MaxDiff (Best-Worst Scaling) Analyzer

Analysis Tool Conjoint Tool →

Ask people to rate ten things and they will call nine of them important. Ask them to pick a best and a worst, and they have to choose. This tool turns those forced choices into a ranked list of priorities on a 0–100 scale that sums to 100 — so you can say one item is worth twice another, see which gaps are real and which are a statistical dead heat, and find the people who disagree with the average.

METHOD OVERVIEW & EQUATIONS

MaxDiff, or best-worst scaling, replaces the question “how important is each of these?” with a question that has a scarce answer: “of these four, which is most important and which is least?” Nobody can call everything important, because naming a best requires naming a loser.

A set of $a$ items is modelled as two linked choices. The best choice is a logit over the whole set; the worst choice is a logit over what remains, on negated utility, because picking the worst item is choosing the most attractive option under a reversed preference order:

$$P(\text{best} = b) = \frac{e^{u_b}}{\sum_{j \in S} e^{u_j}} \qquad P(\text{worst} = w \mid b) = \frac{e^{-u_w}}{\sum_{j \in S \setminus \{b\}} e^{-u_j}}$$

Utilities are reported on a 0–100 scale that sums to 100, which is what makes MaxDiff output usable by someone allocating a budget:

$$\text{score}_i = \frac{e^{u_i}}{e^{u_i} + a - 1} \times \frac{100}{\sum_j \left( e^{u_j} / (e^{u_j} + a - 1) \right)}$$

Key Concepts

Why best-worst beats a rating scale: a rating is denominated in a private currency. Generous respondents live in the 8–10 band and severe respondents never pass 6, so averaging ratings across people adds up quantities that were never in the same units. A best-worst answer records only which item beat which, which has no scale to misuse.
Information per screen: one set of $a$ items yields $a - 1$ comparisons from the best choice plus $a - 2$ from the worst, so $2a - 3$ in total. A set of five gives seven comparisons where “pick your favourite” would give four.
Ratio-scaled: because the scores sum to 100, an item at 20 really is worth about twice an item at 10. Raw utilities are log-odds and do not have that property.
Counting analysis: best count minus worst count, over times shown. The napkin version. It should broadly agree with the model; where it does not, an item probably appeared in unusually strong or weak company.
Relative, never absolute: the scores always sum to 100, so they describe how items rank against each other. Run MaxDiff on ten things nobody wants and you still get a confident-looking winner.

What your data file needs to look like

Long format: one row per item, per set, per respondent.

respondent_id	task_id	item	best	worst
R001	1	No monthly fees, ever	1	0
R001	1	No minimum balance	0	0
R001	1	Credit score monitoring	0	1
R001	2	…	…	…

best and worst are indicator columns — exactly one row per set flagged for each. 1, TRUE and yes are all accepted. Any set that is incomplete or self-contradictory is skipped and reported with a reason rather than quietly dropped.

When to use MaxDiff instead of conjoint

MaxDiff ranks a list of separate items: messages, claims, features, brand associations, benefits, barriers. There is no product being configured.
Conjoint models a product made of attributes that vary together, so it can simulate market share and compute willingness to pay. MaxDiff cannot do either, because nothing has a price and nothing is being built.
Rule of thumb: if the answer you want is a priority list, use MaxDiff. If it is a configuration or a price, use conjoint.

STEP 1: LOAD DATA

Select a marketing scenario:

Select a scenario above to load a preset MaxDiff dataset with context.

Drag & drop your MaxDiff file

Needs respondent, task, item, best and worst columns

No file loaded.

STEP 4: RESULTS

Item scores (0–100, sums to 100)

Error bars are 95% confidence intervals for each item on its own. Resist reading a ranking off overlapping bars in either direction — overlapping intervals do not mean two items are tied, because the gap between them has its own, usually much smaller, standard error. The p vs above column in the table below is the test that actually settles it.

Model scores vs. counting analysis

The counting score is simply best-minus-worst over times shown. It needs no model at all. If the two orderings disagree sharply for an item, that item probably kept appearing alongside unusually strong or unusually weak company.

Full item table

p vs above tests whether each item really outranks the one directly above it. Rows marked too close to call are pairs the data cannot separate at the 5% level, so the ranking between those two is not a finding you should defend.

Rank	Item	Score	95% CI	p vs above	Utility	Times shown	Best	Worst	B−W

APA-Style Report

Managerial Interpretation

Ready for the next step? These scores tell you which items people want most, but not which set to launch: the top three often appeal to the very same people. Open in TURF Analysis and upload the respondent file as-is to find the set that reaches the most people.

STEP 5: FIND SEGMENTS

An aggregate MaxDiff ranking is an average, and an average of two groups who want opposite things describes nobody. This clusters respondents on their individual item utilities to see whether the headline ranking is hiding a disagreement.

Number of segments

MaxDiff (Best-Worst Scaling) Analyzer

👨‍🏫 Professor Mode: Guided Learning Experience

METHOD OVERVIEW & EQUATIONS

STEP 1: LOAD DATA

Use a Case Study

Upload Your Own Data

STEP 2: MAP COLUMNS