Keane (1997) Modeling Heterogeneity and State Dependence in Consumer Choice Behavior

Summary

This paper confronts the classic problem of disentangling preference heterogeneity from true state dependence as competing explanations for the well-documented persistence in consumers' brand choices. Using Nielsen scanner panel data on ketchup purchases, Keane estimates a multinomial-probit brand-choice model that simultaneously admits a rich heterogeneity structure (serially correlated and random-effect errors) and a genuine state-dependence (lagged-choice) term — estimated by the method of simulated moments with the GHK simulator he had developed for panel LDV models. He finds evidence of true state dependence even after controlling for rich heterogeneity, but simulations show the long-run effect of a promotion-induced purchase on future purchase probabilities is positive and small.

Key Claims

• Persistence has two explanations. That a consumer who bought brand A last period is more likely to buy A this period is consistent with two opposite behaviors: heterogeneity — the consumer simply has a stable preference for A (so the correlation is spurious state dependence) — or true state dependence — buying A raises the utility of A next time (habit, loyalty, switching costs). Distinguishing them is Heckman's (1981) identification problem.
• Model both at once. Most prior choice models allowed either heterogeneity or state dependence; Keane's specification includes both — a flexible error structure (random effects plus serial correlation) to capture heterogeneity and a lagged-choice term to capture true state dependence — so their relative importance can be estimated rather than assumed.
• Estimation by MSM + GHK. The high-dimensional panel multinomial-probit choice probabilities are handled by the method of simulated moments using the recursive GHK simulator (Keane 1994), making the rich error structure computationally feasible.
• Both matter, state dependence is real. In the ketchup data, both heterogeneity and true state dependence are statistically significant; controlling for a rich heterogeneity structure does not eliminate the evidence for genuine state dependence.
• But the dynamic marketing effect is small. Simulating the estimated model, a purchase induced by a temporary promotion raises future purchase probabilities only slightly and briefly — the long-run "loyalty" payoff of buying share through promotions is limited.

Concepts Introduced or Extended

• Discrete Choice Models
• Consumer Heterogeneity
• Multinomial Probit

Entities Mentioned

• Michael P. Keane

Quotes

"The fact that brand choices of consumers exhibit persistence is consistent with two diametrically opposed patterns of consumer behavior."

"I find evidence for true state dependence in the choice process, even after controlling for a rich heterogeneity structure."

My Take

The methodological point outlives the ketchup: observed persistence is an analyst's description of the data, not a behavioral fact, and the same pattern can come from stable tastes (heterogeneity) or from history genuinely changing preferences (state dependence) — two mechanisms with opposite policy implications. The contribution is doing what earlier work ducked — putting both in one model and letting the data apportion — which is only feasible because the GHK/MSM machinery from Keane (1994) makes the panel multinomial-probit tractable. For the wiki it grounds the discrete-choice and consumer-heterogeneity pages in a concrete identification problem, and its "small long-run promotion effect" is a nice empirical caution against overstating brand-loyalty dynamics. The caveat is the usual one for this literature: the heterogeneity/state-dependence split is identified largely off functional-form and error-structure assumptions, so the conclusion is conditional on the specification being rich enough.