Scanner Panel Analysis

The loyalty variable is an exponentially smoothed running measure of a household's past brand purchases: each purchase nudges loyalty toward the chosen brand while other brands' loyalty decays, with a smoothing constant controlling the memory length. It was important because it gave the model a compact, household-specific way to capture habit and state dependence — the strong tendency to keep buying what you bought before — without estimating a separate parameter for every household. It let a logit with common coefficients still reflect persistent individual preference, and it became a template reused throughout choice modeling, including for reference-price construction.

They use different data and answer different questions. Store-level models (such as SCAN*PRO) regress aggregate weekly unit sales per store on price and promotion to recover average elasticities and promotion lifts for the store. Scanner panel analysis instead uses individual households' purchase histories to model each choice with a logit, recovering disaggregate elasticities, loyalty, and switching at the household level. The panel approach can distinguish who responds and separate loyal from deal-prone buyers, which the aggregate model cannot; the store model is simpler and well suited to pricing and promotion planning when individual behavior is not the focus.

The chief limitation is the independence-of-irrelevant-alternatives property of plain multinomial logit, which forces unrealistic proportional substitution among brands. A second is limited heterogeneity: with coefficients common across households, differences in taste are captured only through the loyalty variable, not through varying price or promotion sensitivities. The loyalty variable can also soak up the effects of omitted factors and is sensitive to its smoothing constant and initialization. These limitations motivated successors — nested logit to relax IIA, mixed logit and hierarchical Bayes to model full individual-level heterogeneity, and latent-class models to segment — all of which build on the Guadagni-Little foundation.