Double Machine Learning for Static Panel Data with Instrumental Variables: New Method and Applications

1. Problem Statement

The paper addresses a critical gap in econometric causal inference: the inability of standard Instrumental Variable (IV) methods, specifically Two-Stage Least Squares (2SLS), to handle high-dimensional confounders and nonlinear relationships in static panel data settings.

• The Challenge: In many empirical applications (e.g., migration studies using shift-share instruments), instrument validity relies on conditioning on a rich set of covariates. Traditional 2SLS fails here because:
  • It cannot flexibly model nonlinearities without explicit parameterization.
  • It suffers from rank deficiency (singularity) when the number of covariates is large relative to the sample size.
  • It produces biased estimates if the functional form of confounders is misspecified.
• The Specific Context: Existing Double Machine Learning (DML) methods (e.g., Chernozhukov et al., 2018) were designed for cross-sectional data. They do not account for unobserved individual heterogeneity (fixed effects) and serial dependence, which are inherent in panel data. Naively applying cross-sectional DML to panel data leads to inconsistent estimates.
• Weak Identification: Standard inference in IV settings often breaks down when instruments are "weak" (low correlation with the endogenous treatment). Conventional 2SLS diagnostics (like the F-statistic) can be misleading in high-dimensional settings where flexible adjustment reduces effective instrument strength.

2. Methodology: Panel IV DML

The authors propose Panel IV DML, a novel estimator that extends the DML framework to static panel models with endogenous treatments.

A. Econometric Framework

The model is a Partially Linear Panel Regression (PLPR) with instrumental variables:

1. Structural Equation: $Y_{it} = \theta_0 D_{it} + l_0(X_{it}) + \alpha_i + U_{it}$
2. Treatment Equation: $D_{it} = r_0(X_{it}) + \zeta_i + R_{it}$
3. Instrument Equation: $Z_{it} = m_0(X_{it}) + \gamma_i + V_{it}$

Where:

• $Y_{it}$ is the outcome, $D_{it}$ is the endogenous treatment, $Z_{it}$ is the instrument, and $X_{it}$ are high-dimensional controls.
• $\alpha_i, \zeta_i, \gamma_i$ are unobserved individual fixed effects.
• $l_0, r_0, m_0$ are unknown nuisance functions (potentially nonlinear) estimated via Machine Learning (ML).

B. Estimation Strategy

To handle fixed effects and serial correlation, the authors employ a First-Difference (FD) transformation:

1. Transformation: The model is differenced ($\Delta Y_{it} = Y_{it} - Y_{i,t-1}$) to eliminate time-invariant fixed effects ($\alpha_i, \zeta_i, \gamma_i$).
2. Nuisance Learning: The FD-transformed nuisance functions ($\Delta l_0, \Delta r_0, \Delta m_0$) are estimated using flexible ML algorithms (Lasso, Neural Networks, Gradient Boosting).
3. Neyman Orthogonality: The authors derive a Neyman-orthogonal score function tailored to the panel structure. This ensures that the estimator for the structural parameter $\theta_0$ is robust to small errors in the estimation of the nuisance functions (regularization bias).
4. Cross-Fitting: To prevent overfitting, they use block-k-fold cross-fitting, where the entire time series of a specific unit is assigned to the same fold. This preserves the serial dependence structure of the panel data.

C. Weak Identification Diagnostics

A key methodological innovation is the integration of weak-instrument diagnostics into the DML framework:

• First-Stage F-Statistic: Adapted for the Panel IV DML estimator to test instrument relevance.
• Anderson-Rubin (AR) Test: The authors derive an AR test statistic and confidence sets (CS) for the Panel IV DML. Unlike standard 2SLS inference, the AR test remains valid even when instruments are weak, as it relies on the reduced-form relationship rather than the strength of the first stage.

3. Key Contributions

1. Theoretical Extension: Extends DML from cross-sectional to static panel data by incorporating fixed effects via first-differencing and deriving new Neyman-orthogonal score functions that account for unobserved heterogeneity.
2. Weak IV Tools for DML: Develops the first set of weak-identification diagnostics (F-statistic and AR test) specifically for DML procedures in panel settings. This allows researchers to detect when flexible adjustment renders an instrument weak.
3. Empirical Re-evaluation: Re-analyzes three influential migration studies (Tabellini, 2020; Moriconi et al., 2019, 2022) that rely on shift-share instruments, demonstrating how flexible covariate adjustment changes inference.

4. Results

A. Empirical Applications

The authors re-estimate three studies using shift-share instruments:

1. Tabellini (2020) - US Immigration:
   • Baseline: Panel IV DML strengthens the instrument's predictive power (higher F-stat) compared to 2SLS. It confirms the political effect (immigration increases conservatism) but finds no economic effect (contrary to 2SLS).
   • Robustness: When adding many controls, the instrument becomes weak in 2SLS but remains strong in DML for political outcomes. However, causal effects disappear in both methods, suggesting original results were sensitive to confounder selection.
2. Moriconi et al. (2019, 2022) - European Immigration:
   • Weak Identification: In these studies, the shift-share instruments appear moderately strong under 2SLS but become weak under Panel IV DML once high-dimensional and nonlinear controls are flexibly adjusted for.
   • Inference: The AR diagnostics reveal that while some reduced-form effects exist, the instruments are too weak to identify the causal treatment effect reliably. Conventional 2SLS incorrectly suggests strong identification and significant effects, whereas Panel IV DML correctly signals weak identification (unbounded AR confidence sets).

B. Monte Carlo Simulations

• Strong Instruments: Panel IV DML outperforms 2SLS in terms of bias and RMSE, especially when confounding is nonlinear. It achieves root-N convergence.
• Weak Instruments:
  • 2SLS: Systematically fails to detect weak instruments. It produces biased estimates and falsely rejects the null hypothesis of no effect (over-rejection), leading to spurious causal claims.
  • Panel IV DML: Correctly identifies weak identification. The F-statistics drop, and AR confidence sets become unbounded or include zero, providing a more honest assessment of uncertainty.

5. Significance

• Methodological Rigor: The paper provides a rigorous framework for applying causal machine learning to panel data, solving the "fixed effects" problem that previously limited DML applications.
• Policy Implications: It highlights that "flexible adjustment" (controlling for more variables) is necessary for instrument validity but can inadvertently weaken instruments. Researchers must balance the need for rich controls with instrument strength.
• Best Practice: The authors advocate for the routine reporting of Anderson-Rubin diagnostics alongside standard IV estimates. In settings with high-dimensional confounders, AR tests are essential to avoid drawing false causal conclusions when instruments are weak.

In summary, Panel IV DML offers a superior alternative to 2SLS for panel data with endogenous treatments, offering better handling of nonlinearity and high-dimensional data, while providing robust diagnostics to prevent misleading inference in weak-instrument scenarios.