Conditional average treatment effect estimation with marginally constrained models

Wouter A. C. van Amsterdam, Rajesh Ranganath

Abstract

Treatment effect estimates are often available from randomized controlled trials as a single average treatment effect for a certain patient population. Estimates of the conditional average treatment effect (CATE) are more useful for individualized treatment decision making, but randomized trials are often too small to estimate the CATE. Examples in medical literature make use of the relative treatment effect (e.g. an odds-ratio) reported by randomized trials to estimate the CATE using large observational datasets. One approach to estimating these CATE models is by using the relative treatment effect as an offset, while estimating the covariate-specific untreated risk. We observe that the odds-ratios reported in randomized controlled trials are not the odds-ratios that are needed in offset models because trials often report the marginal odds-ratio. We introduce a constraint or regularizer to better use marginal odds-ratios from randomized controlled trials and find that under the standard observational causal inference assumptions this approach provides a consistent estimate of the CATE. Next, we show that the offset approach is not valid for CATE estimation in the presence of unobserved confounding. We study if the offset assumption and the marginal constraint lead to better approximations of the CATE relative to the alternative of using the average treatment effect estimate from the randomized trial. We empirically show that when the underlying CATE has sufficient variation, the constraint and offset approaches lead to closer approximations to the CATE.

Introduction

In modern clinical practice, the "gold standard" for evidence is the Randomized Controlled Trial (RCT). However, RCTs typically report the Average Treatment Effect (ATE), representing the efficacy for a broad population. Clinicians, however, require the Conditional Average Treatment Effect (CATE) to make informed decisions for specific patients. While large observational datasets like Electronic Health Records (EHR) offer the scale needed for CATE estimation, they lack the randomization of trials.

A common technique to bridge this gap is the "offset method," where a relative treatment effect (e.g., an odds ratio) from an RCT is used as a fixed offset in a model trained on observational data. This paper identifies a critical flaw in this practice: non-collapsibility. In logistic models, the marginal odds ratio from a trial is mathematically distinct from the conditional odds ratio required for individualized models. Using them interchangeably introduces bias and inconsistency.

Marginally Constrained Models (MCM)

To address the mismatch between marginal trial results and conditional model requirements, we propose the Marginally Constrained Model (MCM). Rather than fixing a parameter as an offset, the MCM incorporates a constraint or regularizer into the model’s objective function.

The MCM allows the model to learn individual variations from observational data while penalizing deviations from the marginal effect reported in the RCT. We demonstrate that under standard causal inference assumptions—specifically the absence of unobserved confounding—this approach yields a consistent estimator of the true CATE. We also introduce the Constant-Relative Marginally Constrained Model (CR-MCM), which assumes a stable relative treatment effect across patient subgroups while adhering to the marginal constraint.

Experimental Setup

The performance of the MCM was evaluated using synthetic experimental grids. This controlled environment allowed for the precise manipulation of confounding strength, treatment effect variation, and the degree of non-collapsibility.

Table 1: Simulation parameters used to evaluate the MCM and offset models across various scenarios, including baseline risks and sample sizes.

As shown in the parameter table above, simulations covered a wide range of baseline risks ($\beta_0$), treatment effects ($\beta_t$), and heterogeneity ($\beta_x$). Sample sizes ($n$) ranged from 50 to 750 to evaluate statistical efficiency and convergence.

Results and Discussion

Our analysis reveals that the popular offset method is fundamentally flawed for CATE estimation due to its neglect of non-collapsibility. Even with perfect trial data, the offset approach produces incorrect conditional estimates.

Statistical Efficiency

The MCM offers significant gains in statistical efficiency. By leveraging the trial result as a constraint, the model requires substantially less observational data. Our findings indicate that an unconstrained model requires between 71% and 106% more samples to achieve the same level of confidence as the MCM.

Performance under Confounding

In scenarios involving unobserved confounding, we compared the MCM and offset models against a baseline of simply applying the trial's ATE to all patients. When the treatment effect exhibited meaningful variation, the MCM and offset approaches outperformed the ATE-baseline in 90-99% of simulated scenarios. This suggests that even when the "no unobserved confounding" assumption is violated, these models often provide better approximations for patient care than the population average.

Limitations and Future Directions

While promising, several limitations remain. Our current experiments utilized a single binary covariate; the efficiency gains of the marginal constraint may diminish in high-dimensional settings typical of real-world EHR data. Furthermore, the model assumes "transportability"—that the trial and observational populations are sufficiently similar. Finally, the current framework treats the trial's marginal effect as a point estimate; future iterations could incorporate Bayesian methods to account for the inherent uncertainty in RCT results.

Summary updated

Abstract

Treatment effect estimates are often available from randomized controlled trials as a single average treatment effect for a certain patient population. Estimates of the conditional average treatment effect (CATE) are more useful for individualized treatment decision making, but randomized trials are often too small to estimate the CATE. Examples in medical literature make use of the relative treatment effect (e.g. an odds-ratio) reported by randomized trials to estimate the CATE using large observational datasets. One approach to estimating these CATE models is by using the relative treatment effect as an offset, while estimating the covariate-specific untreated risk. We observe that the odds-ratios reported in randomized controlled trials are not the odds-ratios that are needed in offset models because trials often report the marginal odds-ratio. We introduce a constraint or regularizer to better use marginal odds-ratios from randomized controlled trials and find that under the standard observational causal inference assumptions this approach provides a consistent estimate of the CATE. Next, we show that the offset approach is not valid for CATE estimation in the presence of unobserved confounding. We study if the offset assumption and the marginal constraint lead to better approximations of the CATE relative to the alternative of using the average treatment effect estimate from the randomized trial. We empirically show that when the underlying CATE has sufficient variation, the constraint and offset approaches lead to closer approximations to the CATE.
BibTeX
@article{amsterdam2023conditional,
  author = {Wouter A. C. van Amsterdam and Rajesh Ranganath},
  title = {{Conditional average treatment effect estimation with marginally constrained models}},
  journal = {Journal of Causal Inference},
  year = {2023},
  doi = {10.1515/jci-2022-0027},
}

Loading related content...

Figure