During the most recent MBC Webinar, we discovered a bug dealing with combinatorial simulations. Specifically, if you have a dependent variable (DV: a respondent’s choice) that is conditional on another DV being selected (or not selected), then the combinatorial predictions could result in non-feasible menu combinations. For example, a respondent could be predicted to select the left automobile on the screen, but buy a sunroof for the middle (not-selected) automobile.
We have released a new update that fixes this. We also made a minor fix which involves the rare situation in which the tally of first-choice combinatorial predictions across the respondent’s draws results in a tie between two equally likely choice outcomes. We previously were randomly breaking the ties based on system clock. But, that meant that if you repeated the simulation run, a few respondents’ predicted combinatorial choices could change. We’re now seeding the breaking of ties based on respondent-unique information, so the respondent’s predicted choices will not change when you re-run a simulation.
Please download the most recent MBC beta version at: http://www.sawtoothsoftware.com/download/releases/mbc/MBC_v0_9_4.exe
This release of the software is no longer available. Click the button below to download the latest version of MBC. Download MBC
By the way, I am still favoring the idea of simulating using draws rather than point estimates for MBC. For combinatorial choice simulations, the ability to use “weighted draws” to make the combinatorial predictions lean toward those that are actually observed to be made more often by each respondent in the questionnaire, significantly improves predictions. Also, for standard “on the margin” simulations, the magnitude of the shares (scale factor) is more true to counts data (and holdouts) if using the draws rather than point estimates. Of course, in a future release of MBC, you’ll be able to adjust the scale factor in the simulator, so this issue may be addressed if using point estimates.
Stay tuned for more feature improvements…