SPSS CROSSTABS /STATISTICS (交叉表统计量) English: https://www.crosstabs.com/spss/crosstabs-statistics 简体中文: https://www.crosstabs.com/zh-cn/spss/crosstabs-statistics CROSSTABS /STATISTICS tests whether two categorical variables are associated and measures how strongly. In Crosstabs, build the table in Table mode and open All statistics to get the full set: chi-square tests, Fisher's exact tests, nominal and ordinal measures with asymptotic standard errors, correlations, kappa and McNemar, risk estimates, Cochran–Mantel–Haenszel tests across a layer, and cell residuals. When to use: Testing whether answers to a survey question differ by segment, region or wave. Measuring a trend between two ordered variables, such as income band and satisfaction. Comparing two raters' classifications, or before/after categories on the same people. Checking whether an exposure–outcome 2×2 association is confounded by a third variable. When not to: Expected counts below 5 in many cells of a large table and an exact test exceeds the browser workload: merge sparse categories or use the Monte Carlo exact option. Survey data with a complex design (strata, clusters): use Complex Samples Crosstabs for design-based tests. Three or more variables whose joint associations you want to model: use hierarchical loglinear analysis. Multiple-response sets: table-level association statistics are not valid because rows are not mutually exclusive. Validation: The statistics are checked against 31 published CROSSTABS reference cases in the public validation manifest, generated with SciPy 1.18, statsmodels 0.15, R 4.6.1 with DescTools 0.99.60 and stats::fisher.test, and a 50-digit mpmath delta method for standard errors. Every value agrees to a relative tolerance of 1e-10; the cases cover nominal, ordinal, correlation, agreement and risk measures, one-sided and R×C Fisher tests, residuals and stratified 2×2 tests. #### CTABLES Summary Statistics — SPSS CTABLES /COMPARETEST TYPE=MEAN (自定义表汇总统计量) English: https://www.crosstabs.com/spss/ctables-summary-statistics 简体中文: https://www.crosstabs.com/zh-cn/spss/ctables-summary-statistics CTABLES summary statistics put a scale variable's mean, median, spread and counts under each banner column of a table, and COMPARETEST TYPE=MEAN marks which column means differ. In Crosstabs you add these rows from the Summary stats control in Table mode or from the Table designer in a tab book; the column means test uses pooled or pairwise variance, Bonferroni, Holm or Benjamini–Hochberg adjustment, and letters that stay in every export. When to use: A tracker or tab book where every banner (region, age band, segment) needs the mean score of a rating question beneath the category percentages. Reporting medians and quartiles of skewed measures such as spend or time, column by column. Flagging, in a client-ready table, which banner columns have a higher mean than others, with the same letters as the column proportions test. When not to: You need an overall test of whether any group mean differs, effect sizes or post hoc confidence intervals: CTABLES has no omnibus test, so use one-way ANOVA or Compare Means. Data with sampling weights or a stratified, clustered design: the summary rows give weighted point estimates only and the means test is withheld, so use Complex Samples Descriptives or GLM for design-based standard errors. More than 12 banner columns: the column means test is limited to 12 columns, so split the banner or compare the groups in Statistics mode. A categorical rating you want as a distribution: show the category percentages and the column proportions test instead of a mean. Validation: The summary rows and the column means test are checked against five reference scenarios (unweighted and frequency-weighted; pooled and pairwise variance; Bonferroni, Holm, Benjamini–Hochberg and no adjustment; a column dropped by the minimum base) generated independently with NumPy 2.5.4, SciPy 1.18.1 and statsmodels 0.15.0: averaged_inverted_cdf percentiles, scipy.stats.mode, ttest_ind for pairwise variance, an OLS residual mean square for pooled variance and multipletests for the adjustments. Statistics, t and df agree to a relative tolerance of 1e-10, and raw and adjusted p values to 1e-12. ### Describe / 描述统计 #### Frequencies — SPSS FREQUENCIES (频率分析) English: https://www.crosstabs.com/spss/frequencies 简体中文: https://www.crosstabs.com/zh-cn/spss/frequencies FREQUENCIES counts how often each value of a variable occurs and turns the counts into percent, valid percent and cumulative percent. Crosstabs runs it on your device with each user-missing code listed separately, summary statistics, SPSS's default HAVERAGE percentiles and an optional bar chart or histogram. When to use: The first look at any survey question, before crosstabs or modelling. Reporting the distribution of a categorical or rating-scale variable, such as satisfaction from 1 to 5. Checking that recodes, filters and missing-value declarations did what you intended. When not to: Summarising many continuous variables side by side: Descriptives gives one compact table instead of a long frequency table per variable. Comparing the distribution across groups: use a crosstab with column percentages, or Explore with a factor. Multiple-response questions stored as several yes/no columns: a single frequency table per column does not give the share of respondents choosing each option among all options. Validation: The FREQUENCIES statistics share the descriptives engine, which is checked on 13 reference samples generated with NumPy 2.5 and SciPy 1.18: mean within 1e-12, variance and SD within 1e-11, skewness and kurtosis within 1e-10, and all five percentile definitions within 1e-12 of the matching NumPy methods (tolerances scaled by the size of the value). Unit tests cover user-missing percent bases, count ordering, grouped percentiles, multiple modes and weights. #### Descriptives — SPSS DESCRIPTIVES (描述统计) English: https://www.crosstabs.com/spss/descriptive-statistics 简体中文: https://www.crosstabs.com/zh-cn/spss/descriptive-statistics DESCRIPTIVES puts the core summary statistics of several numeric variables into a single table: N, minimum, maximum, mean with its standard error, standard deviation, variance, and skewness and kurtosis with their standard errors. Crosstabs computes it in the browser with SPSS's variable-by-variable missing-value rule and a Valid N (listwise) row. When to use: The descriptive table at the start of a report or thesis: age, income, scores and other scale variables side by side. Screening many numeric variables quickly before modelling. Comparing the scale of variables you intend to combine or standardise. When not to: You need the median, percentiles or normality tests: use Explore, or Frequencies for percentiles. You want statistics within groups: use Means or Explore with a factor. You need z-score variables saved to the data: this panel reports the table only. Categorical codes such as region 1–5: a mean of codes is meaningless, use Frequencies. Validation: On 13 reference samples, including weighted ones, the moments match SciPy 1.18 and NumPy 2.5 (skewness and kurtosis with bias=False, as SPSS reports them): mean within 1e-12, variance, SD and standard error within 1e-11, skewness and kurtosis within 1e-10, and N, minimum and maximum exactly. Unit tests check listwise and variable-by-variable exclusion, sorting and the rejection of string variables. #### Explore — SPSS EXAMINE (探索性分析) English: https://www.crosstabs.com/spss/explore-examine 简体中文: https://www.crosstabs.com/zh-cn/spss/explore-examine Explore (SPSS EXAMINE) is the procedure for looking hard at a scale variable before you test it: centre, spread, shape, outliers and normality, overall or for each level of a factor. Crosstabs runs it in your browser with robust M-estimators, Tukey's hinges, Shapiro-Wilk and Lilliefors-corrected Kolmogorov-Smirnov tests, and Levene's test with four centring options. When to use: Before a t test or ANOVA, to check the normality and equal-variance assumptions group by group. When you suspect outliers or data-entry errors and need the case numbers to find them. To describe skewed variables such as income or response time with robust statistics. When not to: Many variables needing only means and SDs: Descriptives is more compact. Testing a difference between groups: Explore checks assumptions; use a t test, ANOVA or a nonparametric test for the comparison itself. Categorical variables: use Frequencies. Validation: Checked on 13 reference samples from NumPy 2.5, SciPy 1.18 and statsmodels 0.15: confidence intervals match scipy t.interval within 1e-9, Shapiro-Wilk matches scipy.stats.shapiro within 1e-6 (W) and 1e-5 (p), the Lilliefors D statistic matches statsmodels within 1e-12, M-estimators solve the statsmodels psi equations within 1e-9, Tukey's hinges and percentiles agree within 1e-12, and Levene's four variants match SciPy within 1e-10 (F) and 1e-9 (p). ### Compare means / 比较平均值 #### One-Way ANOVA — SPSS ONEWAY (单因素方差分析) English: https://www.crosstabs.com/spss/one-way-anova 简体中文: https://www.crosstabs.com/zh-cn/spss/one-way-anova One-way ANOVA tests whether the mean of a scale variable differs across three or more groups of a single factor. Crosstabs runs the SPSS ONEWAY procedure on your device, with the variance-robust Welch and Brown-Forsythe tests, η² and ω² effect sizes, polynomial trend contrasts and seven post hoc methods. When to use: One continuous outcome (satisfaction score, spend, task time) compared across three or more independent groups. Each case belongs to exactly one group, such as region, plan tier or experimental arm. Group sizes are unequal or variances differ — read the Welch test and Games-Howell instead of the classical F and Tukey. When not to: Only two groups: use the independent-samples t test, which reports the same p value plus a confidence interval for the difference. Two or more factors or a covariate: use GLM Univariate for factorial ANOVA or ANCOVA. The same people measured repeatedly: use repeated-measures ANOVA or the Friedman test. Ordinal ratings or heavily skewed data with small groups: use the Kruskal-Wallis test. Validation: The ONEWAY engine is checked against SciPy 1.18 (f_oneway, tukey_hsd), statsmodels 0.15 (Brown-Forsythe) and pingouin 0.7 (Welch ANOVA, Games-Howell) on reference datasets, agreeing to a relative tolerance of 1e-9; frequency weights are tested to reproduce replicated cases exactly. The Waller–Duncan Bayesian t value agrees within 1e-7 with an independent R solution of the Waller–Kemp equation and with 30-digit mpmath quadrature over 81 combinations of k-ratio, groups, error df and F, and its homogeneous subsets match R agricolae::waller.test. #### Means — SPSS MEANS (分组均值比较) English: https://www.crosstabs.com/spss/compare-means 简体中文: https://www.crosstabs.com/zh-cn/spss/compare-means MEANS breaks a scale variable down by one or more categorical variables and reports the mean, N, standard deviation and other cell statistics for every group and subtotal. Crosstabs runs it in your browser with up to three crossed layers, 19 cell statistics including harmonic and geometric means, and an optional one-way ANOVA with eta and a test for linearity on the first layer. When to use: Reporting an outcome such as spend, score or time by segment, region or wave. Breaking a mean down by two or three variables at once, such as region within gender. A quick check of whether an ordered predictor such as age band relates linearly to the outcome. When not to: Post hoc pairwise comparisons, Welch's test or effect-size intervals: use One-Way ANOVA. Testing interactions between two factors: use GLM Univariate, since the MEANS ANOVA uses the first layer only. A binary or categorical outcome: use a crosstab with row or column percentages. Validation: Cell statistics are checked against SciPy 1.18 and NumPy 2.5, and the ANOVA, eta, linearity and deviation rows against statsmodels 0.15 OLS (anova_lm) on reference data, all to a relative tolerance of 1e-9. Frequency weights are tested to reproduce replicated cases, and user- and system-missing values to be excluded. #### One-Sample T Test — SPSS T-TEST (单样本 t 检验) English: https://www.crosstabs.com/spss/one-sample-t-test 简体中文: https://www.crosstabs.com/zh-cn/spss/one-sample-t-test The one-sample t test asks whether the mean of a scale variable differs from a fixed reference value, such as a target, a norm or a scale midpoint. Crosstabs runs the SPSS T-TEST /TESTVAL procedure in your browser and reports one- and two-sided p values, the confidence interval for the difference, and Cohen's d and Hedges' correction with noncentral-t intervals. When to use: Comparing a sample with a benchmark: a contracted 40-hour week, a published norm, last year's average. Testing whether ratings lean away from the neutral midpoint of a scale. When not to: The reference value is itself an estimate from another sample: use the independent-samples t test on both samples. Small samples from clearly skewed data or ordinal ratings: use the one-sample Wilcoxon signed-rank test in One-Sample Nonparametric Tests. A proportion rather than a mean: use a binomial test or the proportions tests. Validation: On Python-generated reference data, t, df, p and the confidence interval match SciPy 1.18 ttest_1samp to a relative tolerance of 1e-9, including a frequency-weighted case checked by replicating rows. Cohen's d and Hedges' g match to 1e-9, and their confidence limits to 1e-8 against an independent noncentral-t inversion in SciPy. #### Independent-Samples T Test — SPSS T-TEST GROUPS (独立样本 t 检验) English: https://www.crosstabs.com/spss/independent-samples-t-test 简体中文: https://www.crosstabs.com/zh-cn/spss/independent-samples-t-test The independent-samples t test compares the means of a scale variable between two separate groups, such as treatment and control or two customer segments. Crosstabs runs SPSS T-TEST GROUPS in your browser, giving Levene's test, the equal-variances and Welch rows with confidence intervals, and Cohen's d, Hedges' correction and Glass's delta with intervals. When to use: An A/B test or experiment with two arms and a continuous outcome. Comparing two existing groups, such as men and women or two regions, on a score or amount. Splitting a numeric variable at a cut point (for example age ≥ 40 versus younger) and comparing the halves. When not to: The same people measured twice: use the paired-samples t test. Three or more groups: use One-Way ANOVA rather than several t tests. Ordinal ratings or small, skewed samples: use the Mann-Whitney U test. Validation: On Python-generated reference data, Levene's F and p match SciPy 1.18 levene(center="mean"), and the pooled and Welch t, df, p and confidence intervals match SciPy ttest_ind, all to a relative tolerance of 1e-9, for unweighted and frequency-weighted data. Cohen's d also matches pingouin 0.7; effect-size confidence limits agree to 1e-8 with an independent noncentral-t inversion. #### Paired-Samples T Test — SPSS T-TEST PAIRS (配对样本 t 检验) English: https://www.crosstabs.com/spss/paired-samples-t-test 简体中文: https://www.crosstabs.com/zh-cn/spss/paired-samples-t-test The paired-samples t test compares two measurements taken on the same cases — before and after, left and right, or two raters scoring the same items — by testing whether their mean difference is zero. Crosstabs runs SPSS T-TEST PAIRS in your browser for up to 12 pairs, with the pair correlation, the confidence interval for the difference and a choice of three effect-size standardizers. When to use: Pre/post designs: the same respondents before and after a training, campaign or redesign. Matched pairs, such as siblings or matched customers, where each pair is one row. Two ratings of the same product or stimulus by each respondent. When not to: The two sets of values come from different people: use the independent-samples t test. Three or more repeated measurements: use repeated-measures ANOVA or the Friedman test. Ordinal scores or differences that are far from normal in a small sample: use the Wilcoxon signed-rank test. Validation: On Python-generated reference data, t, df, p and the confidence interval match SciPy 1.18 ttest_rel and the pair correlation matches SciPy pearsonr, to a relative tolerance of 1e-9. All three effect sizes (d_z, d_av and d_rm) and their Hedges-corrected versions are checked, with confidence limits within 1e-8 of independent SciPy computations. ### Nonparametric tests / 非参数检验 #### One-Sample Nonparametric Tests — SPSS NPAR TESTS (单样本非参数检验) English: https://www.crosstabs.com/spss/one-sample-nonparametric-tests 简体中文: https://www.crosstabs.com/zh-cn/spss/one-sample-nonparametric-tests One-sample nonparametric tests compare a single variable with a distribution you specify: equal or custom category shares, a hypothesised proportion, a random sequence, or a normal, uniform, Poisson or exponential curve. Crosstabs runs the four classic NPAR TESTS subcommands on your device and reports exact significance where SPSS does. When to use: Checking whether a categorical variable (preferred channel, day of week, answer option) is spread evenly or matches known population shares. Testing a yes/no rate against a benchmark, such as a 50% pass rate or a 20% response rate from last year. Screening a scale variable for normality before choosing between a t test and a rank-based test. Checking a sequence recorded in time order (production runs, daily outcomes) for clustering or alternation. When not to: Two categorical variables measured on the same cases: use the chi-square test of independence in Crosstabs instead of goodness of fit. Comparing a mean with a target value on roughly normal data: the one-sample t test gives a confidence interval for the difference. Very large samples screened for normality: the K-S test flags trivial departures, so look at a P-P or Q-Q plot as well. Validation: The one-sample engines are tested against independent reference values computed with SciPy 1.18.1, statsmodels 0.15.0, pingouin 0.7.0 and scikit-posthocs 0.17.1: scipy.stats.chisquare, binomtest and kstest, statsmodels runstest_1samp and lilliefors, with exact runs probabilities from enumeration. Statistics and p-values agree within 1e-9 (relative to the larger of 1 and the reference value); the Smirnov asymptotic p for Kolmogorov-Smirnov Z is held to 2e-6 because it uses the three-term series rather than SciPy's exact kstwobign. A separate test reproduces the Lilliefors interpolation worked example from the IBM SPSS Statistics Algorithms manual (n = 20, D = 0.170, p = 0.1321469). #### Mann-Whitney U Test — SPSS NPAR TESTS /M-W (Mann-Whitney U 检验) English: https://www.crosstabs.com/spss/mann-whitney-u-test 简体中文: https://www.crosstabs.com/zh-cn/spss/mann-whitney-u-test The Mann-Whitney U test compares two independent groups by ranking all values together and asking whether one group tends to sit higher. Crosstabs runs it as SPSS NPAR TESTS /M-W does, with the tie-corrected Z, an exact significance for small samples and the effect size r = |Z|/√N, alongside three other two-sample tests. When to use: Two independent groups compared on an ordinal rating (1–5 or 1–10 scales) where means are hard to defend. A skewed scale outcome (spend, time on task, counts) with small groups or obvious outliers. An A/B test with two variants where you want a result that does not depend on normality. When not to: The same people measured twice: use the Wilcoxon signed-rank test for paired data. Three or more groups: use the Kruskal-Wallis test, which adds Dunn pairwise comparisons. Roughly normal data with a need for a confidence interval for the mean difference: the independent-samples t test is more informative. Validation: The two-sample engines are tested against independent reference values computed with SciPy 1.18.1, statsmodels 0.15.0, pingouin 0.7.0 and scikit-posthocs 0.17.1. Mann-Whitney U and its exact and asymptotic p-values match scipy.stats.mannwhitneyu (no continuity correction) within 1e-9, |Z| within 1e-7; the two-sample K-S statistic matches scipy.stats.ks_2samp, with the Smirnov asymptotic p held to 2e-6; Moses and Wald-Wolfowitz significance are checked against enumeration of every arrangement of the pooled sample. #### Kruskal-Wallis Test — SPSS NPAR TESTS /K-W (Kruskal-Wallis 检验) English: https://www.crosstabs.com/spss/kruskal-wallis-test 简体中文: https://www.crosstabs.com/zh-cn/spss/kruskal-wallis-test The Kruskal-Wallis H test is the rank-based alternative to one-way ANOVA: it asks whether three or more independent groups come from the same distribution. Crosstabs runs SPSS NPAR TESTS /K-W with a tie-corrected H, the ε² effect size and Dunn pairwise comparisons, and adds the k-sample median test and the Jonckheere-Terpstra test for ordered groups. When to use: Three or more independent groups compared on an ordinal or skewed outcome, such as satisfaction ratings by region or resolution days by team. Group sizes are small or unequal and a normality assumption is hard to defend. The groups have a natural order (dose levels, age bands, plan tiers) and you expect a monotonic trend: use Jonckheere-Terpstra. When not to: Only two groups: use the Mann-Whitney U test, which also gives an exact significance for small samples. The same cases measured under every condition: use the Friedman test for related samples. Roughly normal outcomes with a need for mean differences and confidence intervals: use one-way ANOVA with Welch and Games-Howell. Two or more grouping factors at once: rank-based tests here handle one factor only. Validation: The k-sample engines are tested against independent reference values computed with SciPy 1.18.1, statsmodels 0.15.0, pingouin 0.7.0 and scikit-posthocs 0.17.1: Kruskal-Wallis H and p match scipy.stats.kruskal, Dunn pairwise p-values and Bonferroni adjustments match scikit-posthocs posthoc_dunn, the median test matches scipy.stats.median_test and fisher_exact, and Jonckheere-Terpstra is checked against brute-force pair counting. All agree within 1e-9 (relative to the larger of 1 and the reference value). #### Wilcoxon Signed-Rank Test — SPSS NPAR TESTS /WILCOXON (Wilcoxon 符号秩检验) English: https://www.crosstabs.com/spss/wilcoxon-signed-rank-test 简体中文: https://www.crosstabs.com/zh-cn/spss/wilcoxon-signed-rank-test The Wilcoxon signed-rank test compares two measurements on the same cases by ranking the size of each paired difference and checking whether positive and negative changes balance out. Crosstabs runs it as SPSS NPAR TESTS /WILCOXON, with the sign test and McNemar's test for paired yes/no outcomes in the same panel. When to use: Before/after or pre/post measurements on the same people where the differences are skewed or contain outliers. Two ratings given by the same respondent, such as importance and satisfaction on a 1–5 scale. Matched pairs (twins, matched stores) compared on an ordinal outcome. A yes/no item asked twice of the same people: use McNemar to test whether the proportion changed. When not to: Two independent groups: use the Mann-Whitney U test. Three or more measurements of the same cases: use the Friedman test or repeated-measures ANOVA. Roughly normal differences and a need for a confidence interval for the mean change: the paired-samples t test is more informative. Validation: The paired engines are tested against independent reference values computed with SciPy 1.18.1, statsmodels 0.15.0, pingouin 0.7.0 and scikit-posthocs 0.17.1: Wilcoxon statistics, Z and exact and asymptotic p-values match scipy.stats.wilcoxon, the sign test matches scipy.stats.binomtest for up to 25 pairs and the SPSS normal formula above, and McNemar matches statsmodels mcnemar in exact and continuity-corrected forms, all within 1e-9 (relative to the larger of 1 and the reference value). #### Friedman Test — SPSS NPAR TESTS /FRIEDMAN (Friedman 检验) English: https://www.crosstabs.com/spss/friedman-test 简体中文: https://www.crosstabs.com/zh-cn/spss/friedman-test The Friedman test ranks each case's values across several related variables and checks whether some variables are ranked consistently higher. Crosstabs runs SPSS NPAR TESTS /FRIEDMAN with the tie-corrected chi-square, Kendall's coefficient of concordance W and Cochran's Q for yes/no items, plus Bonferroni-adjusted pairwise comparisons. When to use: The same respondents rate three or more products, concepts or ads on an ordinal scale. A skewed outcome measured at three or more time points on the same cases. Judges rank the same set of items and you want to quantify their agreement (Kendall's W). Several yes/no items answered by the same people, such as awareness of three brands (Cochran's Q). When not to: Independent groups of different people: use the Kruskal-Wallis test. Roughly normal repeated measures where you also need between-subject factors or contrasts: use repeated-measures ANOVA. Agreement between raters assigning categories rather than ranks: use Fleiss' kappa. Validation: The related-samples engines are tested against independent reference values computed with SciPy 1.18.1, statsmodels 0.15.0, pingouin 0.7.0 and scikit-posthocs 0.17.1: Friedman chi-square, p and Kendall's W match scipy.stats.friedmanchisquare, Friedman pairwise comparisons match scikit-posthocs (Siegel-Castellan), and Cochran's Q matches statsmodels cochrans_q, with its pairwise comparisons also checked, all within 1e-9 (relative to the larger of 1 and the reference value). ### Correlate / 相关 #### Bivariate Correlations — SPSS CORRELATIONS (双变量相关) English: https://www.crosstabs.com/spss/pearson-spearman-correlation 简体中文: https://www.crosstabs.com/zh-cn/spss/pearson-spearman-correlation A correlation coefficient measures how closely two variables move together, from −1 to +1. Crosstabs builds SPSS-style CORRELATIONS and NONPAR CORR matrices on your device: Pearson r for linear association, Spearman's rho and Kendall's tau-b for monotonic association, each with significance, N and optional flags. When to use: Screening many survey items or KPIs at once to see which move together. Two scale variables with a roughly linear relationship: Pearson r with its confidence interval. Ordinal ratings, skewed variables or data with outliers: Spearman's rho or Kendall's tau-b. When not to: A third variable may drive both: use partial correlation to control for it. You want to predict one variable from several others: use linear regression. Both variables are nominal categories: use Crosstabs with chi-square and Cramér's V instead. Validation: The correlation engine is tested against independent reference values computed with SciPy 1.18.1, statsmodels 0.15.0, pingouin 0.7.0 and scikit-posthocs 0.17.1: Pearson, Spearman and Kendall tau-b coefficients with two- and one-tailed p-values match scipy.stats under both pairwise and listwise deletion, as do the Pearson Fisher-z intervals; integer frequency weights match SciPy run on replicated cases, and fractional weights for Pearson match statsmodels DescrStatsW with df = W − 2. Tolerance is 1e-9 (relative to the larger of 1 and the reference value). #### Partial Correlation — SPSS PARTIAL CORR (偏相关) English: https://www.crosstabs.com/spss/partial-correlation 简体中文: https://www.crosstabs.com/zh-cn/spss/partial-correlation A partial correlation is the Pearson correlation between two variables after the linear influence of one or more control variables has been removed from both. Crosstabs runs SPSS PARTIAL CORR on your device and can print the zero-order correlations next to the partial ones, so you can see how much of an association a confounder explains. When to use: Checking whether an association survives once an obvious common cause (age, income, temperature, company size) is controlled. Separating the link between two survey scales from a general response tendency measured by a third scale. A quick, symmetric alternative to regression when you care about association rather than prediction. When not to: Ordinal or heavily skewed variables: the partial coefficient is Pearson-based, so rank or transform the variables first. A categorical control variable with more than two levels: enter it as dummy variables in a linear regression instead. You want to know how much one predictor adds to an outcome model: use the R² change from hierarchical linear regression. Validation: Partial correlations are tested against independent reference values computed with SciPy 1.18.1, statsmodels 0.15.0, pingouin 0.7.0 and scikit-posthocs 0.17.1: coefficients and p-values controlling for two variables match pingouin.partial_corr within 1e-9 (relative to the larger of 1 and the reference value), and the frequency-weighted result matches pingouin run on the replicated cases. ### Scale analysis / 度量分析 #### Reliability Analysis — SPSS RELIABILITY (可靠性分析) English: https://www.crosstabs.com/spss/cronbach-alpha 简体中文: https://www.crosstabs.com/zh-cn/spss/cronbach-alpha Reliability analysis checks whether the items of a multi-item scale measure the same thing consistently enough to be summed or averaged. Crosstabs runs SPSS RELIABILITY in your browser with alpha, split-half, Guttman, parallel, strictly parallel and McDonald's omega models, item-total statistics, ANOVA-based tests, Hotelling's T² and intraclass correlation coefficients. When to use: Before summing Likert items into a scale score, to report its reliability. Shortening a questionnaire by spotting items that add little. Measuring agreement among raters scoring the same cases, with an intraclass correlation. When not to: To show that items measure a single construct: alpha assumes it; run Factor analysis to check dimensionality. Agreement between two raters on categories: use Cohen's kappa or weighted kappa. Items with reverse-worded statements that have not been recoded: recode them first or alpha will be understated. Validation: Checked against pingouin 0.7, NumPy 2.5 and R 4.6.1 psych 2.6.9: alpha, standardized alpha, corrected item-total correlations and alpha-if-deleted agree within 1e-12, split-half coefficients and Guttman's lambdas within 1e-12, the ANOVA and Tukey nonadditivity F within 1e-9, Friedman (SciPy 1.18) and Cochran's Q (statsmodels 0.15, pingouin) within 1e-9, and ICCs within 1e-12 of R psych::ICC and pingouin. Frequency weights reproduce replicated cases. #### Factor Analysis — SPSS FACTOR (因子分析) English: https://www.crosstabs.com/spss/factor-analysis 简体中文: https://www.crosstabs.com/zh-cn/spss/factor-analysis Factor analysis looks for a small number of underlying dimensions that explain the correlations among many variables. Crosstabs runs SPSS FACTOR in your browser with principal components, principal axis factoring and maximum likelihood extraction, KMO and Bartlett's test, a scree plot, and varimax, quartimax, equamax, direct oblimin or promax rotation. When to use: Checking that a questionnaire's items group into the subscales they were written for. Reducing many correlated attitude or rating variables to a few components for segmentation or regression. Exploring the dimensions behind a new item pool before a confirmatory study. When not to: Testing a specific, pre-specified factor model: that needs confirmatory factor analysis (structural equation modelling), which this panel does not do. Nominal or ordinal categories with few levels: use categorical PCA or multiple correspondence analysis. Fewer than about five cases per variable: loadings will be unstable. Validation: KMO, MSA, anti-image matrices and Bartlett's test match factor_analyzer 0.5.1 (within 1e-12, chi-square 1e-9); principal components match NumPy eigenvectors within 1e-10; SPSS-style principal axis iteration and varimax, quartimax, equamax and promax rotations match independent NumPy implementations within 1e-10; converged principal axis and oblimin agree with R psych 2.6.9 and GPArotation within 1e-6, and maximum likelihood within 1e-4. Saved factor scores match R psych::factor.scores (regression, Bartlett and Anderson–Rubin) within 1e-8 for principal components, 1e-5 for converged principal axis factoring and 1e-3 for maximum likelihood. ### Regression / 回归 #### Linear Regression — SPSS REGRESSION (线性回归) English: https://www.crosstabs.com/spss/linear-regression 简体中文: https://www.crosstabs.com/zh-cn/spss/linear-regression Linear regression models a scale outcome as a weighted sum of predictors and estimates how much the outcome changes with each one, holding the others fixed. Crosstabs runs SPSS REGRESSION on your device with Enter, Stepwise, Forward, Backward and Remove blocks, R² change between blocks, standardized coefficients, VIF and the usual residual diagnostics. When to use: Explaining or predicting a continuous outcome (spend, satisfaction score, time) from several predictors. Testing whether a set of variables adds explanatory power beyond controls, by entering them in a second block. Key-driver analysis in market research, using Beta and part correlations to compare drivers. When not to: A yes/no outcome: use binary logistic regression; for three or more categories, multinomial or ordinal regression. Counts or strongly skewed positive outcomes: generalized linear models with a Poisson, negative binomial or gamma distribution fit better. Interaction terms or factorial designs: the panel enters main effects only, so build product variables first or use GLM Univariate. Survey data from a stratified or clustered design: use Complex Samples GLM for valid standard errors. Validation: The REGRESSION engine is tested against independent