Introduction
In the field of causal inference, the "gold standard" for policy evaluation typically relies on comparing a treated group to a "never-treated" control group. However, in many real-world scenarios—particularly those involving international treaties or large-scale economic shifts—a true never-treated group does not exist. Eventually, every unit is exposed to the policy. This paper addresses the challenge of retrospective causal inference, where researchers must estimate the impact of a policy using only "later-treated" and "always-treated" units.
The authors apply this framework to the context of European integration. By 2005, most Western European border regions were already part of the Schengen Agreement and the Freedom of Movement. When Eastern European regions joined later, researchers lacked a traditional baseline. This paper provides the mathematical machinery to look backward and ask: "What would have happened if these Eastern regions had been integrated all along?"
The Methodological Challenge
Traditional methods like Difference-in-Differences (DiD) often rely on the "parallel trends" assumption, which can be difficult to justify in complex regional economies. Furthermore, DiD typically requires a control group that remains untreated.
To overcome these hurdles, the authors move toward a Matrix Completion (MC) framework. In this setting, the potential outcomes for the "later-treated" group during their pre-treatment years are treated as missing values in a large data matrix. The algorithm "fills in" these missing values by assuming the underlying data structure is low-rank—meaning the outcomes across different regions and times are driven by a few common, unobserved factors.
Key Methodological Extensions
The paper introduces three critical refinements to standard matrix completion to ensure its validity in causal settings:
- Propensity-Score Weighting: The objective function is weighted to ensure the "always-treated" units used for imputation closely resemble the "later-treated" units in both observed and unobserved characteristics.
- Elapsed-Time Weighting: This accounts for the fact that policy effects are rarely instantaneous, allowing the model to capture the time required to reach a "steady state" equilibrium.
- Endogenous Covariates: The authors propose a procedure to impute covariate values (such as regional GDP) that might themselves be influenced by the treatment, preventing "feedback" loops from biasing the results.
Empirical Application: European Integration
The study utilizes the European Labour Force Survey (ELFS), analyzing quarterly data from 2005 to 2019. The research focuses on working-age individuals (ages 20–59) living in border regions, specifically targeting two groups:
- The Eastern Cluster: Regions in Poland, the Czech Republic, Slovakia, Slovenia, and Hungary.
- The Swiss Cluster: Regions bordering Switzerland.
By observing how long-integrated regions behave, the model predicts how newly integrated regions would have behaved if they had joined earlier, effectively using the "always-treated" group as a mirror for the treated state of the newcomers.
Key Findings
The Impact on Labor Mobility
The results indicate a transformative impact on Eastern European labor markets. The combined effect of the Schengen Agreement and the Freedom of Movement nearly doubled the probability of residents working across the border. This represents an 86% relative increase (0.8 percentage points) compared to the observed baseline.
In the Swiss cluster, the authors found a positive effect (~32% increase), though it was not statistically significant. This is likely due to "top-up" effects, as some labor market access had already been granted to Swiss border regions prior to the study period.
Legal vs. Physical Barriers
A major insight of the study is the distinction between types of integration. The Freedom of Movement (removing work permits) was found to be a much stronger driver of labor mobility than the Schengen Agreement (removing passport checks). Simply making it legal to work across the border mattered significantly more than reducing the physical time of the commute.
Model Performance and Robustness
In placebo tests—where the "treatment" is known to be zero—the matrix completion estimator consistently outperformed standard DiD and Synthetic Control methods, showing a lower Root Mean Squared Error (RMSE). This suggests that MC is more robust to unobserved, time-varying confounders that typically break standard models.
Conclusion
The retrospective causal inference framework offers a powerful tool for analyzing globalized policy shifts where traditional control groups are absent. While the method relies on a low-rank assumption for the outcome matrix, its success in the European context suggests it could be broadly applied to other policy areas, such as trade agreements or environmental regulations, where "never-treated" groups are increasingly rare.
Summary updated