When “Unique Effects” Aren’t Unique: Rethinking Feature Importance in Transdermal Alcohol Sensor Data

Google map

Transdermal alcohol concentration (TAC) sensors provide continuous, passive measures of alcohol exposure, creating new opportunities to study drinking behavior in daily life. But these sensors also create a familiar analytic problem: TAC curves can be summarized in many ways — including TAC levels, rates, and durations — and researchers naturally want to know which features matter most for predicting alcohol-related outcomes.

The conventional answer is to enter these correlated features into a multiple regression model and interpret the partial coefficients as evidence of which features matter most. In our TAC work, we were motivated by a simple concern: the shared alcohol signal among TAC features may be part of what makes those features important, yet multiple regression partials that shared signal out when estimating feature-specific coefficients. We then came to recognize a second problem. Because each observed feature is only an imperfect indicator of the shared signal, partial coefficients may blend common and feature-specific sources of association rather than cleanly separating them. As a result, multiple regression can obscure both a feature’s overall importance and the magnitude of its unique association.

In this talk, I use our TAC application to motivate a bifactor modeling approach for decomposing sensor-derived features into common and feature-specific components. The common factor captures variance shared across TAC features, whereas feature-specific factors capture unique information in level, rise rate, and rise duration after accounting for the common signal. This approach allows us to ask three related but distinct questions: 1) How strongly is the shared TAC signal associated with outcomes? 2) Do individual TAC features carry additional information beyond that shared signal? 3) How important is each TAC feature overall when its common and feature-specific associations are considered together?

I will describe the conceptual motivation, walk through the applied models and results, and discuss why bifactor regression may provide a clearer answer than observed-variable multiple regression when the goal is to interpret the relative importance of correlated behavioral sensor features.