Common mistake 1 with Conjoint analysis

Reading a pooled conjoint as if it described a person. It describes an average that may not exist - where the real question is which segments trade off differently, estimate per respondent and cluster the part-worths.

Common mistake 2 with Conjoint analysis

Treating simulated shares as forecast market shares. They are shares of preference among the profiles you listed, with no distribution, no awareness and no competitor response in them.

Common mistake 3 with Conjoint analysis

Showing levels nobody would believe. An implausible combination makes respondents disengage, and the damage lands on every profile they rated afterwards.

Common mistake 4 with Conjoint analysis

Using a numeric attribute as a number. Price entered as a continuous regressor assumes linear utility; entered as levels it can show the threshold effects that are usually the interesting part.

EconStatsCraft › Conjoint analysis

Conjoint analysis

The technique for "what would people give up to get this". Respondents rate whole product profiles, and the utility of every attribute level is recovered from the trade-offs they made - which is why it beats asking people what matters to them, since everyone says everything matters.

What it needs

Rating or ranking - What each respondent gave each profile.
Profile attributes - Brand, price, size - one column per attribute, holding its levels.

The intuition

Ask what people want and they want everything: more features, higher quality, lower price. Conjoint forces the trade-off instead. Show whole products that differ on several attributes at once and ask which they would choose, and the choices reveal how much each attribute is worth relative to the others - including how much money a feature is worth, which is the number a product manager actually needs.

When to use it

Product design, pricing, and feature prioritisation - any question of the form "which of these would they give up to get that". It is the standard technique in marketing research precisely because direct questioning fails: ask what matters and everyone says everything matters.

Where the estimator comes from

Each profile is a bundle of attribute levels, and utility is additive in part-worths: u_{ij}=\sum_a\beta_{a,\ell(j,a)}+\varepsilon_{ij} where \ell(j,a) is the level of attribute a in profile j. With extreme-value errors the choice probability is multinomial logit, P(j)=\exp(u_j)/\sum_k\exp(u_k), and the part-worths are estimated by maximum likelihood on the choice data. Only differences within an attribute are identified, so one level per attribute is normalised to zero. WILLINGNESS TO PAY follows from the ratio of a part-worth to the price coefficient: \mathrm{WTP}_{a,\ell}=-\frac{\beta_{a,\ell}}{\beta_{\text{price}}}, converting utility into currency because price is the one attribute already measured in it.

The model

U(\text{profile}) = \sum_{a} \beta_{a,\ell(a)} \quad\text{with}\quad \sum_{\ell} \beta_{a,\ell} = 0

What it assumes

That the profiles are a designed set. An orthogonal or near-orthogonal design is what makes the attributes separable; if price and brand always appeared together in the same combinations, no method can tell their effects apart.
Additivity - the utility of a profile is the sum of its parts, with no interaction. A cheap luxury brand may be worth less than the sum suggests, and a full-profile conjoint fitted without interaction terms cannot see that.
That the respondent traded off honestly across a manageable number of profiles. Beyond roughly 20 profiles people simplify, and they simplify by ignoring attributes - which shows up as a spuriously flat part-worth.
The attribute levels span a realistic range. Willingness to pay is computed from the price coefficient, so a price range that is too narrow inflates every money figure and one that is too wide flattens them.

Statistical properties

The design must be orthogonal or near-orthogonal, or attributes are confounded and part-worths cannot be separated. A design where the premium finish only ever appears with the high price cannot tell them apart.
Part-worths are identified only within an attribute, relative to the omitted level. Comparing a part-worth from one attribute against one from another is meaningless without a common reference such as price.
Attribute IMPORTANCE is the range of part-worths within an attribute divided by the total range across attributes. It depends entirely on which levels you chose to show: widen the price range and price becomes more important, by construction.
Willingness to pay is a ratio of two estimates, so its sampling distribution is heavy-tailed and its standard error is unreliable when the price coefficient is imprecise. Use the delta method or bootstrap, and be sceptical of tight intervals.
Aggregate logit assumes everyone has the same part-worths. Hierarchical Bayes estimates them per respondent and is standard practice, because segments with opposite preferences average to indifference.
Stated choice is hypothetical. Respondents are systematically less price-sensitive in a survey than at a till, so WTP from conjoint is usually an overstatement.

