A Nonparametric Density Estimation Method for Brand Choice Using Scanner Data

Nonparametric density estimation using a kernel method is proposed to model consumer brand choice. Recent availability of large scanner panel data allows the use of nonparametric approach, which has few or at least fewer underlying assumptions and affords greater structural flexibility. By removing as many assumptions as possible, the author constructs the “ultimate” nonparametric model, radically departing from the traditional approaches, to highlight the differences in implementation and performance. The proposed model does not involve either parameters that approximate certain distributions as in stochastic models or latent concepts such as utility as in utility maximization models. The performance criteria include prediction of market response and brand choice, share tracking, and robustness under violation of various assumptions involved in parametric choice models, such as correlated disturbance and misspecification. The method is compared with a popular parametric counterpart, the multinomial logit model, on simulated and actual scanner panel data. The paper emphasizes the conceptual importance of the nonparametric approach by discussing its advantages, limitations, and its complementary role in developing, refining, and diagnosing parametric models. This perspective affords insight to modeling philosophy and suggests the possibility of a hybrid approach.