This piece contains my reflections on message optimization research: the two families of quantitative methods, what each can and cannot recover, and why TURF has a structural answer to the more-is-better problem that choice models do not.

It also addresses the behavioral science question, though my reservation there is quite narrow and I approach it from a research design perspective more than anything else.

A recent interaction with Claude highlighted a statistical issue that I had never before considered actively in my work. At first, I was aghast at what I felt was a huge gap in how we conducted and presented market research findings. As the session continued, however, I realized the gap was not quite as stark as it appeared at first blush. What fills that gap, in practice, is domain knowledge: whether a pattern of results is consistent with what you already know or just scatters randomly as a chance finding would. In other words, the lack of statistical rigor, in this particular case, was not as damning as it first sounded.

This article walks through a nested attribute problem from a choice model for a GLP-1 product with multiple dose strengths where the standard fix doesn't work and presents a solution in the form of a Naïve Bayes classifier.

The piece includes a worked example with real numbers, and a plain-language illustration of Bayes' rule before it's applied to the actual problem. One of many illustrations that every project brings its unique challenges and requires out-of-the-box thinking to solve. I hope you find it useful.

Pharma demand assessment using choice and conjoint models is a well-travelled road, but one with many forks and pitfalls. The choice of experimental approach changes significantly depending on the competitive landscape a pipeline product is entering: how many new entrants, how well their profiles are known, how many attributes are in play, and what access hurdles are anticipated. This article maps out those decisions, from the simplest case through to the considerably more complex problem of modeling multiple entrants with uncertain profiles.

It is standard practice to calibrate choice model shares to target shares by pre-specified segments. The standard Train method works beautifully when segments are mutually exclusive. It breaks down when they overlap - you calibrate to one set of segment targets and the others drift. Here is what to do about it.

Mean-centering survey data across markets sounds like the right correction for cross-national scale bias. For segmentation, it is categorically the wrong one — it collapses genuinely different respondents into the same point in space and destroys the signal your clustering algorithm needs to find.

The geometry of why this happens, what to do instead, and why latent methods handle this more gracefully than distance-based approaches — all are addressed in this article.

Choice modeling has been the de facto tool for demand estimation in pharma for decades. Most brand teams have used it. Most research agencies have run it.

And yet it is applied on autopilot more often than it should be — commissioned because it is the expected next step rather than because someone has thought carefully about whether the design is right, whether the scenarios are realistic, and whether it is even the right tool for the question being asked.

I first wrote on this topic in 2008. My views haven't changed much — but I think they are more precisely stated now than they were. The one area where my thinking has genuinely evolved is the multi-stakeholder architecture, where I suspect the field has sometimes added complexity without a proportionate gain in usefulness.

Small samples don't hinder our ability to quantify payer decisions.

This is because Payers aren't making random choices. They operate inside a structured system with clear financial incentives and well-understood levers. When the decision environment is that structured, you don't need hundreds of respondents to detect a signal. You need a disciplined design.

Ordered choice modeling treats formulary outcomes such as tier placement, restrictions, coverage as what they actually are: a natural hierarchy. The result is a probabilistic model whose estimated coefficients become the engine for "what if" scenario simulators that access teams can use to pressure-test pricing, rebates, and contracting strategy before going to market

Chart audits feel incredibly concrete because they’re built from real patient charts, and yet we still run into discrepancies with secondary data that are hard to explain to a client. The reason is simple once you see it — the way we collect chart data quietly shapes the story it tells. And the same is true for claims. Neither one is “the truth.” They’re just different windows into a market we can’t observe directly.

We ask surveys to do things they were never built for. One example: trying to “validate” patient share against market share. It sounds reasonable, but the two measures come from different worlds and behave differently.

I wrote a short piece on what survey‑based demand actually measures, why calibration creates confusion, and how to think about validity in a way that’s more useful for real decision‑making.

If you work with HCP demand data—on either the insights or forecasting side—I think you’ll find it useful.

Demand assessment has always been a balance between clinical realism and the need to aggregate it into a credible market forecast. When I first wrote on this topic in 2007 the industry was wrestling with how to incorporate patient-level nuance without collapsing the ability to forecast at the market level. That tension hasn’t gone away — if anything, it has intensified as clinical complexity, data fragmentation, and therapeutic personalization have increased. In this article I articulate a principle that still guides my work today: methodological realism is only valuable when it preserves — rather than obscures — our ability to make reliable market-level predictions.

What sets a methodologist apart from the rest?
In this article, I share some thoughts on what makes a methodologist truly valuable to insights professionals and decision-makers.

What it takes to be an independent consultant

Consulting, not order taking

The power of simplicity in presenting data