Data-adaptive doubly robust instrumental variable methods for treatment effect heterogeneity

Karla Díaz-Ordaz, Rhian Daniel, Noémi KreifarXiv 2018 · DOI · arXiv

Abstract

We consider the estimation of the average treatment effect in the treated as a function of baseline covariates, where there is a valid (conditional) instrument. We describe two doubly robust (DR) estimators: a locally efficient g-estimator, and a targeted minimum loss-based estimator (TMLE). These two DR estimators can be viewed as generalisations of the two-stage least squares (TSLS) method to semi-parametric models that make weaker assumptions. We exploit recent theoretical results that extend to the g-estimator the use of data-adaptive fits for the nuisance parameters. A simulation study is used to compare standard TSLS with the two DR estimators' finite-sample performance, (1) when fitted using parametric nuisance models, and (2) using data-adaptive nuisance fits, obtained from the Super Learner, an ensemble machine learning method. Data-adaptive DR estimators have lower bias and improved coverage, when compared to incorrectly specified parametric DR estimators and TSLS. When the parametric model for the treatment effect curve is correctly specified, the g-estimator outperforms all others, but when this model is misspecified, TMLE performs best, while TSLS can result in large biases and zero coverage. Finally, we illustrate the methods by reanalysing the COPERS (COping with persistent Pain, Effectiveness Research in Self-management) trial to make inference about the causal effect of treatment actually received, and the extent to which this is modified by depression at baseline.

Introduction

In clinical trials and observational studies, researchers often aim to estimate the "real" effect of a treatment. However, non-adherence—where participants do not follow their assigned treatment protocol—frequently complicates causal inference. To estimate the effect of the treatment actually received (the Average Treatment Effect in the Treated, or ATT), Instrumental Variable (IV) methods are commonly employed. In a Randomized Controlled Trial (RCT), the random assignment serves as a valid instrument.

Standard IV approaches, such as Two-Stage Least Squares (TSLS), rely on rigid parametric assumptions. These assumptions are particularly problematic when treatment effects are heterogeneous, meaning they vary according to baseline characteristics. If the model fails to capture this complexity, TSLS can produce biased results and misleadingly narrow confidence intervals. This paper explores "doubly robust" (DR) IV methods that integrate machine learning to provide more reliable estimates under model misspecification.

The Instrumental Variable Framework

The causal structure considered in this work is represented by a Directed Acyclic Graph (DAG), where $Z$ is the instrument, $A$ is the treatment, $Y$ is the outcome, $W$ represents baseline covariates, and $U$ denotes unobserved confounding.

Causal structure showing the relationships between the instrument (Z), treatment (A), outcome (Y), baseline covariates (W), and unobserved confounding (U).

Methods

The authors focus on two semi-parametric DR estimators:

  1. G-estimation: A locally efficient method that utilizes the lack of residual association between the instrument and the treatment-free potential outcome.
  2. Targeted Minimum Loss-based Estimation (TMLE): A framework that uses a targeting step to optimize the bias-variance tradeoff for the specific parameter of interest.

These estimators are "doubly robust" because they remain consistent if either the model for the exposure or the model for the outcome is correctly specified. To further enhance robustness, the authors employ the Super Learner, an ensemble machine learning method. Super Learner uses cross-validation to find the optimal combination of various algorithms (e.g., GLMs, SVMs, GAMs) to fit the "nuisance parameters"—components of the model necessary for estimation but not of primary interest.

Simulation Study and Real-World Application

The performance of these methods was evaluated through a factorial simulation study comparing TSLS, parametric DR estimators, and data-adaptive DR estimators across different sample sizes ($n=500$ and $n=10,000$) and model specifications.

Additionally, the methods were applied to the COPERS Trial (COping with persistent Pain, Effectiveness Research in Self-management). This trial investigated a cognitive behavioral intervention for chronic back pain, specifically examining whether baseline depression levels modified the treatment effect.

Key Results and Insights

The simulations revealed that TSLS is highly fragile under model misspecification in the presence of effect heterogeneity, often resulting in "zero coverage" where confidence intervals fail to contain the true effect. In contrast, data-adaptive DR estimators consistently showed lower bias and improved coverage.

Key findings include:

  • Efficiency vs. Robustness: The g-estimator is generally more efficient when the shape of the treatment effect curve (e.g., linear) is correctly specified. However, TMLE is more robust when the curve's shape is unknown, as it projects estimates onto the working model of interest.
  • Machine Learning Benefits: Using Super Learner to fit nuisance parameters significantly improved the reliability of results compared to simple parametric models.
  • COPERS Analysis: The re-analysis found no strong evidence that the intervention improved disability scores or that depression modified the effect. However, the DR methods provided more robust confidence intervals that better accounted for non-adherence.

Conclusion and Future Directions

While data-adaptive DR methods offer significant advantages, challenges remain. IV-TMLE can become unstable if the instrument is "weak" (i.e., weakly predictive of treatment). Future work could explore "sample splitting" to further improve the properties of machine learning-based estimators and investigate the theoretical convergence rates of algorithms within the Super Learner library.

Summary updated

Abstract

We consider the estimation of the average treatment effect in the treated as a function of baseline covariates, where there is a valid (conditional) instrument. We describe two doubly robust (DR) estimators: a locally efficient g-estimator, and a targeted minimum loss-based estimator (TMLE). These two DR estimators can be viewed as generalisations of the two-stage least squares (TSLS) method to semi-parametric models that make weaker assumptions. We exploit recent theoretical results that extend to the g-estimator the use of data-adaptive fits for the nuisance parameters. A simulation study is used to compare standard TSLS with the two DR estimators' finite-sample performance, (1) when fitted using parametric nuisance models, and (2) using data-adaptive nuisance fits, obtained from the Super Learner, an ensemble machine learning method. Data-adaptive DR estimators have lower bias and improved coverage, when compared to incorrectly specified parametric DR estimators and TSLS. When the parametric model for the treatment effect curve is correctly specified, the g-estimator outperforms all others, but when this model is misspecified, TMLE performs best, while TSLS can result in large biases and zero coverage. Finally, we illustrate the methods by reanalysing the COPERS (COping with persistent Pain, Effectiveness Research in Self-management) trial to make inference about the causal effect of treatment actually received, and the extent to which this is modified by depression at baseline.
BibTeX
@misc{diazordaz2018data,
  author = {Karla Díaz-Ordaz and Rhian Daniel and Noémi Kreif},
  title = {{Data-adaptive doubly robust instrumental variable methods for treatment effect heterogeneity}},
  year = {2018},
  eprint = {1802.02821},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1802.02821},
}
Figure