Inference approaches for instrumental variable quantile regression

Victor Chernozhukov, Christian Hansen, Michael JanssonEconomics 2007 · DOI

TL;DR

This paper introduces and compares three novel methods for robust statistical inference in instrumental variable quantile regression (IVQR) models, particularly when instruments are weak. The authors demonstrate that traditional asymptotic methods can be highly unreliable in such scenarios, leading to significant differences in confidence intervals compared to their proposed finite-sample and weak-instrument robust approaches.

What Problem Does This Paper Address?

Quantile regression (QR) is a powerful tool for understanding how covariates affect different parts of a distribution, not just the average. However, when the variables of interest (like education) are endogenous (correlated with unobserved factors affecting wages), we need instrumental variables (IV) to get reliable estimates. Combining IV with QR (IVQR) is challenging, especially when the instruments are "weak" – meaning they have a weak correlation with the endogenous variable. Standard IVQR inference methods often rely on asymptotic approximations (assuming large sample sizes), which can break down when instruments are weak, leading to misleading conclusions about the precision of our estimates. This paper addresses the critical need for reliable inference methods in IVQR when faced with weak instruments.

Methods

The authors present three distinct approaches to inference in IVQR:

  1. Asymptotic Approximation (Chernozhukov and Hansen, 2006): This method extends standard QR inference to the IV setting, providing an asymptotically valid approximation. It's the "go-to" method when instruments are strong.
  2. Weak-Instrument Robust Approach (Chernozhukov and Hansen, 2006): This approach is designed to be asymptotically valid even under weak identification. It bases inference on a Wald statistic for instrument relevance, offering a more conservative approach when instrument strength is questionable.
  3. Finite Sample Approach (Chernozhukov, Hansen, and Jansson, 2005): This is the most novel contribution. It leverages a distributional fact about quantile restrictions to construct confidence regions that are exact in finite samples, meaning they do not rely on asymptotic approximations at all. This method involves inverting a GMM objective function and using simulation to obtain critical values.

Datasets

The paper uses data from David Card's (1995) study on the return to schooling, which includes information on wages, years of schooling, demographic controls, and proximity to colleges (used as instruments for schooling). This is a well-known dataset in econometrics, often used to illustrate IV estimation.

Key Insights and Ideas

  • Weak Instruments are a Major Problem: The core takeaway is that weak instruments severely undermine the reliability of standard asymptotic inference in IVQR. The paper provides a stark empirical demonstration of this.
  • Finite Sample Methods Offer Robustness: The finite sample approach, by avoiding asymptotic approximations, provides confidence intervals that are demonstrably more reliable and often much wider when instruments are weak, correctly reflecting the increased uncertainty.
  • Asymptotics Can Be Deceiving: The study highlights how asymptotic methods can produce overly narrow confidence intervals, giving a false sense of precision when instruments are weak. This can lead to incorrect conclusions about the statistical significance of estimated parameters.
  • Heterogeneity Matters: The application to wage returns to schooling reveals significant heterogeneity across different wage quantiles, which is more accurately captured and inferred using the robust IVQR methods.

Key Results

In the case study using Card's (1995) data, the authors found substantial differences between the inference approaches:

  • Panel A (Schooling as Exogenous): Asymptotic and finite sample intervals were qualitatively similar, though finite sample intervals were wider.
  • Panel B (Schooling as Endogenous with IV):
    • Asymptotic IV: Intervals were wide, making statistical evidence difficult to find.
  • Weak-Instrument Robust IV: Intervals were substantially wider than the usual asymptotic IV intervals, often covering almost the entire parameter space.
  • Finite Sample IV: These intervals were the widest of all, covering the entire parameter space in every case, underscoring the uncertainty introduced by weak instruments.

The F-statistic for the excluded instruments was 8.32, which is considered weak in standard IV models. This weakness is identified as the primary driver of the discrepancies.

Limitations and Future Work

  • Computational Demands: The authors note that forming confidence regions by inverting test statistics (especially for the finite sample approach) can be computationally demanding. They mention that their earlier work (Chernozhukov et al., 2005) presents feasible algorithms.
  • Judging Differences: The paper suggests it would be interesting to develop formal procedures for judging the differences between asymptotic and finite sample intervals, which could help researchers decide when asymptotic approximations are acceptable.
  • Generalizability: While the study focuses on a wage model, the implications for inference in other IVQR applications with weak instruments are significant and warrant further investigation. The authors' work provides a strong foundation for developing more robust inference tools across various econometric models.

Summary updated

BibTeX
@inproceedings{chernozhukov2007inference,
  author = {Victor Chernozhukov and Christian Hansen and Michael Jansson},
  title = {{Inference approaches for instrumental variable quantile regression}},
  booktitle = {Economics},
  year = {2007},
  doi = {10.1016/j.econlet.2006.10.016},
}
Figure