reference values (Python 3.13.13, NumPy 2.5.4, SciPy 1.18.1, statsmodels 0.15.0, pandas 2.3.3). With Enter, B, SE, t, p, CI, Beta, zero-order, partial and part correlations, tolerance, VIF, R², adjusted R², F, the ANOVA sums of squares and Durbin-Watson match statsmodels OLS within 1e-9 (relative to the larger of 1 and the reference value); per-case influence measures match OLSInfluence within 1e-9, residual statistics within 1e-8 and Belsley collinearity diagnostics within 1e-8 (variance proportions 1e-7). B, SE, p, VIF and adjusted R² also agree with R 4.6.1 lm and car 3.1.5 vif within 1e-9. Further tests cover frequency weights as case replication, R² change and F change for nested blocks and Remove, and Stepwise, Forward and Backward sequences replayed by a hand-run F-to-enter procedure in the generator. #### Binary Logistic Regression — SPSS LOGISTIC REGRESSION (二元 Logistic 回归) English: https://www.crosstabs.com/spss/binary-logistic-regression 简体中文: https://www.crosstabs.com/zh-cn/spss/binary-logistic-regression Binary logistic regression models the probability of a two-category outcome, such as converted or not, from scale and categorical predictors, and reports each effect as an odds ratio. Crosstabs runs SPSS LOGISTIC REGRESSION on your device with Enter, Forward: LR and Backward: LR, the beginning block, model fit statistics and a classification table at a cutoff you choose. When to use: Predicting a yes/no outcome such as churn, purchase, response or pass/fail from several predictors. Estimating adjusted odds ratios for an exposure while controlling for other variables. Screening which of many candidate predictors matter, with Forward: LR or Backward: LR selection. When not to: Three or more unordered outcome categories: use multinomial logistic regression. Ordered outcome categories (low/medium/high): ordinal regression keeps the ordering. Time until the event matters, with censored cases: use Cox regression. A predictor perfectly separates the outcome: estimates diverge; the output flags separation, but the model needs fewer or merged categories. Validation: The LOGISTIC REGRESSION engine is tested against independent reference values (Python 3.13.13, NumPy 2.5.4, SciPy 1.18.1, statsmodels 0.15.0, pandas 2.3.3). With Enter, B, S.E., Wald, Sig., Exp(B) and its 95% CI, −2LL, the omnibus chi-square, Cox & Snell and Nagelkerke R², the classification table at cut .5 and Hosmer–Lemeshow with SPSS grouping match statsmodels Logit within 1e-8 (relative to the larger of 1 and the reference value), and block 0 score tests within 1e-7. Against R 4.6.1 glm and ResourceSelection 0.3.6 hoslem.test on untied fitted values, B agrees within 1e-6, S.E. within 1e-5, −2LL within 1e-8 and Hosmer–Lemeshow within 1e-6. Forward: LR and Backward: LR are checked against a hand-run score/likelihood-ratio sequence (1e-7), and further tests cover frequency weights, sampling weights and separation detection. #### Mediation and Moderation (PROCESS-style) — SPSS PROCESS-style (中介与调节(PROCESS 风格)) English: https://www.crosstabs.com/spss/mediation-moderation-process 简体中文: https://www.crosstabs.com/zh-cn/spss/mediation-moderation-process Mediation asks whether X affects Y through an intermediate variable M; moderation asks whether the effect of X depends on a third variable W. Crosstabs fits the regression-based models psychologists usually run with Hayes' PROCESS macro (models 4, 1, 6, 7 and 14) on your device, with seeded percentile bootstrap intervals for indirect effects. It is an independent implementation, not affiliated with Andrew F. Hayes or PROCESS. When to use: Testing a theorised mechanism, for example whether service quality raises loyalty by building trust. Checking for whom or under which conditions an effect holds, such as a training effect that grows with prior experience. Replacing a Baron and Kenny causal-steps analysis with a direct test of the indirect effect. When not to: A binary or count outcome: this panel fits linear (OLS) equations only; logistic-outcome PROCESS models are not offered. Latent variables measured by several items, or models with feedback loops: use structural equation modeling software. Cross-sectional data offered as proof of causation: mediation estimates rest on causal assumptions that the data alone cannot verify. Survey data with a sampling weight or a complex design: sampling weights are blocked here; use Complex Samples procedures. Validation: The engine is tested against an independent Python reference (Python 3.13.13, statsmodels 0.15.0, SciPy 1.18.1, NumPy 2.5.4) and an R reference (R 4.6.1 with lm, sandwich 3.1.3, lmtest, car 3.1.5, lavaan 0.7.3 and interactions 1.2.0) on ten scenarios covering models 1, 4, 6, 7 and 14, HC3, centering, covariates, user-missing codes and frequency weights. Coefficients, SEs, t, p, intervals, F and R² change agree within 1e-9; Johnson–Neyman boundaries match SciPy root finding within 1e-9 and interactions::johnson_neyman within 1e-6; lavaan reproduces the indirect and total effects. Both references redraw the bootstrap with identical resample indices from an independent port of the seeded generator, and the bootstrap SEs and percentile limits agree within 1e-8. #### Multinomial Logistic Regression — SPSS NOMREG (多项 Logistic 回归) English: https://www.crosstabs.com/spss/multinomial-logistic-regression 简体中文: https://www.crosstabs.com/zh-cn/spss/multinomial-logistic-regression Multinomial logistic regression predicts which of several unordered categories a case falls into — a brand, a plan, a mode of transport — by fitting one logit for each category against a reference category. Crosstabs runs the SPSS NOMREG procedure on your device and reports model-fitting and likelihood-ratio tests, parameter estimates with Exp(B), goodness of fit, three pseudo R² measures and a classification table. When to use: The outcome is nominal with three or more categories that have no natural order, such as preferred brand or chosen channel. You want odds ratios for each category against a chosen baseline, with categorical factors and numeric covariates in the same model. Survey data with a frequency-weight variable that counts identical respondents. When not to: Only two outcome categories: binary logistic regression gives the same model with simpler output. The categories are ordered (ratings, stages): ordinal regression uses that order and needs far fewer parameters. You need interaction terms: the panel fits main effects of the chosen factors and covariates only. The data come from a complex survey design with sampling weights: use complex-samples logistic regression instead. Validation: Reference values come from statsmodels MNLogit, with the SPSS subpopulation goodness-of-fit statistics re-implemented in NumPy, for an unweighted model with the last category as reference and a frequency-weighted model with the first. Tests require −2 log likelihood, AIC, BIC and pseudo R² to agree within 1e-9, the model and effect chi-squares and Pearson/deviance statistics within 1e-8, B and standard errors within 1e-6, and the classification table exactly. A second fit is cross-checked against R nnet::multinom (−2LL within 1e-7, coefficients within 1e-5). Stepwise selection (forward entry, forward stepwise, backward elimination and backward stepwise with likelihood-ratio, score and Wald tests) matches an independent stepping implementation on statsmodels MNLogit step for step: the same entries and removals, with test statistics within 1e-5. #### Ordinal Regression — SPSS PLUM (有序回归) English: https://www.crosstabs.com/spss/ordinal-regression 简体中文: https://www.crosstabs.com/zh-cn/spss/ordinal-regression Ordinal regression models an outcome whose categories have a natural order — satisfaction ratings, severity grades, education levels — through cumulative probabilities that share one set of slopes. Crosstabs runs the SPSS PLUM procedure on your device with logit, probit, complementary log-log, negative log-log and cauchit links, and tests the proportional-odds assumption with the test of parallel lines. When to use: A Likert item, rating scale or graded outcome with three to about ten ordered categories. You want one effect per predictor rather than a separate logit for each category, as multinomial regression would give. You want to compare link functions — for example complementary log-log when higher categories are more probable. When not to: The categories have no order: use multinomial logistic regression. The outcome has many levels and behaves like a continuous score: linear regression is simpler and easier to report. The test of parallel lines is clearly rejected and the effects differ by threshold: consider multinomial regression or separate binary models. You need scale (heteroscedasticity) terms or interactions: this panel fits a location-only main-effects model. Validation: PLUM is tested against R ordinal::clm for the logit, probit, complementary log-log and negative log-log links and a frequency-weighted logit model: thresholds and location estimates agree within 1e-6, their standard errors within 1e-5, −2 log likelihood within 1e-8 and the general model behind the test of parallel lines within 1e-6. The cauchit link is checked against statsmodels OrderedModel (estimates within 1e-6, standard errors within 1e-4, −2LL within 1e-9). #### Probit Analysis — SPSS PROBIT (概率单位分析) English: https://www.crosstabs.com/spss/probit-analysis 简体中文: https://www.crosstabs.com/zh-cn/spss/probit-analysis Probit analysis fits an S-shaped curve to grouped yes/no responses — how many of n subjects responded at each dose — and turns it into effective doses such as the ED50 with fiducial confidence limits. Crosstabs runs the SPSS PROBIT procedure on your device with probit or logit models, log-transformed doses, an optional natural response rate and relative median potency between groups. When to use: Toxicology, pharmacology or agronomy bioassays where each row is a dose with the number tested and the number responding. You need effective doses (ED50, ED90, LD50) with confidence limits rather than just regression coefficients. Comparing two or more compounds or batches on a common slope (relative median potency). Some subjects respond even at zero dose and you want to estimate or fix that natural response rate. When not to: Each row is one individual with a 0/1 outcome and several predictors: binary logistic regression or GENLIN with a binomial probit link is more natural. The response is a continuous measurement rather than a count of responders: use nonlinear regression for a dose-response curve. You need effective-dose limits with several covariates: the confidence limits table is produced for a single covariate only. Validation: PROBIT is checked against a statsmodels binomial GLM with a probit or logit link and an independent NumPy implementation of the SPSS fiducial-limit and relative-potency formulas. With a grouping factor and log10 dose, estimates agree within 1e-7, standard errors within 1e-6, the Pearson statistic within 1e-7, and fiducial limits and relative median potency within 1e-6; a logit model without a factor also checks the heterogeneity-factor limits, and the estimated natural response rate agrees within 1e-5. #### Curve Estimation — SPSS CURVEFIT (曲线估计) English: https://www.crosstabs.com/spss/curve-estimation 简体中文: https://www.crosstabs.com/zh-cn/spss/curve-estimation Curve estimation fits several simple curves of one predictor — or of time — at once, so you can see which shape describes the relationship best. Crosstabs runs the SPSS CURVEFIT procedure on your device for eleven models, reports R², F and the coefficients of each in one table, and plots every curve over the data. When to use: Exploring the shape of a two-variable relationship before building a fuller model. A quick trend line for a short series, such as weekly adoption or monthly sales, using Time (case sequence). Comparing growth-type models (compound, growth, exponential, power) whose log-linear forms are easy to explain. When not to: More than one predictor: use linear regression, or nonlinear regression for a custom multi-variable formula. A curve that none of the eleven forms captures, such as an asymptotic learning curve with an unknown ceiling: write it in nonlinear regression. Forecasting a time series with seasonality or autocorrelation: use exponential smoothing or ARIMA. Validation: CURVEFIT is checked against NumPy least squares on the linearized form of each model using a fixture with a logistic upper bound of 120: R² agrees within 1e-9, F within 1e-7, coefficients within 1e-8 and both degrees of freedom exactly. A no-constant linear model on the case sequence is tested separately. #### Nonlinear Regression — SPSS NLR / CNLR (非线性回归) English: https://www.crosstabs.com/spss/nonlinear-regression 简体中文: https://www.crosstabs.com/zh-cn/spss/nonlinear-regression Nonlinear regression estimates the parameters of a model you write yourself — a saturation curve, an exponential decay, a Michaelis–Menten rate — by minimizing the sum of squared residuals. Crosstabs runs SPSS NLR on your device with a safe formula parser, and its CNLR options add bounds on the parameters and bootstrap standard errors. When to use: Theory gives a specific nonlinear form, such as y = a·(1 − e^(−b·x)) for learning or y = a·x / (b + x) for saturation. The parameters themselves (a ceiling, a half-life, a rate) are the quantities you want to report. Parameters must stay in a meaningful range, such as a positive rate or a share between 0 and 1 (use the CNLR bounds). When not to: The curve is one of the standard shapes in curve estimation, or becomes linear after a transformation: those fits are simpler and need no starting values. The outcome is a count or a proportion of responders: use generalized linear models or probit analysis, which model the error distribution correctly. You need linear or nonlinear constraints that combine several parameters: CNLR here supports simple per-parameter bounds only. Validation: NLR and CNLR are checked against scipy.optimize (curve_fit and least_squares): estimates agree within 1e-8, standard errors within 1e-6 and the residual sum of squares within 1e-9; a frequency-weighted fit must equal the expanded-data fit, and a CNLR fit with an active upper bound holds that parameter exactly at the bound. Bootstrap standard errors are tested for reproducibility and for staying within 40% of the asymptotic values. #### Weight Estimation (WLS) — SPSS WLS (权重估计(加权最小二乘)) English: https://www.crosstabs.com/spss/weighted-least-squares 简体中文: https://www.crosstabs.com/zh-cn/spss/weighted-least-squares When the spread of the residuals grows with a variable — spending varies more among high earners, errors are larger for bigger firms — ordinary least squares gives too much say to the noisiest cases. Crosstabs runs the SPSS WLS (Weight Estimation) procedure on your device: it weights each case by 1 / source^power and searches a range of powers for the one with the highest log-likelihood, or uses a weight variable you supply. When to use: A residual plot fans out as a predictor (or the size of the unit) increases. Each row is an average of a different number of observations, so its precision is known: use the weight variable directly. You want SPSS-style power estimation rather than guessing a weight such as 1/x or 1/x². When not to: You only need valid standard errors and are content with OLS coefficients: heteroscedasticity-robust standard errors in linear regression are simpler. The weights are survey design weights: these describe sampling, not error variance; use a complex-samples analysis. The outcome is a count or a positive skewed amount whose variance follows the mean: a GENLIN model with the matching distribution handles this directly. Validation: WLS is checked against statsmodels WLS and the SPSS WLS log-likelihood formula: every log-likelihood in the power grid agrees within 1e-9, the selected power matches exactly, coefficients and R² within 1e-9, standard errors and F within 1e-8. The weight-variable mode is tested the same way. #### Two-Stage Least Squares — SPSS 2SLS (两阶段最小二乘) English: https://www.crosstabs.com/spss/two-stage-least-squares 简体中文: https://www.crosstabs.com/zh-cn/spss/two-stage-least-squares Two-stage least squares estimates a regression when an explanatory variable is endogenous — correlated with the error because of omitted factors, simultaneity or measurement error — by replacing it with its prediction from instruments that move it but do not affect the outcome directly. Crosstabs runs the SPSS 2SLS procedure on your device and reports a model description, fit statistics, ANOVA and coefficients based on the structural residuals. When to use: Supply-and-demand or other simultaneous systems where the predictor and outcome are determined together. A key predictor is measured with error, or likely confounded, and you have a variable that shifts it but is unrelated to the outcome's error. Natural experiments, such as distance, policy timing or random encouragement, used as instruments. When not to: No credible instrument: 2SLS with an invalid instrument is often worse than OLS, and weak instruments give large, unstable estimates. All predictors are plausibly exogenous: ordinary linear regression is more precise. You need first-stage diagnostics, overidentification tests or robust standard errors: this panel reports the structural model only. Validation: 2SLS is checked against linearmodels IV2SLS and R AER::ivreg on a fixture with one endogenous predictor, one exogenous predictor and two excluded instruments: coefficients agree within 1e-9, standard errors within 1e-8 (also against AER::ivreg), and R² and the residual sum of squares within 1e-9. A test also confirms the Predictor & Instrumental classification. #### Quantile Regression — SPSS QUANTILE REGRESSION (分位数回归) English: https://www.crosstabs.com/spss/quantile-regression 简体中文: https://www.crosstabs.com/zh-cn/spss/quantile-regression Quantile regression models a chosen percentile of the outcome — the median, the 10th or the 90th — instead of the mean, so it is robust to outliers and shows whether predictors affect the low end, the middle and the high end differently. Crosstabs runs the SPSS QUANTILE REGRESSION procedure on your device for up to nine quantiles at once, with an exact linear-programming fit, i.i.d. or sandwich standard errors and pseudo R². When to use: Skewed outcomes such as income, spend, salaries or response times, where the median is more meaningful than the mean. Outliers in the outcome that would pull an OLS line. Questions about the tails, such as what drives the slowest 10% of delivery times or the top 10% of salaries. When not to: You need the mean, for example to forecast totals: linear regression estimates the conditional mean. Very small samples and extreme quantiles (such as 0.01 with 50 cases): there is too little data in the tail. Outliers in the predictors rather than the outcome: quantile regression is robust to outlying Y, not to high-leverage X values. Validation: Quantile regression is checked against R quantreg::rq with the Barrodale–Roberts method at τ = 0.25, 0.5 and 0.9 with a factor and a covariate: coefficients agree within 1e-8, i.i.d. standard errors (an R implementation of SPSS's sparsity formula with the Bofinger bandwidth) within 1e-8, Hendricks–Koenker sandwich standard errors with the Hall–Sheather bandwidth within 1e-5, and pseudo R² and mean absolute error within 1e-9. #### Partial Least Squares Regression — SPSS PLS (偏最小二乘回归) English: https://www.crosstabs.com/spss/partial-least-squares 简体中文: https://www.crosstabs.com/zh-cn/spss/partial-least-squares Partial least squares regression predicts one or more outcomes from many correlated predictors by first extracting a few latent factors that are built to explain the outcomes. Crosstabs runs SPSS-style PLS on your device with scale or categorical dependents, factors and covariates, and reports the proportion of variance explained per factor, variable importance in the projection (VIP), weights, loadings and coefficients in original units. When to use: Many correlated rating items (brand attributes, satisfaction drivers) used to explain one outcome such as purchase intent or overall satisfaction — classic key-driver analysis. More predictors than you can comfortably fit by least squares, or fewer cases than predictors. Several related outcomes predicted jointly from the same predictor set. When not to: You need significance tests or confidence intervals for each coefficient: the panel reports none; use linear regression when the predictors are not badly collinear. You want to choose a small subset of predictors automatically: use automatic linear modeling or predictor selection. The goal is to summarise predictors without any outcome: that is factor or principal components analysis. Survey weights must be applied: PLS here is unweighted and the panel says so. Validation: PLS is checked against two references on a 120-case file in three designs (one scale dependent with a factor and four covariates; two scale dependents; a categorical plus a scale dependent): an independent NumPy transcription of the SPSS algorithm, which matches weights, loadings, scores, VIP, standardised and raw coefficients and predictions to 1e-9 and the variance-explained and adjusted R² columns to 1e-12; and scikit-learn PLSRegression with scaling, which matches coefficients, intercepts, predictions, X weights, Y weights, X loadings and X scores (up to the sign of each factor) to 1e-10 with one dependent and 1e-6 with several. #### Automatic Linear Modeling — SPSS LINEAR (自动线性建模) English: https://www.crosstabs.com/spss/automatic-linear-modeling 简体中文: https://www.crosstabs.com/zh-cn/spss/automatic-linear-modeling Automatic linear modeling fits a linear regression after cleaning the predictors and choosing which ones to keep. Crosstabs runs SPSS LINEAR on your device: optional automatic data preparation (recasting, outlier trimming, missing-value replacement, merging of categories), forward stepwise or best-subsets selection by AICC, F, adjusted R² or an overfit-prevention error, and an effects table with predictor importance, coefficients and outliers. When to use: A scale outcome with many candidate predictors and messy data — outliers, scattered missing values, categorical predictors with many sparse levels. A quick, defensible first model to see which drivers matter before a hand-built regression. You want selection by an information criterion or a held-out error rather than p-values alone. When not to: The model is fixed in advance by theory: use linear regression or GLM Univariate and report every planned term. A binary, count or ordinal target: LINEAR needs a scale target; use logistic, generalized linear or ordinal regression. You need boosting or bagging ensembles: use the separate Ensembles of linear models procedure. Survey weights: sampling weights are not run; use the complex-samples procedures for design-based inference. Validation: LINEAR is checked against a NumPy hand-run of the SPSS LINEAR and Automated Data Preparation chapters that fits every candidate model with numpy.linalg.lstsq on explicit design matrices (SciPy for the distributions). Eleven cases cover forward stepwise by AICC, F, adjusted R² and ASE, best subsets, include-all with frequency weights, removal steps and the hybrid search beyond 20 effects. Preparation actions, selected effects, step sequences and subset rankings are identical; criteria agree to a relative 1e-9, R², adjusted R², AICC and ASE to 1e-10, coefficients, standard errors, t and confidence limits to 1e-8, p values to 1e-10, effect sums of squares and F to 1e-8, and leverage and Cook's distance to 1e-10. #### Categorical Regression (CATREG) — SPSS CATREG (分类回归(CATREG)) English: https://www.crosstabs.com/spss/categorical-regression 简体中文: https://www.crosstabs.com/zh-cn/spss/categorical-regression Categorical regression fits a linear regression after converting nominal and ordinal categories into optimal numeric values, so a rating outcome or a region predictor needs no arbitrary coding. Crosstabs runs the SPSS CATREG procedure in your browser by alternating least squares, reporting R², standardized coefficients, Pratt importance and the category quantifications. When to use: An ordinal outcome such as a 1–5 satisfaction rating with a mix of nominal, ordinal and numeric predictors. Nominal predictors with many categories where dummy coding would produce a long, hard-to-read coefficient list. Driver analysis: ranking predictors by importance when relationships may be nonlinear. When not to: You need confidence intervals, p values or predicted category probabilities: ordinal or multinomial logistic regression provide them; the panel computes no bootstrap standard errors or F tests. A continuous outcome with linear effects: ordinary linear regression is simpler and gives full inference. You need spline transformations, regularization (ridge, lasso, elastic net) or passive missing-value handling: these are not included. Validation: Reference values come from an independent numpy implementation of the SPSS CATREG alternating-least-squares scheme. With numeric levels it matches statsmodels OLS (R² to about 3e-16, betas to about 1e-8), and its R² for nominal and mixed models agrees with R Gifi::morals to about 3e-11. Tests on numeric, frequency-weighted numeric, nominal and mixed models require R², multiple R and apparent prediction error within 1e-10, betas, correlations, importance and tolerance within 1e-6, and quantifications within 1e-5. #### Ensembles of Linear Models (bagging, boosting) — SPSS LINEAR (线性模型集成(装袋、提升)) English: https://www.crosstabs.com/spss/linear-model-ensembles 简体中文: https://www.crosstabs.com/zh-cn/spss/linear-model-ensembles Ensembles combine many models to make predictions more stable (bagging) or more accurate (boosting). Crosstabs runs the SPSS ensemble algorithms for linear models on your device and reports how the ensemble compares with a single linear model and with the mean. When to use: A linear model's coefficients shift a lot from sample to sample and you want steadier predictions. You want to see whether reweighting hard-to-predict cases improves on a single linear model. When not to: You need interpretable coefficients: use linear regression or automatic linear modeling. Your target is categorical: these ensembles are for continuous targets. Validation: An independent NumPy/scikit-learn implementation reproduces the seeded bootstrap samples exactly and the component, ensemble, reference and naive R², diversity, AdaBoost.R2 model weights and every prediction within 1e-9, with and without frequency weights. ### General linear model / 一般线性模型 #### Generalized Linear Models — SPSS GENLIN (广义线性模型) English: https://www.crosstabs.com/spss/generalized-linear-models 简体中文: https://www.crosstabs.com/zh-cn/spss/generalized-linear-models A generalized linear model links the mean of an outcome to predictors through a link function and lets the outcome follow a distribution other than the normal — counts, positive skewed costs, proportions or yes/no events. Crosstabs runs the SPSS GENLIN procedure on your device with seven distributions, a choice of links and scale methods, Type III Wald or likelihood-ratio tests, robust standard errors, and generalized estimating equations for repeated or clustered data. When to use: Counts such as tickets, visits or defects (Poisson, or negative binomial when the variance exceeds the mean). Positive, right-skewed amounts such as claim costs or durations (gamma or inverse Gaussian, usually with a log link), or amounts with exact zeros (Tweedie). Binary outcomes or events out of trials, when you want a logit, probit, log or complementary log-log link in one framework. Repeated measurements of the same subjects, analysed with GEE and an exchangeable, AR(1) or unstructured working correlation. When not to: A continuous, roughly normal outcome with a few predictors: linear regression or GLM Univariate give the same estimates with familiar output. You need subject-specific random effects rather than population-averaged effects: use linear mixed models. You need interaction terms: the panel fits main effects of factors and covariates only. Frequency weights together with GEE: the repeated-measures model refuses them, so expand or clear the weight first. Validation: GENLIN is compared with statsmodels GLM and, for the negative binomial and Tweedie models, R MASS::glm.nb and statmod/tweedie. The tests cover Poisson (with ML, deviance and Pearson scale, robust SEs and frequency weights), gamma, inverse Gaussian, normal, binary logit, events/trials probit, negative binomial with estimated and fixed k, and Tweedie 1.5: log likelihood and information criteria agree within 1e-9, deviance and Pearson chi-square within 1e-8, and robust standard errors within 1e-6. GEE is checked against statsmodels GEE and a NumPy implementation of SPSS's moment estimators: coefficients within 1e-8, robust SEs within 1e-7, working correlations within 1e-8 and QIC/QICC within 1e-7. #### Hierarchical Loglinear Analysis — SPSS HILOGLINEAR (分层对数线性分析) English: https://www.crosstabs.com/spss/hierarchical-loglinear 简体中文: https://www.crosstabs.com/zh-cn/spss/hierarchical-loglinear Hierarchical loglinear analysis looks for the simplest set of associations that reproduces a cross-tabulation of two to six categorical variables. Crosstabs runs the SPSS HILOGLINEAR procedure on your device: it tests whole orders of interaction (K-way effects), measures each effect's partial association, and removes unnecessary terms by backward elimination until the remaining generating class still fits the table. When to use: Three or more categorical variables and you want to know which pairs are associated after accounting for the rest. Checking for conditional independence, such as whether age and purchase are unrelated once device is known. Survey or audit tables where no variable is singled out as the outcome. When not to: Only two variables: a chi-square test in Crosstabs answers the same question directly. One variable is clearly the response: a logit model in general loglinear analysis or logistic regression is easier to interpret. Many sparse cells, such as six factors with few cases: expected counts become tiny and chi-square tests unreliable. Validation: HILOGLINEAR is checked against Poisson GLM fits in statsmodels and a NumPy iterative-proportional-fitting implementation on a four-factor table: K-way likelihood-ratio and Pearson statistics and partial associations agree within 1e-7, saturated effect-coded estimates and standard errors (delta 0.5) within 1e-8, and backward elimination must reach the same final model, with its fit statistics within 1e-7 and identical df. #### General Loglinear Analysis — SPSS GENLOG (一般对数线性分析) English: https://www.crosstabs.com/spss/general-loglinear 简体中文: https://www.crosstabs.com/zh-cn/spss/general-loglinear General loglinear analysis fits a chosen model to the cell counts of a contingency table and reports a parameter for every effect, so you can read the size and direction of associations — or, in the logit form, how explanatory factors shift the odds of a response category. Crosstabs runs the SPSS GENLOG procedure on your device with Poisson loglinear and multinomial logit models, three design choices, goodness-of-fit tests and cell-level adjusted residuals. When to use: You already know which associations to test and want parameter estimates rather than a search for a model. One categorical variable is the response and the others explain it, but you want to stay within the table framework (multinomial logit). Comparing nested designs, for example main effects against all two-way interactions, by the change in likelihood-ratio chi-square. When not to: You want the data to suggest the model through backward elimination: use hierarchical loglinear analysis. You need custom designs with selected interactions, cell structures or covariates: the panel offers saturated, all two-way and main-effects designs only. A numeric response or continuous predictors: use logistic regression or generalized linear models. Validation: GENLOG is compared with statsmodels Poisson GLM fits for a Poisson main-effects model, a Poisson all-two-way model and a multinomial logit two-way model on three factors: likelihood-ratio and Pearson chi-square agree within 1e-8 with identical df, estimates within 1e-7 (the Poisson constant within 1e-8), standard errors within 1e-6, expected counts within 1e-8 and adjusted residuals within 1e-6. #### GLM Univariate — SPSS UNIANOVA (GLM 单变量) English: https://www.crosstabs.com/spss/glm-univariate 简体中文: https://www.crosstabs.com/zh-cn/spss/glm-univariate GLM Univariate fits one scale outcome against up to four fixed factors and five covariates, which covers two-way and three-way ANOVA, ANCOVA and unbalanced designs. Crosstabs runs the SPSS UNIANOVA procedure on your device and reports the Tests of Between-Subjects Effects with Type I–IV sums of squares and partial eta squared, plus estimated marginal means, Levene's test and post hoc comparisons. When to use: Two or more categorical factors and one scale outcome, such as recall score by ad format and market. You need to control for a continuous covariate (baseline score, age, prior spend) when comparing groups — ANCOVA. Cell sizes are unequal and you need a defined choice of sums of