Skip to content
 

Software for multilevel conjoint analysis in marketing

Someone writes:

The CBC-HB and HB-Reg programs produced by Sawtooth Software are quite popular among marketing researchers and, essentially, introduced hierarchical Bayes to the marketing research community. They have been around for nearly 20 years. More recent versions I don’t have offer jackknifing, and there have been other enhancements. I’m not sure how well-known Sawtooth is outside of marketing research, though, and you might not have heard of them.

About 10 years ago they allowed the user to specify covariates at the upper level of the model to reduce shrinkage to the mean. So the user would not specify different priors for men and women or respondents aged 20s, 30s, 40s, for example, but the covariates in theory at least help us better account for respondent heterogeneity.

More importantly, equations for each individual case (e.g., respondent in a survey) are saved in CSV. They are the means of the draws for each respondent at the lower level of the model. These “equations” can be merged with the original data file for segmentation with cluster analysis or post hoc cross tabs by different kinds of respondents, e.g., income group or purchase frequency.

Sawtooth specializes in what they have dubbed choice-based conjoint, which is an extension of McFadden’s discrete choice model and earlier ratings based conjoint. For your reference, I’ve attached a brief article on “conjoint” as it’s usually called. The forecasts are really what if? simulations in which various product features are varied (e.g., price) to estimate what the impact on share of preference would be.

I was curious if you have heard of anyone using Stan in these ways.

My reply: I’ve never heard of this particular software called Sawtooth. In Stan, there’s this Prophet package for forecasting, developed by Sean Taylor when he was at Facebook but freely available. See also this description I found on the web.

More generally, if people can use Stan to fit more flexible models, maybe starting with something like Prophet that has existing models and then using this as a springboard to building their own custom models, that would be great. We’re also fine with Stan being used within commercial software. Stan and CmdStan have the business-friendly BSD license.

6 Comments

  1. Franz says:

    No familiarity with Sawtooth here, but there may be some resources out there as a starting point:

    (1) Jim Savage has a series of articles on discrete choice models in Stan:
    http://khakieconomics.github.io/2019/03/17/Logit-models-of-discrete-choice.html

    (2) The brms package has some hacks to implement some support for conditional logit models, described here: https://github.com/paul-buerkner/brms/issues/560
    Results can be compared with mlogit package: https://cran.r-project.org/web/packages/mlogit/vignettes/e1mlogit.html

  2. Nick HK says:

    Among R packages, I believe ChoiceModelR is the closest to what Sawtooth is doing. There’s also flipChoice which I understand uses Stan for this purpose, although I haven’t used it myself.

  3. Will Marble says:

    We used Stan to do multilevel conjoint analysis in a political science context in this paper: https://williammarble.co/docs/CityLimits-July2019.pdf

  4. Doug says:

    Within business school marketing programs, there are faculty versed in conjoint, Sawtooth, and Stan. Greg Allenby at Ohio State, Elea Feit at Drexel are a couple that come to mind. At Columbia, maybe Oded Netzer?

  5. Adam says:

    I know Jim Savage has put together a little snippet showing why Sawtooth-style maxdiff doesn’t work: https://discourse.mc-stan.org/t/post-on-ranked-random-coefficients-logit/7136/3

Leave a Reply to Will Marble