Maximum Score Estimation of Preference Parameters for a Binary Choice Model under Uncertainty

Le-Yu Chen, Sokbae Lee, Myung Jae SungarXiv 2013 · DOI · arXiv

Abstract

This paper develops a maximum score estimation framework for preference parameters in binary choice models where decision-making is influenced by conditional expectations under uncertainty. The estimation proceeds in two stages: first, conditional expectations are estimated nonparametrically; second, these estimates are used within Manski’s (1975, 1985) maximum score estimator. The study establishes the consistency and the $N^{-1/3}$ rate of convergence for this two-stage estimator. Furthermore, it identifies sufficient conditions under which the estimator achieves asymptotic equivalence to a single-stage estimator that assumes expectations are known a priori. These findings offer significant insights for maximum score estimation involving nonparametrically generated regressors. Monte Carlo simulations further demonstrate the estimator's finite-sample performance.

Introduction

In structural economic modeling, agents frequently make binary choices—such as whether to attend college, migrate, or enter a specific industry—based on their expectations of future, uncertain outcomes. For the researcher, these expectations are typically unobserved and must be treated as nuisance parameters. Traditionally, researchers have relied on parametric models like Probit or Logit to estimate these choices. However, these models are highly sensitive to distributional assumptions; if the error terms exhibit heteroskedasticity of an unknown form, the resulting parameter estimates are often biased and inconsistent.

This paper addresses this challenge by proposing a robust semiparametric framework. By combining nonparametric first-stage estimation with a second-stage maximum score estimator, the authors provide a method that remains consistent under flexible error distributions while accounting for the uncertainty inherent in the agent's decision-making process.

The Two-Stage Estimation Framework

The proposed methodology utilizes a two-stage approach to handle the unobserved expectations and the potential for misspecification in the error distribution.

Stage 1: Nonparametric Expectation Estimation

In the first stage, the researcher estimates the agent's conditional expectations using nonparametric techniques. Because the functional form of these expectations is often unknown, the paper suggests using flexible tools such as kernel regression (e.g., Nadaraya-Watson) or series estimators. These estimated values serve as "generated regressors" for the subsequent choice model, representing the subjective beliefs of the agents based on observable data.

Stage 2: Maximum Score Estimation

In the second stage, the estimated expectations are plugged into the Maximum Score Estimator developed by Manski. Unlike standard parametric approaches, the maximum score method does not require the error term to follow a specific distribution (like the Normal distribution). Instead, it relies on a much weaker "conditional median restriction" ($Med(\epsilon|z,x) = 0$). This property makes the estimator exceptionally robust to heteroskedasticity of unknown forms, as it only requires that the median of the unobserved shocks, conditional on the covariates, is zero.

Theoretical Contributions

The core of the paper is dedicated to the asymptotic properties of this two-stage estimator. Estimating parameters based on previously estimated "generated regressors" usually complicates the statistical properties of the final estimates.

  1. Consistency and Convergence: The authors prove that the two-stage estimator is consistent. Despite the added noise from the first-stage nonparametric estimation, the estimator converges at a rate of $N^{-1/3}$. This is the same rate achieved by Manski’s original maximum score estimator, which is slower than the standard $\sqrt{N}$ rate but necessary for the level of distributional robustness provided.
  2. The Oracle Property: A significant theoretical finding (Theorem 3) is the "oracle property." The paper establishes regularity conditions under which the two-stage estimator is asymptotically equivalent in distribution to an "infeasible" estimator—one that has perfect knowledge of the true expectations. This suggests that as the sample size grows, the error introduced by the first-stage nonparametric estimation effectively "washes out."

Simulation Results

To validate the theoretical findings, the authors conducted Monte Carlo simulations. The simulations compared the proposed two-stage estimator against an infeasible single-stage estimator across various data-generating processes (DGPs).

  • Linear vs. Nonlinear Expectations: When the agent's expectations were linear, simple OLS first-stage estimation performed adequately. However, when expectations were nonlinear, OLS failed significantly, leading to substantial bias. In contrast, the nonparametric kernel-based first stage remained robust.
  • Sample Size Performance: As the sample size $N$ increased from 300 to 1000, both the bias and the Root Mean Squared Error (RMSE) of the two-stage estimator decreased steadily, confirming the consistency of the method.
  • Kernel Selection: Interestingly, the simulations showed that 2nd-order kernels often outperformed higher-order (8th-order) kernels in smaller samples. This highlights a common trade-off in nonparametric statistics where simpler estimators can sometimes provide better finite-sample stability.

Conclusion and Future Directions

This paper provides a rigorous foundation for estimating binary choice models under uncertainty without relying on restrictive distributional assumptions. While the maximum score method offers superior robustness to heteroskedasticity, it does come with computational challenges. The objective function is a non-differentiable step function, requiring non-gradient-based optimization methods like grid searches, which can be intensive in high-dimensional parameter spaces.

The authors suggest that future research could explore the use of Horowitz’s Smoothed Maximum Score or Laplace estimators. These variations could potentially offer faster convergence rates and easier computation, though they would require the careful selection of additional tuning parameters. Furthermore, the framework established here could be extended to more complex strategic environments, such as games of incomplete information.

Summary updated

Abstract

This paper develops maximum score estimation of preference parameters in the binary choice model under uncertainty in which the decision rule is affected by conditional expectations. The preference parameters are estimated in two stages: we estimate conditional expectations nonparametrically in the first stage and then the preference parameters in the second stage based on Manski (1975, 1985)'s maximum score estimator using the choice data and first stage estimates. The paper establishes consistency and derives rate of convergence of the two-stage maximum score estimator. Moreover, the paper also provides sufficient conditions under which the two-stage estimator is asymptotically equivalent in distribution to the corresponding single-stage estimator that assumes the first stage input is known. These results are of independent interest for maximum score estimation with nonparametrically generated regressors. The paper also presents some Monte Carlo simulation results for finite-sample behavior of the two-stage estimator.
BibTeX
@misc{chen2013maximum,
  author = {Le-Yu Chen and Sokbae Lee and Myung Jae Sung},
  title = {{Maximum Score Estimation of Preference Parameters for a Binary Choice Model under Uncertainty}},
  year = {2013},
  eprint = {1304.5747},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1304.5747},
}
Figure