How to read the output

Within an attribute, only DIFFERENCES matter - the part-worths are centred to sum to zero, so the sign says better or worse than an average profile.
Never compare a utility across attributes. That is what importance is for: each attribute’s utility range as a share of the total range.
Importance depends on the levels you showed. Widen the price range and price becomes more important - a fact about your design, not about the market.
Willingness to pay is the utility converted at the price slope. It is only valid inside the price range you tested; extrapolating it is how conjoint produces confident nonsense.

A worked reading

A design over brand, price and battery life, 1,080 profiles, R² = 0.7369. Price carries the most weight (a utility range of 2.4195, 42.0% importance), then brand (1.7368, 30.1%), then battery life (1.6100, 27.9%). The part-worths convert to money through the price coefficient: the Aurora brand is worth 139.56 against the average, Kestrel is worth -147.59, and going from 8 to 16 hours of battery is worth 133.54 - so on this data the brand name and the battery are worth about the same to a buyer. That conversion is what the whole exercise exists to produce, and it is a RATIO of two estimates rather than one, so its sampling distribution is not symmetric and its standard error is not the sum of two standard errors. Quote a range rather than 139.56, and get it from a bootstrap over respondents rather than the delta method if the price coefficient is at all imprecise. The share simulator at the bottom is the other half: utilities become shares only through the logit transform, so a profile with 40% more utility does not take 40% more of the market.

Symptom → diagnosis → what to do

If you see this	It means	Do this
Price coefficient positive	Confounded design, or a quality signal	Check the design before believing it
One attribute with all the importance	Its level range is much wider than the others	Importance is design-dependent; say so
WTP with an enormous confidence interval	An imprecise price coefficient in the denominator	Normal for a ratio; report the interval
Part-worths not monotonic in an ordered attribute	Noise, or a genuine sweet spot	Check the sample size per level
Aggregate results showing indifference	Opposite segments cancelling	Estimate individual-level part-worths
Predicted shares far from market shares	Hypothetical bias, and no availability or distribution in the model	Calibrate to known shares before forecasting

Common mistakes

Reading a pooled conjoint as if it described a person. It describes an average that may not exist - where the real question is which segments trade off differently, estimate per respondent and cluster the part-worths.
Treating simulated shares as forecast market shares. They are shares of preference among the profiles you listed, with no distribution, no awareness and no competitor response in them.
Showing levels nobody would believe. An implausible combination makes respondents disengage, and the damage lands on every profile they rated afterwards.
Using a numeric attribute as a number. Price entered as a continuous regressor assumes linear utility; entered as levels it can show the threshold effects that are usually the interesting part.

Where to read this properly

Green and Rao (1971), "Conjoint Measurement for Quantifying Judgmental Data", Journal of Marketing Research 8(3).
Louviere, Hensher and Swait, Stated Choice Methods (2000).
Orme, Getting Started with Conjoint Analysis (2019).
Train, Discrete Choice Methods with Simulation (2009), ch. 6, on hierarchical Bayes.

Related Marketing research

MaxDiff (best–worst scaling) · TURF - reach of a product line · Price sensitivity (Van Westendorp) · Market basket analysis · Perceptual map · Scale reliability (Cronbach’s alpha) · Key driver analysis

All analyses

Data & variables

Each example is simulated from a known model, so you can check whether the estimator recovers the truth - and nothing here is anyone's private data.

Filter variables

Drag a variable into a slot on the right - or just click it.

Choose an example dataset above to begin.

Choose a transformation - it opens in the data editor, with a live preview.

Apply to

Multiply by

Columns ctrl-click for several

Group by ctrl-click for several

Within groups

Statistic

Bands

How many

Cut points

Keep rows where

Value

and

Fill with

Value

Also mark which rows were filled

Rules

Base period = 100

Which part

Percent at each tail

Threshold

Survey weight optional

Direction

Names column

Values column

Largest first

Delete them, not just count them

Read this column as

Tidy the text

New name optional

EconStatsCraft rigorous, plain-spoken econometrics

Econometrics that explains itself.

Regression, panel data, time series and the whole battery of tests. Every result is followed by a verdict you can defend.

Load a dataset to see it here.

Conjoint analysis

What it needs

The intuition

When to use it

Where the estimator comes from

The model

What it assumes

Statistical properties

How to read the output

A worked reading

Symptom → diagnosis → what to do

Common mistakes

Where to read this properly

See also

Related Marketing research

All analyses

Charts

Maps

Explore

Regression

Time series - one series

Time series - several series

Panel data

Multivariate statistics

Marketing research

Causal

Finance & risk

Actuarial

Tests & breaks

Machine learning

Econometrics that explains itself.