Post-Selection and Post-Regularization Inference in Linear Models with\n Many Controls and Instruments

Victor Chernozhukov, Christian Hansen, Martin Spindler

TL;DR

This paper introduces a robust "double-selection" method for estimating causal effects in linear models when you have more potential control variables and instruments than observations. By using orthogonal moment functions, the authors provide a way to perform valid statistical inference that is "immunized" against the small mistakes typically made during automated variable selection.

What Problem Does This Paper Address?

In modern empirical research, scientists often face a "curse of dimensionality." Suppose you want to measure the impact of a treatment ($d$) on an outcome ($y$). To get a clean estimate, you must control for other factors ($x$) and potentially use instrumental variables ($z$) to handle endogeneity.

The challenge arises when the number of potential controls and instruments is very large—sometimes even exceeding the sample size ($n$). While machine learning tools like Lasso are great at picking the "best" variables (sparsity), they aren't perfect. If a selection algorithm accidentally drops a variable that has a small but meaningful correlation with the treatment, it introduces a "post-selection bias" that ruins your final estimate and makes standard confidence intervals unreliable.

Methods

The authors extend the "double-selection" strategy to the Instrumental Variables (IV) framework. The core idea is to use orthogonal moment functions. These are estimating equations designed to be locally insensitive to small errors in estimating the "nuisance" parts of the model (the relationship between controls, instruments, and the variables of interest).

Algorithm 1 (The Double-Selection Approach): 1. Select Instruments: Use Lasso to regress the endogenous variable ($d$) on all potential controls ($x$) and instruments ($z$). 2. Select Controls for Outcome: Use Lasso to regress the outcome ($y$) on all potential controls ($x$). 3. Select Controls for Treatment: Use Lasso to regress the predicted treatment ($\hat{d}$) on the controls ($x$). 4. Final Estimation: Construct residuals from these steps to create "immunized" versions of the variables and run a final standard IV regression.

By selecting variables from multiple "directions," the method ensures that any variable relevant to either the treatment or the outcome is retained, protecting the final causal estimate from omitted variable bias.

Datasets

The paper evaluates the method using two primary sources:

  • Synthetic Simulation: A controlled environment where $n=200$, with 300 potential controls and 150 potential instruments. This setup specifically tests the "high-dimensional" case where $p > n$.
  • Automobile Demand (BLP Data): The authors revisit the classic dataset from Berry, Levinsohn, and Pakes (1995). They expand the original set of 5 controls and 10 instruments into a high-dimensional set of 24 controls and 48 instruments by adding time trends, quadratics, cubics, and first-order interactions.

Key Insights and Ideas

  • The "Aha" Moment on Bias: The authors show that "Naive" selection (just picking variables once and running a regression) fails because it doesn't account for the "endogeneity" bias introduced by the selection process itself.
  • Immunization: The use of orthogonal moments acts as a "buffer." Even if the Lasso selection is slightly off, the mathematical structure of the estimator ensures that the error doesn't leak into the estimate of the main parameter ($\alpha$).
  • Practicality for Economists: The method doesn't require "perfect" model selection. As long as the underlying model is "approximately sparse" (meaning a few variables do most of the heavy lifting), the inference remains valid.

Key Results

  • Simulation Performance: The "Double-Selection" estimator achieved a median bias of only .021 and a test "size" (Type I error rate) of .054—very close to the ideal 5%. In contrast, a "Naive" stepwise selection approach had a massive bias of .160 and a 30% error rate, proving it unreliable for scientific inference.
  • Empirical Findings: In the automobile demand example, the new method found that demand was more "elastic" (price-sensitive) than originally thought. The original baseline model suggested many products had "inelastic" demand (which contradicts economic theory for profit-maximizing firms), while the double-selection method reduced the number of these "implausible" cases significantly.

Limitations and Future Work

  • Sparsity Assumption: The method relies on the assumption that the true model is "sparse." If the signal is spread thinly across thousands of variables (dense), Lasso-based selection may still struggle.
  • Linearity: The paper focuses on linear IV models. While the authors mention that extensions to non-linear models are possible, this specific note stays within the linear framework.
  • Weak Instruments: While the method handles many instruments, it still requires that the selected instruments are "strong" enough to identify the effect. The authors suggest using weak-identification robust inference as a complementary tool.

Summary updated

Abstract

In this note, we offer an approach to estimating causal/structural parameters\nin the presence of many instruments and controls based on methods for\nestimating sparse high-dimensional models. We use these high-dimensional\nmethods to select both which instruments and which control variables to use.\nThe approach we take extends BCCH2012, which covers selection of instruments\nfor IV models with a small number of controls, and extends BCH2014, which\ncovers selection of controls in models where the variable of interest is\nexogenous conditional on observables, to accommodate both a large number of\ncontrols and a large number of instruments. We illustrate the approach with a\nsimulation and an empirical example. Technical supporting material is available\nin a supplementary online appendix.\n
BibTeX
@inproceedings{chernozhukov2015post,
  author = {Victor Chernozhukov and Christian Hansen and Martin Spindler},
  title = {{Post-Selection and Post-Regularization Inference in Linear Models with\textbackslash\{\}n Many Controls and Instruments}},
  booktitle = {American},
  year = {2015},
}
Figure