When machine learning misses hidden forces in economic data

The problem left behind by prediction

Suppose researchers want to understand the relationship between minimum wages and wage inequality using repeated observations of U.S. states. They can adjust for measured workforce characteristics, but common economic forces may still influence both policy and wages—and affect each state differently.

These are examples of interactive confounding: shared forces with different strengths across units. Unlike a fixed difference between states or a uniform annual shock, their influence changes across both place and time. Simply subtracting state and year effects does not generally remove them.

In their preprint, Semiparametric Inference in Panel Data with Interactive and High-dimensional Confounding, Bin Chen, Yukun Ma and Qiyang Yu address this problem within double/debiased machine learning, or DML.

DML uses flexible prediction methods to adjust treatment and outcome measurements for observed characteristics. It then estimates a target coefficient from what remains. But even excellent adjustment for recorded variables can leave hidden components in both sets of residuals. If those components are correlated, standard DML can produce misleading estimates and confidence intervals.

Learning two hidden patterns separately

The proposed method, Factor-Augmented DML, adds an explicit step for estimating hidden common patterns. It applies principal component analysis to treatment and outcome residuals, extracting a small number of shared time-series directions.

The outcome-side patterns are estimated from a reduced-form equation: an outcome model that incorporates the treatment equation. Consequently, its hidden-pattern space need not have the same dimension as the treatment-side space or the structural outcome space.

The method combines two projections to remove the estimated factor components from its estimating equation. This double projection makes a key remaining error depend on the product of errors in the two estimated spaces, rather than leaving an uncontrolled contribution from one space alone.

Sample splitting is equally important. Researchers divide complete unit histories into three groups. One trains the treatment-side estimates, another trains the outcome-side estimates, and the third evaluates the coefficient. The procedure averages results across all six assignments of these roles.

Why not train both sides on the same group? The authors show that doing so can pair a unit’s treatment disturbance with its own outcome-side disturbance inside the factor-estimation errors. Separate training groups exclude those same-unit pairings.

Testing whether uncertainty estimates work

The simulations use panels with 100 or 200 units and between 50 and 300 periods. They include strong interactive confounding, linear or quadratic relationships with observed covariates, and regularized first-stage learners.

Across the reported designs, nominal 95% confidence intervals from three-fold FA-DML covered the true coefficient between 93.4% and 95.9% of the time for linear relationships, and between 92.6% and 95.8% for quadratic relationships. Both known and estimated factor dimensions are included in those ranges.

Standard DML had zero empirical coverage throughout this particular simulation grid. A two-fold factor-augmented variant also substantially undercovered. These results concern the authors’ deliberately confounded designs, not every possible use of DML. Moreover, the two-fold comparison changes training-sample size as well as separation, so it does not isolate separation alone.

The authors also propose a procedure for choosing the number of factors by penalizing and thresholding singular values. Each factor dimension was selected correctly in at least 95.0% of replications in every reported design.

What the state data show

The main application covers all 50 U.S. states from 1979 to 2012, totaling 1,700 state-year observations. The outcome compares median wages with wages at the 10th percentile: the 50/10 wage ratio.

At the selected factor dimensions, the estimates associate a 10% increase in the real minimum wage with a roughly 1.1% lower ratio. Results were similar across Lasso and random-forest learners and two control dictionaries; all four primary 95% intervals excluded zero.

This is an association, not an unconditional demonstration of causation. A causal interpretation requires a structural outcome equation and the paper’s strict-exogeneity assumptions.

Strong results under strong conditions

The theory establishes that FA-DML has the same first-order limiting behavior as an infeasible estimator that already knows the nuisance functions and factor spaces. That result requires strong, pervasive factor signals, fixed factor dimensions, adequate remaining treatment variation, and specified learning and moment conditions as both panel dimensions grow.

Units must be independent conditional on the common factors; temporal dependence is allowed only within the stated restrictions. Generic dynamic panels with predetermined regressors and weak-factor settings are not covered.

The paper is an arXiv preprint. Its contribution is a conditional framework for more reliable inference—not a guarantee that hidden confounding can always be recovered from data.