Model Selection in Panel Data Models: A Generalization of the Vuong Test

This paper generalizes the classical Vuong (1989) test to panel data models by employing modified profile likelihoods and the Kullback-Leibler information criterion. Unlike the standard likelihood function, the profile likelihood lacks certain regular properties, making modification necessary. We adopt a generalized panel data framework that incorporates group fixed effects for time and individual pairs, rather than traditional individual fixed effects. Applications of our approach include linear models with non-nested specifications of individual-time effects.

💡 Research Summary

The paper “Model Selection in Panel Data Models: A Generalization of the Vuong Test” extends the classic Vuong (1989) likelihood‑ratio test to the context of panel data, where the presence of individual and time fixed effects creates an incidental‑parameters problem that invalidates the standard properties of the likelihood function. The authors propose a framework that replaces the usual individual fixed‑effects specification with a more flexible grouping structure: each observation is assigned to a group for the individual dimension (via a function g(·)) and a group for the time dimension (via a function m(·)). This allows for both the traditional “one‑individual‑one‑group” case and more general clustered fixed‑effects specifications, thereby encompassing a wide class of linear and nonlinear panel models.

The central object of the Vuong test is the Kullback–Leibler (KL) distance between the true data‑generating process and each candidate model. In a cross‑sectional setting, the Vuong statistic is based on the quasi‑likelihood ratio (QLR) = √(n) · (ℓ̂₁ − ℓ̂₂), where ℓ̂ⱼ denotes the maximized log‑likelihood for model j. In panel data, however, the log‑likelihood ℓⱼ is biased because the incidental parameters (the fixed effects) are estimated with error that does not vanish as n → ∞ when T is fixed or grows slowly. The authors derive an explicit first‑order bias term for ℓⱼ that takes the form (2⁻¹) ∑ᵢ Ψ²_{γ,i}, where Ψ_{γ,i} is a function of the score for the individual fixed effect γᵢ. This bias is of the same order as the stochastic fluctuation of the log‑likelihood itself, so it must be removed before a valid test can be constructed.

To correct the bias, the paper introduces a “modified profile likelihood” LM*ⱼ = ℓ̂ⱼ − \hat{B}_j, where \hat{B}j is a consistent estimator of the bias term derived from the sample analog of Ψ²{γ,i}. The modified Vuong statistic becomes

 QLR* = (nT)^{‑½} · (LM₁ − LM₂).

The authors then analyze the asymptotic distribution of QLR* under three scenarios: (i) the two models are non‑nested, (ii) one model is nested within the other, and (iii) the models overlap (share some but not all parameters). In the non‑nested case, both the stochastic component of the log‑likelihood and the bias‑correction component contribute to the variance. In the nested case, the stochastic components cancel, leaving the bias‑correction term as the dominant source of variation. The paper provides rigorous proofs (Theorem 1 for the infeasible, bias‑corrected statistic; Theorem 5 for the feasible version) that QLR* converges to a normal distribution with a variance that can be consistently estimated using a cluster‑robust covariance matrix.

A practical algorithm is outlined in Sections 2.5–2.6. First, each model’s parameters are estimated by maximizing the original (unmodified) likelihood. Second, the bias term is estimated by averaging the squared, centered individual‑score contributions across individuals and time periods. Third, the modified log‑likelihoods are computed, and the variance of QLR* is obtained by a sandwich estimator that accounts for within‑individual and within‑time correlation. The resulting test statistic is straightforward to implement in standard statistical software, requiring only a few additional lines of code to compute the bias correction and robust variance.

The paper also supplies a general asymptotic theory for panel estimators with both individual and time fixed effects (Appendix C). This theory relaxes the global concavity assumption used in Fernández‑Val and Weidner (2016) and instead imposes restrictions on the grouping structure, thereby broadening the applicability of the results to a wider class of nonlinear panel models.

Although the empirical illustrations are not reproduced in the excerpt, the authors claim that Monte‑Carlo simulations demonstrate substantial improvements in size control and power relative to the unadjusted Vuong test, especially when n is large and T is moderate—a regime common in modern micro‑panel datasets.

In summary, the paper makes three major contributions: (1) it identifies and quantifies the first‑order bias in panel log‑likelihoods caused by incidental parameters; (2) it constructs a bias‑corrected Vuong statistic that remains valid for both nested and non‑nested model comparisons; and (3) it provides a feasible estimation procedure and asymptotic justification that can be readily applied by practitioners. This work thus fills a notable gap in the econometric literature on model selection for high‑dimensional panel data, offering a theoretically sound and practically implementable tool for researchers confronting competing specifications of individual‑time heterogeneity.