Breaking the curse of dimensionality in conditional moment inequalities for discrete choice models

Le‐Yu Chen, Sokbae Lee

TL;DR

This paper introduces a powerful dimension-reduction technique for discrete choice models where parameters are only "partially identified." By proving that high-dimensional conditional moment inequalities can be simplified to depend on just two indexing variables, the authors "break" the curse of dimensionality, allowing for much more precise and powerful statistical inference in complex models.

What Problem Does This Paper Address?

In econometrics and statistics, we often use discrete choice models to understand how people make decisions (e.g., "Why did this person choose brand A over brand B?"). Often, we want to avoid making strict assumptions about the "noise" or error terms in these models, leading us to semiparametric approaches.

However, a major hurdle arises: sometimes the data isn't "rich" enough to identify a single "true" parameter value. Instead, we find an identified set—a range of values that are all consistent with the data. Characterizing this set usually involves conditional moment inequalities (e.g., "The probability of choosing A given covariates $X$ must be greater than 0.5 if the utility index is positive").

The "Curse of Dimensionality" hits when the number of covariates ($X$) is large. Nonparametric methods used to estimate these probabilities become incredibly imprecise as the number of variables increases, requiring astronomical amounts of data to produce reliable results. This paper seeks a way to perform inference that doesn't fall apart when the number of covariates grows.

Methods

The authors' core contribution is a mathematical proof showing that you don't actually need to condition on the entire high-dimensional vector of covariates $X$.

  1. Dimension Reduction (The "Two-Index" Result): They prove (Lemma 1 and Theorem 1) that the identified set can be equivalently characterized by conditioning on only two continuous indexing variables, regardless of how many covariates are in the original model.
  2. Sign Equivalence: The method exploits the structure of discrete choice models where the sign of a conditional expectation is tied to the sign of a linear index ($X'\beta$).
  3. Inversion of Tests: To construct a confidence set, they use the CLR method (Chernozhukov, Lee, and Rosen, 2013). Instead of plugging in a high-dimensional nonparametric estimator, they plug in a 2D kernel estimator based on their new characterization.
  4. Generalization: They show this approach works for a wide variety of models, including:
  5. Binary choice models.
  6. Ordered choice models (where choices have a natural ranking).
  7. Multinomial choice (multiple unordered options).
  8. Panel data models with fixed effects.
  9. Monotone single-index models.

Datasets

The paper does not use real-world empirical datasets. Instead, it relies on Monte Carlo simulation experiments to validate the theory.

  • Setup: A binary response model with a covariate dimension $d$ ranging from 3 to 10.
  • Goal: Compare the "Index" approach (the authors' new method) against the "Full" approach (conditioning on all $d$ variables).

Key Insights and Ideas

  • The Power of Two: The most striking "aha moment" is that the complexity of the nonparametric part of the problem is fixed at a dimension of 2. Whether you have 5 covariates or 500, the statistical "difficulty" of the nonparametric estimation remains the same.
  • Conditional Moment Analogs: This work provides a conditional moment inequality version of the "index restrictions" commonly used in simpler, point-identified semiparametric models.
  • Trading Complexity for Power: While the new method requires searching over a slightly larger space (an extra "index" variable $\gamma$), it gains massive amounts of statistical power because the kernel density estimation doesn't get "diluted" by high-dimensional space.

Key Results

The simulation results (Tables 2 and 3) demonstrate a clear advantage:

  • Size Control: Both the new "Index" approach and the standard "Full" approach correctly maintain the nominal significance level (usually 5%), meaning they don't falsely reject the true parameter too often.
  • Substantial Power Gains: As the number of covariates ($d$) increases, the "Full" approach loses power rapidly. For example, with 10 covariates and 1000 samples, the "Index" approach had a test power of 0.520, while the "Full" approach lagged behind at 0.225.
  • Alleviating the Curse: The results show that the "Index" approach is much more sensitive to detecting incorrect parameter values in high-dimensional settings, effectively bypassing the sparsity issues that plague standard nonparametric methods.

Limitations and Future Work

  • Computational Demand: The authors acknowledge that their algorithm is computationally intensive. Because it involves inverting a pointwise test over a grid and calculating infimums over multiple variables, it can be slow as the parameter space grows.
  • Bandwidth Selection: While they provide a "rule-of-thumb" for bandwidth selection in their kernel estimators, they note that choosing the optimal bandwidth for this specific two-index setup is a complex task left for future research.
  • Integration with Other Methods: The paper focuses on the CLR test statistic. The authors suggest that applying their dimension-reduction characterization to other popular inference methods (like the instrumental variable approach of Andrews and Shi) would be a valuable next step.

Summary updated

Abstract

This paper studies inference of preference parameters in semiparametric discrete choice models when these parameters are not point-identified and the identified set is characterized by a class of conditional moment inequalities. Exploring the semiparametric modeling restrictions, we show that the identified set can be equivalently formulated by moment inequalities conditional on only two continuous indexing variables. Such formulation holds regardless of the covariate dimension, thereby breaking the curse of dimensionality for nonparametric inference based on the underlying conditional moment inequalities. We further apply this dimension reducing characterization approach to the monotone single index model and to a variety of semiparametric models under which the sign of conditional expectation of a certain transformation of the outcome is the same as that of the indexing variable.
BibTeX
@article{chen2019breaking,
  author = {Le‐Yu Chen and Sokbae Lee},
  title = {{Breaking the curse of dimensionality in conditional moment inequalities for discrete choice models}},
  journal = {Journal of Econometrics},
  year = {2019},
  doi = {10.1016/j.jeconom.2018.12.024},
}

Loading related content...

Figure