Powerful Inference.

Xiaohong Chen, Sokbae Lee, Myung Hwan Seo

Abstract

We develop an inference method for a (sub)vector of parameters identified by conditional moment restrictions, which are implied by economic models such as rational behavior and Euler equations. Building on Bierens (1990), we propose penalized maximum statistics and combine bootstrap inference with model selection. Our method is optimized to be powerful against a set of local alternatives of interest by solving a data-dependent max-min problem for tuning parameter selection. We demonstrate the efficacy of our method by a proof of concept using two empirical examples: rational unbiased reporting of ability status and the elasticity of intertemporal substitution.

Introduction

In modern econometrics, many structural models—ranging from Euler equations in macroeconomics to models of rational behavior—are characterized by conditional moment restrictions (CMR). These models posit that the expected error of a structural function, conditional on a set of instruments or conditioning variables ($W$), is zero.

However, researchers often face a dilemma: including more instruments can provide more information, but it also risks "instrument proliferation." When the dimension of $W$ is large or when instruments are weak, standard statistical tests frequently lose their power. This leads to the "uninformative interval" problem, where confidence intervals for key parameters (like the elasticity of intertemporal substitution) become so wide that they encompass all possible values, rendering the analysis useless for policy or theory verification. This paper introduces a novel penalized maximum statistic designed to maintain power even in the presence of many irrelevant or weak instruments.

Methodology: Penalization and Max-Min Calibration

The proposed method builds upon the Bierens (1990) maximum statistic. Bierens’ original approach transforms a conditional moment into an infinite family of unconditional moments using an exponential weight function $\exp(W' \gamma)$. The test identifies the "worst-case" $\gamma$ that deviates from zero to detect model violations.

The $\ell_1$ Innovation

To address the noise introduced by high-dimensional instruments, the authors introduce an $\ell_1$ penalty ($-\lambda \|\gamma\|_1$) to the Bierens statistic. Similar to the LASSO in regression analysis, this penalty performs internal model selection. It effectively "sifts" through the conditioning information, ignoring irrelevant instruments and focusing the test's power on the most informative directions in the data.

Data-Driven Tuning

A critical challenge in penalized methods is the selection of the tuning parameter $\lambda$. The authors propose a max-min calibration strategy. They select $\lambda$ by solving a data-dependent problem that maximizes the "least-favorable" local power. This ensures that the test is optimized to detect the specific alternative hypotheses that are most relevant to the researcher. To calculate $p$-values and critical values, the method employs a computationally efficient multiplier bootstrap procedure.

Empirical Applications

The efficacy of the penalized maximum statistic is demonstrated through two classic empirical challenges where traditional methods often fail.

1. Elasticity of Intertemporal Substitution (EIS)

Using US annual data (1891–1995) from Yogo (2004), the authors test the EIS, a parameter notorious for being weakly identified. Traditional tests like the Lagrange Multiplier (LM) and Conditional Likelihood Ratio (LR) tests yield completely uninformative 95% confidence intervals ($[-\infty, \infty]$).

In contrast, the penalized method produces a narrow, informative interval of $[-0.30, 0.15]$. This result suggests that by properly penalizing irrelevant information, researchers can extract meaningful signals from datasets previously thought to be too "weak" for inference.

2. Rational Unbiased Reporting (RUR)

The paper re-examines the work of Benítez-Silva et al. (2004) regarding whether individuals report their disability status truthfully compared to actual Social Security Administration (SSA) awards. While the original study—using an unpenalized test—failed to reject the null hypothesis of rational reporting ($p$-value $= 0.09$), the penalized test rejects the null at the 5% level ($p$-value $\approx 0.02$). This indicates that the new method is more sensitive to deviations from rational behavior that were previously masked by noise.

Simulation and Computational Performance

Monte Carlo experiments designed to mimic the Yogo (2004) data confirm these findings. As the number of irrelevant instruments increases, the power of unpenalized tests decays rapidly. The penalized test, however, maintains high power, demonstrating its robustness to "noise" in the instrument set.

Furthermore, the addition of the $\ell_1$ penalty actually improves computational efficiency. By shrinking the search space for the weight parameters ($\gamma$), the optimization problem becomes more tractable for modern solvers like Particle Swarm Optimization (PSO), making the method practical for large-scale empirical research.

Conclusion and Future Directions

This research demonstrates that penalization is not just a tool for point estimation, but a potent instrument for inference. By integrating model selection directly into the hypothesis testing framework, the penalized maximum statistic provides a path forward for models with high-dimensional conditional information.

Future work remains to extend this framework to non-parametric functions, where the parameter of interest is itself a function, and to adapt the theory for the specific dependencies found in large-scale panel data. Currently, the $\ell_1$ penalty is the primary focus, but exploring other penalty structures (such as Elastic Net) may offer further refinements in specific econometric contexts.

Summary updated

Abstract

We develop an inference method for a (sub)vector of parameters identified by conditional moment restrictions, which are implied by economic models such as rational behavior and Euler equations. Building on Bierens (1990), we propose penalized maximum statistics and combine bootstrap inference with model selection. Our method is optimized to be powerful against a set of local alternatives of interest by solving a data-dependent max-min problem for tuning parameter selection. We demonstrate the efficacy of our method by a proof of concept using two empirical examples: rational unbiased reporting of ability status and the elasticity of intertemporal substitution.
BibTeX
@article{chen2020powerful,
  author = {Xiaohong Chen and Sokbae Lee and Myung Hwan Seo},
  title = {{Powerful Inference.}},
  journal = {The Review of Economics and Statistics},
  year = {2020},
}

Loading related content...

Figure