In 1985, Green and Srinivasan, pioneers of modern conjoint analysis, published an emblematic case that revolutionized product decisions in consumer goods companies. Procter & Gamble had invested millions in the development of a detergent with advanced stain-proof formula. The internal focus groups considered it excellent, but the market test was a failure. The reason? Consumers claimed they wanted to pay more for the new formula, but at the supermarket they chose the cheapest product.
Conjoint analysis addresses this paradox: instead of asking directly what consumers want, it puts them in front of realistic choices with explicit trade-offs. Their decisions reveal real preferences, not declared.
This method is adopted by companies like Apple, Airbnb and Spotify to optimize products and prices. It is accessible: with a spreadsheet and Python you can conduct a serious study with a budget content. Let’s see how it works and how to apply it for concrete decisions.
Real problem
Traditional market questions ask how important every feature is, but without forcing to choose between alternatives. Thus all attributes are “very important” making the answers useless for practical decisions.
In reality, each purchase is a compromise: you pay more for a better feature or you save less accepting. The joint analysis simulates this process, showing options with different combinations and asking to choose. The aggregate choices reveal the real importance of each attribute.
Conceptual model
Imagine having to choose a smartphone from three options with different brands, batteries, cameras and prices. The choice you make communicates what features you value most, without needing verbal explanations.
Strict formalisation
There are various types of joint analysis:
-
**Choice-Based Conjoint (CBC):**The most popular, shows 2-4 profiles per task and asks to choose your favorite. Simulate the actual decision and allow statistical models such as the Multinomial Logit.
-
MaxDiff (Best-Worst Scaling): suitable for many attributes, asks to indicate the most and least important among a few task items.
-
Adaptive Conjoint (ACA): customize questions according to the previous answers, estimating individual preferences.
-
Rating-Based Conjoint: Evaluate each profile individually, less realistic but simple.
Example or case study
For a SaaS project management service, you choose 5 attributes with 3 levels each (price, users, integrations, support, storage). The experimental design selects an orthogonal subset of profiles to be tested to avoid correlations between attributes.
A logit model estimates part-worth utilities, i.e. the contribution of each level to perceived utility. For example, the low price has a high utility, while the high price a low.
The relative importance of attributes is calculated as the range of utilities divided by the total sum. In case, the price weighs 68%, the storage 15%, and the support only 3%.
Lab / exercise
Basic level: Define 4-5 attributes for a product of your choice, with 2-3 credible levels each.
** Intermediate level:** Designs a reduced orthogonal design for a CBC studio with at least 12 profiles.
Level research-grade: Simulate CBC data for 300 respondents, estimate a logit model and interpret part-worth utilities.
Dataset and recommended materials: Simulated Python datasets available on GitHub GinnyTech, pyDOE2 libraries for experimental design, scikit-learn for logit models.
Typical error to avoid
- Too many attributes in the CBC cause cognitive overload and random responses., Unbelievable levels distort responses., Non-representative samples invalidate results., Ignore segmentation hides different preferences., Non-orthogonal experimental designs confuse analysis.
Quiz or checkpoint
- Why is the joint analysis more reliable than direct questions about preferences?, What is the main advantage of the Choice-Based Conjoint over Rating-Based?, How is the relative importance of an attribute calculated?, What common errors can compromise a joint study?