squares: Type III is the default, Type I gives sequential tests. When not to: Only one factor and no covariates: one-way ANOVA gives the same F plus Welch and Brown-Forsythe tests. Several correlated outcomes analysed together: use GLM Multivariate (MANOVA). The same respondents measured more than once: use repeated-measures ANOVA or a linear mixed model. A binary, count or ordinal outcome: use logistic regression, generalized linear models or ordinal regression. Validation: The UNIANOVA engine is tested against statsmodels 0.15 (anova_lm Type I–III, OLS parameter estimates, pairwise_tukeyhsd), SciPy 1.18 (Levene, noncentral F for observed power) and R 4.6 with car 3.1.5 (Type II and III with sum-to-zero contrasts) and emmeans 2.0.4. Sums of squares agree to a relative 1e-9, F to 1e-8 and p to 1e-7; marginal means and their standard errors to 1e-10; post hoc p values to 1e-8, or 2e-6 for the studentized-range tests. Frequency weights are tested to reproduce the case-replicated analysis. #### GLM Multivariate (MANOVA) — SPSS GLM (GLM 多变量(MANOVA)) English: https://www.crosstabs.com/spss/manova 简体中文: https://www.crosstabs.com/zh-cn/spss/manova MANOVA tests whether groups differ on several correlated outcomes at once, controlling the error rate that separate ANOVAs would inflate. Crosstabs runs SPSS GLM with two to ten dependent variables on your device and reports the four multivariate statistics, Box's test of equal covariance matrices, Bartlett's test of sphericity and a univariate F for each outcome. When to use: Two or more related scale outcomes measured on the same cases, such as speed and accuracy, or several satisfaction dimensions. You want one overall test before looking at outcomes one by one, to limit false positives. One to four fixed factors, optionally with covariates (MANCOVA). When not to: The outcomes are the same measure at different times: use repeated-measures ANOVA, which models the time structure. The outcomes are essentially uncorrelated: separate GLM Univariate analyses are simpler to report. You want to predict group membership from the outcomes: use discriminant analysis. Very small groups relative to the number of outcomes: the error matrix becomes unstable and Box's M cannot be computed. Validation: The four multivariate statistics, their F approximations, df and p values match R 4.6 car 3.1.5 Type III multivariate tests (values and F to a relative 1e-9, p to 1e-7); Pillai, Wilks and Roy also match statsmodels 0.15 mv_test. Box's M and Bartlett's test are checked against independent NumPy implementations of the SPSS formulas (1e-9), and the univariate follow-ups against the reference Type III sums of squares. #### Repeated Measures ANOVA — SPSS GLM WSFACTOR (重复测量方差分析) English: https://www.crosstabs.com/spss/repeated-measures-anova 简体中文: https://www.crosstabs.com/zh-cn/spss/repeated-measures-anova Repeated measures ANOVA tests whether the mean of the same measure changes across time points or conditions within the same people. Crosstabs runs SPSS GLM with WSFACTOR on wide columns, with one or two within-subject factors and up to two between-subject factors, and reports Mauchly's test, the sphericity-corrected F tests, multivariate tests, polynomial contrasts and Bonferroni- or Šidák-adjusted comparisons. When to use: The same respondents rated the same item at three or more waves, or under several conditions, stored as one column per measurement. You also want to compare groups (for example treatment vs control) on their trajectories. Every subject has all measurements; the design is complete after listwise deletion. When not to: Only two measurements and no groups: the paired-samples t test is simpler and gives the same p value. Many respondents miss some waves, or time points differ by person: a linear mixed model uses all available data. Ordinal ratings or strongly non-normal differences with small samples: use the Friedman test. Data are in long format (one row per measurement): restructure to wide columns first, or use MIXED. Validation: One within factor is checked against pingouin 0.7 rm_anova (sums of squares to a relative 1e-10, F, Mauchly's W and the epsilons to 1e-9, Greenhouse-Geisser p to 1e-6); the mixed design against R 4.6 car 3.1.5 Anova with idata (between, averaged within and multivariate tests; p to 1e-7) and pingouin mixed_anova; two within factors against statsmodels 0.15 AnovaRM. Mauchly's significance follows the SPSS chi-square formula, and marginal means with Bonferroni comparisons match an independent implementation to 1e-10. #### Variance Components — SPSS VARCOMP (方差成分) English: https://www.crosstabs.com/spss/variance-components 简体中文: https://www.crosstabs.com/zh-cn/spss/variance-components Variance components analysis splits the variability of a measurement into parts due to random factors — batches, raters, sites, operators — and residual error. Crosstabs runs SPSS VARCOMP on your device with six estimation methods, mixed models that include fixed factors, and the ANOVA and expected mean squares tables for the ANOVA method. When to use: Gauge or measurement-system studies: how much variation is due to operators, parts and repeat measurements. Multi-site or multi-rater data where you want to know how much sites or raters differ, not which ones. Planning: deciding whether to sample more batches or more measurements per batch. When not to: You care about specific levels (these three stores, these two formats): treat them as fixed factors in GLM Univariate. You need fixed-effect tests, random slopes or covariates alongside the random effects: use linear mixed models. Only a handful of levels per random factor: the component will be estimated very imprecisely whatever the method. Validation: MINQUE(0), MINQUE(1) and the ANOVA Type I and III methods match an independent dense NumPy implementation of the VARCOMP algorithms to a relative 1e-8 on a crossed operator × part design, and the ANOVA method equals the classical closed form on its balanced subset (1e-9). ML and REML match R 4.6 lme4 2.0.6 to a relative 2e-5 for every component and 1e-6 for the log likelihood, including a model with a fixed factor and a boundary estimate. #### Linear Mixed Models — SPSS MIXED (线性混合模型) English: https://www.crosstabs.com/spss/linear-mixed-models 简体中文: https://www.crosstabs.com/zh-cn/spss/linear-mixed-models Linear mixed models analyse clustered and longitudinal data — respondents within regions, repeated waves within people — by combining fixed effects with random effects and a residual covariance structure. Crosstabs runs SPSS MIXED on your device with up to three random-effects blocks, a repeated-measures block with five covariance types, REML or ML estimation and Type III tests with Satterthwaite degrees of freedom. When to use: Long-format longitudinal data where people have different numbers of waves — mixed models use every observed row. Respondents nested in interviewers, stores, schools or regions, where ignoring the clustering would overstate significance. Growth-curve questions: does the slope over time differ between people, and does a fixed factor change the average slope? When not to: No clustering or repeated measurements: GLM Univariate or linear regression gives the same fixed effects more simply. A binary or count outcome: this panel fits linear models with a normal residual only. Survey data with sampling weights: the panel does not run likelihood tests under a sampling weight; use Complex Samples GLM. Very few clusters (say under five): random-effect variances are poorly estimated; treat the cluster as a fixed factor. Validation: Thirteen models — random intercepts and slopes, nested, crossed, factorial, and AR1, CS, UN and diagonal repeated structures — are checked against R 4.6 lme4 2.0.6 with lmerTest 3.2.1 (−2LL to a relative 1e-8, fixed effects 1e-6, standard errors 1e-5, Satterthwaite df and Type III F 1e-4) and nlme 3.1.170 for the repeated structures, and against an independent dense SciPy implementation with numerical derivatives (information criteria 1e-8, covariance parameters 2e-4). A statsmodels MixedLM fit cross-checks the random-slope model. #### Generalized Linear Mixed Models (GENLINMIXED) — SPSS GENLINMIXED (广义线性混合模型(GENLINMIXED)) English: https://www.crosstabs.com/spss/generalized-linear-mixed-models 简体中文: https://www.crosstabs.com/zh-cn/spss/generalized-linear-mixed-models Generalized linear mixed models handle binary, count or continuous outcomes measured repeatedly within subjects. Crosstabs fits the SPSS GENLINMIXED model on your device with random intercepts and slopes per subject, using the restricted pseudo-likelihood with penalized quasi-likelihood linearization (REPL-PQL) that SPSS uses. When to use: Respondents answer the same yes/no question several times, or survey waves are nested in panelists. Counts per store or per respondent, with stores or respondents measured repeatedly. When not to: You need repeated-measures (R-side) covariance structures or several random-effect blocks: only one subject block of G-side effects is supported. You need Laplace or adaptive quadrature estimates: the panel uses REPL-PQL, as SPSS does. Validation: A dense-matrix R implementation of REPL-PQL run in Docker matches fixed effects and their standard errors within 1e-5 and covariance parameters within 1e-5 for binomial logit and probit and Poisson models; the normal model matches nlme and lme4 REML. lme4::glmer Laplace estimates are recorded for comparison only. #### Loglinear and Logit Models (LOGLINEAR) — SPSS LOGLINEAR (对数线性与 logit 模型(LOGLINEAR)) English: https://www.crosstabs.com/spss/loglinear-logit-models 简体中文: https://www.crosstabs.com/zh-cn/spss/loglinear-logit-models The legacy LOGLINEAR command fits loglinear and logit models to cross-tabulations with any design you specify. Crosstabs runs it on your device with deviation contrasts, goodness-of-fit tests, standardized and adjusted residuals and, for logit models, Haberman's analysis of dispersion. When to use: You want a specific non-hierarchical or custom design rather than K-way searches. You model a categorical response from categorical factors as a logit model and want dispersion measures. When not to: You want automatic backward elimination: use hierarchical loglinear analysis. Your predictors are continuous: use multinomial logistic regression. Validation: R glm Poisson fits with one constant per factor combination, run in Docker, with deviation codes cross-checked against contr.sum, match estimates, standard errors, fitted counts, adjusted residuals and both chi-squares within 1e-7, and Haberman's dispersion within 1e-7. ### Classify / 分类 #### Attribute Importance (AIM) — SPSS AIM (属性重要性(AIM)) English: https://www.crosstabs.com/spss/attribute-importance 简体中文: https://www.crosstabs.com/zh-cn/spss/attribute-importance Attribute importance profiles each group — typically each cluster from a segmentation — by testing which attributes make it different from the sample as a whole. Crosstabs runs SPSS AIM on your device: a chi-square test of each categorical attribute's category mix against the overall mix, a t test of each continuous attribute's group mean against the overall mean, and a −log10(Sig.) chart per group with an optional Bonferroni adjustment. When to use: Describing segments after a TwoStep, K-means or hierarchical cluster analysis, to write cluster profiles and names. Comparing any existing grouping (region, persona, plan) with the total sample on many attributes at once. Choosing which variables to feature in a segment summary. When not to: You want to compare groups with each other rather than with the whole sample: use one-way ANOVA or a crosstab chi-square. The clusters were formed from the same attributes: the tests will be significant by construction, so read them as descriptions, not as evidence. You want to predict group membership: use discriminant analysis or multinomial logistic regression. Case weights must be applied: the AIM algorithm defines none, so results are unweighted. Validation: AIM is checked against SciPy on a 300-case file with three groups, two categorical and three continuous attributes and scattered missing values: the per-group chi-square statistic (with SciPy's chi-square survival function) and scipy.stats.ttest_1samp against the overall mean. Every statistic agrees to 10 decimal places, every df exactly and every p value to 1e-12. #### Decision Tree (CHAID, CRT) — SPSS TREE (决策树(CHAID、CRT)) English: https://www.crosstabs.com/spss/chaid-decision-tree 简体中文: https://www.crosstabs.com/zh-cn/spss/chaid-decision-tree A decision tree splits cases into segments that differ as much as possible on a target, one predictor at a time. Crosstabs runs the SPSS TREE procedure with three growing methods: CHAID and Exhaustive CHAID, which make multiway splits backed by Bonferroni-adjusted tests, CRT, which makes binary splits by impurity reduction and can be pruned, and QUEST, which picks the split variable with ANOVA F, chi-square and Levene tests and places binary splits by quadratic discriminant analysis. When to use: Segmenting customers or respondents into rule-based groups you can describe in a sentence, such as Basic plan and under a year of tenure. Screening many categorical and scale predictors for the few that matter, including their interactions, before building a regression model. Targeting: ranking nodes by response rate to decide which segments to contact first. Nominal, ordinal or scale targets: CHAID uses chi-square, likelihood-ratio or F tests, CRT uses Gini, twoing or least squares. When not to: You need effect sizes with confidence intervals for each predictor: use binary or multinomial logistic regression. The sample is small: with the default minimum of 100 cases per parent and 50 per child node, a few hundred cases may give only one or two splits. You want groups without a target variable: that is cluster analysis (TwoStep, k-means or hierarchical). You need custom misclassification costs or profits: the panel uses equal costs and empirical priors. Validation: CHAID and Exhaustive CHAID are checked node by node against the R package chaidr 0.1.0 (R 4.6.1) on a 1,200-case synthetic market dataset — nominal, ordinal and scale targets, the likelihood-ratio statistic, frequency weights and a deeper tree — with identical node sizes and merged categories, and adjusted p values and statistics within a relative 1e-6; 10-fold cross-validated risk matches chaidr to 12 decimals, and the root split agrees with the R-Forge CHAID package 0.1.2. CRT matches rpart 4.1.27 on Gini and least-squares trees and surrogate routing, a brute-force search for Gini, twoing and least-squares splits to 12 decimals, and the manual's cost-complexity pruning sequence. #### TwoStep Cluster — SPSS TWOSTEP CLUSTER (两步聚类) English: https://www.crosstabs.com/spss/twostep-cluster 简体中文: https://www.crosstabs.com/zh-cn/spss/twostep-cluster TwoStep cluster finds natural groups in data that mixes continuous and categorical variables, and can choose the number of clusters for you. Crosstabs runs the SPSS TWOSTEP CLUSTER procedure on your device: CF-tree pre-clustering, then agglomerative clustering with the log-likelihood or Euclidean distance, with the cluster count picked by BIC or AIC. When to use: Segmenting customers on a mix of numbers (age, spend) and categories (channel, region) without converting categories to dummies. You do not know how many segments to expect and want an information-criterion answer as a starting point. Files too large for a full case-by-case distance matrix, since the CF tree summarises cases before clustering. When not to: Only continuous variables on comparable scales and a known number of groups: k-means is simpler and its centers are easier to report. A small file where you want to see every merge: hierarchical clustering shows the full agglomeration schedule and dendrogram. You already have a known group variable and want to predict it: use discriminant analysis or a decision tree. Validation: TwoStep is checked against an independent NumPy implementation of the TWOSTEP CLUSTER chapter of the IBM SPSS Statistics Algorithms manual: on 120 mixed continuous and categorical records, BIC and AIC for every cluster count and the distance-measure ratios agree within 1e-9, and the chosen number of clusters and every case assignment are identical (BIC, AIC and a fixed four-cluster solution, plus a continuous-only Euclidean set of 90 records). A 1,500-case test confirms the CF tree stays within 512 sub-clusters and repeats exactly; silhouette and predictor importance match an independent implementation of the Cluster Evaluation formulas. #### K-Means Cluster — SPSS QUICK CLUSTER (K-均值聚类) English: https://www.crosstabs.com/spss/k-means-cluster 简体中文: https://www.crosstabs.com/zh-cn/spss/k-means-cluster K-means assigns every case to the nearest of k cluster centers and moves the centers until assignments settle. Crosstabs runs SPSS QUICK CLUSTER: SPSS's single-pass rule for well-separated initial centers, optional running means, then nearest-center iterations, with no random starts, so the same file always gives the same clusters. When to use: Continuous variables on comparable scales, such as several 0–100 importance ratings, and a number of segments fixed in advance. Larger files where a full distance matrix would be slow: k-means only compares cases with k centers. Refining a solution found by hierarchical clustering, using its number of clusters. When not to: Categorical variables: k-means needs numeric distances; use TwoStep cluster for mixed data. Variables in very different units, unless you standardize them first, because the largest-scale variable would dominate. You do not know k at all: compare several values, or start with TwoStep or hierarchical clustering. Validation: QUICK CLUSTER is checked on 140 cases with k = 3 and k = 4: initial centers against an independent NumPy implementation of the SPSS initial-center rule, and iterations against scikit-learn 1.9.1 Lloyd k-means started from those centers. Initial and final centers agree to 10 decimal places, memberships are identical and the ANOVA F values agree with scipy.stats.f_oneway to 9 decimals; the running-means option is checked against a NumPy implementation of the update step, and the default 10-iteration limit is tested. #### Hierarchical Cluster — SPSS CLUSTER (系统聚类) English: https://www.crosstabs.com/spss/hierarchical-cluster 简体中文: https://www.crosstabs.com/zh-cn/spss/hierarchical-cluster Hierarchical clustering starts with every case as its own cluster and merges the two closest clusters at each stage until one remains, so you can see the whole grouping history before choosing how many clusters to keep. Crosstabs runs SPSS CLUSTER for cases with seven linkage methods and the interval, count and binary measures of PROXIMITIES. When to use: Small to medium files (up to 2,500 cases) where you want to inspect the grouping rather than fix k in advance. Grouping stores, products, countries or respondents on a handful of profile variables, with a dendrogram to show stakeholders. Binary or count data, using measures such as Jaccard, simple matching or the chi-square measure. When not to: Thousands of cases: the case-by-case matrix grows with the square of n; use k-means or TwoStep instead. Grouping variables rather than cases: this panel clusters cases; for variables use factor analysis or the Distances matrix between variables. Mixed continuous and categorical variables: TwoStep handles both directly. Validation: Agglomeration coefficients for all seven linkage methods on squared Euclidean distance, between-groups linkage on a Pearson correlation similarity and complete linkage on Z-scored Euclidean distance agree with SciPy 1.18.1 scipy.cluster.hierarchy within 1e-9 on a 36-case reference set (within-groups linkage, which SciPy lacks, against a brute-force NumPy implementation), and cluster memberships are identical. The silhouette matches an independent implementation of the SPSS Cluster Evaluation formulas to 12 decimals, and the dendrogram is checked to span the 0–25 rescaled range. #### Discriminant Analysis — SPSS DISCRIMINANT (判别分析) English: https://www.crosstabs.com/spss/discriminant-analysis 简体中文: https://www.crosstabs.com/zh-cn/spss/discriminant-analysis Discriminant analysis finds the linear combinations of numeric predictors that best separate known groups, then uses them to classify cases. Crosstabs runs SPSS DISCRIMINANT with direct or stepwise Wilks' lambda entry, Box's M, canonical function coefficients, Fisher's classification functions and both original and leave-one-out classification results. When to use: A categorical outcome with known groups (repaid, late, default) and several continuous predictors measured on every case. Describing how groups differ in a low-dimensional space, for example placing segments found by a cluster analysis on two functions. Predictors that are roughly normal within groups, where LDA is efficient and its posterior probabilities are well calibrated. When not to: Categorical predictors or clearly non-normal ones: use multinomial or binary logistic regression. Group covariance matrices that differ strongly (Box's M significant in a large sample): the linear rule is no longer optimal, and quadratic classification is not offered. You only need a predictive classifier and do not care about the functions: k-nearest neighbors or a decision tree may fit better. Validation: DISCRIMINANT is checked against an independent NumPy/SciPy implementation of the SPSS algorithms (NumPy 2.5.4, SciPy 1.18.1, statsmodels 0.15.0) on Fisher's iris data with equal and size-based priors, a two-class wine subset and a frequency-weighted set: eigenvalues and univariate Wilks' lambda and F agree within 1e-9, coefficients, structure matrix, centroids, Fisher functions, posteriors and Box's M within 1e-8, with identical original and leave-one-out predictions. R's MASS 7.3.66 lda and biotools 4.3 boxM agree within 1e-7; stepwise selection matches brute-force determinant search and statsmodels MANOVA, and weights equal replicated rows. #### Distances (Proximities) — SPSS PROXIMITIES (距离(近似值)) English: https://www.crosstabs.com/spss/distance-matrix 简体中文: https://www.crosstabs.com/zh-cn/spss/distance-matrix PROXIMITIES turns a set of variables into a matrix that says how alike or different every pair of cases, or every pair of variables, is. Crosstabs offers the SPSS interval, count and binary measures, six standardizations applied by variable or by case, and the absolute-value, sign-change and 0–1 rescale transforms. When to use: Preparing or inspecting the input to clustering or multidimensional scaling. Comparing a handful of items on several attributes, such as competitor products on price and ratings. Measuring agreement between binary profiles with Jaccard, Dice, simple matching or Yule's Q. When not to: You need significance tests for correlations: use bivariate correlations, which report p values and N. You want the groups themselves: run hierarchical clustering, which computes the same measures internally. More than 2,500 cases in a between-cases matrix: the panel stops; compute between variables or use a sample. Validation: On reference data, the eight interval measures between cases (Euclidean, squared Euclidean, block, Chebychev, Minkowski, customized power, cosine and Pearson correlation), correlation between variables, six standardizations by variable and two by case, the chi-square and phi-square count measures and all 27 binary measures agree within 1e-9 with values computed by SciPy 1.18.1 or independent NumPy implementations of the IBM SPSS PROXIMITIES formulas. A test confirms the absolute-value, sign-change and rescale transforms are applied in SPSS's order. #### Orthogonal Design (ORTHOPLAN) — SPSS ORTHOPLAN (正交设计(ORTHOPLAN)) English: https://www.crosstabs.com/spss/orthogonal-design 简体中文: https://www.crosstabs.com/zh-cn/spss/orthogonal-design An orthogonal design picks a small fraction of all attribute combinations so that every attribute's main effect can still be estimated independently. Crosstabs runs the SPSS ORTHOPLAN procedure in your browser: describe the factors and levels, and it builds the plan cards, adds holdout cards and verifies orthogonality. When to use: Planning a full-profile conjoint study: price, brand and feature attributes with two to nine levels each. The full factorial is too large for one respondent to rate (for example 3 × 3 × 3 × 2 × 2 = 108 profiles). You want holdout cards to test how well the estimated utilities predict new profiles. When not to: You need interactions between attributes: a main-effects plan cannot estimate them. Choice-based (discrete choice) conjoint with choice sets: this generates single profiles for rating or ranking, not choice tasks. Some level combinations are impossible and must be prohibited: ORTHOPLAN has no prohibitions, so edit the plan or use another design method. Validation: Tests check that every plan in the built-in library — each Plackett–Burman plan, each fixed plan and each column split — is an orthogonal array, that the SPSS decision rules choose the expected base plan for common designs (the 3 × 3 × 3 × 2 × 2 carpet-cleaner design gives 16 cards), that 300 random designs are all orthogonal and meet the requested minimum whenever the library allows, and that a seed reproduces the same cards and CSV layout. #### Conjoint Analysis — SPSS CONJOINT (联合分析) English: https://www.crosstabs.com/spss/conjoint-analysis 简体中文: https://www.crosstabs.com/zh-cn/spss/conjoint-analysis Conjoint analysis estimates how much each attribute level adds to a product's appeal from respondents' ratings or rankings of whole profiles. Crosstabs runs the SPSS CONJOINT procedure in your browser on a plan from ORTHOPLAN or a plan file, giving part-worth utilities with standard errors, importance values, fit statistics and preference simulations per respondent and pooled. When to use: Respondents rated or ranked full product profiles built from an orthogonal plan. Pricing and packaging decisions: how much a feature is worth relative to price, and which configuration wins share. Segmenting customers on their individual utilities (export them and cluster). When not to: Choice-based conjoint (respondents choose one option from each choice set): that needs a multinomial logit or hierarchical Bayes model, not OLS on ratings. You need interactions between attributes; the model is additive main effects only. The design cards are not orthogonal or are too few for the parameters: the panel refuses a design with no residual degrees of freedom. Validation: Reference values come from statsmodels OLS on SPSS-style effect-coded designs and SciPy correlations; for all-discrete factors, the utilities also agree with R's conjoint::caUtilities to about 4e-14. For score, rank and sequence data, tests require every respondent's and the group's utilities, standard errors, constant, linear and ideal-point coefficients, Pearson and Kendall correlations (including holdouts) and simulated shares to agree within 1e-9, and importance values within 1e-8. ### Dimension reduction / 降维 #### Correspondence Analysis — SPSS CORRESPONDENCE (对应分析) English: https://www.crosstabs.com/spss/correspondence-analysis 简体中文: https://www.crosstabs.com/zh-cn/spss/correspondence-analysis Correspondence analysis decomposes the chi-square association in a two-way table into a few dimensions and plots row and column categories on one map, so you can see which categories go together. Crosstabs runs the SPSS CORRESPONDENCE (ANACOR) procedure on your device, from two categorical variables or from a table of counts already stored in the data. When to use: A crosstab with many rows or columns — brand by attribute, region by preferred product, age band by news source — that is too large to read cell by cell. You already have an aggregated table of counts (one row per category, one column per category) rather than raw responses. You need a brand or positioning map for a report and want the numbers behind it (masses, contributions, quality). When not to: Three or more categorical variables at once: use multiple correspondence analysis. You only need to test whether two variables are associated: the crosstab chi-square test answers that directly. The table holds measurements or percentages rather than nonnegative counts or frequencies. A tiny table (2 × 2 or 2 × k) has only one dimension, so a map adds little beyond the row and column profiles. Validation: Reference values come from an independent numpy implementation of the SPSS CORRESPONDENCE formulas, cross-checked against R's ca package and Python's prince (largest difference about 1e-15). On three tables — the smoke data from ca, a random 6 × 5 and a wide 3 × 7 table — tests require singular values and inertias to agree within 1e-12, chi-square within 1e-9, masses, point inertias and contributions within 1e-10, and row and column scores under all four named normalizations within 1e-10. #### Multiple Correspondence Analysis — SPSS MULTIPLE CORRESPONDENCE (多重对应分析) English: https://www.crosstabs.com/spss/multiple-correspondence-analysis 简体中文: https://www.crosstabs.com/zh-cn/spss/multiple-correspondence-analysis Multiple correspondence analysis (homogeneity analysis) places the categories of several nominal variables in one low-dimensional space, so categories chosen by the same respondents land close together. Crosstabs runs the SPSS MULTIPLE CORRESPONDENCE (HOMALS) procedure in your browser, with passive or listwise missing values and four normalizations. When to use: Three or more nominal variables — plan type, channel, region, usage band — whose joint pattern you want to see. Exploring a survey before clustering: object scores give each respondent continuous coordinates. Some answers are missing and you want respondents to keep contributing through the questions they did answer (passive missing values). When not to: Exactly two variables: simple correspondence analysis gives the classic row-and-column map and a chi-square test. Ordinal or numeric variables whose order should be respected: use categorical principal components (CATPCA). Variables with dozens of rare categories: merge sparse categories first, or they dominate the solution (the panel lists variables with 2–30 categories). Validation: Reference values come from correspondence analysis of the indicator matrix in numpy, checked against R FactoMineR::MCA (agreement to about 1e-14), and, for passive missing values, from a direct implementation of the SPSS alternating-least-squares iteration. Tests on complete, frequency-weighted and passive-missing datasets require eigenvalues, Cronbach's alpha and discrimination measures to agree within 1e-10, and object scores, category coordinates and contributions within 1e-10 (1e-9 with passive missing values); object principal scores match FactoMineR within 1e-9. #### Multidimensional Scaling (PROXSCAL) — SPSS PROXSCAL (多维尺度分析(PROXSCAL)) English: https://www.crosstabs.com/spss/multidimensional-scaling 简体中文: https://www.crosstabs.com/zh-cn/spss/multidimensional-scaling Multidimensional scaling finds a low-dimensional map in which the distances between objects reproduce their dissimilarities as closely as possible. Crosstabs runs the SPSS PROXSCAL procedure on your device with SMACOF majorization, from a proximity matrix in your data or from Euclidean distances it computes between variables or cases. When to use: You have judged similarities or confusions between products, brands or stimuli and want a map of how people perceive them. You want to see which survey items or variables behave alike, from distances between variables. Only the order of the dissimilarities is trustworthy — use the ordinal transformation (nonmetric MDS). When not to: Several proximity matrices, one per judge, that should share a common space with individual weights (INDSCAL): the panel fits one matrix with the identity model. You want ALSCAL's S-stress algorithm specifically: use the separate Multidimensional scaling (ALSCAL) procedure. Your data are a contingency table of counts: correspondence analysis is designed for that. You need groups rather than a continuous map: use hierarchical or k-means clustering on the same distances. Validation: Reference configurations come from an independent SMACOF implementation of the PROXSCAL rules, cross-checked against R smacof::smacofSym (Torgerson start, secondary ties), stats::cmdscale and, for the ratio case, scikit-learn's SMACOF. On two 12-object matrices (one with tied proximities), tests require Stress-I within 1e-7, normalized raw Stress and D.A.F. within 1e-8, Stress-II and S-Stress within 1e-6 and every inter-point distance within 1e-5, for ratio, interval and ordinal transformations; classical scaling matches cmdscale within 1e-9. #### Categorical Principal Components (CATPCA) — SPSS CATPCA (分类主成分分析(CATPCA)) English: https://www.crosstabs.com/spss/categorical-pca 简体中文: https://www.crosstabs.com/zh-cn/spss/categorical-pca Categorical principal components analysis is PCA for variables that are not interval: it assigns optimal numeric values to the categories of nominal and ordinal items while extracting components. Crosstabs runs the SPSS CATPCA procedure in your browser by alternating least squares, with nominal, ordinal and numeric scaling levels per variable. When to use: Likert-type or 0–10 rating items where you doubt that the distance between adjacent points is