Complier General Causal Effect in Randomized Controlled Trials with One-Sided Noncompliance

1. Problem Statement

The paper addresses the challenge of estimating causal effects in Randomized Controlled Trials (RCTs) when one-sided noncompliance occurs.

• Context: In many RCTs (e.g., medical drug trials or training programs), individuals assigned to the treatment group may choose not to take the treatment, while those assigned to the control group cannot access the treatment.
• The Issue: Standard Intention-to-Treat (ITT) analysis estimates the effect of assignment, not the effect of receiving the treatment. Furthermore, the causal effect for the entire population is generally not identifiable because the compliance status (whether a subject is a "complier" or a "never-taker") is unobserved for those assigned to the control group.
• Limitations of Existing Methods: Previous literature has focused on specific estimands like the Complier Average Causal Effect (CACE/LATE) or Complier Quantile Causal Effect (CQCE). Existing efficient estimators often rely on strong assumptions regarding the convergence rates of nuisance parameter estimators (e.g., requiring specific rates for kernel or sieve methods), which limits the use of modern, flexible machine learning (ML) algorithms.

2. Methodology

A. Framework and Identifiability

The authors establish a rigorous likelihood-based identifiability framework for one-sided noncompliance.

• Data Structure: They define the observed data as $(X, Z, T, Y)$, where $Z$ is assignment, $T$ is treatment received, and $W$ is the unobserved compliance status ($W=1$ for compliers, $W=0$ for never-takers). Under one-sided noncompliance, $T = ZW$.
• Nonparametric Likelihood: By constructing the nonparametric likelihood function, they prove that the model is identifiable. Specifically, they show that the joint distribution of the observed data uniquely determines the underlying components: the density of outcomes for compliers ($f_1$), the density for never-takers ($f_0$), and the compliance probability ($q(x)$).
• New Estimand (CGCE): Based on this identifiability, they propose the Complier General Causal Effect (CGCE), denoted as $\tau = \tau_1 - \tau_0$.
  • $\tau_k$ is defined as the solution to $E[u(Y_k, \tau_k) | W=1] = 0$.
  • This is a unified framework: choosing $u(y, \tau) = y - \tau$ yields the CACE; choosing $u(y, \tau) = \mathbb{I}(y \le \tau) - \alpha$ yields the CQCE.

B. Estimation Strategies

The paper develops two estimators for the CGCE:

1. Simple Estimator ($\hat{\tau}_s$):

   • A root-$n$ consistent estimator that does not require nonparametric estimation or machine learning.
   • It solves estimating equations based on marginal probabilities and observed outcomes, treating the nuisance components as zero or using simple moment conditions.
   • While simple to implement, it is shown to be inefficient.

2. Efficient Estimator ($\hat{\tau}$):

   • Derivation: The authors derive the Efficient Influence Function (EIF) by projecting the influence function of the simple estimator onto the tangent space of the semiparametric model. This yields a transparent decomposition involving interpretable nuisance components ($\mu_1, \mu_2, \mu_3, q$) tied directly to the data structure.
   • Implementation: The estimator solves the empirical EIF equation. It utilizes sample splitting (cross-fitting) to estimate the nuisance functions ($\mu$ and $q$) and the target parameter.
   • Key Theoretical Breakthrough (Rate-Free Condition): Theorem 2 establishes that achieving semiparametric efficiency requires only that the nuisance estimators converge in $L_2$-norm (consistency). No specific convergence rates (e.g., $n^{-1/4}$) are required.
   • Implication: This "rate-free" property allows the use of modern, complex ML methods (Deep Neural Networks, Random Forests, etc.) that may not satisfy traditional smoothness or rate conditions but are consistent, without sacrificing the efficiency of the final causal estimate.

3. Key Contributions

1. Systematic Identifiability Analysis: The paper provides the first systematic treatment of likelihood-based identifiability in RCTs with one-sided noncompliance, clearly delineating which causal effects are identifiable and which are not.
2. Unified Estimand (CGCE): It introduces the CGCE, a general framework that subsumes CACE and CQCE, allowing for flexible causal inference beyond just mean or quantile effects.
3. Rate-Free Efficiency: The most significant theoretical contribution is the proof that semiparametric efficiency can be achieved with nuisance estimators that only satisfy $L_2$-consistency. This removes the "product-rate" bottleneck common in semiparametric literature, enabling the robust integration of modern ML.
4. Transparent EIF Derivation: Unlike approaches using optimal instrumental variables, the authors derive the EIF by projecting the simple estimator's influence function, resulting in a decomposition with clear statistical interpretations linked to the data structure.

4. Results

Simulation Studies

• Setup: Simulations were conducted with varying covariate dimensions ($d=1, 4, 9$) and different data generating processes (linear and non-linear).
• Performance:
  • Bias: All estimators (Simple, Efficient-Kernel, Efficient-NN) showed negligible bias, confirming consistency.
  • Variance: The Simple estimator had significantly higher variance (inefficient). The Efficient estimator achieved the semiparametric efficiency bound.
  • ML Integration: In high-dimensional settings ($d=9$), the Efficient estimator combined with Deep Neural Networks (Eff-NN) outperformed the kernel-based version, demonstrating lower variability and better coverage. The Eff-NN results were nearly identical to the "Oracle" estimator (which uses true nuisance functions).
  • Inference: The 95% confidence intervals constructed using the Efficient estimator achieved empirical coverage rates close to the nominal 95%.

Real Data Application

• Dataset: A microcredit experiment in Morocco ($n=4,934$ households) analyzing the effect of credit on self-employment output.
• Findings:
  • The Efficient-NN estimator yielded a point estimate of $\hat{\tau} = 1.113$ with a 95% CI of $[-0.071, 2.296]$.
  • The Simple estimator yielded $\hat{\tau} = 1.425$ with a 95% CI of $[-0.029, 2.878]$.
  • Both intervals included zero, suggesting no statistically significant causal effect of microcredit on self-employment output in this specific subsample.
  • The results differed from a previous parametric study (Sawada, 2019), highlighting the importance of the nonparametric, flexible approach taken here.

5. Significance

This paper bridges the gap between rigorous causal inference theory and modern machine learning. By proving that semiparametric efficiency does not depend on the convergence rates of nuisance estimators, the authors remove a major barrier to using flexible ML algorithms in causal inference. This allows researchers to model complex, high-dimensional relationships in compliance and outcomes without compromising the statistical validity of the causal effect estimates. The proposed CGCE framework further generalizes the scope of causal questions that can be answered in noncompliant RCTs.