Heterogeneous Responses to Continuous Treatments: A Cluster-Based Causal Framework

1. Problem Statement

The paper addresses a critical gap in causal inference literature regarding continuous treatments (e.g., funding levels, dosage, intensity) in observational settings.

• Limitations of Existing Methods: Standard estimators for continuous treatments, such as the Average Dose-Response Function (ADRF) proposed by Hirano & Imbens (2004) or recent non-parametric approaches, assume a single population-level causal relationship. These methods rely on two stringent assumptions:
  1. Weak Unconfoundedness: Potential outcomes are independent of treatment assignment conditional on covariates.
  2. Positivity (Common Support): Every unit has a non-zero probability of receiving any level of treatment within the observed covariate space.
• The Core Issue: In many real-world scenarios (e.g., EU regional funds), treatment assignment is non-random, and treatment effect heterogeneity (HTE) exists not just due to observed covariates but due to unobserved "regimes" or subgroups.
  • When HTE is present, the global positivity assumption is often violated because different subgroups naturally receive different ranges of treatment intensities (limited common support).
  • Estimating a single ADRF in such contexts leads to biased estimates and misleading policy conclusions because it averages out distinct, potentially opposing, causal mechanisms across subgroups.

2. Methodology: The Clustered Dose-Response Function (Cl-DRF)

The authors propose a novel estimator, the Clustered Dose-Response Function (Cl-DRF), which shifts the focus from a population average to cluster-specific causal relationships.

A. Causal Framework and Identification

• Target Estimand: Instead of $\mu(t) = E[Y(t)]$, the target is the cluster-specific DRF: $\mu_c(t) = E[Y(t) | C_i = c]$, where $C_i$ is an unobserved cluster indicator.
• Relaxed Assumptions: The framework relaxes the global identification assumptions to local assumptions within clusters:
  • Assumption B1 (Existence of Regimes): The population can be partitioned into $C$ clusters with distinct treatment-outcome relationships.
  • Assumption B2 (Local Unconfoundedness): Within each cluster $c$, $Y(t) \perp T | X, C=c$.
  • Assumption B3 (Local Positivity): Within each cluster, the conditional density of treatment is bounded away from zero over the cluster's specific support ($\mathcal{T}_c$).
• Identification: Under these assumptions, the cluster-specific DRF is identified via the Cluster-Specific Generalized Propensity Score (GPS), defined as the conditional density of treatment given covariates and cluster membership: $R_{i,c}(t) = f(T|X, C=c)$.

B. Estimation Algorithm

The authors develop an iterative, data-driven algorithm that simultaneously estimates cluster membership and the cluster-specific DRFs. This mirrors clusterwise regression (or $k$-means regression) but adapted for causal inference with continuous treatments.

1. Parametric Models:
   • Treatment Model: Assumes a cluster-specific Gaussian distribution for treatment: $T_i | X_i, C_i=c \sim N(\beta_c' X_i, \sigma_c^2)$. This generates the cluster-specific GPS ($R_{i,c}$).
   • Outcome Model: Assumes a parametric regression of the outcome on treatment and the GPS: $E[Y_i | T_i, R_{i,c}, C_i=c] = f(T_i, R_{i,c}; \alpha_c)$.
2. Iterative Procedure (Block-Coordinate Descent):
   • Step 1: Initialize cluster assignments.
   • Step 2: Estimate cluster-specific treatment models to compute GPS.
   • Step 3: Estimate cluster-specific outcome regressions using the GPS as a control variable.
   • Step 4: Update cluster assignments by assigning each unit to the cluster that minimizes the squared residual of the outcome model.
   • Step 5: Repeat until convergence.
3. Model Selection (Choosing $C$):
   • Since the number of clusters is unknown, the authors propose a BIC-like penalized empirical risk criterion:
     $$IC(C) = \sum (y_i - \hat{\alpha}_{c_i}' z_i)^2 + \zeta_n \sum \text{dim}(\hat{\alpha}_c)$$
   • They use an "Elbow Rule" to select the optimal $C$ by identifying the point where adding more clusters yields diminishing returns in fit relative to model complexity.

3. Key Contributions

1. Theoretical Innovation: Introduces the Cl-DRF as a new causal estimand that explicitly models heterogeneous responses in continuous treatment settings, moving beyond the limitations of the population ADRF.
2. Methodological Advancement: Develops a unified estimator that jointly discovers the clustering structure (data-driven) and estimates causal effects, relaxing the global positivity assumption to a more plausible local version.
3. Handling Limited Common Support: Demonstrates that by restricting analysis to clusters with similar covariate profiles and treatment ranges, the method avoids the extrapolation bias inherent in global estimators when common support is limited.
4. Integration of Clustering and Causal Inference: Bridges the gap between grouped regression literature (e.g., $k$-means) and causal inference frameworks (GPS), providing a tractable parametric solution for complex observational data.

4. Empirical Results

The authors apply the Cl-DRF to evaluate the impact of European Union (EU) Cohesion Funds on regional economic growth (1994–2015).

• Standard Approach (Global ADRF): Using Hirano & Imbens (2004) and Huling et al. (2024) methods, the estimated ADRF shows a modest positive effect that flattens and turns negative for high funding levels. This suggests a "diminishing returns" or "over-subsidization" hypothesis.
• Cl-DRF Approach: The algorithm identifies three distinct clusters of regions:
  • Cluster 1 (Wealthier regions, e.g., Ireland, Southern Germany): Shows a strong, positive, and linear relationship between funds and growth. No evidence of diminishing returns or negative effects at high intensities.
  • Cluster 2 (Mixed regions, e.g., Spain, Northern Germany): Shows a relationship similar to the global average (positive but flattening).
  • Cluster 3 (Less developed regions, e.g., Italy, Greece): Shows a slightly negative relationship. The authors attribute this to limited absorptive capacity and the widening regional inequalities during the Great Recession (2008–2015), where funds failed to stimulate growth.
• Simulation Evidence: Extensive simulations confirm that the Cl-DRF estimator accurately recovers the true number of clusters (high Rand Index) and the true cluster-specific DRFs, even under conditions of limited common support and confounding.

5. Significance and Policy Implications

• Policy Relevance: The study demonstrates that a "one-size-fits-all" policy recommendation based on a global ADRF is misleading.
  • For Cluster 1: Policymakers should avoid arbitrary funding caps, as these regions can absorb and benefit from higher intensities.
  • For Cluster 3: Simply increasing funding is ineffective; policy should focus on building absorptive capacity (human capital, institutions) before increasing financial transfers.
• Methodological Impact: The paper provides a robust toolkit for researchers dealing with continuous treatments in observational data where heterogeneity is structural rather than just a function of observed covariates. It offers a solution to the "positivity violation" problem by localizing the analysis.
• Future Directions: The authors suggest future work could explore non-parametric outcome models and formal statistical tests for the validity of the clustering assumption.

In summary, the paper argues that heterogeneity in continuous treatment effects is often structural and regime-based. By adopting a cluster-based framework, researchers can uncover distinct causal mechanisms that are otherwise obscured, leading to more accurate causal estimates and nuanced, effective policy design.