equal. A mix of nominal (plan, region), ordinal (agreement) and numeric (counts) variables that should be summarized together. Building a small number of index scores from survey items with nonlinear relationships. When not to: All variables are continuous and roughly linear: ordinary factor analysis or PCA is simpler and gives rotation options. Every variable is nominal: multiple correspondence analysis is the special case designed for that. You need spline scaling levels, rotation, supplementary variables or passive missing-value imputation: these are not in the panel. Validation: Reference values come from an independent numpy implementation of the SPSS CATPCA alternating-least-squares scheme. Its eigenvalues match numpy PCA on all-numeric data (difference about 4e-16) and agree with R Gifi::princals within about 2e-5 on mixed data. Tests on a numeric and a mixed nominal/ordinal/numeric dataset require eigenvalues within 1e-9, Cronbach's alpha within 1e-8, loadings and quantifications within 1e-5 and object scores within 1e-4. #### Multidimensional Scaling (ALSCAL) — SPSS ALSCAL (多维尺度分析(ALSCAL)) English: https://www.crosstabs.com/spss/alscal-multidimensional-scaling 简体中文: https://www.crosstabs.com/zh-cn/spss/alscal-multidimensional-scaling ALSCAL builds a map whose distances reproduce dissimilarities, fitting squared distances by alternating least squares. Crosstabs runs the SPSS ALSCAL algorithm on your device for one dissimilarity matrix under the Euclidean model, at the ordinal or ratio level. When to use: You want to reproduce or compare an ALSCAL analysis from older SPSS syntax. Only the rank order of the dissimilarities is trustworthy (ordinal level). When not to: You have several matrices with individual weights (INDSCAL): only one matrix with the simple Euclidean model is fitted. You need the interval level: use PROXSCAL (Multidimensional scaling). Validation: An independent NumPy implementation of the ALSCAL chapter reproduces the additive constant, every S-stress in the iteration history, coordinates, disparities, Kruskal stress and RSQ within 1e-7 to 1e-9 for ordinal (ties kept or untied) and ratio levels; SciPy confirms the ratio solution is a stationary point of S-stress. #### Nonlinear Principal Components (PRINCALS) — SPSS PRINCALS (非线性主成分(PRINCALS)) English: https://www.crosstabs.com/spss/princals-nonlinear-pca 简体中文: https://www.crosstabs.com/zh-cn/spss/princals-nonlinear-pca PRINCALS finds a few components that summarise categorical variables while it also chooses the best numeric values (quantifications) for their categories. Crosstabs runs the SPSS PRINCALS algorithm on your device: alternating least squares with multiple nominal, single nominal, ordinal and numeric levels and passive missing values. When to use: You want to reproduce a PRINCALS analysis from older SPSS syntax. Some variables should keep their category order (ordinal) while others are free (nominal or multiple nominal). Objects with a missing value should stay in the analysis for their other variables (passive missing values). When not to: You need case weights, spline levels, supplementary variables or bootstrap confidence ellipses: use Categorical principal components (CATPCA). All variables are numeric and linear: ordinary factor or principal components analysis is simpler. Validation: When every variable is multiple nominal or numeric the solution has a closed form: an independent reference solves it as an eigenproblem with NumPy, and the eigenvalues, losses, loadings and object scores agree within 1e-6 (numeric PRINCALS equals PCA of the correlation matrix divided by the number of variables). Ordinal and single nominal solutions match an independent NumPy alternating least squares (best of 12 starts) within 1e-6, and ordinal PRINCALS matches R Gifi::princals within 1e-5 for the eigenvalues. #### Nonlinear Canonical Correlation (OVERALS) — SPSS OVERALS (非线性典型相关(OVERALS)) English: https://www.crosstabs.com/spss/overals-nonlinear-canonical-correlation 简体中文: https://www.crosstabs.com/zh-cn/spss/overals-nonlinear-canonical-correlation OVERALS asks what two or more sets of variables have in common. Each set is combined into scores that agree as closely as possible with the scores of the other sets, while the categories of each variable get optimal numeric values. Crosstabs runs the SPSS OVERALS algorithm on your device with multiple nominal, single nominal, ordinal and numeric levels. When to use: You have blocks of survey questions (for example attitudes and behaviours) and want the dimensions they share. The variables are categorical or ordinal, so linear canonical correlation would force arbitrary codes. When not to: All variables are numeric and only two sets are compared: linear canonical correlation gives the same answer with tests. You need significance tests or case weights: OVERALS is descriptive and every case counts once. Validation: When every variable is multiple nominal or numeric, an independent reference solves OVERALS in closed form as an eigenproblem of the average projector (NumPy), including listwise deletion per set; eigenvalues, losses per set, loadings and object scores agree within 1e-6, and two numeric sets reproduce SciPy-based canonical correlations. Single nominal and ordinal solutions match an independent NumPy alternating least squares (best of 12 starts, confirmed as a fixed point) within 1e-6. ### Survival / 生存分析 #### Kaplan-Meier — SPSS KM (Kaplan-Meier 生存分析) English: https://www.crosstabs.com/spss/kaplan-meier 简体中文: https://www.crosstabs.com/zh-cn/spss/kaplan-meier Kaplan-Meier estimates the probability of remaining event-free over time when some cases are censored, that is, followed for a while without the event happening. Crosstabs runs the SPSS KM procedure on your device: a survival table per group, mean and median survival with standard errors, quartiles, and log rank, Breslow and Tarone-Ware comparisons, optionally within strata. When to use: Time-to-event data where not everyone has had the event yet: churn, relapse, device failure, time to first purchase. Describing survival for the whole sample or comparing a handful of groups without adjusting for covariates. Comparing groups within strata, such as treatment arms within study sites, using the pooled-over-strata tests. When not to: You need to adjust for several covariates or estimate hazard ratios: use Cox regression. Event times are only known by interval (month 0–6, 6–12) or the sample is very large: actuarial life tables summarise more compactly. No case is censored and you only need to compare typical times: a Mann-Whitney U test on the times is simpler. Data come from a complex survey design: use Complex Samples Cox regression, because sampling weights are not accepted here. Validation: On a 150-case reference file with frequency weights, the KM engine matches R survival 3.8.12 (survfit: survival, Greenwood SE, restricted mean and its SE) to 1e-10, R survdiff log rank (with and without strata) and lifelines 0.30.3 log rank, Breslow and Tarone-Ware to 1e-9, and an independent NumPy implementation of the SPSS KM algorithms for medians, percentiles, trend and pairwise tests to 1e-9. #### Life Tables — SPSS SURVIVAL (寿命表) English: https://www.crosstabs.com/spss/life-tables 简体中文: https://www.crosstabs.com/zh-cn/spss/life-tables A life table groups survival times into equal intervals and reports, for each interval, how many cases entered, withdrew, were exposed to risk and had the event. Crosstabs runs the SPSS SURVIVAL procedure on your device, with cumulative survival, probability density and hazard rates with standard errors, the life-table median, and the Wilcoxon (Gehan) comparison of groups. When to use: Large samples where a row per event time would be unreadable, such as thousands of customers or policyholders. Times recorded in coarse units (months, years) with many ties, where interval summaries are the natural presentation. Reporting hazard by period, for example to show that cancellation risk peaks in the second quarter. When not to: Small samples with precise times: Kaplan-Meier uses every event time and loses no information to grouping. You need to adjust for covariates: use Cox regression. You need a second, controlling factor: this panel takes one factor, so split the file by the control variable or use Kaplan-Meier with strata. Validation: No mainstream library reproduces SPSS's actuarial table, so the engine is checked against an independent NumPy implementation of the SURVIVAL algorithm on a weighted 150-case reference file: entering, exposed, proportion terminating, cumulative surviving and its SE, density, hazard and their SEs, the median, and the Gehan (Lee-Desu) statistic and p value all agree to 1e-9. #### Cox Regression — SPSS COXREG (Cox 回归) English: https://www.crosstabs.com/spss/cox-regression 简体中文: https://www.crosstabs.com/zh-cn/spss/cox-regression Cox regression models how covariates change the hazard, the instantaneous risk of the event, without assuming a shape for the baseline survival curve. Crosstabs runs the SPSS COXREG procedure on your device with Breslow handling of tied times, indicator-coded categorical covariates, optional strata, hazard ratios Exp(B) with 95% intervals and the survival function at the covariate means. When to use: Time-to-event outcomes with censoring where you want to adjust a group comparison for age, severity or other covariates. Estimating the effect of a continuous predictor, such as dose or monthly spend, on the risk of an event. Baseline risk differs by site or cohort: put that variable in Strata so each stratum gets its own baseline hazard. When not to: Hazards clearly cross over time (one group is worse early and better late): the proportional hazards assumption fails; compare Kaplan-Meier curves or split follow-up into periods. You need interaction terms between ordinary covariates: compute the product first and enter it as a covariate. There is no time dimension, only whether the event happened: use binary logistic regression. Survey data with sampling weights, strata and clusters: use Complex Samples Cox regression. Validation: On a 150-case reference file with a continuous and a three-level covariate, coefficients and standard errors match statsmodels 0.15 PHReg and R survival 3.8.12 coxph (Breslow ties) to 1e-7, unweighted and with case weights; -2LL and the score chi-square match to 1e-9; the Kalbfleisch-Prentice survival at the covariate means and its SE match R to 1e-7; a stratified fit matches coxph with strata() to 1e-7. Forward and backward stepwise selection with conditional, LR and Wald removal follows an independent R survival stepping implementation step for step (statistics within 1e-6), and a time-dependent covariate x·ln(T_) matches coxph with tt() to 1e-6. ### Agreement and diagnostics / 一致性与诊断 #### ROC Curve — SPSS ROC (ROC 曲线) English: https://www.crosstabs.com/spss/roc-curve 简体中文: https://www.crosstabs.com/zh-cn/spss/roc-curve A ROC curve shows how well a score separates positive from negative cases by plotting sensitivity against 1 − specificity at every possible cutoff. Crosstabs runs the SPSS ROC procedure on your device for up to four test variables, with the area under the curve, its nonparametric or bi-negative exponential standard error, a test against AUC = .5 and a full table of cutoffs. When to use: Evaluating a diagnostic test, risk score or model probability against a known binary outcome. Choosing a cutoff that balances missed positives against false alarms. Comparing the discrimination of up to four scores side by side on the same cases. When not to: You need a significance test of the difference between two AUCs: use ROC Analysis, which runs the DeLong or independent-groups comparison. You care about calibration (are predicted probabilities correct?), not ranking: AUC ignores calibration; check a Hosmer-Lemeshow test in logistic regression. Positives are very rare and false positives are costly: also look at the precision column, since AUC can look good while precision is poor. Validation: The engine reproduces the worked example in the SPSS ROC algorithms documentation (area .64, SE .183). On a 180-case weighted reference file the area matches scikit-learn 1.9.1 roc_auc_score and R pROC 1.19.1 to 1e-12, and the nonparametric (Hanley-McNeil with ties) standard error matches an independent NumPy implementation to 1e-12 for both test directions, with the p value to 1e-9. #### ROC Analysis — SPSS ROC ANALYSIS (ROC 分析) English: https://www.crosstabs.com/spss/roc-analysis 简体中文: https://www.crosstabs.com/zh-cn/spss/roc-analysis ROC Analysis tests whether two areas under ROC curves differ. Crosstabs runs it on your device in both designs SPSS offers: two test variables measured on the same cases, compared with DeLong's paired test, or one test variable in two independent groups. It also draws precision-recall curves and lists every cutoff. When to use: Replacing an expensive test with a cheaper one and asking whether discrimination drops. Checking whether a risk model discriminates as well in a new population as in the development sample. Comparing two classifiers' scores on the same validation set. When not to: You only need one AUC or want up to four curves without a formal test: ROC Curve is simpler. You want to compare more than two tests at once: run pairwise comparisons and correct for multiplicity. The two tests were measured on different people but you treat them as paired: the DeLong test would be invalid; use the grouping-variable design. Validation: The paired DeLong comparison is checked against R pROC 1.19.1 roc.test on a 180-case weighted reference file: both areas agree to 1e-12, their variances and covariance to 1e-10, and z and p to 1e-9. The independent-groups z test combines the Hanley-McNeil standard errors validated on the ROC Curve page; it has no separate external reference test. #### Weighted Kappa — SPSS WEIGHTED KAPPA (加权 Kappa) English: https://www.crosstabs.com/spss/weighted-kappa 简体中文: https://www.crosstabs.com/zh-cn/spss/weighted-kappa Weighted kappa measures agreement between two raters on an ordered scale, giving partial credit when they disagree by one category instead of three. Crosstabs runs the SPSS WEIGHTED KAPPA procedure on your device with linear or quadratic weights, the asymptotic standard error, a z test against chance agreement, a confidence interval and the full rating crosstabulation. When to use: Two raters score the same items on an ordinal scale: severity grades, essay marks, Likert-type ratings. Test-retest agreement of one rater on an ordered scale, with the two occasions as the two variables. Reporting inter-rater reliability where quadratic weighted kappa is the expected statistic (it approximates the intraclass correlation). When not to: Categories have no order (diagnosis A, B, C): use unweighted Cohen's kappa in the Crosstabs statistics panel. Three or more raters: use Fleiss' kappa. Continuous measurements: use an intraclass correlation from Reliability Analysis or a Bland-Altman style analysis. Validation: On a 140-case weighted reference table, linear and quadratic weighted kappa, the asymptotic standard error and the null standard error match statsmodels 0.15 cohens_kappa to 1e-12 and z to 1e-10; kappa also matches R irr 0.85 to 1e-12. #### Fleiss' Kappa — SPSS FLEISS MULTIRATER KAPPA (Fleiss 多评分者 Kappa) English: https://www.crosstabs.com/spss/fleiss-kappa 简体中文: https://www.crosstabs.com/zh-cn/spss/fleiss-kappa Fleiss' kappa measures how consistently a group of raters assign items to nominal categories, beyond what chance would produce. Crosstabs runs SPSS's Fleiss multirater kappa on your device, with an overall kappa and a kappa for every rating category, each with an asymptotic standard error, z test and confidence interval. When to use: Three or more raters classify the same set of items into unordered categories, such as coders labelling open-ended answers. Different raters rate different items but every item gets the same number of ratings, which is what Fleiss' kappa assumes. Auditing a coding frame: low category kappas point to labels that need clearer definitions. When not to: Ordered categories with two raters: weighted kappa credits near misses. Exactly two raters on nominal categories: Cohen's kappa in the Crosstabs statistics panel is the usual choice. Items have different numbers of ratings: rows with a missing rating are dropped, so consider Krippendorff's alpha instead. Validation: On a reference file of 45 subjects rated by five raters, with frequency weights, the overall kappa matches statsmodels 0.15 fleiss_kappa and R irr 0.85 kappam.fleiss to 1e-12 and the overall z matches irr to 1e-9; category kappas and z values agree with irr's detail table within 6e-4, the precision irr prints. #### Ratio Statistics — SPSS RATIO STATISTICS (比率统计) English: https://www.crosstabs.com/spss/ratio-statistics 简体中文: https://www.crosstabs.com/zh-cn/spss/ratio-statistics Ratio Statistics summarises the ratio of two positive variables, most often assessed value to sale price, with the uniformity and vertical-equity measures assessors report. Crosstabs runs the SPSS RATIO STATISTICS procedure on your device: median, mean and weighted mean with confidence intervals, COD, PRD, PRB with its interval, coefficients of concentration and the coefficient of variation, overall and by group. When to use: Property tax assessment studies comparing appraised values with sale prices. Any audit of an estimate against an observed value: insurance reserves versus paid claims, forecast versus actual spend. Comparing uniformity across districts or classes with one column per group. When not to: Either variable can be zero or negative: those cases are excluded, so a ratio is not meaningful for your data. You want to model the ratio on several predictors: use linear regression on the log ratio. You need general descriptives of one variable: use Descriptives or Explore. Validation: On a 60-case weighted reference file, every statistic (median and mean with intervals, weighted mean and its delta-method interval, AAD, COD, COV, PRD, PRB with its interval, COC at 10% and 20%, minimum and maximum) matches an independent NumPy implementation of the SPSS RATIO STATISTICS algorithm on replicated data, with statsmodels OLS for PRB, to 1e-9. #### Proportions — SPSS PROPORTIONS (比例检验) English: https://www.crosstabs.com/spss/proportions-tests 简体中文: https://www.crosstabs.com/zh-cn/spss/proportions-tests The PROPORTIONS procedure tests and estimates binomial proportions in three designs: one sample against a test value, two paired measurements on the same cases, and two independent groups. Crosstabs runs it on your device and prints every test and interval method side by side, from Wald and score tests to exact binomial, Wilson, Clopper-Pearson, Newcombe and Agresti-Caffo. When to use: A share from a survey or audit compared with a benchmark: 62% satisfied against a 50% target. Before-and-after yes/no outcomes on the same respondents, such as awareness before and after a campaign. A/B tests and two-group comparisons of conversion, response or defect rates. When not to: More than two groups or a multi-category outcome: use a chi-square test in Crosstabs. You need to adjust for covariates: use binary logistic regression. Data come from a weighted, clustered survey: use Complex Samples Crosstabs, since sampling weights are not accepted here. Validation: One-sample score and Wald z match a direct computation to 1e-12; exact and mid-p one-sided p values match SciPy 1.18 binomial probabilities to 1e-10; Agresti-Coull, Clopper-Pearson, Jeffreys, Wilson and Wald intervals match statsmodels 0.15 proportion_confint to 1e-9; Wilson with and without continuity correction match R prop.test to 1e-12. Independent-sample Wald, pooled z, Agresti-Caffo, Newcombe and Wald intervals match statsmodels and R PropCIs 0.3.0 to 1e-9 or better; paired McNemar z, exact p and the Newcombe interval match an independent implementation. #### Meta-Analysis: Binary Outcomes — SPSS META BINARY (Meta 分析:二分类结局) English: https://www.crosstabs.com/spss/meta-analysis-binary 简体中文: https://www.crosstabs.com/zh-cn/spss/meta-analysis-binary A binary-outcome meta-analysis combines studies that each report events and non-events in a treatment and a control group. Crosstabs runs the SPSS META BINARY procedure on your device: odds ratio, risk ratio, risk difference or Peto's odds ratio, Mantel-Haenszel or inverse-variance fixed effects, random effects with seven τ² estimators and optional Knapp-Hartung adjustment, heterogeneity statistics, Egger's test and a forest plot. When to use: Systematic reviews of trials or cohort studies reporting a yes/no outcome per arm. Combining several A/B tests of the same change that each report conversions out of visitors. Rare events with zero cells: Peto's odds ratio handles them without adding 0.5. When not to: Studies report means and SDs, not counts: use the continuous-outcome meta-analysis. You want to explain heterogeneity with study characteristics: compute log odds ratios and their SEs, then use meta-regression. You have individual-level data from all studies: a stratified or mixed logistic model uses more information. Validation: On an eight-study reference set, odds ratio, risk ratio and risk difference effect sizes and variances match R metafor 5.2.1 escalc to 1e-12; Mantel-Haenszel estimates match rma.mh to 1e-10 (the MH risk-difference SE, where metafor uses a different variance formula, matches an independent implementation of the SPSS form); inverse-variance fixed effects and Q match to 1e-10; all seven τ² estimators match rma to 1e-6 for interior solutions; Peto matches rma.peto to 1e-10. #### Meta-Analysis: Continuous Outcomes — SPSS META CONTINUOUS (Meta 分析:连续结局) English: https://www.crosstabs.com/spss/meta-analysis-continuous 简体中文: https://www.crosstabs.com/zh-cn/spss/meta-analysis-continuous A continuous-outcome meta-analysis pools studies that compare the mean of a treatment and a control group. Crosstabs runs the SPSS META CONTINUOUS procedure on your device, computing Hedges' g, Cohen's d or raw mean differences from group summaries, or taking precomputed effects and standard errors, then pooling them with inverse-variance fixed or random effects, heterogeneity statistics, Egger's test and a forest plot. When to use: Studies report group sizes, means and SDs of the same outcome, possibly on different scales (use Hedges' g). All studies use the same instrument and you want the effect in its natural units (use Mean difference). You already have effect sizes and standard errors, e.g. from regression or published tables. When not to: Outcomes are counts of events: use the binary-outcome meta-analysis. You want to relate effect sizes to study characteristics: use meta-regression. You have raw data for one study only: use an independent-samples t test. Validation: On a seven-study reference set, Hedges' g and mean-difference effect sizes and variances match R metafor 5.2.1 escalc to 1e-12, and all seven τ² estimators with their pooled estimates match rma to 1e-6 for interior solutions; Hedges' g also agrees with statsmodels 0.15 effectsize_smd, and the DerSimonian-Laird τ² with statsmodels combine_effects, to 1e-3. Egger's test matches metafor regtest to 1e-8. #### Meta-Regression — SPSS META REGRESSION (Meta 回归) English: https://www.crosstabs.com/spss/meta-regression 简体中文: https://www.crosstabs.com/zh-cn/spss/meta-regression Meta-regression asks whether study-level characteristics, such as dose, year or design, explain why effect sizes differ between studies. Crosstabs runs the SPSS META REGRESSION procedure on your device with continuous and categorical moderators, random effects under seven τ² estimators or fixed effects with or without a dispersion parameter, Knapp-Hartung tests, residual heterogeneity and R². When to use: A meta-analysis shows substantial heterogeneity and you have a prespecified hypothesis about its source. Testing a dose-response relationship across trials that used different doses. Comparing subgroups of studies (e.g. randomized versus observational) with a categorical moderator. When not to: Fewer than about ten studies per moderator: estimates are unstable and false positives are common. You only need the pooled effect: use the binary or continuous meta-analysis. You want to infer individual-level effects from study averages: that is the ecological fallacy; you need individual data. Validation: On a seven-study reference set with a continuous and a three-level moderator, fixed-effect coefficients, SEs, the moderator test and residual Q match R metafor 5.2.1 rma(mods=) to 1e-8; REML, DL and ML random-effects τ², coefficients and SEs match to 1e-6, and the moderator test, I² and H² to 1e-5; Knapp-Hartung SEs match to 1e-6 and the F test and p to 1e-5. The multiplicative-dispersion model and R² have no separate reference test. ### Missing values / 缺失值 #### Missing Value Analysis — SPSS MVA (缺失值分析) English: https://www.crosstabs.com/spss/missing-value-analysis 简体中文: https://www.crosstabs.com/zh-cn/spss/missing-value-analysis Missing Value Analysis tells you how much data is missing, in which combinations, and whether the gaps look random. Crosstabs runs the SPSS MVA procedure on your device: univariate missing counts with extremes, tabulated patterns, separate-variance t tests, listwise, pairwise and EM estimates, and Little's MCAR test. When to use: Before any model on survey or panel data, to decide whether complete-case analysis is defensible or imputation is needed. When a reviewer or client asks whether the missingness is random: report Little's MCAR test and the separate-variance t tests. To get EM means and correlations as a quick MAR-based summary when a full multiple imputation is more than you need. When not to: To produce a filled-in dataset for analysis: EM estimates are summary statistics, not imputed values. Use Multiple Imputation and pool the results. To fill gaps in a time series by neighbouring values: use Replace Missing Values. With a sampling (survey) weight when you need tests: only the descriptive tables are shown, so use Complex Samples for design-based inference. Validation: The MVA engine is checked against independent Python references: univariate, listwise and pairwise statistics match numpy/pandas to 1e-11 relative, separate-variance t tests match scipy.stats.ttest_ind (Welch) to 1e-10, EM with the SPSS n − 1 denominator matches an independent numpy implementation to 1e-9 and Little's test to 1e-8, and maximum-likelihood EM with Little's test matches R norm::em.norm through naniar::mcar_test to 1e-7. Frequency weights are tested against the expanded data. #### Analyze Patterns — SPSS MULTIPLE IMPUTATION /IMPUTE METHOD=NONE (分析模式) English: https://www.crosstabs.com/spss/missing-data-patterns 简体中文: https://www.crosstabs.com/zh-cn/spss/missing-data-patterns Analyze Patterns is the first step of SPSS Multiple Imputation: it summarizes how much is missing at three levels and lists the patterns before anything is imputed. Crosstabs shows the overall summary, a variable summary, the pattern chart, the ten most frequent patterns and whether the pattern is monotone. When to use: Before Multiple Imputation, to choose the variables for the imputation model and see whether the monotone method applies. Panel or longitudinal studies with drop-out, where you expect later waves to be missing for people who left. Quick data-quality checks on a new survey file, including string and categorical variables. When not to: To test whether data are missing at random: use Missing Value Analysis for separate-variance t tests and Little's MCAR test. To fill in values: this step imputes nothing. Use Impute Missing Data Values or Replace Missing Values. Validation: The test suite checks that a constructed monotone pattern is reported as monotone and then imputed with the monotone method, and that a non-monotone pattern sends Multiple Imputation to FCS. The pattern counts of the companion MVA procedure are checked exactly against an independent numpy/pandas reference. #### Multiple Imputation — SPSS MULTIPLE IMPUTATION (多重插补) English: https://www.crosstabs.com/spss/multiple-imputation 简体中文: https://www.crosstabs.com/zh-cn/spss/multiple-imputation Multiple imputation replaces each missing value with several plausible draws, analyses every completed dataset, and combines the results so standard errors reflect the missing information. Crosstabs runs the SPSS MULTIPLE IMPUTATION approach on your device, with FCS or monotone imputation, linear regression, predictive mean matching or logistic models, and pooled descriptives and linear regression. When to use: Data are plausibly missing at random given other variables you have, and complete-case analysis would lose cases or bias results. You need honest standard errors for means or linear regression coefficients, not just a single filled-in dataset. You want the completed datasets as a file: the stacked CSV has an Imputation_ column, as in SPSS. When not to: Data missing not at random (for example, high earners skipping the income question for reasons unrelated to anything observed): imputation cannot correct that. With a sampling weight: imputation draws assume independent cases, so it is blocked. Clear the weight or declare a frequency weight. A gap-filled time series: use Replace Missing Values for interpolation or trend. Validation: Rubin's rules match R mice::pool (Barnard–Rubin df) to 1e-10 and an independent implementation of the SPSS formulas (Rubin 1987 df, fraction of missing information, relative efficiency, D1 test) to 1e-10. Imputations are checked structurally: with 60 imputations, pooled regression coefficients fall within 0.35 standard errors of R mice (norm) and 0.6 of statsmodels MICE, pooled means within 0.35 standard errors of scikit-learn IterativeImputer and R mice, and predictive mean matching imputes observed values only. #### Replace Missing Values — SPSS RMV (替换缺失值) English: https://www.crosstabs.com/spss/replace-missing-values 简体中文: https://www.crosstabs.com/zh-cn/spss/replace-missing-values Replace Missing Values fills gaps in a variable treated as a series in case order, creating a new variable and leaving the original untouched. Crosstabs offers the five SPSS RMV functions (SMEAN, MEAN, MEDIAN, LINT and TREND), a preview of every replaced value and a CSV download with the new columns. When to use: A time-ordered series (daily sales, monthly counts) with occasional gaps before forecasting or charting. A quick, transparent fill where a neighbour-based value is defensible and you want the method written next to the new variable. When not to: Cross-sectional survey data where row order means nothing: neighbour-based methods are meaningless there; use Multiple Imputation. When you need valid standard errors after filling: single imputation understates uncertainty, so pool multiple imputations instead. Long runs of missing values: interpolation across them invents a straight line; consider a time-series model. Validation: LINT, TREND, SMEAN, and the mean and median of nearby points (spans 1, 2, 3 and all) are checked against independent Python references — pandas interpolate for LINT, numpy polyfit for TREND and independent code for the nearby-point methods — and agree to 1e-12. Tests also confirm SPSS-style result names and that the source column is never changed. ### Bayesian statistics / 贝叶斯统计 #### Bayesian One-Sample Inference — SPSS BAYES ONESAMPLE (贝叶斯单样本推断) English: https://www.crosstabs.com/spss/bayesian-one-sample 简体中文: https://www.crosstabs.com/zh-cn/spss/bayesian-one-sample Bayesian one-sample inference asks how strongly the data support a test value for a mean, a proportion or a count rate, and what values of that parameter remain plausible. Crosstabs runs the SPSS BAYES ONESAMPLE procedure on your device for normal, binomial and Poisson data, reporting BF10 and BF01 next to the classical t test and an equal-tailed credible interval. When to use: Comparing a scale variable with a benchmark, target or published norm (satisfaction against 7, weight against a label value). Testing whether a yes/no rate differs from a reference proportion, such as a 50% preference or a 5% defect rate. Count data per unit (complaints per day, errors per page) tested against an expected Poisson rate. You want to state evidence for "no difference from the benchmark", not only fail to reject it. When not to: Two measurements on the same people: use Bayesian paired samples, which tests the mean of the differences. Two separate groups compared with each other: use Bayesian independent samples. Data collected with a survey design weight: the panel refuses a sampling weight, so use Complex Samples or report unweighted results. Validation: Checked against an independent Python reference (mpmath 30-digit quadrature of the BAYES formulas, SciPy, pingouin) and R BayesFactor 0.9.12.4.8: the JZS factor matches ttestBF(rscale = 1) to a relative 1e-8 and pingouin to 1e-7; fixed-g, known-variance, binomial and Poisson factors match to 1e-12; posteriors under all nine normal priors match to 1e-11 with credible limits to 1e-9; frequency weights reproduce replicated cases. #### Bayesian Paired Samples — SPSS BAYES RELATED (贝叶斯配对样本) English: https://www.crosstabs.com/spss/bayesian-paired-samples 简体中文: https://www.crosstabs.com/zh-cn/spss/bayesian-paired-samples The Bayesian paired-samples test weighs the