UX Researcher @ Perceptual User Experience Lab | Human-AI Interaction Researcher @ University of Arkansas at Little Rock
Most customer research still asks people to rate products, yet in the real world people choose one option from a small set. Choice Based Conjoint turns that reality into data by showing repeated sets of plausible products and recording the selected alternative. Each set is a clean trade off. By varying brand size and price over tasks you learn how people swap one thing for another when they must pick one. A practical workflow starts with scoping the decision. List the attributes that actually move choice and keep levels realistic. Include a none option so people can walk away. Build 8 to 12 choice sets per respondent with 3 or 4 alternatives per set. Randomize order and avoid obviously dominated options. In your data, store one row per alternative per task per respondent with a chosen flag, and keep factors coded with clear base levels. Begin estimation with a multinomial logit. Treat price as continuous so the sign should be negative. Inspect the output. Coefficients are part worths on the logit scale and are interpreted relative to the base level. Signs tell direction and larger absolute values mean stronger effects. Standard errors and z values tell you which effects are clearly different from zero. Before trusting results, run quick checks that each task has exactly one chosen alternative and that there is no strong position bias. Turn coefficients into business language. Odds ratios show how much a level raises or lowers the odds of choice. To talk money, divide an attribute effect by the absolute price slope to get a rough willingness to pay. Then simulate. Create a table with the alternatives you plan to sell, compute utilities, convert to predicted shares, and compare scenarios such as a small price cut or a feature change. Most audiences care more about these simulations than the coefficient table. People do not all think alike, so model heterogeneity when the stakes warrant it. Mixed logit lets coefficients vary across people. The model reports mean effects and standard deviations of those effects. When a standard deviation is similar to or larger than the mean you likely have preference reversals in the population and that is a signal to consider multiple variants in the lineup or targeted offers. When respondents answer only a handful of tasks, reach for Hierarchical Bayes. HB shrinks noisy individual estimates toward a population distribution so you recover stable person level utilities without needing long surveys. You also get posterior draws that let you show uncertainty bands around shares and willingness to pay. Common pitfalls are easy to avoid Do not compute per dollar willingness to pay if you dummy coded price levels Do not mix long and wide formats in the same pipeline Do not overload the design with too many attributes for a fixed survey length Do not report aggregate means alone when the mixed logit says heterogeneity is large Tie every finding to a simulated shelf or price test and make a concrete recommendation.