Sawtooth Software: The Survey Software of Choice

Hierarchical Bayes Estimation and ICE Revisited

As you have probably noticed, hierarchical Bayes (HB) has been the subject of several favorable journal articles and many presentations at the more technical market research conferences. We expect that the amazing advances in the speed of computers will hasten the adoption of hierarchical Bayes algorithms not only for conjoint problems, but for other market research applications as well.

Until now, there have been two problems with HB: it can take a lot of computer time, and user-friendly software has not been generally available. We think that we have, by in large, solved both of those problems with the CBC/HB Module we recently released. CBC/HB is easy to use and runs rapidly, so answers for medium-sized problems are available in a few hours.

Those familiar with our ICE (Individual Choice Estimation) software may wonder how HB compares to ICE. What is our position on these two pieces of software that both achieve individual-level estimates from CBC data?

At the 1998 Advanced Research Techniques (ART) conference, Joel Huber of Duke University presented results from three different studies. He and co-authors Neeraj Arora (Virginia Tech) and Rich Johnson (Chairman, Sawtooth Software) found that HB and ICE both performed about equally in terms of hit rates and share predictions for holdout choice tasks.

The three data sets that Huber et al. examined had a minimum of 18 tasks each. In the ICE documentation, we suggested that about 20 tasks or more should be available for ICE estimation. Since Huber's presentation at ART, we have seen examples involving data sets that contain less information than suggested for ICE in which ICE has performed poorly, but for which HB has performed favorably. Indeed, the superiority of HB in achieving useful individual-level estimates with as few as six choice tasks is a point of differentiation we had not previously recognized.

ICE offers at least two unique benefits. It can be significantly faster than HB for very large problems. Also, if respondents conform to the assumptions of the latent class model (high degree of homogeneity within groups with large separation between groups) and enough choice tasks are available, ICE has the potential to produce results superior to HB.

We will continue to market ICE, though for most users we will recommend CBC/HB. Those who own either ICE or CBC/HB receive a $500 discount on the purchase of the other (ICE and CBC/HB each cost $2,000).