evidence that the mean difference between two related measurements is zero against the evidence that it is not, and describes the plausible size of that difference. Crosstabs runs the SPSS BAYES RELATED procedure on your device for any number of variable pairs, with the same Bayes factor methods and nine posterior priors as the one-sample procedure. When to use: The same respondents measured twice, such as satisfaction before and after a service change. Matched pairs or two ratings of the same item where each row holds both values. A small pilot where you want to know whether the evidence already favours "no change" before collecting more data. When not to: The two scores come from different people: use Bayesian independent samples. Three or more repeated measurements: use Bayesian repeated-measures ANOVA. Ordinal ratings or differences with heavy outliers: consider the Wilcoxon signed-rank test. Validation: Checked against R BayesFactor 0.9.12.4.8 ttestBF(paired = TRUE, rscale = 1), agreeing to a relative 1e-8, and pingouin's paired Bayes factor to 1e-7; the t statistic matches the Python reference to 1e-12 and the credible limit under the diffuse prior to 1e-9. #### Bayesian Independent Samples — SPSS BAYES INDEPENDENT (贝叶斯独立样本) English: https://www.crosstabs.com/spss/bayesian-independent-samples 简体中文: https://www.crosstabs.com/zh-cn/spss/bayesian-independent-samples The Bayesian independent-samples t test compares the means of two groups by asking how much more likely the data are with a group difference than without one. Crosstabs runs the SPSS BAYES INDEPENDENT procedure on your device with three Bayes factor methods and five priors for the posterior of the difference, including the unequal-variance Behrens–Fisher case. When to use: Two independent groups (treatment vs control, two regions, two product versions) compared on a scale outcome. An A/B test where you want to report evidence for "no difference" when the variants perform alike. A grouping variable with more than two codes from which you pick the two groups to compare. When not to: Three or more groups at once: use Bayesian one-way ANOVA. The same people in both conditions: use Bayesian paired samples. A binary outcome such as converted/not converted: compare proportions with the Bayesian test of independence instead. Validation: The Rouder factor matches R BayesFactor 0.9.12.4.8 ttestBF(rscale = 1) to a relative 1e-8 and pingouin to 1e-7; the Gönen factor matches the mpmath reference to 1e-12 (1e-8 with a shifted effect-size mean), and the hyper-prior factor matches 30-digit quadrature to 1e-10 and its closed form to 1e-12. The posterior of the difference under all five priors matches to 1e-11 with credible limits to 1e-9. #### Bayesian Pearson Correlation — SPSS BAYES CORRELATION (贝叶斯 Pearson 相关) English: https://www.crosstabs.com/spss/bayesian-correlation 简体中文: https://www.crosstabs.com/zh-cn/spss/bayesian-correlation Bayesian Pearson correlation tests whether two scale variables are linearly related and gives a posterior distribution for the population correlation ρ. Crosstabs runs the SPSS BAYES CORRELATION procedure on your device for every pair of the variables you choose, computing the posterior moments and the highest-density interval by deterministic quadrature rather than sampling. When to use: Two or more scale variables whose pairwise linear association you want to quantify, such as study time and grades. Replication studies where demonstrating the absence of a correlation matters as much as finding one. Small samples where a credible interval for ρ communicates uncertainty better than a p value. When not to: Ordinal or heavily skewed variables: the classical Spearman correlation is more appropriate. You need to control for a third variable: use partial correlation or Bayesian linear regression. The relationship is clearly curved: Pearson's ρ only measures linear association. Validation: For every pair of a three-variable reference dataset, the JZS factor matches 30-digit mpmath quadrature to a relative 1e-8 and pingouin's bayesfactor_pearson(method = "wetzels") to 1e-7; posterior mode, mean, variance and interval bounds match the mpmath reference to 1e-8 under the uniform and c = −1.5 priors, and the fractional factor matches to 1e-8. #### Bayesian One-Way ANOVA — SPSS BAYES ANOVA (贝叶斯单因素方差分析) English: https://www.crosstabs.com/spss/bayesian-anova 简体中文: https://www.crosstabs.com/zh-cn/spss/bayesian-anova Bayesian one-way ANOVA compares a model in which group means differ with the intercept-only model and reports the Bayes factor between them. Crosstabs runs the SPSS BAYES ANOVA procedure on your device with four g-prior Bayes factors, the classical ANOVA table, and posterior distributions for every group mean and the error variance. When to use: One scale outcome across three or more independent groups, such as yield by fertilizer or score by teaching method. You need to argue that groups do not differ, for example equivalence of regional samples. You want interval estimates for each group mean that can be read as probabilities. When not to: Only two groups: Bayesian independent samples gives the posterior of the difference directly. Several factors or covariates: use Bayesian linear regression with factors, or GLM Univariate. Repeated measurements of the same people: use Bayesian repeated-measures ANOVA. Validation: The JZS and Rouder (s = 0.5) factors match 30-digit mpmath quadrature to a relative 1e-8 and R BayesFactor 0.9.12.4.8 lmBF on k − 1 dummy columns to 1e-8; Zellner and hyper-g factors match closed forms to 1e-12 and 1e-8. Group-mean and error-variance posteriors match the reference to 1e-10 (credible limits 1e-9) under both the reference and conjugate priors. #### Bayesian Linear Regression — SPSS BAYES REGRESSION (贝叶斯线性回归) English: https://www.crosstabs.com/spss/bayesian-linear-regression 简体中文: https://www.crosstabs.com/zh-cn/spss/bayesian-linear-regression Bayesian linear regression estimates how an outcome depends on several predictors and gives a full posterior distribution for every coefficient, plus a Bayes factor for the model against the intercept-only model. Crosstabs runs the SPSS BAYES REGRESSION procedure on your device with scale covariates and indicator-coded factors, four g-prior Bayes factors and a choice of reference or conjugate prior. When to use: A scale outcome predicted by several scale or categorical predictors, such as rent from area, distance and district. You want coefficient intervals that can be read as probabilities, or want to incorporate prior estimates. You need evidence that a set of predictors is useless (large BF01), not just a non-significant F. When not to: A binary or count outcome: use binary logistic regression or generalized linear models. You need per-predictor Bayes factors or model averaging: the panel reports one Bayes factor for the whole model. Stepwise variable selection, residual diagnostics or collinearity statistics: use Linear Regression. Validation: On a reference model with two covariates and a three-level factor, the JZS and Rouder (s = 0.5) factors match 30-digit mpmath quadrature to a relative 1e-8 and R BayesFactor 0.9.12.4.8 lmBF to 1e-8, Zellner and hyper-g factors match to 1e-12 and 1e-8, and coefficient posteriors match to 1e-10 (credible limits 1e-9) under both priors. A frequency-weighted fit matches BayesFactor on replicated rows to 1e-8. #### Bayesian Test of Independence — SPSS BAYES LOGLINEAR (贝叶斯独立性检验) English: https://www.crosstabs.com/spss/bayesian-test-of-independence 简体中文: https://www.crosstabs.com/zh-cn/spss/bayesian-test-of-independence The Bayesian test of independence asks whether two categorical variables are associated by comparing an independence model with a model that lets every cell vary. Crosstabs runs the SPSS BAYES LOGLINEAR procedure on your device with Gunel–Dickey, Good and Quintana Bayes factors under the sampling model that matches your design, and locates the association with posterior log odds ratios. When to use: Two categorical variables, such as plan type by acquisition channel or treatment by outcome category. You want evidence for independence, for example that a randomization produced balanced groups. Aggregated tables, entered as one row per cell with a count used as a frequency weight. When not to: More than two classification variables, or loglinear models with selected terms: use hierarchical or general loglinear analysis. You need Fisher's exact test, odds ratios with confidence intervals or Cramér's V: use the Crosstabs statistics in Analyze mode. Tables larger than 400 cells, which the panel refuses. Validation: On 2 × 2, 2 × 3 and 3 × 3 reference tables, the Gunel–Dickey factor matches R BayesFactor 0.9.12.4.8 contingencyTableBF(priorConcentration = 1) to a relative 1e-9 and the Python reference to 1e-10 under all four sampling models; Good's mixture matches to 1e-8 and Quintana's factor to 1e-12. Interaction posterior means match to 1e-12 and SDs to 1e-10, cross-checked against 200 000 NumPy Monte Carlo draws within 0.01. #### Bayesian Repeated-Measures ANOVA — SPSS BAYES REPEATED (贝叶斯重复测量方差分析) English: https://www.crosstabs.com/spss/bayesian-repeated-measures-anova 简体中文: https://www.crosstabs.com/zh-cn/spss/bayesian-repeated-measures-anova Bayesian repeated-measures ANOVA tests whether the mean of the same people changes across two or more conditions while separating stable differences between people from the condition effect. Crosstabs runs the SPSS BAYES REPEATED procedure on your device with the Nathoo–Masson BIC approximation or Rouder's mixed-design Bayes factor, the latter integrated by deterministic quadrature. When to use: The same participants measured under three or more conditions or time points, stored as one column per measure. Usability or within-subject experiments where you need evidence that versions perform alike. Pilot studies where a BIC-based Bayes factor gives a quick, prior-light summary. When not to: A between-subjects factor in addition to the repeated measures: use GLM Repeated Measures or linear mixed models. Subjects with missing time points you want to keep: the panel excludes incomplete subjects; consider linear mixed models. Ordinal ratings across conditions: use the Friedman test. Validation: The sums of squares and ΔBIC10 match the Python reference to 1e-11 (ΔBIC10 also with a frequency weight), and the posterior means and variance to 1e-6 and 1e-12. Rouder's factor, weighted and unweighted, matches an explicit-matrix quadrature to a relative 1e-6 and R BayesFactor 0.9.12.4.8 lmBF within four times BayesFactor's own Monte Carlo error. ### Forecasting / 预测 #### Time Series Forecasting (Expert Modeler) — SPSS TSMODEL (时间序列预测(专家建模器)) English: https://www.crosstabs.com/spss/time-series-forecasting 简体中文: https://www.crosstabs.com/zh-cn/spss/time-series-forecasting TSMODEL fits a forecasting model to one series and projects it forward with prediction intervals. In Crosstabs you can let an Expert Modeler compare exponential smoothing and ARIMA candidates by normalized BIC, or name one of seven smoothing models yourself; the model statistics follow the SPSS TSMODEL definitions. When to use: You track a metric over waves or months (awareness, NPS, share of purchase intent) and need a defensible forecast for the next quarter or year. You do not know whether a trend or seasonal model suits the data and want candidates compared on one criterion. Your file has one row per respondent with an interview date: the time-axis step aggregates rows into period means, sums or percentages before modelling. When not to: You need predictors, interventions or transfer functions: TSMODEL here is univariate, and the Expert Modeler does not detect outliers (ARIMA with a chosen order can). You already know the ARIMA orders you want: use the ARIMA procedure to set p, d, q and the seasonal orders directly. You only need to describe the seasonal pattern, not forecast: seasonal decomposition gives the seasonal factors directly. There are only a handful of periods; with fewer than six values the Expert Modeler will not run, and short series make any forecast fragile. Validation: The smoothing recursions for the Simple, Holt, damped, Winters' additive and Winters' multiplicative models are checked against statsmodels ExponentialSmoothing with the same initial states, agreeing to 1e-9; the ARIMA candidates use the ARIMA engine tested against statsmodels. Tests also confirm that the optimised Winters' additive weights sit at a local SSE minimum and that short series use exponential smoothing only. SPSS does not publish its ARIMA identification rules, so the Expert Modeler's ARIMA search (KPSS and seasonal-strength differencing, then a stepwise normalized-BIC search) is a documented substitute and can pick a different model than SPSS. #### Exponential Smoothing — SPSS EXSMOOTH (指数平滑) English: https://www.crosstabs.com/spss/exponential-smoothing 简体中文: https://www.crosstabs.com/zh-cn/spss/exponential-smoothing Exponential smoothing forecasts a series from weighted averages of its past, with separate weights for level, trend and season. Crosstabs runs the SPSS EXSMOOTH procedure: pick a trend (none, linear, exponential, damped) and a season (none, additive, multiplicative), and it tries every combination on the SPSS default grid to find the weights with the smallest sum of squared errors. When to use: You want a transparent forecast whose behaviour you can explain: recent periods weigh more, older ones less. You know the shape of the series (e.g. linear trend plus a stable monthly pattern) and want to reproduce an EXSMOOTH specification. You want to fix the weights yourself to compare with an earlier SPSS run: untick the grid search and type them in. When not to: You are unsure which trend and season to choose: let TSMODEL's Expert Modeler compare models by normalized BIC. The series has gaps: EXSMOOTH needs a complete series, so fill missing time points by interpolation in step 1 or use ARIMA. Autocorrelation remains after smoothing or you need coefficient tests: an ARIMA model is the better tool. Validation: For all twelve trend × season combinations, the initial level and trend (to 10 decimals), the fitted values at fixed weights (to 1e-10) and the grid-search best SSE and alpha are checked against an independent Python implementation of the SPSS manual's recursions. For the additive and non-seasonal trend models, those recursions are in turn matched to statsmodels ExponentialSmoothing with the same initial states to 1e-10. #### ARIMA — SPSS ARIMA (ARIMA 模型) English: https://www.crosstabs.com/spss/arima-forecasting 简体中文: https://www.crosstabs.com/zh-cn/spss/arima-forecasting ARIMA models a series through its own past values (AR), past shocks (MA) and differencing (I), with optional seasonal terms. Crosstabs runs the SPSS ARIMA procedure by exact maximum likelihood with a Kalman filter, then reports coefficient tests, SPSS-style AIC and SBC, residual autocorrelations and forecasts with t-based limits. When to use: The ACF and PACF suggest a specific order, e.g. an AR(1) cut-off or a seasonal spike at lag 12. You want to reproduce a published or earlier SPSS model such as the airline model ARIMA(0,1,1)(0,1,1)12. You need formal tests on the coefficients rather than only a forecast. When not to: You do not know which order to use: let TSMODEL's Expert Modeler search, or start from the ACF and PACF. You need predictors, intervention events or transfer functions: those parts of SPSS ARIMA and TSMODEL are not included. The series is very short: after differencing there must be more usable values than parameters plus three. Validation: Estimates are compared with statsmodels ARIMA (exact ML, Kalman filter) on ARMA(1,1) with a constant, the log airline model, ARIMA(2,1,0) with a constant and a seasonal AR: the maximised log-likelihood agrees within 1e-3 and is never lower by more than 1e-4, coefficients agree to about three decimals, Hessian standard errors within 2% and forecasts within 1e-3. At fixed coefficients, log-likelihoods match to 7–9 decimals, including a seasonal model and a series with missing values, and forecasts and their standard errors to 1e-6. Outlier detection is checked against R tsoutliers: Chen–Liu statistics agree to 1e-8 for additive, innovational, level-shift, transient and seasonal additive outliers (tsoutliers' seasonal level shift has the TSMODEL seasonal-additive pattern), and the same outliers are found on simulated series. Local-trend and additive-patch statistics are checked to 1e-8 against the same definitions computed in R. On simulated series with planted outliers of every type, the search reports the planted type at the planted time, and the joint fit matches R arima with the outlier regressors (coefficients within 2e-4, log-likelihood within 1e-4). #### Seasonal Decomposition — SPSS SEASON (季节分解) English: https://www.crosstabs.com/spss/seasonal-decomposition 简体中文: https://www.crosstabs.com/zh-cn/spss/seasonal-decomposition Seasonal decomposition separates a series into a repeating seasonal pattern, a smooth trend-cycle and what is left over. Crosstabs runs the classical ratio-to-moving-average method of SPSS SEASON, multiplicative or additive, and gives the seasonal factors plus the SAS, SAF, STC and ERR series as a CSV download. When to use: A monthly tracker jumps every December and you want to know whether this December was better than usual. You need seasonally adjusted figures for reporting or as input to another model. You want to see whether the seasonal swing is constant (additive) or grows with the level (multiplicative). When not to: There is no clear periodicity or fewer than two full cycles of data. You need forecasts: use exponential smoothing with a seasonal component or a seasonal ARIMA model. You need X-13ARIMA-SEATS style adjustment with trading-day and holiday effects; SEASON is the simpler classical method. Validation: The additive model with centred weights matches statsmodels seasonal_decompose (seasonal and trend components) to 1e-11. The multiplicative model, equal weights and an odd periodicity are checked against an independent numpy implementation of the SPSS manual's formulas, with factors, SAS, STC and ERR agreeing to 1e-11. #### Autocorrelations (ACF and PACF) — SPSS ACF (自相关与偏自相关(ACF/PACF)) English: https://www.crosstabs.com/spss/autocorrelation-acf-pacf 简体中文: https://www.crosstabs.com/zh-cn/spss/autocorrelation-acf-pacf The autocorrelation function (ACF) measures how strongly a series correlates with its own past at each lag; the partial autocorrelation function (PACF) removes the effect of the shorter lags in between. Crosstabs runs SPSS ACF and PACF with white-noise or Bartlett standard errors, the Box-Ljung test and optional log, regular and seasonal differencing. When to use: Before fitting an ARIMA model, to choose differencing and the AR and MA orders. To check whether period-to-period changes in a tracker are just noise or carry over from one wave to the next. To test model residuals for remaining autocorrelation (ARIMA shows its own residual ACF). When not to: You want the relationship between two different series: use cross-correlations. You want to find cycles of unknown length: spectral analysis looks at all frequencies at once. You have only a handful of periods: with fewer than about 20 points the correlations are too noisy to read reliably. Validation: On a 72-month seasonal series and an ARMA(1,1) series, autocorrelations and white-noise standard errors match statsmodels acf to 1e-12, Bartlett standard errors and Box-Ljung statistics (acorr_ljungbox) to 1e-10, p values to 1e-8 and the PACF (Levinson-Durbin) to 1e-10. With missing values, the Cryer-formula autocorrelations match both an independent numpy implementation and statsmodels to 1e-12, and the Box-Ljung statistics match the numpy implementation to 1e-10. #### Cross-Correlations — SPSS CCF (互相关) English: https://www.crosstabs.com/spss/cross-correlation 简体中文: https://www.crosstabs.com/zh-cn/spss/cross-correlation Cross-correlation measures how two series move together when one is shifted in time against the other. Crosstabs runs SPSS CCF on two aligned series, reporting the correlation and its standard error at every lag from −k to +k, so you can see whether ad spend leads sales or awareness lags media. When to use: You want to know how many periods it takes for a campaign, price change or news event to show up in a tracked metric. You are choosing which lag of a predictor to put into a regression or AREG model. Two trackers move together and you want to check whether one is a leading indicator for the other. When not to: Both series trend: shared trends create large spurious cross-correlations, so difference them first. You need a model with significance tests for the effect: use AREG or a regression on lagged variables. Either series has gaps inside the shared span: CCF needs both series complete there. Validation: Cross-correlations at lags −7 to +7 and their standard errors match statsmodels ccf, computed in both directions, to 1e-12 on a reference pair of series. #### Autoregression (AREG) — SPSS AREG (自回归(AREG)) English: https://www.crosstabs.com/spss/autoregression-areg 简体中文: https://www.crosstabs.com/zh-cn/spss/autoregression-areg When a regression is run on time series, the errors are often correlated from one period to the next, which makes ordinary standard errors too small. AREG estimates the first-order autocorrelation rho together with the coefficients by the Prais-Winsten or Cochrane-Orcutt method, as SPSS AREG does, and reports tests you can trust. When to use: You regress a monthly outcome (sales, awareness) on drivers (price, spend) and the Durbin-Watson is well below 2. You need valid significance tests for a driver in time-series data, not just a forecast. You want to reproduce an SPSS AREG result by Prais-Winsten (keeps the first observation) or Cochrane-Orcutt (drops it). When not to: The errors follow a higher-order or seasonal pattern: AREG corrects only AR(1). The data are cross-sectional, not ordered in time: use ordinary linear regression. You need forecasts beyond the data: AREG here reports fitted values and intervals for observed periods only. Validation: Prais-Winsten and Cochrane-Orcutt iterations are checked against an independent Python implementation of the SPSS manual: the rho at each iteration, final rho and its standard error, coefficients, mean squared error and Durbin-Watson agree to 1e-10, coefficient standard errors to 1e-9, with the same residual df. Cochrane-Orcutt coefficients and standard errors also match statsmodels GLSAR. #### Spectral Analysis — SPSS SPECTRA (谱分析) English: https://www.crosstabs.com/spss/spectral-analysis 简体中文: https://www.crosstabs.com/zh-cn/spss/spectral-analysis Spectral analysis breaks a series into sine and cosine waves and shows how much of its variance sits at each frequency, so cycles of any length show up as peaks. Crosstabs runs univariate SPSS SPECTRA: the periodogram at the Fourier frequencies and a spectral density smoothed with a Tukey-Hamming, Tukey, Parzen, Bartlett or Daniell window. When to use: You suspect a cycle but do not know its length, or want to confirm the periodicity before a seasonal model. A long daily or weekly series may contain several overlapping rhythms. You want a frequency-domain view to complement the ACF. When not to: The series has a strong trend: low-frequency power swamps the cycles, so difference or detrend it first (e.g. with Create time series). You need the coherence or phase between two series: bivariate SPECTRA is not included. The series is short or has gaps: SPECTRA needs a complete series, and with few points the frequency resolution is coarse. Validation: On even- and odd-length reference series, the Fourier cosine and sine coefficients and the periodogram match a numpy FFT to 1e-10 and scipy.signal.periodogram to 1e-9 (allowing for scipy's halved Nyquist term). Smoothed densities for all five windows at span 5, and Tukey-Hamming at span 9, match an independent Python implementation of the SPSS window formulas to 1e-10. #### Create Time Series — SPSS CREATE (创建时间序列) English: https://www.crosstabs.com/spss/create-time-series 简体中文: https://www.crosstabs.com/zh-cn/spss/create-time-series CREATE builds a new series from an existing one with a time-series function: a difference to remove a trend, a lag or lead to line up a predictor, or a moving average or T4253H smoother to reveal the underlying movement. Crosstabs offers nine SPSS CREATE functions, previews the result and lets you download it as CSV without changing your file. When to use: You need a lagged or differenced variable for a regression, AREG or FIT run. A monthly tracker is noisy and you want a 3-month moving average or a robust smooth for a chart. You want year-on-year changes: a seasonal difference of order 1 with periodicity 12. When not to: You want the new variable inside your open dataset: CREATE here produces a preview and CSV, not a new column. You want to fill missing values: use Replace missing values, or the time axis's linear interpolation option. You want a model-based trend or seasonal adjustment: use seasonal decomposition or exponential smoothing. Validation: Twelve function settings (cumulative sum, first and second differences, a 12-period seasonal difference, lag 2, lead 3, centred moving averages of span 5 and 4, a prior moving average of span 3, running medians of span 5 and 4, and T4253H) match numpy and pandas rolling computations to 1e-11; T4253H is checked against an independent Python implementation of the SPSS manual's description. #### Forecast Goodness of Fit — SPSS FIT (预测拟合优度) English: https://www.crosstabs.com/spss/forecast-goodness-of-fit 简体中文: https://www.crosstabs.com/zh-cn/spss/forecast-goodness-of-fit FIT compares an observed series with a fitted or forecast series and summarises the errors. Use it to score a model from another tool, a hold-out forecast or a simple benchmark: Crosstabs reports the SPSS FIT error statistics, from mean error and MAPE to RMSE and Durbin-Watson. When to use: You have forecasts from another source (a spreadsheet model, a vendor, a past report) next to the actual values. You want to compare several models on the same hold-out periods with identical error definitions. You need a naive benchmark (last value or same month last year) to judge whether a model adds anything. When not to: You have not built the forecast yet: TSMODEL, ARIMA and exponential smoothing already report their own fit statistics. The observed series contains zeros and you need MAPE: percent errors are undefined there, so rely on MAE or RMSE. You need a significance test comparing two forecasts; FIT reports descriptive error statistics only. Validation: Mean error, mean percent error, MAE, MAPE and Durbin-Watson match a numpy reference to 12 decimals, MSE and RMSE to 10 and SSE to 9, including the MSE with a hypothesis df of 2. ### Machine learning / 机器学习 #### Nearest Neighbor Analysis (KNN) — SPSS KNN (最近邻元素分析(KNN)) English: https://www.crosstabs.com/spss/nearest-neighbor-knn 简体中文: https://www.crosstabs.com/zh-cn/spss/nearest-neighbor-knn Nearest neighbor analysis classifies or predicts each case from the k most similar cases in a training sample. Crosstabs runs SPSS KNN on your device with adjusted normalization, Euclidean or city-block distance, k chosen by V-fold cross-validation, optional forward feature selection, and error rates on a random holdout sample. When to use: Assigning new respondents to existing segments from a few profile variables, without assuming a linear boundary. A quick, assumption-light benchmark before fitting logistic or discriminant models. Predicting a scale value (spend, score) as the mean or median of similar cases. When not to: You need to explain how each predictor changes the outcome: KNN gives no coefficients; use logistic regression or a decision tree. Many irrelevant features: distances become noisy; turn on feature selection or reduce dimensions first. Survey weights must be respected: KNN ignores case weights and says so. Very small training samples, especially when some categories have only a few cases. Validation: KNN is checked against a Python reference that finds neighbours with scikit-learn's brute-force NearestNeighbors on an independently encoded matrix and re-implements the SPSS voting, tie-breaking, Laplace probability, k-selection and feature-selection rules. Across fixed k, cross-validated k with city-block distance, both stopping rules, importance weighting and mean or median scale prediction, the neighbour sets and predicted categories are identical and probabilities and errors agree to at least 8 decimal places. #### Naive Bayes — SPSS NAIVEBAYES (朴素贝叶斯) English: https://www.crosstabs.com/spss/naive-bayes 简体中文: https://www.crosstabs.com/zh-cn/spss/naive-bayes Naive Bayes classifies cases by combining each predictor's evidence under the assumption that predictors are independent within each class. Crosstabs runs SPSS NAIVEBAYES on your device: scale predictors are cut into equal-width bins, probabilities are smoothed, partially missing predictors are kept, and forward selection can pick the best subset by pseudo-BIC or a held-out test sample. When to use: Many categorical survey answers used to predict a categorical outcome, such as renewal, churn or preferred brand. You want a fast, transparent classifier whose probability tables can be read directly. Predictors have scattered missing values and you do not want to drop whole cases. When not to: Predictors are strongly related to each other (several items measuring the same thing): the independence assumption double-counts their evidence. You need effect sizes or significance tests for predictors: use logistic or multinomial logistic regression. A scale target: Naive Bayes needs a categorical target; use KNN or regression for numeric prediction. Case weights must be applied: the panel ignores them and notes this. Validation: Posterior probabilities match scikit-learn CategoricalNB (alpha = 1/N) on complete cases to 12 decimal places. An independent NumPy implementation checks the equal-width binning with empty-bin merging, the handling of missing predictors, the priors and the forward-selection sequence under both the pseudo-BIC and the test-data criterion, with log-likelihoods and criteria agreeing to 12 decimals and the same predictors selected. #### Multilayer Perceptron — SPSS MLP (多层感知器) English: https://www.crosstabs.com/spss/multilayer-perceptron 简体中文: https://www.crosstabs.com/zh-cn/spss/multilayer-perceptron A multilayer perceptron is a feed-forward neural network that learns nonlinear relations and interactions between predictors and a categorical or scale target. Crosstabs trains it on your device following the SPSS MLP algorithm — training, testing and holdout partitions, one or two hidden layers, scaled conjugate gradient or gradient descent, the SPSS stopping rules — and reports classification tables, ROC, gains, lift and sensitivity-based predictor importance. When to use: Predicting churn, purchase or a segment where effects are nonlinear or interact and a logistic model underfits. You have enough cases to set aside testing and holdout samples — typically several hundred or more. Accuracy matters more than an equation you can explain coefficient by coefficient. When not to: You need interpretable effects or significance tests: use logistic, multinomial or linear regression. Small samples: a network with dozens of weights can easily overfit a few hundred cases; try logistic regression or a decision tree first. You must reproduce weights from SPSS exactly: SPSS's random starting values are undocumented, so the weights differ even when the settings match. Survey weights must be applied: the neural-network algorithms define no case weights, so results are unweighted. Validation: The MLP engine is checked against an independent, vectorised NumPy implementation written from the SPSS algorithm chapter that shares only the documented mulberry32 random-number protocol. Backpropagated gradients match the NumPy matrix gradients to 1e-10 and central finite differences to 1e-6 for four network shapes; five steps of scaled conjugate gradient and batch, online and mini-batch gradient descent match to 1e-10. Seven full runs (custom one- and two-layer networks, all training types, automatic architecture selection) reach the same step count and stopping rule, with final weights within 1e-6 for short runs and 1e-3 for longer ones, errors within 1e-8 (1e-6), importance within 1e-7 (1e-4), and AUC equal to scikit-learn roc_auc_score to 1e-12. Weights are not compared with SPSS, whose random stream is undocumented. #### Radial Basis Function Network — SPSS RBF (径向基函数网络) English: https://www.crosstabs.com/spss/radial-basis-function-network 简体中文: https://www.crosstabs.com/zh-cn/spss/radial-basis-function-network A radial basis function network predicts a categorical or scale target from a layer of Gaussian units, each centred on a cluster of similar training cases. Crosstabs follows the SPSS RBF procedure on your device: TwoStep clustering places the hidden units, output weights are fitted by least squares, and the number of units is chosen by testing error or BIC — no iterative training, so the same partition always gives the same network. When to use: A smooth nonlinear relation — saturation, a peak, a threshold — between a few predictors and an