Statistical interactions in studies of physician utilization. Promise and pitfalls

Abstract

It has been suggested that use of interactive statistical models would greatly increase the proportion of variance accounted for by studies of physician utilization. The purpose of this paper is to evaluate and describe the benefits and pitfalls of using interactive statistical models of physician utilization. The paper presents Monte Carlo simulation data and real world data to determine how much more of the variance in physician utilization can be accounted for by interactive regression models. Results indicate that adding interaction terms is unlikely to produce large increases in variance accounted for. The usefulness of interactive models is particularly low when there is substantial measurement error in the predictor variables. Other advantages and disadvantages of interactive models are discussed, including 1) improved understanding, 2) inflation of alpha, 3) sensitivity to transformations and scale of measurement, and 4) confounding of interaction effects with nonlinear effects.