Doubly robust uniform confidence band for the conditional average treatment effect function

Sokbae Lee, Ryo Okui, Yoon‐Jae Whang

This paper, published in 2016, addresses a fundamental challenge in causal inference: how to visualize and statistically validate how a treatment's effect varies across different groups (heterogeneity) when you have a mountain of control variables but only care about a few specific factors.

TL;DR

The authors develop a "doubly robust" framework to estimate the Conditional Average Treatment Effect (CATE) function while controlling for high-dimensional covariates. Their primary contribution is an easy-to-compute uniform confidence band that allows researchers to perform global inference—such as testing if a treatment effect is constant across all ages—without falling victim to the "curse of dimensionality."

What Problem Does This Paper Address?

In modern causal inference, researchers often face a dilemma. To satisfy the "unconfoundedness" assumption (the idea that treatment is effectively random once you control for enough factors), they must include a large number of covariates ($Z$). However, they are usually only interested in how the treatment effect changes with respect to a few variables ($X$), like age or income.

Standard nonparametric methods (like basic kernel smoothing) fail when $Z$ is large because the data becomes too sparse—the infamous "curse of dimensionality." Conversely, purely parametric models are often too rigid and prone to misspecification. Furthermore, most existing tools only provide "pointwise" confidence intervals, which tell you the uncertainty at a single point but can't tell you if the entire shape of the resulting curve is statistically significant.

Methods

The authors propose a two-stage approach:

  1. Doubly Robust (DR) Estimation: They use parametric models for two "nuisance" functions: the propensity score (probability of receiving treatment) and the outcome regressions (expected outcome given covariates). The estimator is "doubly robust" because it remains consistent if either the propensity score model or the outcome regression models are correctly specified.
  2. Local Linear Smoothing: They take the "pseudo-outcomes" generated by the DR step and apply local linear regression to estimate the CATE function $g(x)$ over the low-dimensional variables of interest.
  3. Uniform Confidence Bands: The core innovation is the construction of a $(1-\alpha)$ uniform confidence band. Unlike pointwise intervals, this band covers the entire true function across a range with a specified probability. They use a Gaussian approximation and a higher-order expansion (based on Piterbarg’s method) to calculate critical values that are more accurate in small samples than the standard Gumbel approximation.

Datasets

The authors evaluate their method using:

  • Monte Carlo Simulations: Synthetic data with 10 and 30 covariates to test bias and coverage under various misspecification scenarios.
  • Pennsylvania Birth Weight Data: A dataset of 3,754 white, non-Hispanic mothers used to study the effect of smoking on infant birth weight.
  • North Carolina Birth Weight Data: A much larger dataset ($N = 433,558$) used to demonstrate how the method performs with massive samples and to provide more precise estimates of the smoking effect.

Key Insights and Ideas

  • Decoupling Identification from Analysis: You can use high-dimensional parametric models to "clean" the data for causal identification, then use low-dimensional nonparametric models to "analyze" the heterogeneity.
  • First-Stage "Invisibility": Because the first stage (parametric) converges much faster than the second stage (nonparametric smoothing), the errors from estimating the initial parameters do not affect the first-order asymptotic distribution of the final CATE estimate.
  • Global vs. Local Inference: Pointwise intervals are often too narrow for testing hypotheses like "is the effect of smoking the same for a 20-year-old as it is for a 30-year-old?" The uniform band provides a rigorous way to reject the null hypothesis of a constant treatment effect visually.
  • Precision over Gumbel: The authors show that the traditional Gumbel distribution approach to confidence bands is often too conservative (too wide). Their analytic correction provides a tighter, more accurate fit.

Key Results

  • Simulation Performance: The DR estimator effectively removed bias as long as one of the two nuisance models was correct. The proposed uniform bands maintained the nominal coverage (e.g., 95%) much better than simpler approximations.
  • Smoking and Birth Weight:
    • In the small Pennsylvania dataset, the CATE appeared to vary with age, but the uniform bands were wide enough that a "constant effect" (a straight line) couldn't be strictly rejected.
  • In the massive North Carolina dataset, the bands narrowed significantly. The results clearly showed that maternal smoking has a negative effect on birth weight across all ages (15–35), and the authors were able to statistically reject the hypothesis that this effect is constant across different maternal ages.

Limitations and Future Work

  • Undersmoothing Requirement: The theoretical validity of the bands relies on "undersmoothing" (choosing a smaller-than-optimal bandwidth) to make bias negligible. Choosing the "right" amount of undersmoothing in practice remains an open challenge.
  • Dimensionality of X: While $Z$ can be high-dimensional, the variables of interest $X$ must remain low-dimensional (typically $d < 4$) for the smoothing to remain effective.
  • Extensions: The authors suggest that future work could extend this "doubly robust uniform band" approach to:
  • Quantile Treatment Effects (looking at the bottom 10% of birth weights rather than the average).
  • Local Average Treatment Effects (LATE) using instrumental variables.
  • Marginal Treatment Effects (MTE).

Summary updated

Abstract

Summary In this paper, we propose a doubly robust method to estimate the heterogeneity of the average treatment effect with respect to observed covariates of interest. We consider a situation where a large number of covariates are needed for identifying the average treatment effect but the covariates of interest for analyzing heterogeneity are of much lower dimension. Our proposed estimator is doubly robust and avoids the curse of dimensionality. We propose a uniform confidence band that is easy to compute, and we illustrate its usefulness via Monte Carlo experiments and an application to the effects of smoking on birth weights.
BibTeX
@article{lee2017doubly,
  author = {Sokbae Lee and Ryo Okui and Yoon‐Jae Whang},
  title = {{Doubly robust uniform confidence band for the conditional average treatment effect function}},
  journal = {Journal of Applied Econometrics},
  year = {2017},
  doi = {10.1002/jae.2574},
}

Loading related content...

Figure