outcome. You want a fast, reproducible network without tuning learning rates or training steps. Comparing against a multilayer perceptron on the same partition. When not to: Many predictors: Gaussian units cover the input space poorly in high dimensions; reduce predictors first or use a multilayer perceptron. You need interpretable coefficients and tests: use linear or logistic regression. Very small training samples: the number of hidden units cannot exceed the number of training cases. Survey weights must be applied: results are unweighted. Validation: RBF is checked against an independent NumPy implementation of the SPSS RBF chapter that takes cluster assignments from the repository's separate TwoStep reference, places centres and widths from the clusters and fits output weights with numpy.linalg.lstsq; ROC areas come from scikit-learn. Across four configurations (ordinary and normalized units, categorical and scale targets, automatic, range and fixed unit counts, testing-error and BIC selection, a holdout sample and a custom overlap) the TwoStep cluster count and the candidate sequence are identical, testing errors and BIC agree to 1e-8, centres and widths to 1e-10, partition errors to 1e-8, relative errors to 1e-9, importance to 1e-9 and AUC to 1e-12. #### Predictor Selection — SPSS SELECTPRED (预测变量选择) English: https://www.crosstabs.com/spss/predictor-selection 简体中文: https://www.crosstabs.com/zh-cn/spss/predictor-selection Predictor selection narrows a long list of candidate variables before modelling. Crosstabs runs SPSS SELECTPRED on your device: it first screens out variables that are mostly missing, nearly constant or have too many categories, then ranks every remaining predictor by how well it alone predicts a categorical or scale target, and marks the top of the list as selected. When to use: Hundreds of survey or CRM variables and you need a defensible shortlist before modelling. A quick data-quality pass that flags near-empty and near-constant variables. Ranking many categorical predictors of a categorical target by Cramér's V or lambda. When not to: You need each predictor's effect adjusted for the others: ranking is one variable at a time; use a regression or automatic linear modeling. Predictors that matter only in combination (interactions) can rank low here; check them in a multivariable model. You want a final model: the list is a screening step, not a model. Case weights must be applied: the SELECTPRED algorithm defines none, so results are unweighted. Validation: SELECTPRED is checked against SciPy on a 300-case file with missing values in seven configurations — the four categorical criteria, continuous predictors of a categorical target, mixed predictors and a scale target: scipy.stats.chi2_contingency (Pearson and log-likelihood, no continuity correction), f_oneway and pearsonr, with Cramér's V and Goodman-Kruskal lambda written out and the chapter's ranking keys. The ranking order and the number of removed cases are identical, statistics agree to 10 decimal places and p values to 1e-12; the automatic list-length rule is checked separately. #### Anomaly Detection (DETECTANOMALY) — SPSS DETECTANOMALY (异常检测(DETECTANOMALY)) English: https://www.crosstabs.com/spss/anomaly-detection 简体中文: https://www.crosstabs.com/zh-cn/spss/anomaly-detection Anomaly detection flags cases that look unusual compared with similar cases. Crosstabs runs the SPSS DETECTANOMALY procedure on your device: it forms peer groups with TwoStep clustering, scores each case's deviation from its group, and lists the most anomalous cases with the variables that explain them. When to use: Auditing a survey or customer file for suspicious records before analysis. Spotting respondents whose combination of answers is rare even when each answer alone is not. When not to: You need a single variable's outliers: a box plot or the Explore procedure is simpler. You need survey or frequency weights: the SPSS algorithm defines no case weights, so results are unweighted. Validation: An independent NumPy implementation of the DETECTANOMALY chapter reproduces peer groups, BIC values, every variable deviation index, group deviation index and anomaly index within 1e-9, and the anomaly and reason lists exactly, with listwise deletion, missing-value handling and the missing-proportion variable. #### Automated Data Preparation (ADP) — SPSS ADP (自动数据准备(ADP)) English: https://www.crosstabs.com/spss/automated-data-preparation 简体中文: https://www.crosstabs.com/zh-cn/spss/automated-data-preparation Automated data preparation recommends how to clean and transform fields before modeling. Crosstabs runs the SPSS ADP algorithms on your device as a preview: it never changes your data, but shows every recommended transformation and lets you download the transformed fields. When to use: Before building a regression, tree or neural network on a messy survey or customer file. To see quickly which candidate predictors carry signal for a target. When not to: You need an ordinal target or date and time handling: these parts of SPSS ADP are not included. You need survey sampling weights: only frequency weights are supported. Validation: An independent NumPy/SciPy implementation of the ADP chapter, cross-checked with SciPy f_oneway, chi2_contingency and boxcox_llf, reproduces screening, recasting, outlier bounds, replacements, merges, Scheffé cut points, constructed features, transformed values and predictive power within 1e-9 for continuous, nominal and no target, with and without frequency weights. ### Complex samples / 复杂抽样 #### Complex Samples Plan — SPSS CSPLAN (复杂抽样计划) English: https://www.crosstabs.com/spss/complex-samples-plan 简体中文: https://www.crosstabs.com/zh-cn/spss/complex-samples-plan A sampling plan tells every Complex Samples procedure how your survey was drawn: which variable is the sampling weight, which define the strata and the clusters, and whether variances use with-replacement or without-replacement formulas. Crosstabs saves the plan for the dataset in this browser, reports the design degrees of freedom, and nothing is guessed from column names. When to use: Your data come from a stratified and/or clustered survey with a weight variable, and you want design-based standard errors instead of simple-random-sampling ones. Before running Complex Samples Descriptives, Crosstabs, GLM, Logistic, Ordinal or Cox Regression: they all read the same saved plan. The sampling fractions are large enough to matter and you know the population sizes or sampling rates per stratum, so a finite population correction applies. When not to: Your weight comes from raking or post-stratification of a simple random sample and you only need weighted tables: use the weighting and crosstab tools instead. Your weight is a frequency count of identical cases: declare it as a frequency weight in the workspace, not as a sampling weight. The design needs replicate weights (jackknife, BRR, bootstrap) or unequal-probability without-replacement variance formulas: this panel offers only Taylor linearization with WR or equal-probability WOR estimation. Validation: The plan builds the design frame that every Complex Samples procedure uses, so it is validated through them: on a synthetic two-stage stratified survey (235 cases, 4 strata, 15 PSUs) the estimates agree with the R survey package (svydesign) to a relative 1e-9 under three plans — stratified clusters with replacement, two stages without replacement with population sizes, and stratified with sampling rates. Unit tests also check that stored plans are parsed defensively, that without-replacement stages require population sizes or rates, that a single-PSU stratum adds zero variance with df = PSUs − strata, and that population sizes below the sampled count are rejected. #### Complex Samples Descriptives — SPSS CSDESCRIPTIVES (复杂抽样描述统计) English: https://www.crosstabs.com/spss/complex-samples-descriptives 简体中文: https://www.crosstabs.com/zh-cn/spss/complex-samples-descriptives Complex Samples Descriptives estimates population means, sums, ratios and population sizes from a weighted, stratified or clustered sample, with standard errors that respect the design. Crosstabs runs the SPSS CSDESCRIPTIVES procedure on your device, reporting t-based confidence intervals on the design degrees of freedom, coefficients of variation, design effects and unweighted counts, overall or by subpopulation. When to use: Reporting a survey mean or prevalence (a mean of a 0/1 variable) with an honest margin of error under clustering. Estimating population totals (households, spend, hours) by weighting up a sample. Comparing estimates across regions, age bands or other subgroups without discarding the cases outside each subgroup from the variance. When not to: Your data are a simple random sample with no weight: Descriptives or Frequencies give the same answers more directly. You want percentages of a categorical variable: use Complex Samples Crosstabs, which also tests associations. You want to test whether subgroup means differ after adjusting for other variables: use Complex Samples GLM. Validation: On a synthetic two-stage stratified survey (235 cases) under three designs — stratified clusters with replacement, two stages without replacement with population sizes, and stratified with sampling rates — totals, means, ratios and by-region domain means and totals match the R survey package (svytotal, svymean, svyratio, svyby) to a relative 1e-9. Design effects and confidence-interval half-widths are checked to the same tolerance against an independent NumPy/SciPy implementation of the SPSS formulas. #### Complex Samples Crosstabs — SPSS CSTABULATE (复杂抽样交叉表) English: https://www.crosstabs.com/spss/complex-samples-crosstabs 简体中文: https://www.crosstabs.com/zh-cn/spss/complex-samples-crosstabs Complex Samples Crosstabs estimates population counts and row, column and table percentages from a weighted, stratified or clustered survey, and tests association with a chi-square that is corrected for the design. Crosstabs runs the SPSS CSTABULATE procedure on your device: one-way frequency tables or two-way crosstabs, Rao–Scott second-order adjusted Pearson and likelihood-ratio F tests, and odds ratios, relative risks and risk differences for 2×2 tables. When to use: Weighted survey percentages for a single question, such as vote intention or service use, with margins of error that reflect clustering. Testing whether an outcome differs between groups in a stratified or clustered survey, where an ordinary chi-square would be too liberal. Comparing the risk of a binary outcome between two groups in survey data, with design-based odds ratios and risk differences. When not to: Banner tables with many columns, significance letters and export to a report: build them in Analyze mode, where saved tables can also use a complex design. More than two variables at once (layers or control variables): CSTABULATE here crosses one row and one optional column variable. Adjusting an association for other predictors: use Complex Samples Logistic Regression. Validation: On a synthetic two-stage stratified survey (235 cases) under three designs, population sizes, cell totals, table and row proportions and their standard errors match the R survey package (svytotal, svymean, svyby) to a relative 1e-9; the two-way Rao–Scott F matches svychisq (statistic to 1e-9, p to 1e-8), and the 2×2 odds ratio, relative risks and risk differences match svycontrast. Design effects, logit intervals, the one-way homogeneity test and the adjusted likelihood-ratio test are checked against an independent NumPy/SciPy implementation of the SPSS formulas. #### Complex Samples General Linear Model — SPSS CSGLM (复杂抽样一般线性模型) English: https://www.crosstabs.com/spss/complex-samples-glm 简体中文: https://www.crosstabs.com/zh-cn/spss/complex-samples-glm Complex Samples GLM fits a linear model to a scale outcome from a weighted, stratified or clustered survey, with coefficients that describe the population and standard errors that respect the design. Crosstabs runs the SPSS CSGLM procedure on your device: weighted least-squares estimates, Taylor-linearized (Binder) standard errors, Fellegi's Wald F and Rao–Scott adjusted F tests on the design degrees of freedom, and design effects for every parameter. When to use: A scale outcome (spend, hours, a score) modelled on several predictors in survey data with strata, clusters and weights. Comparing group means after adjusting for covariates (ANCOVA) in a national or multi-stage survey. Checking whether a regression result from unweighted OLS still holds once the design is taken into account. When not to: A binary, nominal or ordinal outcome: use Complex Samples Logistic or Ordinal Regression. You need interaction terms, estimated marginal means or custom contrasts: this panel fits main effects only, with parameter tests. Data from a simple random sample with no weights: Linear Regression or GLM Univariate gives the same estimates with more diagnostics. Validation: On a synthetic two-stage stratified survey (235 cases) under three designs, coefficients and standard errors for a model with two factors and a covariate match the R survey package's svyglm to a relative 1e-9, and the Wald chi-square for a factor matches R's. Design effects, R², Fellegi's Wald F and the Rao–Scott adjusted tests (including the corrected model) are checked against an independent NumPy/SciPy implementation of the SPSS formulas to 1e-8. #### Complex Samples Logistic Regression — SPSS CSLOGISTIC (复杂抽样 Logistic 回归) English: https://www.crosstabs.com/spss/complex-samples-logistic 简体中文: https://www.crosstabs.com/zh-cn/spss/complex-samples-logistic Complex Samples Logistic Regression models a categorical survey outcome — binary or multinomial — on factors and covariates, with odds ratios that describe the population and standard errors that respect strata, clusters and weights. Crosstabs runs the SPSS CSLOGISTIC procedure on your device by pseudo-maximum likelihood, with design-based (Binder) variances, Wald tests on the design degrees of freedom, three pseudo R² measures and a weighted classification table. When to use: A yes/no survey outcome (owns a home, voted, has a condition) with predictors, in data with strata, clusters and weights. An unordered outcome with three or more categories (employment status, preferred brand): the model becomes baseline-category multinomial logistic. Reporting adjusted odds ratios from a national survey, where unweighted logistic regression would misstate both estimates and errors. When not to: An ordered outcome such as a five-point rating: Complex Samples Ordinal Regression uses the ordering and needs fewer parameters. A crude 2×2 comparison with no adjustment: Complex Samples Crosstabs gives the odds ratio and risks directly. You need stepwise selection, interactions or ROC-based diagnostics: this panel fits a main-effects model you specify. Validation: On a synthetic two-stage stratified survey (235 cases) under three designs, the binary model's coefficients, standard errors and odds ratios match the R survey package's svyglm (quasibinomial) to a relative 1e-8, and its pseudo −2 log likelihoods and pseudo R² match an independent implementation of the SPSS formulas. The multinomial model is checked against an independent NumPy/SciPy pseudo-likelihood fit and sandwich variance: coefficients to 1e-8, standard errors, design effects and Wald tests to 1e-7. #### Complex Samples Ordinal Regression — SPSS CSORDINAL (复杂抽样有序回归) English: https://www.crosstabs.com/spss/complex-samples-ordinal 简体中文: https://www.crosstabs.com/zh-cn/spss/complex-samples-ordinal Complex Samples Ordinal Regression models an ordered survey outcome, such as a satisfaction rating, with a cumulative link model whose thresholds and coefficients describe the population. Crosstabs runs the SPSS CSORDINAL procedure on your device, with logit, probit, complementary log-log, negative log-log or Cauchit links, pseudo-likelihood estimates, design-based standard errors and Wald tests on the design degrees of freedom. When to use: Likert-type ratings, satisfaction scales or ordered bands (income bracket, severity) as the outcome in a weighted, clustered survey. You want one coefficient per predictor rather than a separate set for every category, as multinomial logistic would give. Comparing link functions: probit for a normally distributed latent variable, complementary log-log when higher categories are more likely, negative log-log when lower categories are. When not to: The categories have no natural order: use Complex Samples Logistic Regression (multinomial). You need a test of parallel lines (proportional odds): it is not offered here; compare the category-specific coefficients of a multinomial fit instead. A rating with many points that behaves like a scale: Complex Samples GLM may be simpler to interpret. Validation: On a synthetic two-stage stratified survey (235 cases) under three designs, all five links are checked against an independent full-precision NumPy/SciPy cumulative link fit with sandwich variance: coefficients to a relative 1e-7, standard errors and design effects to 1e-6, −2 log likelihoods to 1e-10. Three links are also compared with the R survey package's svyolr: estimates agree to 1e-7 and standard errors to 2e-3, because svyolr uses a numerical Hessian. #### Complex Samples Cox Regression — SPSS CSCOXREG (复杂抽样 Cox 回归) English: https://www.crosstabs.com/spss/complex-samples-cox-regression 简体中文: https://www.crosstabs.com/zh-cn/spss/complex-samples-cox-regression Complex Samples Cox Regression models time to an event — leaving a programme, a first diagnosis, a customer cancelling — in survey data with strata, clusters and weights. Crosstabs runs the SPSS CSCOXREG procedure on your device: a weighted pseudo partial likelihood with Efron or Breslow ties, hazard ratios with intervals on the design degrees of freedom, and a robust design-based variance from weighted score residuals. When to use: Survey or panel data with a follow-up time and an event indicator, drawn with strata, clusters and unequal weights. Population-level hazard ratios are the goal, for example from a national health survey linked to later outcomes. Many tied event times (months, years): pick the Efron approximation, which handles ties more accurately than Breslow. When not to: Data from a simple random sample or a trial: Cox Regression offers more diagnostics and plots. You need survival curves, time-dependent covariates, late entry or a stratified baseline hazard: each record here is one subject entering at time 0, and only coefficient tables are produced. You only want to describe survival without predictors: use Kaplan-Meier or Life Tables. Validation: On a synthetic two-stage stratified survey (235 cases) under three designs, coefficients, robust standard errors and hazard ratios for a model with a factor and a covariate match the R survey package's svycoxph to a relative 1e-9, with both Breslow and Efron ties, on continuous survival times and on heavily tied ones. #### Select a Complex Sample — SPSS CSSELECT (选择复杂样本) English: https://www.crosstabs.com/spss/complex-sample-selection 简体中文: https://www.crosstabs.com/zh-cn/spss/complex-sample-selection Select a Complex Sample treats your dataset as a sampling frame and draws a probability sample from it: simple random, systematic or probability proportional to size (PPS), within strata and by clusters. Crosstabs runs the SPSS CSSELECT procedure on your device with a seed, reports each selected case's inclusion probability and sampling weight, and offers the sample as a CSV download without changing your data. When to use: Planning fieldwork: drawing schools, villages, stores or households from a list, with strata to guarantee coverage of each region. Selecting larger units with higher probability (PPS) using a size measure such as enrollment or population. Teaching or simulating sampling designs reproducibly: the same seed always gives the same sample here. When not to: You only need a random subset of cases: Random Sample of Cases is simpler. A multi-stage selection in one run (PSUs, then units inside them): this panel draws a single stage; run it again on the selected clusters' list for a second stage. You need to reproduce a specific SPSS-drawn sample: the random number stream differs, so a sample drawn here will differ from SPSS's even with the same seed. Validation: Unit tests draw from a stratified 40-unit frame and check that each seed reproduces the same sample, that without-replacement methods return exactly the requested sizes, and that inclusion probabilities and weights follow the stated formulas. For all six methods, 4,000 seeded draws give empirical selection frequencies within five standard errors of each unit's stated inclusion probability. Samples are not compared with SPSS, whose random stream differs. ### Utilities and charts / 实用工具与图表 #### Bootstrapping — SPSS BOOTSTRAP (Bootstrap(自助法)) English: https://www.crosstabs.com/spss/bootstrapping 简体中文: https://www.crosstabs.com/zh-cn/spss/bootstrapping Bootstrapping resamples your cases with replacement many times and reads the bias, standard error and confidence interval of a statistic from how it varies across resamples. Crosstabs runs the SPSS BOOTSTRAP approach on your device for descriptives, Pearson correlations and linear regression coefficients, with percentile or BCa intervals, optional strata and a reproducible seed. When to use: Skewed data such as spend or time, where a normal-theory interval for the mean is doubtful. Medians and standard deviations, whose sampling distributions have no convenient closed form. Checking that a regression or correlation result does not depend on normality or equal-variance assumptions. When not to: Very small samples (say, under 20 cases): resamples can only recombine the few values you have, so intervals are too narrow. Survey data with a sampling weight, strata and clusters: case resampling ignores the design and is blocked; use Complex Samples. Time series or other dependent observations, where resampling single cases breaks the dependence. Validation: The independent references use a Python port of the same seeded generator, so resample indices are identical. On 1,000 resamples, bias and standard errors match to 1e-9 and percentile and BCa limits to 1e-10 against an independent implementation of the SPSS rules ((B + 1)-order statistics, jackknife acceleration), and scipy.stats.bootstrap fed the same resamples agrees to 1e-9. Regression coefficients, stratified resampling and frequency weights (against expanded data) are also tested. #### Rank Cases — SPSS RANK (个案排秩) English: https://www.crosstabs.com/spss/rank-cases 简体中文: https://www.crosstabs.com/zh-cn/spss/rank-cases Rank Cases turns a numeric variable into ranks and rank-based scores as new variables. Crosstabs implements the SPSS RANK procedure: eight rank types, four tie rules, Blom, Tukey, Rankit or Van der Waerden proportion estimates, optional ranking within groups and frequency weights, all as a preview and download that never changes your data. When to use: Creating quartile or decile groups from a scale variable, for example spend tiers. A rank-based inverse normal transformation of a skewed variable before a parametric analysis. Percentile ranks within groups, such as a student's standing within their class. When not to: To test differences between groups on ranks: the Mann-Whitney U and Kruskal-Wallis tests rank the data for you. To cut a variable at points that best predict an outcome: use Optimal Binning. Validation: RANK is checked against an independent implementation of the SPSS RANK formulas cross-checked with scipy.stats.rankdata. All eight rank functions agree to 1e-12 for every reference case, covering weighted and unweighted data, ascending and descending order, the four tie rules and the four proportion formulas. #### Optimal Binning — SPSS OPTIMAL BINNING (最优分箱) English: https://www.crosstabs.com/spss/optimal-binning 简体中文: https://www.crosstabs.com/zh-cn/spss/optimal-binning Optimal Binning cuts a scale variable into bins that best separate the categories of a guide variable, such as default versus repaid. Crosstabs implements the SPSS OPTIMAL BINNING procedure with supervised MDLP (Fayyad and Irani), optional merging of sparse bins, and an unsupervised equal-frequency alternative. When to use: Scorecards and risk models where a continuous predictor should become a few interpretable bands. Preparing a predictor for a decision tree, Naive Bayes or crosstab that needs categories. Equal-frequency groups when there is no outcome to guide the cuts. When not to: When you will test the binned variable against the same outcome on the same data: the cuts were chosen to fit it, so p values are too optimistic. When the continuous variable itself fits the model well: binning throws information away. Validation: Cut points are checked against an independent Python reference, which contains a separate MDLP implementation (Fayyad and Irani) and the worked example printed in the SPSS OPTIMAL BINNING documentation. Cut points match to 1e-12, and bin counts per category and model entropy match to 1e-12, for the manual example, four random datasets and equal-frequency binning with 3, 4 and 7 bins. #### Sample Cases — SPSS SAMPLE (个案抽样) English: https://www.crosstabs.com/spss/random-sample-cases 简体中文: https://www.crosstabs.com/zh-cn/spss/random-sample-cases Sample Cases selects a random subset of rows, either keeping each case with a given probability or taking exactly n cases from the first N. Crosstabs follows the SPSS SAMPLE algorithm with a visible seed, so the same seed always returns the same rows, and the sample downloads as a CSV. When to use: A holdout or validation subset for checking a model built on the rest. A random subsample for manual review, coding open ends or quality checks. Cutting a very large file down to a quick working sample. When not to: Stratified, cluster or PPS sampling for a survey design: use Complex Sample Selection. Weighted selection: SAMPLE ignores case weights, so every row has the same chance. Validation: SAMPLE has no external fixture; its tests check the algorithm's properties. Exact sampling returns exactly n cases, never selects beyond the first N and repeats for a seed. Over 2,000 seeds, each of 20 cases is chosen with probability within 5 standard errors of 0.25 for 5-of-20 sampling, and a 30% proportion on 100,000 cases lands within 0.006 of 0.30. #### Error Bars — SPSS GRAPH /ERRORBAR (误差条) English: https://www.crosstabs.com/spss/error-bar-chart 简体中文: https://www.crosstabs.com/zh-cn/spss/error-bar-chart An error bar chart shows each group's mean with a bar for its uncertainty or spread. Crosstabs draws SPSS-style error bars for a confidence interval of the mean, a multiple of the standard error, or a multiple of the standard deviation, and lists the numbers behind every point. When to use: Presenting group means from a survey or experiment alongside an ANOVA or t test. A quick visual check of which group means are estimated precisely and which rest on few cases. When not to: As a significance test: overlapping 95% bars do not prove means are equal. Run One-Way ANOVA or a t test. Survey data with a sampling weight when you need intervals: CI and SE bars are withheld; use Complex Samples Descriptives. Heavily skewed data, where a mean ± SD bar can run below the minimum; show medians in Explore instead. Validation: Error-bar statistics are checked against an independent Python reference: weighted mean, SD and SE agree to 1e-12 and confidence limits, computed with scipy.stats.t quantiles, agree to 1e-10 at each tested level. The engine's median and its Hettmansperger–Sheather interval are also checked against scipy.stats.binom to 1e-10, though the panel currently charts means only. #### P-P and Q-Q Plots — SPSS PPLOT (P-P 图与 Q-Q 图) English: https://www.crosstabs.com/spss/p-p-q-q-plots 简体中文: https://www.crosstabs.com/zh-cn/spss/p-p-q-q-plots P-P and Q-Q plots show whether a variable follows a chosen distribution: points on the diagonal mean a good fit. Crosstabs follows SPSS PPLOT, estimating the distribution's parameters from the data and plotting against the normal, lognormal, exponential, uniform, logistic, Laplace, Weibull, gamma, beta, Pareto, chi-square or Student's t. When to use: Checking normality of a variable or of saved residuals before a t test, ANOVA or regression. Choosing a distribution for a simulation input, such as lognormal for durations or Weibull for lifetimes. When not to: When you need a formal test with a p value: use the Kolmogorov-Smirnov or Shapiro-Wilk tests in Explore or One-Sample Nonparametric Tests. Weighted data: PPLOT does not use case weights, so the plot shows the unweighted values. Validation: PPLOT is checked against an independent Python reference, which implements the SPSS fractional-rank formulas and parameter estimates and takes quantiles and CDFs from scipy.stats for all 12 distributions. Estimated parameters agree to 1e-8, fractional ranks to 1e-12, expected values to 1e-7 and cumulative probabilities to 1e-8. #### Control Charts — SPSS SPCHART (控制图) English: https://www.crosstabs.com/spss/control-charts 简体中文: https://www.crosstabs.com/zh-cn/spss/control-charts Control charts plot a process measure over time against a centre line and control limits, so you can tell routine variation from a real change. Crosstabs follows SPSS SPCHART for variables charts (X-bar/R, X-bar/s, I-MR) and attribute charts (p, np, c, u), and checks every point against 11 run and zone rules. When to use: Monitoring a measured quality characteristic (fill weight, cycle time) sampled in subgroups or one at a time. Tracking defect rates or counts per batch, day or unit with p, np, c or u charts. Checking whether a change to a process produced a real shift rather than noise. When not to: Data with strong autocorrelation or seasonality: limits will be wrong; model the series with time-series tools. Comparing a few fixed groups once: that is a job for ANOVA or a chi-square test, not a control chart. Validation: SPCHART is checked against an independent Python reference: the d2 and d3 constants, computed by numerical integration in scipy.integrate, agree to 1e-8 and 1e-7, and c4 from scipy.special.gamma to 1e-12, all cross-checked against the published three-decimal tables. Limits for X-bar/R and X-bar/s with unequal subgroups, I-MR, p, np, c and u charts agree to between 1e-8 and 1e-10. Rule detection is covered by unit tests. #### Simulation — SPSS SIMPLAN / SIMRUN (模拟) English: https://www.crosstabs.com/spss/monte-carlo-simulation 简体中文: https://www.crosstabs.com/zh-cn/spss/monte-carlo-simulation Monte Carlo simulation draws thousands of random values for uncertain inputs, evaluates a formula for each draw and shows the spread of possible outcomes. Crosstabs runs the simulation step of SPSS SIMULATION on your device: eleven input distributions, a safe expression language, percentiles, a histogram and SPSS's three sensitivity measures. When to use: Budget, pricing or capacity questions where several inputs are uncertain and the formula is known. Communicating risk as a range instead of a single best guess. Deciding which input is worth measuring more precisely. When not to: Inputs that are strongly correlated with each other: inputs are drawn independently here, which can understate or overstate the spread. Fitting distributions to a dataset or simulating from a fitted predictive model: SPSS SIMPLAN can do that, this panel works from distributions you type in. Validation: Simulation has no external fixture; its tests check known answers. With 100,000 iterations of a + 2b (a normal, b uniform) the mean falls within 5 Monte Carlo standard errors of 16, the SD within 0.05 of 4, the correlations within 0.01 of their theoretical values and the one-at-a-time changes match exactly to 1e-9. Each sampler's mean and variance are checked on 200,000 draws, runs are reproducible by seed, and the expression parser rejects anything outside its grammar. ## Browser-local market-research tools Free survey tools that run entirely in the visitor's browser: files are parsed in the tab and never uploaded, and no account is needed. Each page has a synthetic sample dataset, the method and formula, a worked example computed from that sample, and references. They are web pages, not API or MCP endpoints. - [Survey Data Cleaner — Speeders, Straight-liners, Duplicates, Fraud Checks](https://www.crosstabs.com/tools/data-cleaner): Flag speeders, straight-liners, duplicates, gibberish open ends, near-duplicates, repeated IPs and failed attention checks, then export a cleaned file. Simplified Chinese: https://www.crosstabs.com/zh-cn/data-cleaner - [TURF Analysis Calculator — Reach and Frequency](https://www.crosstabs.com/tools/turf-analysis): Find the combination of products, flavors or features that reaches the most respondents, with incremental reach ladders. Simplified Chinese: https://www.crosstabs.com/zh-cn/turf-analysis - [Penalty Analysis — Just-About-Right (JAR) Scales](https://www.crosstabs.com/tools/penalty-analysis): Mean drops and weighted penalties for too-little and too-much JAR responses, with the 20% action threshold. Simplified Chinese: https://www.crosstabs.com/zh-cn/penalty-analysis - [Key Driver Analysis — Relative Importance (Shapley, Johnson)](https://www.crosstabs.com/tools/relative-importance-analysis): Decompose R² across correlated