Nonparametric Identification of Finite Mixture Models of Dynamic Discrete Choices

Citation

• Authors: Hiroyuki Kasahara, Katsumi Shimotsu
• Year: 2009
• Venue: Econometrica 77(1), 135-175
• arXiv: None
• URL: https://doi.org/10.3982/ECTA6763

One-Line Takeaway

Kasahara and Shimotsu show that finite mixtures of dynamic discrete-choice models can be nonparametrically identified from panel structure, covariate variation, and exclusion-style heterogeneity, without committing to a low-dimensional parametric specification for each latent type.

Technical Summary

• Setup: the data are panel choices $(Y_t, X_t)_{t=1}^{T}$ generated by a finite latent-type variable $Z \in \{1,\ldots,K\}$, with type-specific dynamic choice behavior and unknown type probabilities $P(Z=k)$.
• Identification mechanism: nonparametric recovery uses repeated time structure, covariate support, and exclusion-style variation so that different latent types induce distinguishable conditional choice laws across periods.
• Abstract-level theorem content: the identifiable object is the finite mixing distribution over types together with the type-specific choice components; the sufficient conditions depend on panel length $T$, the support size of covariates, and whether state dependence is present, with simple regimes already identified at $T=3$ and stronger dynamic regimes requiring longer panels such as $T \ge 6$.

Assumptions And Limits

The results rely on panel or repeated-choice structure and on sufficiently informative covariate variation across latent types. This is not generic one-shot mixture identifiability; the identification comes from temporal and exclusion-style structure in the observed data.

Relevance To This Project

This is directly relevant to the repo's x -> z -> y framing because it is a clean example of restoring latent-type identifiability by enriching the observation process. If single outputs do not identify latent strategies, repeated interactions, covariate interventions, or exclusion variables may make the latent routing variable identifiable even when the component laws are nonparametric.

Comparison Notes

Compared with Allman et al. (2009), the extra structure is panel dynamics and exclusion rather than conditional independence across simultaneous views. Compared with Vandermeulen and Scott (2019), the paper uses economically meaningful covariate structure rather than grouped-sample symmetry.

Related Notes

• Background on identification and learning of mixture distributions
• Identifiability of parameters in latent structure models with many observed variables
• An Operator Theoretic Approach to Nonparametric Mixture Models