Conjoint and MaxDiff Analysis

Is share of preference the same as market share?

No, and the difference is the most consequential misreading in the technique. A share of preference is the proportion of modelled choices going to each alternative in a set you defined, from an exercise where people chose between hypothetical profiles with full attention on the attributes shown. Real market share depends on awareness, distribution, sales effort, promotion, habit, switching costs and competitor response, none of which the model has seen. A simulated share can be a legitimate input to a forecast alongside those factors. It is not a forecast.

Why does attribute importance change when I change the levels?

Because derived importance is calculated from the range between an attribute's best and worst levels, divided by the sum of ranges across attributes. Test price from 100 to 120 and price looks unimportant. Test it from 100 to 300 and it dominates. Nothing about the market changed. This is why importance should never be reported without the range it was derived from in the same line, and why two correctly analysed studies of the same category can disagree completely.

Can I compare utilities between attributes?

No. Part-worths are typically zero-centred within each attribute on an arbitrary interval scale, so a value of 40 on colour and 40 on price have no common meaning. Only differences within an attribute are interpretable. Comparing across respondents also requires rescaling, because the scale factor absorbs how consistently a person chose rather than how strongly they preferred.

What is the red bus, blue bus problem?

It is the practical consequence of the independence of irrelevant alternatives property in logit simulation. The rule assumes the ratio of shares between two alternatives is unaffected by what else is in the set. Add a near-duplicate of an existing product and the rule splits share proportionally across everything, so the two near-identical items together gain share that in reality they would take almost entirely from each other. This makes plain logit actively misleading for line extensions and portfolio questions, which are exactly the questions clients ask most. Use a rule that handles similarity, such as randomised first choice.

What do MaxDiff scores actually tell me?

The relative order and distance of the items on the list you supplied, and nothing more. Everything on the list can be critically important or entirely trivial, and the scores look the same either way, because the exercise never asked whether any item mattered in absolute terms. If you need an absolute reading, it has to come from an anchored design fielded alongside. A low-scoring item is less preferred than the others present, not unimportant.

How many attributes and tasks can a conjoint carry?

There is no fixed number, but complexity is roughly attributes times levels times alternatives times tasks, and beyond a moderate load respondents simplify: they narrow attention to one or two attributes and ignore the rest. That biases derived importance toward whatever is easiest to process, usually price and brand. Diagnose it by estimating importance on early tasks and late tasks separately and comparing. A study with four attributes that matter beats one with nine, of which five were added for completeness.

How do I know whether a conjoint model is any good?

Fit to the data it was estimated on is not evidence of prediction. Field holdout tasks, keep them out of estimation, and report the hit rate against two benchmarks: the chance rate given the number of alternatives, and a naive rule such as always choosing the cheapest. Also report the aggregate share prediction error on the holdouts, which matters more than the hit rate when the output is a share. Where no holdouts exist, say predictive validity was not assessed.

Can a conjoint tell me what price to charge?

It can tell you how choice moves across the price range you tested, relative to the alternatives you showed, which is genuinely useful. It cannot tell you demand at a price in absolute terms, because it never observed a purchase and never presented the whole market. Read the sensitivity curve for where the response is flat and where it is steep, never extrapolate beyond the tested range, and treat any willingness-to-pay figure as stated preference with all the limits that carries.

Should I remove respondents who always chose the cheapest option?

Usually not, and the decision needs evidence. Non-trading on price is often genuine behaviour rather than inattention, and removing those respondents can move derived price importance substantially, which means the exclusion is not a cleaning decision but an analytical one. Check their response times and individual fit statistics first. Speeders whose choices are indistinguishable from random are a different case. Whatever you decide, log the criterion, the count and the effect on the headline numbers.

Can AI analyse a conjoint reliably?

It can run the estimation and the simulation consistently, which is a real advantage over hand-built spreadsheets. The specific risks are producing a utility table when no level wording was ever supplied, simulating configurations containing levels outside the tested range, running a line extension under a plain logit rule without flagging what that does, quoting importance without its range, and moving the not-a-forecast caveat to the appendix where nobody reads it. Require the design sheet before any estimate, a range check on every simulated profile, and the caveat on the same page as the number.