Calibrating a Choice Simulator When Your Segments Overlap

There is a step in deploying a conjoint-based simulator that most practitioners take: calibrating the model to a base case target. Kenneth Train described the method decades ago — adjust the alternative specific constants iteratively by the log ratio of actual to predicted until the differences vanish. It works cleanly, it converges fast, and most simulator builders implement it as part of their process.

This works well, until the segments overlap.

The problem of overlapping target segments

The standard calibration procedure assumes your target shares map to mutually exclusive groups. Every respondent belongs to exactly one group, you calibrate one constant per group, and everything converges simultaneously. Gastroenterologists and primary care physicians, for example — clean, exhaustive, no overlap.

But pharmaceutical market research rarely stops there. The same study that segments by specialty also segments by prescribing behavior: loyalists, switchers, prospects. And by practice setting. And by volume tier. These segmentations cut across each other: a gastroenterologist can be a loyalist or a prospect; a switcher can practice in an academic center or in a community setting.

In this scenario, when you calibrate to one set of targets, you displace the others. There is no set of group-level constant adjustments that can simultaneously satisfy overlapping constraints. You pick your anchor segmentation, accept the residual error in the rest, and move on. Most of us have done exactly this, accepting the residual error as the cost of making unreasonable demands on the model.

The solution to calibrating across overlapping segments

The solution is conceptually simple, if practically complicated: instead of calibrating to group-level targets, calibrate to individual-level targets.

If every respondent's predicted share matches their personal target, then any aggregate grouping — however defined, however overlapping — will automatically match its implied target by construction. A group mean is just a weighted average of individually calibrated respondents. Get the individuals right and the groups take care of themselves.

This requires two things. First, respondent-level target shares for each product, for each person in the study. Second, one correction factor per respondent rather than one per group.

Where do individual-level targets come from?

This is where the pharma context makes the method not just possible but natural.

The correct calibration target in a patient-allocation-based choice model is not actual market share. Conjoint-based patient shares and real-world market shares are measuring different constructs — one reflects relative valuations under survey conditions, the other reflects actual prescribing behavior in the market. Calibrating one to the other is a category error. The right anchor is a set of respondent-level corrected shares derived from stated likelihood responses in the same survey — one for each product, corrected for overstatement bias using an established methodology, and scaled so they sum to one at the individual level.

These corrected shares yield exactly what individual-level calibration needs: a personal target for every respondent in the study, for every product in the model. And because they come from the survey itself, they can be aggregated to any subgroup the survey supports — including behaviorally defined segments like loyalists and prospects that external market data could never provide.

What this looks like in practice

The update rule at the respondent level is the same log-ratio adjustment Train described, just applied individually. The simulator steps through respondents one at a time anyway — that is how individual-coefficients choice simulators work. Each respondent carries their own correction factor, looked up by index. The calibration loop updates each factor at every pass until the sum of squared differences across all respondents falls below a threshold.

Once converged, run the simulator once more to harvest the calibrated predictions. Every subgroup will match its implied target. You no longer have to choose which segmentation to anchor to.

The complete methodology, including worked example and implementation notes, is available in the full technical note.

Next
Next

When the Correction Becomes the Problem