drivers with Shapley values and Johnson's relative weights, plus a quadrant map. Simplified Chinese: https://www.crosstabs.com/zh-cn/relative-importance-analysis - [NPS Calculator — Net Promoter Score with Margin of Error](https://www.crosstabs.com/tools/nps-calculator): Net Promoter Score with a confidence interval, a two-group significance test, and NPS by segment from your data. Simplified Chinese: https://www.crosstabs.com/zh-cn/nps-calculator - [Survey Sample Size Calculator — Margin of Error](https://www.crosstabs.com/tools/sample-size-calculator): Sample size for a target margin of error, finite population correction, design effect, and margin of error from n. Simplified Chinese: https://www.crosstabs.com/zh-cn/sample-size-calculator - [Survey Label Mapper — Apply Value Labels to Codes](https://www.crosstabs.com/tools/label-mapper): Turn numeric codes into readable labels, merge categories, and export a labeled file ready for crosstabs. Simplified Chinese: https://www.crosstabs.com/zh-cn/label-mapper - [Segment Builder — K-Means Survey Segmentation](https://www.crosstabs.com/tools/segment-builder): Cluster respondents on standardized attitudes, compare solutions by silhouette, and profile each segment. Simplified Chinese: https://www.crosstabs.com/zh-cn/segment-builder - [Latent Class Analysis — Segment Categorical Survey Data](https://www.crosstabs.com/tools/latent-class-analysis): Segment respondents on categorical, ordinal and numeric answers with latent class models, compare k by BIC, and export a typing tool. Simplified Chinese: https://www.crosstabs.com/zh-cn/latent-class-analysis - [Van Westendorp Calculator — Price Sensitivity Meter](https://www.crosstabs.com/tools/van-westendorp): Price sensitivity curves, the acceptable price range, OPP and IPP from respondent data, with the Newton-Miller-Smith revenue extension. Simplified Chinese: https://www.crosstabs.com/zh-cn/van-westendorp - [Gabor-Granger Calculator — Demand and Revenue by Price](https://www.crosstabs.com/tools/gabor-granger): Demand curve, revenue index, revenue-maximising price and price elasticity from purchase intent at each tested price. Simplified Chinese: https://www.crosstabs.com/zh-cn/gabor-granger - [MaxDiff Analysis — Best-Worst Scaling with HB Scores](https://www.crosstabs.com/tools/maxdiff-analysis): Counts, aggregate logit and individual hierarchical Bayes utilities from best-worst data, rescaled to 0–100, with a share simulator. Simplified Chinese: https://www.crosstabs.com/zh-cn/maxdiff-analysis - [Choice-Based Conjoint Analysis — Part-Worths, HB and Simulator](https://www.crosstabs.com/tools/conjoint-analysis): Effects-coded logit and hierarchical Bayes part-worths, attribute importance, willingness to pay and a market share simulator. Simplified Chinese: https://www.crosstabs.com/zh-cn/conjoint-analysis - [CSAT and CES Calculator — Satisfaction and Effort Scores](https://www.crosstabs.com/tools/csat-ces-calculator): Top-2-box CSAT and Customer Effort Score with Wilson and t intervals, two-group significance tests, and results by segment from your data. Simplified Chinese: https://www.crosstabs.com/zh-cn/csat-ces-calculator - [Kano Model Analysis — Categories and CS+/CS− Coefficients](https://www.crosstabs.com/tools/kano-analysis): Classify features as attractive, one-dimensional, must-be or indifferent, with Fong's test, CS+/CS− coefficients and a quadrant plot. Simplified Chinese: https://www.crosstabs.com/zh-cn/kano-analysis - [Brand Funnel Analysis — Conversion and Leak Finder](https://www.crosstabs.com/tools/brand-funnel): Stage percentages, stage-to-stage conversion with confidence intervals, tests against competitors and each brand's biggest leak. Simplified Chinese: https://www.crosstabs.com/zh-cn/brand-funnel - [CBC Conjoint Design Generator — D-Efficiency](https://www.crosstabs.com/tools/conjoint-design): Minimal or balanced overlap CBC designs with prohibitions, versions and holdouts, scored by D-efficiency, with expected standard errors. Simplified Chinese: https://www.crosstabs.com/zh-cn/conjoint-design - [MaxDiff Design Generator — Balanced Best-Worst Sets](https://www.crosstabs.com/tools/maxdiff-design): Balanced best-worst sets and versions with item, pair and position checks, prohibited pairs and a data template for analysis. Simplified Chinese: https://www.crosstabs.com/zh-cn/maxdiff-design - [Open-End Word Cloud & Word Frequency Analysis](https://www.crosstabs.com/tools/word-cloud): Word frequencies and a word cloud from open-ended answers in any language, with keyword-in-context and distinctive words by segment. Simplified Chinese: https://www.crosstabs.com/zh-cn/word-cloud - [Brand Perceptual Map — Correspondence Analysis](https://www.crosstabs.com/tools/perceptual-map): Correspondence analysis of a brand × attribute grid: symmetric and asymmetric biplots, inertia, contributions, cos² and supplementary points. Simplified Chinese: https://www.crosstabs.com/zh-cn/perceptual-map All tools: https://www.crosstabs.com/tools (Simplified Chinese: https://www.crosstabs.com/zh-cn/tools) - [Crosstabs for Researchers](https://www.crosstabs.com/solutions): How agency analysts, independent researchers and graduate students use Crosstabs for free survey analysis: crosstabs, tab books, key drivers and segments. - [Crosstabs for Market Research Agency Analysts](https://www.crosstabs.com/solutions/agency-analysts): Banner tables, weighting, significance letters and editable reports for client deliverables, without uploading client data. - [Crosstabs for Independent Researchers and Consultants](https://www.crosstabs.com/solutions/independent-researchers): A free, full survey-analysis toolkit for freelancers and small consultancies: tab books, TURF, key drivers and segmentation, all in your browser. - [Crosstabs for Graduate Students](https://www.crosstabs.com/solutions/graduate-students): Free thesis and dissertation survey analysis: open SPSS files, run crosstabs and significance tests, and cite the method. No install, no account. - [Forsta Alternative — Free Browser Survey Tabulation](https://www.crosstabs.com/forsta-alternative): Where a free, browser-local tool can replace Forsta-style survey tabulation workflows such as banner tables, tests and exports, and where it cannot. - [Supported Import Formats](https://www.crosstabs.com/help/import-formats): CSV, TSV, Excel, SPSS, Stata, SAS, R, Parquet, Arrow, SQLite, JSON and Triple-S import: encodings, headers, value labels, limits and troubleshooting. - [Export Formats](https://www.crosstabs.com/help/export-formats): CSV, XLSX, PDF, PNG, SVG, editable PPTX and DOCX exports and SPSS .sav data files: what each contains and when to use it. - [In-Browser Performance Benchmark](https://www.crosstabs.com/benchmark): Run a live crosstab benchmark on synthetic survey data in your own browser and see how long each step takes on your device. Nothing is uploaded. - [Browser Support](https://www.crosstabs.com/browser-support): Supported browsers and devices for Crosstabs, how each engine is tested, and which features depend on browser capabilities such as workers and IndexedDB. ## Public API and machine-readable discovery Public REST v1 compatibility policy: breaking changes use a new major URL prefix rather than silently changing /api/v1. Additive response fields may be introduced; clients should tolerate unknown fields. Current v1 endpoints are not deprecated and have no scheduled sunset. If a deprecation is scheduled, the public changelog and developer documentation will publish the affected endpoints, migration guidance and planned dates. This policy is not an availability SLA or a promise of indefinite service. Published changes: https://www.crosstabs.com/changelog. Migration guidance: https://www.crosstabs.com/developers#compatibility when a deprecation is scheduled. Unknown /api paths return HTTP 404 JSON with error, code=endpoint_not_found, hint and documentation; HEAD has no body. REST errors include error, code and hint. GET/POST responses expose RateLimit-Limit, RateLimit-Remaining and RateLimit-Reset (seconds until the next sliding-window slot, not an epoch or full-quota reset). HTTP 429 adds Retry-After seconds. Limits are shared across statistics routes within one process; serverless counters are not global and reset on cold starts. - [OpenAPI 3.1 specification](https://www.crosstabs.com/openapi.json): Existing GET/POST aggregate contingency, chi-square, Fisher exact and odds-ratio endpoints, plus GET/HEAD health. No API key. Integer frequency counts only: 2–12 rows and columns, positive margins, grand total at most 1,000,000,000,000; POST body at most 16,384 bytes. Consult the specification for error shapes and unavailable-result conventions. - [Developer documentation](https://www.crosstabs.com/developers): REST examples, the public aggregate MCP, and local MCP/CLI setup. - [Aggregate-only MCP](https://www.crosstabs.com/api/headless-mcp): Five tools for public evidence and bounded counts; no respondent uploads or private project storage. Initialize MCP and call tools/list for schemas. - [Local package](https://pypi.org/project/crosstabs/1.5.3/): Install with uv tool install --python 3.12 crosstabs==1.5.3 (uv brings its own Python; or pipx install crosstabs==1.5.3 with Python 3.10+, or run uvx crosstabs). crosstabs-headless also needs Node.js 22+. Add crosstabs-headless and crosstabs as stdio MCP commands using their full paths (uv tool dir --bin prints the folder). The project server reads only the folders listed in CROSSTABS_MCP_ALLOWED_ROOTS (set it in that server's env in the client configuration) plus a private exchange folder; get_runtime_status reports them. This is a published package, not a browser upload API. - Statistics procedures (since crosstabs 1.5.0): the local headless MCP tools list_statistical_procedures and run_statistical_procedure run 110 SPSS-equivalent Statistics procedures on a local project file with the same engines as the browser Statistics mode. Respondent rows stay on the device. Weights must be a saved frequency weight (ordinary inference) or sampling weight (weighted estimates only; ordinary tests withheld); untyped weights are rejected. Complex Samples procedures (CSDESCRIPTIVES, CSTABULATE, CSGLM, CSLOGISTIC, CSORDINAL, CSCOXREG) run only on a survey design approved through define_survey_design. Seeded procedures report their seed; every result cites its reference-parity tests. - Local data pulls (crosstabs 1.5 or later, run on the user's own machine): crosstabs pull runs one read-only SELECT/WITH query (database drivers as extras, e.g. pip install "crosstabs[postgres]>=1.5") and crosstabs pull fetches responses from Qualtrics, SurveyMonkey, Medallia, Alchemer, Typeform, Google Forms, Forsta/Decipher, QuestionPro, Jotform, SurveySparrow, REDCap, KoboToolbox, ODK Central or LimeSurvey (pip install "crosstabs[surveys]>=1.5"); both write a labelled .sav/.dta/.csv/.xlsx/.parquet file to open in the browser workspace and never apply or guess weights. crosstabs connect starts a loopback-only local bridge that hands a result to the workspace tab that opened it, paired with a one-time code. Credentials stay on the user's machine and rows never reach a Crosstabs server. Details: https://www.crosstabs.com/developers#local-data ## Trust, support, and review status Named statistical review is pending. crosstabs.com does not claim that an anonymous methods team reviewed the calculations. Selected automated SciPy parity cases are public, but they are not independent human review or blanket software certification. Methods and assumptions: https://www.crosstabs.com/methods Selected validation evidence: https://www.crosstabs.com/methods/validation Machine-readable reference cases: https://www.crosstabs.com/validation/statistical-reference-cases.json Support and corrections: support@crosstabs.com (https://www.crosstabs.com/support) Live self-check: https://www.crosstabs.com/status Health JSON: https://www.crosstabs.com/api/health ## Sharing aggregate crosstabs The workspace can create immutable, view-only snapshot links for review and citation. Snapshots contain aggregate counts, labels, visible display settings, warnings, and optional provenance; source rows and the original filename are not included. The payload is stored in the URL fragment rather than sent in the HTTP request. Anyone with the complete link can read it, and this version has no expiry or revocation control. Sharing guide: https://www.crosstabs.com/share-crosstabs Shared viewer: https://www.crosstabs.com/s ## Crosstab Atlas The Crosstab Atlas is a reproducible collection of sourced and clearly labeled illustrative contingency tables. Each record links aggregate counts, provenance, statistical analysis, effect size, assumptions, and interpretation. ### Titanic survival by sex Question: Was survival on the Titanic associated with sex? Takeaway: Females survived at a much higher rate than males in the aggregate passenger and crew counts. Table: Sex by Outcome; 2 by 2; N = 2201. Provenance: British Board of Trade inquiry (1912), as distributed in R's Titanic dataset. Analysis: https://www.crosstabs.com/examples/titanic-survival-by-sex CSV: https://www.crosstabs.com/examples/titanic-survival-by-sex/download/csv JSON with provenance: https://www.crosstabs.com/examples/titanic-survival-by-sex/download/json ### Titanic survival by passenger class Question: Was survival on the Titanic associated with passenger class? Takeaway: Survival was highest in first class and fell sharply across second class, third class, and crew. Table: Passenger class by Outcome; 4 by 2; N = 2201. Provenance: British Board of Trade inquiry (1912), as distributed in R's Titanic dataset. Analysis: https://www.crosstabs.com/examples/titanic-survival-by-class CSV: https://www.crosstabs.com/examples/titanic-survival-by-class/download/csv JSON with provenance: https://www.crosstabs.com/examples/titanic-survival-by-class/download/json ### Berkeley admissions by sex (1973) Question: Were aggregate Berkeley admission decisions associated with sex? Takeaway: The pooled association is strong evidence of Simpson's paradox, not a stand-alone causal estimate of bias. Table: Sex by Admission; 2 by 2; N = 4526. Provenance: Bickel, P. J., Hammel, E. A., & O'Connell, J. W. (1975). Sex bias in graduate admissions: Data from Berkeley. Science, 187(4175), 398–404. Analysis: https://www.crosstabs.com/examples/berkeley-admissions-by-sex CSV: https://www.crosstabs.com/examples/berkeley-admissions-by-sex/download/csv JSON with provenance: https://www.crosstabs.com/examples/berkeley-admissions-by-sex/download/json ### Smoking and lung cancer (Doll & Hill, 1950) Question: Did smoking status differ between lung-cancer cases and controls? Takeaway: Smoking was far more common among cases, but the case-control design supports an odds ratio rather than a risk estimate. Table: Patient group by Smoking status; 2 by 2; N = 1298. Provenance: Doll, R., & Hill, A. B. (1950). Smoking and carcinoma of the lung. British Medical Journal, 2(4682), 739–748. Analysis: https://www.crosstabs.com/examples/smoking-and-lung-cancer-doll-hill CSV: https://www.crosstabs.com/examples/smoking-and-lung-cancer-doll-hill/download/csv JSON with provenance: https://www.crosstabs.com/examples/smoking-and-lung-cancer-doll-hill/download/json ### Aspirin and heart attack (Physicians' Health Study) Question: Did low-dose aspirin reduce first heart attacks versus placebo? Takeaway: Heart attacks were less frequent in the randomized aspirin arm, although the absolute difference was small. Table: Treatment by Myocardial infarction; 2 by 2; N = 22071. Provenance: Steering Committee of the Physicians' Health Study Research Group (1989). Final report on the aspirin component of the ongoing Physicians' Health Study. New England Journal of Medicine, 321(3), 129–135. Analysis: https://www.crosstabs.com/examples/aspirin-and-heart-attack CSV: https://www.crosstabs.com/examples/aspirin-and-heart-attack/download/csv JSON with provenance: https://www.crosstabs.com/examples/aspirin-and-heart-attack/download/json ### Snoring and heart disease Question: Was snoring frequency associated with heart disease? Takeaway: Heart-disease prevalence rose with snoring frequency, but the observational table cannot establish causation. Table: Snoring frequency by Heart disease; 4 by 2; N = 2484. Provenance: Norton, P. G., & Dunn, E. V. (1985). Snoring as a risk factor for disease. British Medical Journal, 291, 630–632. Reproduced in Agresti, A., Categorical Data Analysis. Analysis: https://www.crosstabs.com/examples/snoring-and-heart-disease CSV: https://www.crosstabs.com/examples/snoring-and-heart-disease/download/csv JSON with provenance: https://www.crosstabs.com/examples/snoring-and-heart-disease/download/json ### Product preference by age group Question: Does product preference differ by age group? Takeaway: In this illustrative table, preference shifts from Product A toward Product C across older age groups. Table: Age group by Preferred product; 3 by 3; N = 300. Provenance: Illustrative counts constructed for this example. Illustrative data constructed for this example — not from a real study. Analysis: https://www.crosstabs.com/examples/product-preference-by-age-group CSV: https://www.crosstabs.com/examples/product-preference-by-age-group/download/csv JSON with provenance: https://www.crosstabs.com/examples/product-preference-by-age-group/download/json ### Customer churn by billing plan Question: Is customer churn associated with billing plan? Takeaway: In this illustrative table, annual-plan customers churn less often, but plan choice is not randomized. Table: Billing plan by Outcome; 2 by 2; N = 1000. Provenance: Illustrative counts constructed for this example. Illustrative data constructed for this example — not from a real study. Analysis: https://www.crosstabs.com/examples/customer-churn-by-plan CSV: https://www.crosstabs.com/examples/customer-churn-by-plan/download/csv JSON with provenance: https://www.crosstabs.com/examples/customer-churn-by-plan/download/json ## Cross-tabulation (crosstab / contingency table) Stable ID: cross-tabulation Locale: en Topic cluster: survey_workflow A cross-tabulation shows the joint frequency distribution of two categorical variables: one variable's categories are rows, the other's are columns, and each cell contains a count or percentage. Interpretation: Use row percentages to compare column outcomes across row groups; use column percentages to compare row composition across column groups. Limitations: A crosstab describes association, not causation. Very sparse or high-cardinality tables can be unstable and difficult to interpret. Primary page: https://www.crosstabs.com/contingency-table-calculator Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation ## Chi-square test of independence Stable ID: chi-square-independence Locale: en Topic cluster: test_selection The chi-square test of independence compares observed cell counts with the counts expected if two categorical variables were unrelated. Formula: χ² = Σ (O − E)² / E; E = row total × column total / grand total; df = (rows − 1) × (columns − 1). Interpretation: A small p-value is evidence against independence; inspect effect size, cells, assumptions, multiplicity, and study design before drawing a substantive conclusion. Limitations: The p-value is asymptotic. Warnings appear if any expected count is below 1 or more than 20% of expected counts are below 5. Fisher exact coverage is limited; larger sparse tables need an appropriate exact or fixed-margin Monte Carlo method. Ordinary weighted inference is limited to a saved, validated non-negative safe-integer frequency weight with weighted test bases. Untyped, calibration, and analytic weights fail closed; sampling weights require an approved design-aware execution. Primary page: https://www.crosstabs.com/chi-square-calculator Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation Evidence: https://www.crosstabs.com/validation/statistical-reference-cases.json ## Fisher's exact test Stable ID: fisher-exact Locale: en Topic cluster: test_selection For an unweighted 2×2 table of non-negative integer counts, Fisher's exact test computes the two-sided Fisher–Irwin fixed-margin p-value by summing table probabilities no greater than the observed table's probability. Interpretation: Use it when an exact fixed-margin 2×2 test matches the design and sparse expected counts make an asymptotic approximation questionable. Limitations: Unavailable for fractional weighted cells, larger tables, or supports above 100,000 feasible fixed-margin tables. It does not solve confounding, dependence, or sampling-design problems. Primary page: https://www.crosstabs.com/fishers-exact-test Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation Evidence: https://www.crosstabs.com/validation/statistical-reference-cases.json ## Cramér's V Stable ID: cramers-v Locale: en Topic cluster: association_measure Cramér's V measures the strength of association between two categorical variables on a 0-to-1 scale derived from Pearson's chi-square statistic. Formula: V = sqrt(χ² / (n × (min(rows, columns) − 1))). Interpretation: Larger values indicate stronger association; contextual and degrees-of-freedom-aware interpretation is preferable to universal cutoffs. Limitations: It does not provide direction or identify the cells driving the association. Weighted results inherit the table's weight-design caveats. Primary page: https://www.crosstabs.com/cramers-v-calculator Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation ## Phi coefficient Stable ID: phi-coefficient Locale: en Topic cluster: association_measure Phi is a signed measure of association for a 2×2 table; its magnitude equals Cramér's V for that table. Formula: φ = (ad − bc) / sqrt((a+b)(c+d)(a+c)(b+d)). Interpretation: The sign depends on row and column ordering; the absolute value expresses strength. Limitations: Signed phi is a 2×2 measure and should not be confused with generalized sqrt(χ²/n) for larger tables. Primary page: https://www.crosstabs.com/phi-coefficient-calculator Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation ## Pearson's contingency coefficient Stable ID: pearson-contingency-coefficient Locale: en Topic cluster: association_measure Pearson's contingency coefficient is a chi-square-derived association measure whose attainable maximum depends on table size. Formula: C = sqrt(χ² / (χ² + n)). Limitations: Values from differently sized tables are not directly comparable because the maximum is below 1 and depends on shape. Primary page: https://www.crosstabs.com/contingency-coefficient-calculator Evidence: https://www.crosstabs.com/methods ## Odds ratio Stable ID: odds-ratio Locale: en Topic cluster: association_measure The odds ratio compares the odds of an outcome between two groups in a 2×2 table. Formula: OR = (a × d) / (b × c) for [[a,b],[c,d]]. Interpretation: OR = 1 means equal odds; values above or below 1 indicate higher or lower odds under the chosen orientation. Limitations: The displayed 95% interval is the large-sample Woolf log interval, not exact. A zero cell produces an uninformative [0, infinity] interval because no continuity correction is added. An odds ratio is not a risk ratio except as an approximation for rare outcomes. Primary page: https://www.crosstabs.com/odds-ratio-calculator Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/odds-ratio-vs-risk-ratio ## G-test (likelihood-ratio chi-square) Stable ID: g-test Locale: en Topic cluster: test_selection The G-test is a likelihood-ratio alternative to Pearson's chi-square statistic for a contingency table. Formula: G = 2 × Σ O × ln(O / E); df = (rows − 1) × (columns − 1). Limitations: Its p-value uses the same asymptotic chi-square reference distribution, so switching statistics does not solve sparse-cell concerns. Primary page: https://www.crosstabs.com/g-test-calculator Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation ## Goodman and Kruskal's lambda Stable ID: lambda Locale: en Topic cluster: association_measure Lambda is a proportional-reduction-in-error measure: it asks how much knowing one categorical variable reduces errors when predicting the other. Interpretation: It ranges from 0 for no error reduction to 1 for complete error reduction. Limitations: A value of zero can occur even when variables are associated if knowledge does not change the modal prediction. Primary page: https://www.crosstabs.com/goodman-kruskal-lambda-calculator Evidence: https://www.crosstabs.com/methods ## Goodman and Kruskal's gamma Stable ID: gamma Locale: en Topic cluster: association_measure Gamma measures ordinal association from concordant and discordant pairs while ignoring ties. Formula: γ = (C − D) / (C + D). Limitations: Ignoring ties can make the magnitude appear stronger when ties are common. Primary page: https://www.crosstabs.com/goodman-kruskal-gamma-calculator Evidence: https://www.crosstabs.com/methods ## Kendall's tau-b and tau-c Stable ID: kendall-tau Locale: en Topic cluster: association_measure Kendall's tau-b measures ordinal association with tie correction and is suited to square tables; tau-c adjusts the scale for rectangular tables. Formula: tau-c = 2m(C − D) / (n²(m − 1)), where m = min(rows, columns). Limitations: Category order must be substantively meaningful and specified correctly. Primary page: https://www.crosstabs.com/kendalls-tau-calculator Evidence: https://www.crosstabs.com/methods ## Somers' d Stable ID: somers-d Locale: en Topic cluster: association_measure Somers' d is an asymmetric ordinal association measure that treats one variable as the outcome and accounts for ties on the chosen side. Interpretation: Choose the direction that matches the research question; reversing outcome and predictor can change the result. Limitations: It is inappropriate when no defensible dependent-variable direction exists. Primary page: https://www.crosstabs.com/somers-d-calculator Evidence: https://www.crosstabs.com/methods ## Theil's U (uncertainty coefficient) Stable ID: theils-u Locale: en Topic cluster: association_measure Theil's U is an entropy-based, directional measure of how much knowing one variable reduces uncertainty about another. Interpretation: It ranges from 0 for no uncertainty reduction to 1 when the predictor fully determines the outcome. Limitations: Direction matters, and sparse category estimates may be unstable. Primary page: https://www.crosstabs.com/uncertainty-coefficient-calculator Evidence: https://www.crosstabs.com/methods ## Adjusted standardized residual Stable ID: adjusted-residuals Locale: en Topic cluster: association_measure An adjusted residual standardizes a cell's observed-minus-expected difference to show which cells contribute most to a table-level association. Interpretation: An absolute value above 1.96 receives a nominal two-sided 0.05 flag in the workspace. Limitations: Cellwise flags are nominal and are not adjusted for multiple comparisons. Primary page: https://www.crosstabs.com/crosstab-guide Evidence: https://www.crosstabs.com/methods ## McNemar's test for paired binary change Stable ID: mcnemar-paired-change Locale: en Topic cluster: test_selection McNemar's test evaluates marginal change in paired binary outcomes by comparing the two discordant directions in a paired 2×2 table; concordant pairs do not enter the statistic. Formula: Uncorrected χ² = (b − c)² / (b + c), df = 1; continuity-corrected χ² = max(|b − c| − 1, 0)² / (b + c); the exact two-sided result uses Binomial(b + c, 0.5). Interpretation: A small p-value indicates asymmetric movement between the two categories, not the magnitude or cause of that movement. Limitations: Observations must be paired, binary, and independent across subjects or matched pairs. The public calculator requires non-negative integer counts and does not accept percentages or survey weights. Exact binomial tails are computed for at most 10,000 discordant pairs; larger inputs are explicitly labeled as a continuity-corrected normal approximation. A significant result does not by itself establish a causal intervention effect. Primary page: https://www.crosstabs.com/mcnemar-test-calculator Evidence: https://www.crosstabs.com/which-statistical-test Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation ## t test (Welch, Student, paired, one-sample) Stable ID: t-test Locale: en Topic cluster: test_selection A t test asks whether a mean, or a difference between two means, is larger than sampling noise. Welch's independent-samples test does not assume equal variances and is the default; the paired test analyses within-pair differences; the one-sample test compares a mean with a fixed value. Formula: Welch t = (x̄₁ − x̄₂) / √(s₁²/n₁ + s₂²/n₂) with Welch–Satterthwaite df; Student t uses the pooled SD and df = n₁ + n₂ − 2; paired t = d̄ / (s_d/√n), df = n − 1. Interpretation: Report the mean difference with its confidence interval and Cohen's d or Hedges' g; a small p-value alone does not show a large or important difference. Limitations: Assumes independent observations and roughly normal data or moderate samples. Levene and Brown–Forsythe variance tests need raw data; summary-statistic input reports the SD ratio instead. Weighted survey data and complex sample designs are outside the calculator's assumptions. Primary page: https://www.crosstabs.com/t-test-calculator Evidence: https://www.crosstabs.com/spss/independent-samples-t-test Evidence: https://www.crosstabs.com/methods ## One-way ANOVA with Welch's F and Tukey HSD Stable ID: one-way-anova Locale: en Topic cluster: test_selection One-way ANOVA tests whether three or more independent group means are equal by comparing between-group and within-group variance; Welch's F drops the equal-variance assumption, Tukey's HSD identifies which pairs differ, and η² or ω² measure the share of variance explained. Formula: F = MS_between / MS_within with df (k − 1, N − k); η² = SS_between / SS_total; ω² = (SS_between − (k − 1)MS_within) / (SS_total + MS_within). Interpretation: A significant F means not all means are equal; Tukey's adjusted p-values and intervals show which pairs differ while controlling the familywise error rate. Limitations: Each respondent must belong to exactly one group. Levene's test and Kruskal–Wallis are reported for raw data only. Repeated-measures and multi-factor designs need other models. Primary page: https://www.crosstabs.com/anova-calculator Evidence: https://www.crosstabs.com/spss/one-way-anova Evidence: https://www.crosstabs.com/methods ## Pearson, Spearman and Kendall correlation Stable ID: correlation Locale: en Topic cluster: association_measure Pearson's r measures linear association between two numeric variables; Spearman's rho and Kendall's tau-b measure monotonic association on ranks. Fisher's z transformation gives the confidence interval for r and a test comparing two independent correlations. Formula: r = Σ(x − x̄)(y − ȳ) / √[Σ(x − x̄)² Σ(y − ȳ)²]; t = r√(n − 2)/√(1 − r²); CI = tanh(atanh r ± z/√(n − 3)). Interpretation: The sign gives direction and the magnitude strength; r² is the share of variance the variables share. Correlation does not establish causation. Limitations: Pearson's r misses curved relationships and is sensitive to outliers. Fisher's two-correlation test requires independent samples. Kendall's tau p-values use the normal approximation. Primary page: https://www.crosstabs.com/correlation-calculator Evidence: https://www.crosstabs.com/spss/pearson-spearman-correlation Evidence: https://www.crosstabs.com/kendalls-tau-calculator Evidence: https://www.crosstabs.com/methods ## How to choose a categorical statistical test Stable ID: statistical-test-selection Locale: en Topic cluster: test_selection Choose from the design first: independent nominal groups usually imply chi-square or an exact alternative; paired binary data imply McNemar; agreement implies kappa; ordered categories imply an ordinal association measure. Limitations: No decision tree can repair dependence, confounding, invalid sampling, or a mismatch between the table and the estimand. Primary page: https://www.crosstabs.com/which-statistical-test Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation ## Sparse expected cells Stable ID: sparse-cells Locale: en Topic cluster: test_selection Sparse expected counts make chi-square reference approximations less reliable; inspect the pattern, consider defensible category consolidation, or use a design-appropriate exact or Monte Carlo method. Limitations: Do not merge categories solely to obtain significance. Fisher's exact test on this site is limited to eligible 2×2 integer-frequency tables. Primary page: https://www.crosstabs.com/chi-square-expected-count-less-than-5 Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation ## Survey weighting and raking for crosstabs Stable ID: survey-weights Locale: en Topic cluster: survey_workflow A weighted crosstab multiplies each eligible respondent by a positive finite weight before calculating counts and percentages, while retaining the unweighted base for diagnostics; raking creates calibration weights by iteratively matching trusted population margins. Limitations: The quick workspace rejects an untyped numeric weight instead of assuming frequency semantics. Saved projects can create a raked sampling weight with guided survey weighting from explicit population targets after preview and human confirmation; that weight does not authorize ordinary