Survey data quality
Survey data cleaner
How results are checked
Named statistical review pending. Selected automated reference cases are published on the validation page.Found in the sample
80of 600 rows
520 interviews kept after cleaning
26 speeders · 38 straight-liners · 4 duplicates · 17 bad open ends · 11 repeated IPs/devices · 46 failed attention checks
Survey data cleaning flags interviews that should not count: people who finished far faster than everyone else (speeders), gave the same answer to every item in a grid (straight-liners), submitted twice (duplicates), or left most of the questionnaire blank. It also runs the fraud checks usually found only in commercial survey platforms: gibberish or copy-pasted open ends, near-duplicate respondents, repeated IP addresses, device IDs or emails, and failed attention checks. Every check is a transparent, deterministic rule — no AI model — run on your file in the browser; the tool shows why each row was flagged and lets you download a cleaned file — free, with no account, and the file is never uploaded.
Clean a survey file online
Open a CSV, Excel, SPSS, Stata, SAS or JSON file (or the synthetic sample), pick the duration, grid, ID, open-text, identifier and attention-check columns, and untick any check you do not want. Every rule is reported separately with its threshold, rows caught by more than one rule are counted, and nothing is removed until you download the cleaned file — the flagged download keeps every row and adds flag_* columns so you can review the decisions in a spreadsheet.
CSV, Excel, SPSS, Stata, SAS, R, Parquet, SQLite or JSON. The file is read in this browser tab and never uploaded. For Triple-S, select the .sss or .xml metadata file and its data file together.
The eight checks
Speeders. The median interview length is a robust benchmark for how long a careful respondent needs. A row is flagged when its duration is below a chosen share of that median (33% by default) or below an absolute floor in seconds, whichever is larger. Rows with a blank or non-numeric duration are never flagged as speeders.
Straight-liners. Across the grid items you select, a row is flagged when every answered item has the same value and at least three items (configurable) were answered. The longest run of identical consecutive answers is shown as a softer secondary signal, because partial straight-lining is common in long grids.
Duplicates. A respondent ID that appears again is flagged on every repeat after the first. Exact copies of an entire row are also flagged, which catches double submissions even when the ID column was regenerated.
Incomplete interviews. A row is flagged when the share of blank cells across all columns is above the threshold (50% by default). Routed-out questions are legitimately blank, so raise the threshold for heavily filtered questionnaires.
Low-quality open ends. Each answer in the open-text columns you pick is scored with fixed rules, in this order: a stock non-answer (“n/a”, “idk”, “nothing”, “.”, “good”); no letters (“12345”); too short (fewer than 2 letters); keyboard mash (a run of 5 neighbouring keys on one QWERTY row inside any word, such as “asdfg”, or a one-word answer that is a 4-key run, such as “hjkl” — so “property” or “Bewertung” are not mash); repeated characters (one character repeated, like “kkkk”; two copies of a keyboard run or vowel-less unit, like “asdasd” or “sdsd”; or three or more copies of any unit, like “jkjkjkjk”); too few vowels (under 10% vowels across 6+ letters, or 7 consonant sounds in a row, where “sch”, “ch”, “th”, “sz” and similar clusters count as one sound and a syllabic r or l counts as a vowel, so “Rechtsschutzversicherung” and “Strč prst skrz krk” pass); and unusually low character variety (an answer of 16+ characters whose character entropy is a robust z-score of −3.5 or lower within the column and below 0.45 — short real answers such as “Mississippi” or “banana banana” are never judged by it). The Latin-script rules — keyboard mash, vowels and length — only run on answers written mainly in Latin letters, so Chinese, Japanese, Korean, Cyrillic, Arabic, Hebrew, Greek, Thai or Devanagari answers are never judged by them. Copy-paste groups are identical answers (ignoring case, spacing, punctuation and invisible characters, at least 30 characters long) given by at least 3 respondents, or 0.5% of the answers in large files — a common sign of bot farms; every member of the group is flagged. Non-answers are a soft signal: they are listed with a reason but kept in the cleaned file unless you tick the option to remove them. Low and high sensitivity presets loosen or tighten each threshold; even high sensitivity flags no word in a 236,000-word English dictionary as keyboard mash or too few vowels.
