About Hierarchical Bayes for CBC
Researchers and academics have argued that respondents are unique. The world does not consist of clean market segments, and aggregate models that neglect respondent differences cannot be optimal. The Latent Class Module takes an important step toward recognizing respondent heterogeneity, but stops short of achieving strong individual-level predictions.
The generally preferred method for analyzing CBC data is hierarchical Bayes (HB) estimation. Importantly, HB develops individual-level part worth from choice data. Having individual-level estimates improves the accuracy of market simulations and leads to better understanding of market structure and attribute importances than aggregate logit modeling.
The CBC/HB module leverages information from all respondents to estimate results for each individual. The individual-level part worths are estimated by a statistical simulation technique called Gibbs Sampling. HB uses each individual's choices along with information about the distribution of part worths for all respondents to estimate individual-level parameters. If the market is truly segmented, separating respondents first into groups using Latent Class analysis and then running HB within those groups might be a useful approach (especially if your sample size is relatively large), though many attempts to do this have demonstrated that HB already works very well, even if the population includes mixtures of relatively distinct segments.
The analysis of many data sets has confirmed that HB generally improves predictions for holdout concepts relative to aggregate logit and Latent Class. A long-standing complaint against aggregate logit has been its IIA assumption, often referred to as the red bus/blue bus problem. Very similar products tend to capture too much net share in competitive aggregate logit simulations. With individual-level modeling, this effect is less problematic. If you use the First Choice model, it is entirely resolved.
The CBC/HB Module features estimation of individual-level part worths or linear functions (main effects and/or first-order interactions) for standard CBC, best-worst CBC, or allocation-based (constant-sum) CBC questionnaires. Typical run times for market research data sets are from 5 minutes to an hour. CBC/HB reads and writes data to/from text-only files. You do not need to have used CBC to collect the data as long as you arrange the data in the CBC-compatible text-only format. The part worths from CBC/HB can be used within Sawtooth Software's market simulator.
The CBC/HB Module benefits from a fast computer and plenty of available memory.