TL;DR
This paper introduces a robust framework for using Lasso and Post-Lasso to select "optimal instruments" in linear models where the number of potential instruments ($p$) may far exceed the sample size ($n$). By assuming "approximate sparsity," the authors prove that their estimator is consistent and asymptotically normal even under heteroscedasticity and non-Gaussian noise, without requiring the restrictive "beta-min" conditions common in earlier high-dimensional theory.
What Problem Does This Paper Address?
In causal inference, Instrumental Variables (IV) are the gold standard for dealing with endogeneity (when your treatment variable is correlated with the error term). To get the most precise estimates, researchers want to use "optimal instruments"—essentially the best possible prediction of the treatment variable using the instruments.
However, modern datasets often provide hundreds or thousands of potential instruments (e.g., many different demographic interactions or series expansions). Standard IV methods like Two-Stage Least Squares (2SLS) break down in this "many instruments" setting; they overfit the first stage, leading to biased results and invalid inference. Previous attempts to use machine learning (like Lasso) to select instruments often relied on "beta-min" conditions—the assumption that all relevant instruments have effects large enough to be easily detected. In the real world, many important signals are small, making these conditions unrealistic.
Methods
The authors propose using Lasso and Post-Lasso for the first-stage regression (predicting the endogenous variable using instruments).
- Lasso Selection: They solve a least-squares problem with an $\ell_1$-penalty to select a sparse set of instruments.
- Data-Driven Penalty: A key innovation is a new, data-weighted penalty loadings scheme. Unlike standard Lasso, which assumes homoscedastic Gaussian noise, this method uses a self-normalization approach to handle non-Gaussian and heteroscedastic (varying variance) errors.
- Post-Lasso Refinement: After Lasso selects the instruments, they refit the model using Ordinary Least Squares (OLS) on just those selected variables. This "Post-Lasso" step removes the shrinkage bias inherent in Lasso, leading to better performance in the second stage.
- Sup-Score Testing: For cases where instruments might be weak (individually having very little predictive power), they provide a "sup-score" test that remains valid even when $p \gg n$.
Datasets
The paper evaluates the method through two primary lenses:
- Extensive Simulations: They test various "Exponential" and "Cut-off" designs where the signal is either concentrated in a few instruments or decays across many. They compare their method against the Fuller estimator, a traditional many-instrument robust procedure.
- Eminent Domain Case Study: They apply the method to a real-world economic problem: the effect of judicial decisions regarding eminent domain (the government's power to take private property) on house prices and state-level GDP. They use a dataset of judge demographics and panel characteristics as potential instruments, originally compiled by Chen and Yeh (2010).
Key Insights and Ideas
- Approximate Sparsity is Sufficient: You don't need "exact" sparsity (where only $s$ variables matter and the rest are zero). As long as the signal can be well-approximated by a small set of variables, Lasso-based IV works.
- Goodbye, Beta-Min: This is a major theoretical "aha" moment. The authors show that even if Lasso fails to pick up some "small but relevant" instruments, the resulting IV estimator is still root-$n$ consistent and asymptotically normal. Model selection mistakes don't necessarily ruin the final causal estimate.
- Handling Real-World Noise: By using moderate deviation theory for self-normalized sums, they provide a way to choose the Lasso penalty level that is robust to the "messy" noise typical of economic data (non-normality and heteroscedasticity).
Key Results
- Efficiency: When errors are homoscedastic, the Lasso-based IV estimator achieves the semi-parametric efficiency bound—meaning it's as good as the best possible theoretical estimator.
- Simulation Performance: In settings with $n=100$ and $p=100$, Lasso-based procedures significantly outperformed traditional 2SLS and the Fuller estimator in terms of median bias and absolute deviation.
- Empirical Impact: In the eminent domain application, Lasso selected a very small, powerful set of instruments (often just 1 to 4 variables). This resulted in first-stage Wald statistics (a measure of instrument strength) that were 2 to 3 times larger than the baseline intuitive instruments used in previous literature, leading to much tighter confidence intervals for the economic effects.
Limitations and Future Work
The authors acknowledge that their method relies heavily on the approximate sparsity assumption. If the true relationship is "dense"—meaning every instrument contributes a tiny, equal amount of signal—Lasso will perform poorly. In such cases, traditional many-instrument methods like Fuller might be more appropriate.
Future research directions suggested include exploring the interplay between variable selection and "many-weak-instrument" asymptotics more deeply, as well as extending these sparse selection methods to non-linear IV models.
Summary updated