Near-duplicate respondents. Beyond exact copies, a later row is flagged when its answers match an earlier row on at least 95% (configurable) of the selected columns both rows answered; both rows must share at least half the selected columns (minimum 3), and the summary says so if you select fewer. Comparing every pair would be too slow on large files, so rows are hashed into buckets: with k columns and at most d = ⌊(1 − threshold)·k⌋ differences, d + 1 disjoint bands of columns guarantee that a true match shares a whole band; with few columns every set of k − d columns is used as a key instead. A blank cell is matched against every value seen in its column, so blanks never hide a pair; rows with too many blanks for that are compared with every row directly. Only rows that share a key are compared. If a bucket exceeds 400 rows it is compared by sorted neighbours instead, and comparisons stop at 2 million; the result then says it was capped or approximate and may have missed some pairs.
Repeated technical identifiers. For IP address, device or fingerprint ID, email or panel ID columns, each later repeat of a value is flagged per column (first occurrence kept), optionally ignoring case and spaces. Placeholders such as “N/A”, “-”, “unknown”, “0.0.0.0” or a loopback address are never counted, and IP addresses are compared in canonical form (leading zeros and ports dropped, “::1” equals “0:0:0:0:0:0:0:1”). A looser option also groups IPv4 addresses by their /24 block and IPv6 addresses by their /64. Offices, universities and mobile carriers share addresses, so treat this as a reason to look, not proof of fraud.
Attention checks. Name one or more trap questions with the answers that pass (e.g. “select Strongly agree”), as codes or value labels — “5” and “Strongly agree” both work when the file carries value labels. Separate several accepted answers with commas or semicolons, and put an answer that contains a comma in double quotes. The summary warns when an expected answer never occurs in the column. A row is flagged when it fails at least the chosen number of checks; blank answers are not judged unless you ask for that, because routing may have skipped the question.
The method
Formula
speeder if duration < max(p × median(duration), floor); straight-liner if answered ≥ m and SD(answers) = 0
where
- p
- share of the median below which an interview is too fast (default 0.33)
- floor
- optional absolute minimum duration in seconds
- m
- minimum answered grid items before a row can be judged (default 3)
- missing share
- blank cells ÷ all columns; flagged when it exceeds the threshold (default 0.50)
- near-duplicate
- identical answers ÷ columns both rows answered ≥ s (default 0.95), with at least half the selected columns answered by both
- copy-paste
- the same normalized open end (≥ 30 characters) from at least max(3, 0.5% of the answers) respondents in one column
- entropy outlier
- character entropy ÷ its maximum for the answer's length (answers of 16+ characters), with a robust z ≤ −3.5 within the column and below 0.45
Worked example
The built-in synthetic tracker has 600 rows (596 generated respondents plus 4 exact duplicate submissions). Running the defaults with duration_sec, the eight brand_grid_* items and respondent_id:
- Median duration is 520 s, so the speeder threshold is 0.33 × 520 = 171.6 s; 26 rows are faster.
- 38 rows gave the same answer to all 8 brand-grid items.
- 4 rows repeat an earlier respondent ID (each is also an exact copy of its first row).
- 0 rows are more than 50% blank.
With those four checks alone, 54 rows are flagged; 14 of them are both speeders and straight-liners, which is why the rule counts (26 + 38 + 4 = 68) add up to more than 54. The sample also has an open end (like_most_oe), ip_address, device_id and an attention_check that asks for 5. With the fraud checks on as well:
- 17 rows have a low-quality open end: 10 gibberish answers (4 keyboard mash, 3 repeated characters, 2 with too few vowels, 1 with no letters) and a 7-row copy-paste group of one sentence. The 6 non-answers are listed as a soft signal but kept. The Chinese, Japanese, Korean, Russian, Arabic and Spanish answers in the sample are not flagged.
- 4 rows are near-duplicates (28 of 28 answers identical) — the same 4 resubmissions the duplicate check found.
- 11 rows repeat an earlier IP address (8) or device ID (7); ticking the /24 option adds 5 more IPs from the address block shared by the copy-paste group.
- 46 rows fail the attention check, most of them straight-liners who carried their grid answer into it.
All eight checks together flag 80 rows, 47 of them by two or more rules, and the cleaned file keeps 520 rows; 5 of the kept rows carry only the soft non-answer signal. Also ticking “Also remove rows whose only open-end issue is a non-answer” flags 85 rows and keeps 515.
Using it well
When to use it
Use it when
- Use: Before tabulating an online panel or self-completion survey, where speeding and straight-lining are most common.
- Use: When a field partner delivers a file that may contain re-entries or test interviews.
- Use: To document a reproducible cleaning rule (thresholds and reasons per row) for a methods appendix or client report.
- Use: To screen an online panel or social-media sample for bots and survey farms: gibberish or copy-pasted open ends, near-identical answer patterns and shared IPs or devices.
Not the right tool when
- Avoid: As an automatic deletion step — review the flagged rows, especially respondents caught by only one rule.