inference. Saved tab-book projects can assign a supported Taylor-linearized or replicate-weight survey design, with explicit strata, PSU, FPC, lonely-PSU, confidence, and replicate policies plus Rao–Scott or adjusted-Wald association; unsupported or invalid designs fail closed, and independent statistical review remains pending. Primary page: https://www.crosstabs.com/survey-weighting-raking Evidence: https://www.crosstabs.com/survey-crosstab-tool Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/capacity Evidence: https://www.crosstabs.com/local-mode ## Multi-select crosstabs Stable ID: multi-select Locale: en Topic cluster: survey_workflow A select-all-that-apply question should become auditable option-level binary outcomes and be crossed with an independent grouping variable; report percent of cases to show each option's incidence among included respondents rather than treating combined answer strings as exclusive categories. Interpretation: Percent-of-cases rows can sum above 100% because one respondent may select multiple options; percent of responses uses a different denominator and answers a different question. Limitations: An omnibus Pearson chi-square across overlapping option rows is invalid; option-level binary comparisons answer a narrower question. Bonferroni correction covers banner-column pairs within each option but not the complete family across options, tables, or exploratory cuts. All-blank skipped sets are excluded while explicit all-zero answered cases remain in the respondent base. An approved survey design uses respondent-case item estimates and separate selected/not-selected Rao–Scott tests with adjusted-Wald alternatives; response-denominator percentages remain descriptive, and independent statistical review remains pending. Primary page: https://www.crosstabs.com/multi-select-crosstab Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/local-mode Evidence: https://www.crosstabs.com/crosstab-significance-letters ## Crosstab significance letters Stable ID: significance-letters Locale: en Topic cluster: survey_workflow Column significance letters encode pairwise column-proportion tests by labeling a cell with the columns it differs from under the configured threshold and correction policy. Limitations: Small bases and multiple comparisons can produce misleading flags. Untyped weighting cannot produce ordinary significance letters. Primary page: https://www.crosstabs.com/crosstab-significance-letters Evidence: https://www.crosstabs.com/methods ## Tracker wave schema drift Stable ID: tracker-schema-drift Locale: en Topic cluster: tracker_workflow Treat each imported tracker wave as immutable: fingerprint both schemas, match stable identifiers before labels, classify structural and semantic drift, preview and approve explicit transforms, apply them to a copy, validate bases and distributions, and retain the source, audit record, and guarded inverse before comparing waves. Interpretation: A successful structural repair does not prove longitudinal comparability; question wording, universe, routing, recall period, concept, and representation must still remain defensibly comparable. Limitations: Automatic suggestions and label similarity are not evidence of semantic equivalence. Exact reversal is refused after later project changes because restoring the old snapshot would erase subsequent work. The current Workspace UI exposes compatible-wave import, rename mapping, retained waves, and comparison, but not the engine's full approval and reversal interface. Primary page: https://www.crosstabs.com/tracker-schema-drift Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/local-mode Evidence: https://www.crosstabs.com/capacity Evidence: https://www.crosstabs.com/survey-tab-book ## Tab-book project model Stable ID: tab-book-project Locale: en Topic cluster: survey_workflow A tab-book project keeps row sets, multibanners, filters, weights, tests, table order, waves, coding variables, findings, and the export recipe in one versioned graph rather than as unrelated outputs. Limitations: A definition-only manifest intentionally omits respondent rows; it rebuilds a working analysis only when applied to a new data file, and definitions that do not fit that file are listed and skipped. Primary page: https://www.crosstabs.com/survey-tab-book Evidence: https://www.crosstabs.com/local-mode Evidence: https://www.crosstabs.com/data-processing ## Auditable open-end coding Stable ID: open-end-coding Locale: en Topic cluster: ai_governance Crosstabs provides a browser-local manual workflow for open ends: define or reuse themes, assign responses, attest that review is complete, approve the coding, apply it as project variables, and retain decision history with a guarded undo path. Limitations: The manual workflow does not send responses to a model. Optional AI open-end coding sends the answers to one chosen question through the crosstabs.com server to OpenAI only after the researcher confirms that run, and returns proposed themes and codes for review. A researcher remains responsible for theme definitions, assignment coverage, approval, interpretation, and any sensitive-content handling. Supplying responses to a separate AI assistant is governed by that assistant's data policy, not the Crosstabs local boundary. Primary page: https://www.crosstabs.com/open-end-coding Evidence: https://www.crosstabs.com/privacy Evidence: https://www.crosstabs.com/data-processing Evidence: https://www.crosstabs.com/local-mode Evidence: https://www.crosstabs.com/subprocessors ## Outcome metrics and key-driver modeling Stable ID: outcome-modeling Locale: en Topic cluster: survey_workflow The outcome studio fits a regularized logistic model to an explicitly selected binary outcome and drivers, then reports deterministic holdout AUC, accuracy, Brier score, calibration, observed segments, and reviewable actions. Limitations: Association and prediction do not establish causality. At least 60 complete rows, 25 successes, and 25 other outcomes are required. Synthetic scenarios are labeled modeled estimates rather than observed evidence. Primary page: https://www.crosstabs.com/key-driver-analysis Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/capacity ## Editable report refresh Stable ID: editable-report-refresh Locale: en Topic cluster: reporting A compatible Crosstabs PPTX or DOCX can refresh current generated values without replacing the whole edited package: stable anchors identify selected tables, metadata, statistics text, and PowerPoint charts; supported equal-shape content is patched in place while manual regions, unselected objects, and unrelated package parts remain under analyst control. Interpretation: Preservation is regional, not universal: the existing Office package remains the target, and the refresh audit distinguishes updated anchors, preserved anchors, blocking reasons, and package entries that were not rewritten. Limitations: Selected generated table text is replaced by current approved Crosstabs values; selected PowerPoint chart series and its embedded workbook are also replaced. Equal-shape table formatting and geometry are supported, but arbitrary structural edits inside generated objects are not. A missing or duplicate stable anchor, changed table shape, incompatible chart, corrupt package, or unanchored third-party file fails atomically before a modified file is returned. The current Workspace Export menu refreshes the open tab-current table; it does not expose arbitrary multi-table selection or automatic schema-changing reflow. Primary page: https://www.crosstabs.com/editable-pptx-docx-refresh Evidence: https://www.crosstabs.com/workspace Evidence: https://www.crosstabs.com/local-mode Evidence: https://www.crosstabs.com/data-processing Evidence: https://www.crosstabs.com/methods/validation Evidence: https://www.crosstabs.com/survey-tab-book ## Local browser analysis Stable ID: local-mode Locale: en Topic cluster: trust Browser import, table computation, filters, supported survey designs, tests, project state, review, manual open-end coding, and exports run locally without uploading source files. Local MCP is a separate on-device option. Every table and statistic is computed deterministically; optional AI insights send data to OpenAI only when the person clicks an AI button. Limitations: Browser memory, storage quotas, device performance, and file width can constrain work before row count alone does. A full project package contains rows and is not encrypted. Primary page: https://www.crosstabs.com/local-mode Evidence: https://www.crosstabs.com/privacy Evidence: https://www.crosstabs.com/data-processing Evidence: https://www.crosstabs.com/capacity ## Multidimensional workload envelope Stable ID: workload-envelope Locale: en Topic cluster: trust Crosstabs publishes tested workload envelopes across rows, file representation, category shape, output cells, memory, filters, weights, statistical tests, cancellation, and browser execution mode instead of claiming unlimited rows. Limitations: Published fixtures are evidence for tested dimensions and environments, not a universal device guarantee. Primary page: https://www.crosstabs.com/capacity Evidence: https://www.crosstabs.com/capacity Evidence: https://www.crosstabs.com/methods/validation ## Evidence-linked AI analysis Stable ID: ai-evidence-governance Locale: en Topic cluster: ai_governance Crosstabs computes every table deterministically; its optional hosted AI insights (OpenAI) receive only a suppression-checked table, questionnaire metadata, or, after confirmation, one question's open-ended answers, and never respondent rows. An external agent can use public evidence and bounded aggregate tools without uploading a file, or the local MCP, which keeps files on the device. Evidence validation fails closed: AI findings with numbers must quote the computed table verbatim or are dropped. Deterministic evidence, typed operations, project revisions, approval requirements, audit history, and guarded undo are distinct from model-generated interpretation. Interpretation: Provider routing and model agreement can improve availability or offer another opinion, but neither validates statistical claims nor grants authority to change a research project. Limitations: AI-client policies apply independently to anything shared with an assistant. Exact undo is refused after later project changes because restoring the earlier snapshot would overwrite subsequent work. Model providers can still make errors, and independent statistical review and third-party certification evidence remain pending where explicitly disclosed. Primary page: https://www.crosstabs.com/ai-analysis-governance Evidence: https://www.crosstabs.com/security Evidence: https://www.crosstabs.com/privacy Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/open-end-coding Evidence: https://www.crosstabs.com/survey-tab-book ## Crosstabs Codex plugin and MCP server Stable ID: codex-plugin Locale: en Topic cluster: ai_governance The published crosstabs 1.5.3 package passed fresh public installation checks: 39 statistical tools, six market-research tools, database and survey-platform pull tools plus runtime status on the statistical server, and 37 local project tools including portable Plan → Validate → Run → Review workflows, explicit guided weighting preview/apply and Statistics procedures. Immutable PyPI artifacts, exact source provenance and the MCP Registry record were verified. The 1.5.0, 1.4.0, 1.3.0 and 1.2.2 evidence remains historical. Plugin source installation, native-host review and directory approval are separate states; see the exact source binding on the MCP page. Interpretation: Install the local package (pip install crosstabs) and add its stdio servers to your MCP client; files remain on the device. Any AI client has its own data policies. Limitations: The browser worker and local MCP enter the same HeadlessCrosstabsService and share TabBookProject, initial-analysis, deterministic calculation, project revision, audit, replay, and undo orchestration. Browser interaction and local MCP transport remain separate release-tested surfaces. Package publication and source installation do not establish native-host widget review, directory submission, approval or listing. The local plugin does not reproduce browser interaction state or browser-local manual open-end coding. Any file supplied to an AI assistant remains subject to that assistant's data boundary and the user's configuration. Primary page: https://www.crosstabs.com/mcp Evidence: https://www.crosstabs.com/mcp Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation Evidence: https://www.crosstabs.com/local-mode Evidence: https://www.crosstabs.com/privacy ## Shapley value regression vs Johnson's relative weights Stable ID: relative-importance-shapley-johnson Locale: en Topic cluster: survey_workflow Shapley value regression and Johnson's relative weights both divide a regression model's R² among correlated drivers: Shapley averages each driver's R² gain over every order of entry (2^p subset regressions); Johnson's method approximates the same split in one step from an orthogonal transformation of the drivers. Formula: Shapley_j = Σ_{S ⊆ D∖{j}} |S|!(p − |S| − 1)!/p! × [R²(S ∪ {j}) − R²(S)]; RW_j = Σ_k λ_jk² β*_k² with β* = Λ⁻¹ r_xy and Λ = VΛ^½V′ from R_xx = VΛV′. Interpretation: Shares sum to 100% of the explained variance; drivers within a point or two of each other should be reported as tied. Limitations: Importance is associative, not causal. The relative importance tool decomposes a linear R²; it is not a logistic model for yes/no outcomes. Exact Shapley needs 2^p regressions, so the tool accepts at most 15 drivers. Primary page: https://www.crosstabs.com/shapley-vs-johnson-relative-importance Evidence: https://www.crosstabs.com/tools/relative-importance-analysis Evidence: https://www.crosstabs.com/methods ## Van Westendorp price sensitivity meter Stable ID: van-westendorp-psm Locale: en Topic cluster: survey_workflow The Van Westendorp price sensitivity meter asks four open price questions (too cheap, cheap, expensive, too expensive), plots their cumulative shares, and reads the point of marginal cheapness, optimal price point, indifference price point and point of marginal expensiveness from the curve crossings; PMC to PME is the range of acceptable prices. Formula: PMC: too cheap = 1 − cheap; OPP: too cheap = too expensive; IPP: cheap = expensive; PME: 1 − expensive = too expensive; crossings by linear interpolation between answered prices. Limitations: The PSM measures price acceptance, not revenue; revenue needs the Newton-Miller-Smith purchase questions or a Gabor-Granger test. Respondents whose four answers are out of order are excluded by default and counted. Primary page: https://www.crosstabs.com/van-westendorp-guide Evidence: https://www.crosstabs.com/tools/van-westendorp Evidence: https://www.crosstabs.com/methods ## Gabor-Granger demand curve Stable ID: gabor-granger-demand Locale: en Topic cluster: survey_workflow The Gabor-Granger method asks purchase intent at several specific prices; the weighted share who would buy at each price is a demand curve, price × demand is expected revenue per respondent, and the tested price with the highest revenue is the revenue-maximising tested price. Formula: demand(p) = Σ wᵢ buyᵢ(p) / Σ wᵢ; revenue(p) = p × demand(p); arc elasticity E = [(d₂ − d₁)/((d₁ + d₂)/2)] / [(p₂ − p₁)/((p₁ + p₂)/2)]. Limitations: Stated intent overstates purchase, so the level of the curve is optimistic. No value between tested prices is assumed; competitors and features are not modelled. Primary page: https://www.crosstabs.com/gabor-granger-guide Evidence: https://www.crosstabs.com/tools/gabor-granger Evidence: https://www.crosstabs.com/methods ## NPS margin of error and significance testing Stable ID: nps-margin-of-error Locale: en Topic cluster: survey_workflow Because NPS is the promoter share minus the detractor share from the same respondents, its standard error is 100 × √{[p_P(1 − p_P) + p_D(1 − p_D) + 2·p_P·p_D] / n}; two independent NPS results are compared with z = (NPS₂ − NPS₁) / √(SE₁² + SE₂²). Formula: MOE = z × SE; n_required = z² × [p_P(1 − p_P) + p_D(1 − p_D) + 2·p_P·p_D] / MOE². Limitations: The independent-samples test is not valid for a panel that answered both waves. Weighted data use the Kish effective sample size; the interval covers sampling error only. Primary page: https://www.crosstabs.com/nps-margin-of-error Evidence: https://www.crosstabs.com/tools/nps-calculator Evidence: https://www.crosstabs.com/methods ## Survey data cleaning: speeders, straight-liners and duplicates Stable ID: survey-data-cleaning Locale: en Topic cluster: survey_workflow Survey data cleaning flags interviews with predefined rules before tabulation and weighting: speeders below a fraction of the median duration, straight-liners with one answer across a rating grid, repeated IDs or exact duplicate rows, and interviews with too many blanks; flagged rows are listed with reasons, counted per rule and kept in an audit file. Limitations: A flag is evidence of low effort, not proof; short grids can produce sincere identical answers. Rules should be fixed before results are seen and applied identically to every wave. Primary page: https://www.crosstabs.com/how-to-clean-survey-data Evidence: https://www.crosstabs.com/tools/data-cleaner Evidence: https://www.crosstabs.com/methods ## K-means vs latent class segmentation Stable ID: kmeans-vs-latent-class Locale: en Topic cluster: survey_workflow K-means assigns respondents to the nearest of k centres on standardised numeric variables and is compared across k by the average silhouette; latent class analysis fits a finite mixture model by maximum likelihood, handles categorical, ordinal and numeric indicators and missing answers, gives class membership probabilities, and is compared across class counts by BIC. Formula: Silhouette s(i) = (bᵢ − aᵢ) / max(aᵢ, bᵢ); latent class P(yᵢ) = Σ_c π_c Π_j P(y_ij | c); BIC = −2 log L + q ln n. Limitations: Both methods return the number of segments requested even without real group structure. K-means drops respondents with any blank clustering variable; latent class estimation skips missing indicators under a missing-at-random assumption. Primary page: https://www.crosstabs.com/kmeans-vs-latent-class-segmentation Evidence: https://www.crosstabs.com/tools/segment-builder Evidence: https://www.crosstabs.com/tools/latent-class-analysis Evidence: https://www.crosstabs.com/methods ## Linear and logistic regression on survey data Stable ID: survey-regression Locale: en Topic cluster: survey_workflow Linear regression models a numeric survey outcome and reports B per unit of each predictor; binary logistic regression models a yes/no outcome and reports odds ratios Exp(B). A declared sampling weight changes the point estimates, but ordinary regression standard errors are not valid under a sampling design, so design-based inference needs Complex Samples procedures. Formula: y = b₀ + Σ b_j x_j + e; log[p/(1 − p)] = b₀ + Σ b_j x_j; Beta_j = b_j s_xj / s_y; odds ratio = e^{b_j}. Limitations: In Statistics mode a declared sampling weight yields weighted point estimates only; tests, standard errors and intervals are withheld, and stepwise selection is refused. Cross-sectional regression adjusts for measured variables only and does not establish causation. Primary page: https://www.crosstabs.com/survey-regression-guide Evidence: https://www.crosstabs.com/spss/linear-regression Evidence: https://www.crosstabs.com/spss/binary-logistic-regression Evidence: https://www.crosstabs.com/spss/complex-samples-glm Evidence: https://www.crosstabs.com/methods ## Bonferroni and Holm corrections for crosstab column tests Stable ID: bonferroni-column-tests Locale: en Topic cluster: test_selection Column-proportion letters compare every pair of banner columns within a row, so the family-wise false-positive rate grows with the number of pairs; Bonferroni multiplies each pairwise p-value by the number of pairs m, and Holm's step-down procedure controls the same family-wise error rate with at least as much power. Formula: Bonferroni p̃ᵢ = min(1, m·pᵢ); Holm p̃₍ᵢ₎ = min(1, max_{j ≤ i} (m − j + 1)·p₍ⱼ₎); m = k(k − 1)/2 column pairs per row. Limitations: Crosstabs corrects across the column pairs within each row; comparisons across rows, tables and banners are not corrected together. Holm, Benjamini–Hochberg or no correction can be chosen in saved tab-book projects; Bonferroni is the default. Primary page: https://www.crosstabs.com/bonferroni-correction-crosstabs Evidence: https://www.crosstabs.com/crosstab-significance-letters Evidence: https://www.crosstabs.com/survey-crosstab-tool Evidence: https://www.crosstabs.com/methods ## Design effect and effective sample size Stable ID: design-effect Locale: en Topic cluster: survey_workflow The design effect is the variance of an estimate under the survey design divided by its variance under a simple random sample of the same size; the effective sample size is n / deff. For unequal weights Kish's approximation is n_eff = (Σw)² / Σw², so deff = 1 + CV² of the weights, and margins of error grow by √deff. Formula: n_eff = (Σwᵢ)² / Σwᵢ²; deff_w = n / n_eff = 1 + CV_w²; clustering deff ≈ 1 + (m̄ − 1)ρ; MOE = z√deff √(p(1 − p)/n). Limitations: The Kish design effect covers unequal weighting only; clustering and stratification need design-based estimation. Each subgroup has its own effective base; one overall design effect does not fit every column. Primary page: https://www.crosstabs.com/design-effect-effective-sample-size Evidence: https://www.crosstabs.com/margin-of-error-calculator Evidence: https://www.crosstabs.com/survey-weighting-raking Evidence: https://www.crosstabs.com/methods ## 卡方独立性检验 Stable ID: zh-chi-square Locale: zh-CN Topic cluster: test_selection 卡方独立性检验先按行列边际计算独立假设下的期望频数,再用 χ² = Σ(O − E)² / E 汇总观察值与期望值的差异,用于判断两个相互独立观测的分类变量是否存在统计关联。 Formula: χ² = Σ (O − E)² / E;E = 行合计 × 列合计 / 总样本量。 Interpretation: 较小的 P 值反对独立性原假设;还应报告 Cramér's V、样本基数、期望频数诊断、关键单元格和研究设计,不能把统计关联解释为因果。 Limitations: P 值使用渐近卡方分布;最小期望频数小于 1 或超过 20% 的期望频数小于 5 时,近似可能较差。 观察应相互独立;同一对象前后两次的配对二分类表应使用 McNemar 等配对方法。 输入应是互斥类别的观察频数,百分比不能替代检验所需的样本基数。 本网站的 Fisher 精确结果只覆盖合格的非加权 2×2 非负整数频数表和有界工作量。 设计或校准权重、分层和整群抽样需要设计型方差与 Rao–Scott 等方法。 调整标准化残差是诊断量;跨多个单元格、列对、问题或表格的推断仍需定义检验族并处理多重比较。 Primary page: https://www.crosstabs.com/zh-cn/chi-square-calculator Evidence: https://www.crosstabs.com/chi-square-calculator Evidence: https://www.crosstabs.com/zh-cn/which-statistical-test Evidence: https://www.crosstabs.com/zh-cn/chi-square-expected-count-less-than-5 Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation ## Crosstabs Codex 插件与 MCP 统计服务器 Stable ID: zh-codex-plugin Locale: zh-CN Topic cluster: ai_governance 已发布的 crosstabs 1.5.3 通过了公开索引全新安装验证:统计服务器提供 39 个统计工具、6 个市场研究工具、数据库与调查平台拉取工具及运行时状态,项目服务器提供 37 个本地工具,包括可移植的计划 → 验证 → 运行 → 审阅流程、经明确批准的引导式加权预览及应用和统计过程工具。PyPI 不可变发布文件、准确源码来源和 MCP Registry 记录均已验证;1.5.0、1.4.0、1.3.0 和 1.2.2 的证据保留为历史记录。插件源码安装、原生宿主审查和目录批准属于不同状态;准确的源码绑定见 MCP 页面。 Interpretation: 安装本地包(pip install crosstabs),并把其 stdio 服务器添加到 MCP 客户端;文件留在设备上。AI 客户端的数据政策另行适用。 Limitations: 浏览器 Worker 与本地 MCP 进入同一个 HeadlessCrosstabsService,并共享 TabBookProject、初始分析、确定性计算、项目修订、审计、重放与撤销编排。浏览器交互与本地 MCP 传输仍是分别测试的发布界面。 包发布和源码安装不代表原生宿主组件审查、目录提交、批准或上架。 本地插件不复制浏览器交互状态,也不提供浏览器本地的开放题人工编码。 提供给 AI 助手的任何文件仍受该助手的数据边界与用户配置约束。 Primary page: https://www.crosstabs.com/zh-cn/mcp Evidence: https://www.crosstabs.com/zh-cn/mcp Evidence: https://www.crosstabs.com/mcp Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation Evidence: https://www.crosstabs.com/local-mode Evidence: https://www.crosstabs.com/privacy ## McNemar 配对二分类变化检验 Stable ID: zh-mcnemar-paired-change Locale: zh-CN Topic cluster: test_selection McNemar 检验通过比较配对 2×2 表中两个不一致方向 b 与 c,判断同一对象前后或一一匹配对象的二分类边际比例是否发生不对称变化;一致配对 a 与 d 不进入统计量。 Formula: 未校正 χ² = (b − c)² / (b + c),df = 1;连续性校正 χ² = max(|b − c| − 1, 0)² / (b + c);精确双侧结果使用 Binomial(b + c, 0.5)。 Interpretation: 较小的 P 值表示两个变化方向不对称,不表示变化幅度,也不会单独证明变化由干预造成。 Limitations: 对象或匹配对之间应相互独立,且每次结果只有两个互斥类别。 公开计算器只接受非负整数频数,不接受百分比、小数或调查权重。 不一致配对不超过 10,000 对时计算精确二项尾概率;更大输入明确标为连续性校正正态近似。 复杂抽样、失访、混杂和多时点重复测量需要适合完整设计的方法。 Primary page: https://www.crosstabs.com/zh-cn/mcnemar-test-calculator Evidence: https://www.crosstabs.com/mcnemar-test-calculator Evidence: https://www.crosstabs.com/zh-cn/which-statistical-test Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation ## 分类数据统计检验选择 Stable ID: zh-statistical-test-selection Locale: zh-CN Topic cluster: test_selection 分类数据检验应从研究设计开始:独立分类变量通常使用 Pearson 卡方或适合的精确方法;合格的稀疏 2×2 整数频数表可用 Fisher;配对二分类结果使用 McNemar;对同一对象的分类一致性使用 Cohen's kappa;有序变量还应报告保留顺序信息的关联指标。 Limitations: P 值和关联不能单独证明因果。 普通交叉表方法不适合连续结局、复杂抽样、聚类或未建模的重复测量结构。 本网站的 Fisher 精确检验只覆盖合格的非加权 2×2 非负整数频数表。 Primary page: https://www.crosstabs.com/zh-cn/which-statistical-test Evidence: https://www.crosstabs.com/which-statistical-test Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation Evidence: https://www.crosstabs.com/mcnemar-test-calculator Evidence: https://www.crosstabs.com/cohens-kappa-calculator ## 期望频数小于 5 Stable ID: zh-sparse-cells Locale: zh-CN Topic cluster: test_selection 期望频数过小会削弱卡方渐近近似;应先检查稀疏分布,再考虑有理论依据的类别合并、2×2 Fisher 精确检验或适合设计的精确/蒙特卡洛方法。 Limitations: 不要只为获得显著结果而合并类别。 本网站的 Fisher 精确检验仅适用于合格的 2×2 非加权整数频数表。 Primary page: https://www.crosstabs.com/zh-cn/chi-square-expected-count-less-than-5 Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation ## 调查加权与 Raking(迭代比例拟合) Stable ID: zh-survey-weights Locale: zh-CN Topic cluster: survey_workflow 加权交叉表让每位合格受访者按正数有限权重贡献到单元格;应使用加权百分比表达估计,同时保留未加权样本基数用于样本量和数据质量诊断。Raking 通过迭代调整权重,使指定加权边际接近可信总体目标。 Formula: 加权单元格 = Σ wᵢI(rowᵢ = r, columnᵢ = c);权重离散诊断 n_eff = (Σwᵢ)² / Σwᵢ²。 Limitations: 快速工作区会拒绝未声明语义的数值权重,不会把它默认为频数权重。 已保存项目可通过引导式调查加权,按明确的总体目标并经预览和人工确认生成 Raking 抽样权重;该权重不会授权普通推断。 已保存项目可执行受支持的 Taylor 线性化或重复权重设计,并报告分层、PSU、FPC、设计自由度以及 Rao–Scott 或调整 Wald 关联;无效或不受支持的设计会关闭失败。 Primary page: https://www.crosstabs.com/zh-cn/survey-weighting-raking Evidence: https://www.crosstabs.com/survey-weighting-raking Evidence: https://www.crosstabs.com/survey-crosstab-tool Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/capacity ## Cramér's V 效应量 Stable ID: zh-cramers-v Locale: zh-CN Topic cluster: association_measure Cramér's V 将卡方统计量转换为 0 到 1 的分类变量关联强度指标。 Formula: V = sqrt(χ² / (n × (min(行数, 列数) − 1)))。 Limitations: 它不表示方向,也不说明哪些单元格推动了关联。 Primary page: https://www.crosstabs.com/zh-cn/cramers-v-calculator Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/methods/validation ## 多选题(可多选)交叉分析 Stable ID: zh-multi-select Locale: zh-CN Topic cluster: survey_workflow “选择所有适用项”应定义或整理为选项级二分类结果,再与独立分组变量交叉;通常报告基于个案的百分比,表示每个 Banner 组中选择该选项的受访者比例,而不是把答案组合当成互斥类别。 Interpretation: 同一受访者可选择多个选项,所以基于个案的各行百分比总和可以超过 100%;基于响应的百分比使用全部选择次数作为不同分母。 Limitations: 不能对相互重叠的所有选项行运行一个普通总体 Pearson 卡方检验。 每个选项内部的 Banner 列对使用 Bonferroni 校正,但不会跨所有选项、表格或探索性切分控制完整推断族。 所有选项均为空的跳题记录默认排除;明确作答且全为 0 的个案仍保留在受访者基数。 设计或校准权重下的普通列比例检验只能视为近似结果。 Primary page: https://www.crosstabs.com/zh-cn/multi-select-crosstab Evidence: https://www.crosstabs.com/multi-select-crosstab Evidence: https://www.crosstabs.com/crosstab-significance-letters Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/local-mode ## 交叉表显著性字母(A、B、C) Stable ID: zh-significance-letters Locale: zh-CN Topic cluster: survey_workflow 交叉表中的显著性字母表示同一结果行内的成对列比例检验:每个 Banner 列先分配一个字母,某单元格中的字母标出该单元格所在列的比例显著高于哪些列。 Interpretation: 例如 A 列单元格显示“42% B”,表示 A 列的 42% 在设定阈值和多重比较校正下显著高于 B 列同一行的比例;字母不表示差异大小、总体关联强度或因果关系。 Limitations: Crosstabs 对每行内所有合格 Banner 列对使用合并方差双比例 z 检验和 Bonferroni 校正;校正范围不跨其他行、表格或探索性切分。 检验列合计小于 30 的列不参与字母检验;普通表超过 12 个 Banner 列或 250 个行类别时不生成字母;没有字母不等于两列相同。 所选权重时不生成普通列比例字母;受支持的已保存调查设计也不会回退到普通加权列比例检验。 显著性字母应与百分比、未加权基数、置信阈值、效应量和研究设计一起报告。 Primary page: https://www.crosstabs.com/zh-cn/crosstab-significance-letters Evidence: https://www.crosstabs.com/crosstab-significance-letters Evidence: https://www.crosstabs.com/zh-cn/multi-select-crosstab Evidence: https://www.crosstabs.com/zh-cn/survey-weighting-raking Evidence: https://www.crosstabs.com/methods ## 调查追踪项目的波次结构漂移 Stable ID: zh-tracker-schema-drift Locale: zh-CN Topic cluster: tracker_workflow 把每个导入波次视为不可变源数据:比较结构指纹,优先使用稳定内部 ID,分类记录结构与语义漂移,预览并审批明确转换,只在副本上执行,验证基数与分布,并保留来源、审计记录和受保护的逆操作后再比较波次。 Interpretation: 结构修复成功不等于纵向可比;题目措辞、目标总体、跳转逻辑、回忆期、概念和表示方式仍须由研究人员确认。 Limitations: 自动建议和标签相似不能证明语义等价。 项目在修复后已改变时必须拒绝精确回退,以免覆盖后续工作。 当前工作区支持兼容波次导入、改名映射、保留旧波次与比较,但尚未提供底层引擎完整的审批与回退界面。 Primary page: https://www.crosstabs.com/zh-cn/tracker-schema-drift Evidence: https://www.crosstabs.com/tracker-schema-drift Evidence: https://www.crosstabs.com/methods Evidence: https://www.crosstabs.com/local-mode Evidence: https://www.crosstabs.com/capacity Evidence: https://www.crosstabs.com/survey-tab-book ## 交叉表 AI 分析治理 Stable ID: zh-ai-evidence-governance Locale: zh-CN Topic cluster: ai_governance Crosstabs 的所有表格与统计均由确定性引擎计算;可选的托管 AI 解读(OpenAI)只接收经小格抑制的汇总表、问卷元数据,或经确认后某一道开放题的回答,从不接收受访者行数据。外部智能体可使用公开证据和有界汇总计算,或通过本地 MCP 处理只保存在设备上的私有项目。Crosstabs 把结果绑定到确定性证据,并在输入无效时默认拒绝;项目变更使用类型化操作、当前修订、本地审计记录与受保护撤销,而不把模型文字当成授权。 Interpretation: 多模型路由或模型间一致只能改善可用性或提供另一种意见,不能验证统计结论,也不能授予模型修改研究项目的权限。 Limitations: 私有项目应留在浏览器或本地 MCP;交给外部智能体的任何文件仍受该智能体的数据边界约束。 项目后来发生变化时会拒绝精确回退,以免覆盖后续工作。 模型仍可能出错;页面明确披露的独立统计审查与第三方认证证据仍处于待完成状态。 Primary page: https://www.crosstabs.com/zh-cn/ai-analysis-governance Evidence: https://www.crosstabs.com/ai-analysis-governance Evidence: https://www.crosstabs.com/security Evidence: https://www.crosstabs.com/privacy Evidence: https://www.crosstabs.com/open-end-coding Evidence: https://www.crosstabs.com/survey-tab-book ## 可编辑 PPTX 与 DOCX 选择性刷新 Stable ID: zh-editable-report-refresh Locale: zh-CN Topic cluster: reporting 兼容的 Crosstabs PPTX 或 DOCX 可以通过稳定锚点识别选中表格、元数据、统计文字和 PowerPoint 图表,只在支持的等形生成区域内更新当前值,同时保留人工区域、未选中对象和无关 Office 包部件。 Interpretation: 保留是按区域定义的,而不是任意往返编辑承诺;现有 Office 文件保持为目标,刷新审计分别记录已更新锚点、已保留锚点、阻止原因与未重写部件。 Limitations: 选中生成表格内的文字会替换为当前获批 Crosstabs 值;选中 PowerPoint 图表的序列和嵌入工作簿也会替换。 仅支持等形表格的格式与几何保留,不支持生成对象内任意结构修改。 锚点缺失或重复、表格形状变化、图表不兼容、文件损坏或第三方文件没有锚点时会原子失败,不会返回部分修改的文件。 当前 Workspace Export 菜单刷新打开的 tab-current 表格,尚未提供任意多表选择或结构变化后的自动重排。 Primary page: https://www.crosstabs.com/zh-cn/editable-pptx-docx-refresh Evidence: https://www.crosstabs.com/editable-pptx-docx-refresh Evidence: https://www.crosstabs.com/workspace Evidence: https://www.crosstabs.com/local-mode Evidence: https://www.crosstabs.com/data-processing Evidence: https://www.crosstabs.com/methods/validation Evidence: https://www.crosstabs.com/survey-tab-book ## 可审计的开放题编码 Stable ID: zh-open-end-coding Locale: zh-CN Topic cluster: ai_governance Crosstabs 在浏览器本地提供开放题人工编码流程:定义或复用主题、逐条分配、确认审核完成、批准编码表、把已批准主题作为 0/1 项目变量应用,并保留决策历史和受保护的撤销。 Limitations: 人工流程不会把回答发送给模型。可选 AI 开放题编码只在研究者确认本次运行后,经 crosstabs.com 服务器把所选一道题的回答发送给 OpenAI,并返回待审核的主题与编码提案。 主题定义、分配覆盖、批准、解释和敏感内容处理始终由研究者负责。 应用后数据如有其他修改,撤销会被拒绝,以免覆盖后续工作。 把回答提供给其他 AI 助手时,适用该助手的数据政策,而不是 Crosstabs 的本地边界。 Primary page: https://www.crosstabs.com/zh-cn/open-end-coding Evidence: https://www.crosstabs.com/open-end-coding Evidence: https://www.crosstabs.com/privacy Evidence: https://www.crosstabs.com/data-processing Evidence: https://www.crosstabs.com/local-mode Evidence: https://www.crosstabs.com/subprocessors ## 交叉表多重比较 Bonferroni 校正 Stable ID: zh-multiple-comparisons Locale: zh-CN Topic cluster: test_selection k 个 Banner 列两两比较共有 k(k − 1) ÷ 2 次;Bonferroni 校正把每次比较的阈值降为 α 除以比较次数,使同一组比较中至少出现一次假阳性的概率不超过 α。 Formula: m = k(k − 1) / 2;α′ = α / m;校正 P = min(1, m × P)。 Limitations: Crosstabs 的显著性字母默认只在每个结果行内的合格列对之间校正,不跨其他行、题目、表格或筛选切分。 列基数低于 30 的列不参与检验;没有字母不证明两列相同。 已保存项目的表格设计器可改用 Holm、Benjamini–Hochberg 或不校正。 Primary page: https://www.crosstabs.com/zh-cn/crosstab-bonferroni-correction Evidence: https://www.crosstabs.com/crosstab-significance-letters Evidence: https://www.crosstabs.com/zh-cn/crosstab-significance-letters Evidence: https://www.crosstabs.com/methods ## 设计效应与有效样本量 Stable ID: zh-design-effect Locale: zh-CN Topic cluster: survey_workflow 设计效应是实际抽样设计下估计量方差与同样本量简单随机抽样方差之比;有效样本量为 n 除以设计效应,误差范围随设计效应的平方根扩大。 Formula: Kish:deff = n·Σw² / (Σw)² = 1 + CV²;n_eff = (Σw)² / Σw²;整群:deff ≈ 1 + (m − 1)ρ。 Limitations: Kish 近似只反映权重不等造成的方差膨胀,不包含分层与整群,也不能代替设计型方差估计。 误差范围计算器接受输入的设计效应或由权重列计算 Kish 设计效应;完整调查设计推断需要已保存项目中受支持的 Taylor 或重复权重设计。 报告时应分别给出未加权基数和有效基数。 Primary page: https://www.crosstabs.com/zh-cn/design-effect-effective-sample-size Evidence: https://www.crosstabs.com/margin-of-error-calculator Evidence: https://www.crosstabs.com/zh-cn/margin-of-error-calculator Evidence: https://www.crosstabs.com/zh-cn/survey-weighting-raking Evidence: https://www.crosstabs.com/methods ## 问卷表格册项目模型 Stable ID: zh-tab-book-project Locale: zh-CN Topic cluster: survey_workflow 表格册项目把题目组、多组表头、筛选、权重、检验设置、表格顺序、波次、编码变量、发现和导出设置保存在一个有版本的模型中,而不是一组互不关联的输出文件。 Limitations: 仅定义清单刻意不含受访者数据行;只有应用到新的数据文件时才会重建分析,与该文件不匹配的定义会被列出并跳过。 Primary page: https://www.crosstabs.com/zh-cn/survey-tab-book Evidence: https://www.crosstabs.com/survey-tab-book Evidence: https://www.crosstabs.com/local-mode Evidence: https://www.crosstabs.com/data-processing ## 问卷数据的线性回归与设计型回归 Stable ID: zh-survey-regression Locale: zh-CN Topic cluster: survey_workflow 线性回归为数值结果给出每个预测变量的 B;声明抽样权重会改变点估计,但普通回归的标准误在抽样设计下不成立,设计型推断需要复杂抽样 GLM 的 Taylor 线性化标准误和设计自由度。 Formula: y = b₀ + Σ b_j x_j + e;设计自由度 = PSU 数 − 分层数。 Limitations: 统计模式中声明抽样权重时只给出加权点估计,不显示检验、标准误和区间,并拒绝逐步法。 横截面回归只能控制已测量的变量,不能证明因果关系。 Primary page: https://www.crosstabs.com/zh-cn/survey-regression-guide Evidence: https://www.crosstabs.com/survey-regression-guide Evidence: https://www.crosstabs.com/spss/complex-samples-glm Evidence: https://www.crosstabs.com/methods