Market research method
MaxDiff analysis: how best–worst scaling ranks preferences
MaxDiff (maximum difference scaling, or best–worst scaling) ranks a list of items — claims, features, messages, benefits — by showing respondents small sets of them and asking for the best and worst in each set. A logit model turns those picks into a score for every item on a common ratio scale, so you can say not only which item wins but by how much.
How results are checked
Named statistical review pending. Selected automated reference cases are published on the validation page.Analyze a MaxDiff — free, no signup
Open your best/worst picks (one row per task or one row per item shown) and get counts scores, aggregate logit utilities, individual HB utilities, probability-rescaled scores and a share-of-preference simulator. Free and browser-local.
Analyze a MaxDiff →Why best–worst instead of a rating scale
Ask people to rate ten claims on a 1–10 scale and most of them land at 7, 8 or 9: everything is “important”, and some respondents use the top of the scale for everything. MaxDiff forces a trade-off. Picking one best and one worst from five items is easy, has no scale to interpret, and is far less affected by the scale-use habits that differ between people and between countries.
Each set gives a lot of information: best and worst in a set of five imply 7 of the 10 possible pairwise comparisons. Across a balanced design every item meets every other item, and the model recovers a full ranking with distances.
The method
Formula
P(best = j | S) = e^(uⱼ) / Σₖ∈S e^(uₖ) P(worst = m | S∖{j}) = e^(−uₘ) / Σₖ∈S∖{j} e^(−uₖ)
where
- S
- the items shown in one task; j is the best pick, m the worst
- uⱼ
- the item's utility; the worst pick is a logit choice among the items left, with utility −u
- score
- probability-rescaled score: e^u / (e^u + a − 1) for a items per set, normalised so all items sum to 100
- counts
- the quick check: (times best − times worst) / times shown, from −1 to 1
Worked example
The synthetic snack-claims MaxDiff that the tool loads as sample data has 300 respondents and 10 claims. Each respondent saw 8 sets of 5 claims (2,400 tasks) and picked the most and least important claim in each. Aggregate logit results, best to worst:
| Claim | Score | Utility | Times best | Times worst | Counts score |
|---|---|---|---|---|---|
| Great taste | 22.2 | 1.38 | 590 | 27 | 0.47 |
| Low price | 14.3 | 0.63 | 355 | 101 | 0.21 |
| Natural ingredients | 14.2 | 0.63 | 357 | 110 | 0.21 |
| Low sugar | 12.6 | 0.46 | 316 | 132 | 0.15 |
| High protein | 9.2 | 0.03 | 206 | 192 | 0.01 |
| Trusted brand | 7.7 | -0.17 | 165 | 236 | -0.06 |
| Organic | 7.0 | -0.29 | 154 | 264 | -0.09 |
| Resealable pack | 5.5 | -0.58 | 109 | 335 | -0.19 |
| Recyclable packaging | 4.3 | -0.85 | 88 | 430 | -0.28 |
| Locally made | 3.0 | -1.24 | 60 | 573 | -0.43 |
Synthetic snack-claims MaxDiff · 300 respondents · the same data as “Use sample data”. Each claim was shown 1,200 times.
Great taste leads with a score of 22.2, well ahead of Low price (14.3). Because the scores are on a ratio scale, Great taste is about 7.4 times as likely to be chosen as Locally made (3.0) when each faces an average set — a statement a rating scale cannot support. The counts score gives the same order here, as it usually does with a balanced design; the logit adds distances and standard errors.
If the pack can carry only the top three claims, the simulator splits preference among them Great taste 51.4%, Low price 24.4%, Natural ingredients 24.3% — the last two are a near tie (utilities 0.63 and 0.63).
Using it well
When to use it
Use it when
- Use: Prioritising 8–30 claims, features, benefits or messages when a rating grid would not discriminate.
- Use: Comparing preference across segments or markets, where rating-scale habits differ.
- Use: Feeding a shortlist into TURF, concept tests or conjoint.
Not the right tool when
- Avoid: When items have levels that combine into products (price, brand, size): use choice-based conjoint.
- Avoid: When you need absolute appeal — MaxDiff is relative. An item can top the list and still be unappealing; add an anchor question if that matters.
- Avoid: With fewer than about six items, where a simple ranking question is cheaper.
How to interpret it
Rule of thumb
Read probability-rescaled scores as shares: a score of 20 is twice as preferred as a score of 10, and the average item scores 100 ÷ number of items. Look at individual (HB) utilities before reporting segment differences, because an average can hide two groups that want opposite things.
Designing the sets
A good design shows every item equally often, pairs every two items about equally often, and spreads each item over the positions within a set. A common rule of thumb is to show each item to each respondent about three times; with 10 items in sets of 5, that is six sets, and the sample uses 8. Several versions of the questionnaire reduce order and context effects.
The free MaxDiff design generator builds balanced versions, reports one- and two-way balance, and exports the design as CSV for your survey platform.
Counts, aggregate logit or HB?
Counts are transparent and good for a first look. The aggregate logit gives one utility per item for the whole sample with standard errors. Hierarchical Bayes estimates utilities for each respondent, which you need for segmentation, for TURF on MaxDiff scores and for honest simulator shares. The Crosstabs tool runs all three in the browser; HB runs in a background worker.
Questions
Frequently asked questions
What is the difference between MaxDiff and best–worst scaling?
In market research they are the same thing. Best–worst scaling is the academic name; MaxDiff (maximum difference scaling) is the name most survey software uses for the object-case version, where respondents pick the best and worst items from a set.
How many items and sets does a MaxDiff need?
Most studies use 8 to 30 items in sets of 4 or 5, with enough sets that each respondent sees each item about three times. More sets per respondent make individual-level (HB) utilities more stable.
What are probability-rescaled MaxDiff scores?
They convert logit utilities into the chance that an item is picked best against an average set, then normalise them so all items sum to 100. A score twice as high means about twice as preferred, which makes them easy to explain to stakeholders.
Can I analyse MaxDiff with counts only?
Yes for a quick look: (times best − times worst) / times shown ranks items well when the design is balanced. A logit or HB model is better for distances between items, standard errors, individual scores and simulations.
Is MaxDiff the same as conjoint?
No. MaxDiff ranks whole items against each other. Choice-based conjoint shows complete product profiles built from attributes and levels and estimates the value of each level, including price.
Sources and next steps
References & further reading
- [1]Louviere, J. J., Flynn, T. N., & Marley, A. A. J. (2015). Best-Worst Scaling: Theory, Methods and Applications. Cambridge University Press.
- [2]Finn, A., & Louviere, J. J. (1992). Determining the appropriate response to evidence of public concern: The case of food safety. Journal of Public Policy & Marketing, 11(2), 12–25.
Analyze a MaxDiff — free, no signup
Open your best/worst picks (one row per task or one row per item shown) and get counts scores, aggregate logit utilities, individual HB utilities, probability-rescaled scores and a share-of-preference simulator. Free and browser-local.
Analyze a MaxDiff →