- Avoid: On grids where uniform answers are plausible (short grids, factual items, or all-positive brands); a straight line there may be honest.
- Avoid: On interviewer-administered or CATI data where duration reflects the interviewer, not the respondent.
- Avoid: As a substitute for platform-side fraud prevention (CAPTCHA, digital fingerprinting, panel verification); these checks only see what is in the file.
How to interpret it
Rule of thumb
Treat a single flag as a reason to look, and two or more flags as strong evidence of careless responding. Report how many interviews were removed under each rule so readers can judge whether cleaning changed the sample profile.
After cleaning
Load the cleaned CSV into the Crosstabs workspace to build banner tables, or turn numeric codes into readable answers first with the Label Mapper. If your fieldwork file is an SPSS .sav, see how to open a .sav file without SPSS.
Questions
Frequently asked questions
Why compare duration to the median instead of the mean?
Interview lengths are right-skewed: a few people leave the survey open for hours, which drags the mean upward and would flag too many ordinary respondents. The median is not affected by those long tails, so a fixed share of it (commonly a third to a half) is a steadier speeding benchmark.
Is every straight-liner a careless respondent?
No. Someone who genuinely likes every brand can answer 5 to all of them. Straight-lining is more suspicious on long grids, on grids that mix positively and negatively worded items, and when it coincides with speeding. Use the longest identical run and the overlap count to decide, and consider a higher minimum of answered items for short grids.
Which duplicate is kept?
The first occurrence in file order is kept and every later repeat of the same respondent ID (or of an identical full row) is flagged. If your file is sorted differently from the order of submission, sort it by submission time before cleaning so the earliest response is the one kept.
Does the tool delete anything from my file?
No. Your original file is untouched and stays on your device. The cleaned download omits the rows flagged by the rules you left ticked (rows with only a soft signal, an open-end non-answer, are kept unless you choose to remove them); the flagged download keeps every row and adds flag_speeder, flag_straightliner, flag_duplicate and flag_incomplete columns, a flag_open_end, flag_near_duplicate, flag_identifier or flag_attention column for each newer check that ran (plus flag_rule_count), and flag_reasons.
How are blank cells handled?
Blank cells are ignored by the speeder and straight-liner rules (a blank duration is never flagged, and only answered grid items are compared). They are what the incomplete rule counts, and they are compared exactly in the full-row duplicate check. Blank open ends, identifiers and attention-check answers are skipped, and the near-duplicate rule compares only columns both rows answered.
How does the gibberish check avoid flagging other languages?
Each answer's letters are classified by Unicode script. The keyboard-mash, vowel and length rules only run when most letters are Latin, so Chinese, Japanese, Korean, Cyrillic, Arabic, Hebrew, Greek, Thai or Devanagari text is judged only by the script-neutral checks: non-answers, no letters, whole-answer repetition, entropy outliers and copy-paste. For Latin-script languages the vowel rule counts consonant clusters such as German “sch” or Polish “sz” as one sound and syllabic r, l or w as vowels, so German compounds and Czech or Welsh words such as “Strč prst skrz krk” or “cwm” are not flagged. Low sensitivity is the most conservative setting if your answers use other unusual spellings.
Is this an AI or a bot detector?
No. Every check is a fixed, documented rule that gives the same answer every time and runs in your browser; no text is sent to a model or a server. The rules catch the common signatures of careless and automated responding, but well-written AI-generated answers will not look like gibberish, so pair these checks with platform-side protections.
How large a file can the near-duplicate check handle?
It avoids comparing every pair of rows by hashing bands of columns, so it scales roughly with rows × columns: 50,000 rows × 30 columns takes about a second in a typical browser. Very large buckets of identical answers are compared with sorted neighbours only and the total is capped at 2 million comparisons; when that happens the summary says the result was capped.
Sources and next steps
References & further reading
- [1]Meade, A. W., & Craig, S. B. (2012). Identifying careless responses in survey data. Psychological Methods, 17(3), 437–455.doi.org
- [2]Curran, P. G. (2016). Methods for the detection of carelessly invalid responses in survey data. Journal of Experimental Social Psychology, 66, 4–19.doi.org
- [3]DeSimone, J. A., Harms, P. D., & DeSimone, A. J. (2015). Best practice recommendations for data screening. Journal of Organizational Behavior, 36(2), 171–181.doi.org
Try it on your own data — free, no signup
Upload a CSV, Excel, SPSS, Stata, SAS, R, Parquet or JSON file. Everything runs in your browser; your file never leaves your device.
Open the workspace →