Estimating Utilities with HB

Top  Previous  Next

Note: HB is the most commonly used utility estimation approach among our users for developing the utilities used in market simulators.  We strongly recommend you use HB estimation for your final models.  Counts and Logit are useful for obtaining quick top-line summary results.  Latent Class is especially useful for segmenting respondents.

Lighthouse Studio includes a built-in HB estimation routine.  As an alternative, you may choose to analyze data separately using our standalone CBC/HB software.

When you run HB (Analysis | Analysis Manager...  then by selecting HB as the Analysis Type then by clicking Run), the results are displayed in the report window and utilities are saved and stored in the internal database file.  You can select subsets of respondents or tasks to process.


Note: if you want to examine more advanced information such as draws and covariances, prior to estimating the utilities you must request to write out that information on the File Options dialog accessed from the CBC/HB Settings gear_blue menu.



The earliest methods for analyzing choice-based conjoint data (e.g. the 70s and 80s) usually did so by combining data across individuals (e.g. counting and aggregate logit). Although many researchers realized that aggregate analyses could obscure important aspects of the data, methods for estimating robust individual-level part-worth utilities using a reasonable number of choice sets didn't become available until the 90s.

The Latent Class Segmentation Module was offered as the first add-on to CBC in the mid-90s, permitting the discovery of groups of individuals who respond similarly to choice questions.

Landmark articles by Allenby and Ginter (1995) and Lenk, DeSarbo, Green, and Young (1996) described the estimation of individual part worths using Hierarchical Bayes (HB) models. This approach seemed extremely promising, since it could estimate reasonable individual part worths even with relatively little data from each respondent. However, it was very intensive computationally. The first applications required as much as two weeks of computational effort using the most powerful computers available to early academics!

In 1997 Sawtooth Software introduced the ICE Module for Individual Choice Estimation (now since retired), which also permitted the estimation of part worths for individuals, and was much faster than HB. In a 1997 paper describing ICE, we compared ICE solutions to those of HB, observing:


"In the next few years computers may become fast enough that Hierarchical Bayes becomes the method of choice; but until that time, ICE may be the best method available for other than very small data sets."

Over the next few years, computers indeed became faster, and our CBC/HB software soon could handle even relatively large-sized problems in an hour or less. Today, most datasets will take about 15 minutes or less for HB estimation.

HB has been described favorably in many journal articles. Its strongest point of differentiation is its ability to provide estimates of individual part worths given only a few choices by each individual. It does this by "borrowing" information from population information (means and covariances) describing the preferences of other respondents in the same dataset. Although ICE also made use of information from other individuals, HB did so more effectively and required fewer choices from each individual.

Latent Class analysis is also a valuable method for analyzing choice data. Because Latent Class can identify segments of respondents with similar preferences, it is an additional valuable method. Recent research suggests that default HB is actually faster for researchers to use than LC, when one considers the decisions that should be made to fine-tune Latent Class models and select an appropriate number of classes to use (McCullough 2009)

Our software estimates an HB model using a Monte Carlo Markov Chain algorithm.

We at Sawtooth Software are not experts in Bayesian data analysis. In producing this software we have been helped by several sources listed in the References. We have benefited particularly from the materials provided by Professor Greg Allenby in connection with his tutorials at the American Marketing Association's Advanced Research Technique Forum, and from correspondences with Professor Peter Lenk.

For more information about HB estimation, please see the CBC/HB Manual.

Page link: