Alleviating the constant stochastic variance assumption in decision research: Theory, measurement, and experimental test

Analysts often rely on methods that presume constant stochastic variance, even though its degree can differ markedly across experimental and field settings. This reliance can lead to misestimation of effect sizes or unjustified theoretical or behavioral inferences. Classic utility-based discrete-choice theory makes sharp, testable predictions about how observed choice patterns should change when stochastic variance differs across items, brands, or conditions. We derive and examine the implications of assuming constant stochastic variance for choices made under different conditions or at different times, in particular, whether substantive effects can arise purely as artifacts. These implications are tested via an experiment designed to isolate the effects of stochastic variation in choice behavior. Results strongly suggest that the stochastic component should be carefully modeled to differ across both available brands and temporal conditions, and that its variance may be relatively greater for choices made for the future. The experimental design controls for several alternative mechanisms (e.g., flexibility seeking), and a series of related models suggest that several econometrically detectable explanations like correlated error, state dependence, and variety seeking add no explanatory power. A series of simulations argues for appropriate flexibility in discrete-choice specification when attempting to detect temporal stochastic inflation effects.