Optimizing precision in stepped-wedge designs via machine learning and quadratic inference functions

Stepped-wedge designs are increasingly used in randomized experiments to accommodate logistical and ethical constraints by staggering treatment roll-out over time. Despite their popularity, existing analytical methods largely rely on parametric models with linear covariate adjustment and prespecified correlation structures, which may limit achievable precision in practice. We propose a new class of estimators for the causal average treatment effect in stepped-wedge designs that optimizes precision through flexible, machine-learning-based covariate adjustment to capture complex outcome-covariate relationships, together with quadratic inference functions to adaptively learn the correlation structure. We establish consistency and asymptotic normality under mild conditions requiring only $L_2$ convergence of nuisance estimators, even under model misspecification, and characterize when the estimator attains the minimal asymptotic variance. Moreover, we prove that the proposed estimator never reduces efficiency relative to an independence working correlation. The proposed method further accommodates treatment-effect heterogeneity across both exposure duration and calendar time. Finally, we demonstrate our methods through simulation studies and reanalyses of two empirical studies that differ substantially in research area and key design parameters.

💡 Research Summary

Stepped‑wedge designs have become a popular solution for randomized experiments that must stagger treatment rollout due to logistical or ethical constraints. Traditional analyses of such designs rely heavily on parametric mixed‑effects models or generalized estimating equations (GEE) that assume linear covariate adjustment and a pre‑specified correlation structure (e.g., exchangeable, AR‑1). These assumptions can severely limit statistical precision when the true outcome‑covariate relationship is nonlinear or when the correlation structure is misspecified.

The authors propose a novel class of estimators that simultaneously leverages flexible machine‑learning‑based covariate adjustment and quadratic inference functions (QIF) to learn the correlation structure adaptively. The core model is a partial‑linear regression: the treatment effect enters linearly (or via a vector of period‑specific indicators) while the nuisance function (g_j(X)) captures the conditional mean of the outcome given baseline covariates. To estimate (g_j), the paper adopts a cross‑fitting scheme: clusters are partitioned into (M) folds, and for each fold a machine‑learning algorithm (e.g., random forests, gradient boosting, neural networks) is trained on the remaining folds and then used to predict outcomes for the held‑out fold. This procedure yields an estimator of the nuisance function that converges in (L_2) norm without over‑fitting, even when the algorithm is highly adaptive.

For the correlation structure, QIF replaces the traditional working correlation matrix. QIF constructs a set of basis matrices representing multiple candidate correlation patterns and estimates optimal weights from the data, thereby achieving a data‑driven compromise among competing structures. The authors prove that, under mild regularity conditions requiring only (L_2) convergence of the nuisance estimators, the resulting estimator of the average treatment effect is consistent and asymptotically normal. Moreover, when the mean model is correctly specified and the nuisance estimator is consistent, the estimator attains the semiparametric efficiency bound for individually randomized stepped‑wedge trials. Importantly, they establish a “no‑loss” theorem: the QIF‑augmented estimator never has larger asymptotic variance than the estimator that assumes independence, guaranteeing that the adaptive correlation learning cannot hurt precision.

The framework accommodates four common treatment‑effect structures: (i) constant effect, (ii) duration‑specific effect, (iii) period‑specific effect, and (iv) saturated effect that varies both by exposure duration and calendar time. By embedding the appropriate design matrix (D^) and parameter vector (\beta^) into the partial‑linear model, the same estimation algorithm yields unbiased estimates for any of these structures, allowing researchers to explore heterogeneity without re‑deriving estimators.

Simulation studies explore a wide range of realistic scenarios: unequal cluster sizes, various true correlation patterns, non‑Gaussian outcomes (binary, count), and misspecified working correlations. The proposed method consistently outperforms standard mixed‑effects models, ordinary GEE, and recent robust standardization approaches in terms of mean‑squared error and confidence‑interval width. Gains are especially pronounced when the outcome‑covariate relationship is nonlinear or when the true correlation deviates from the working assumption.

Two empirical applications illustrate practical impact. The first re‑analyzes a cluster‑randomized stepped‑wedge trial in Swiss nursing homes evaluating an Integrated Palliative Care Outcome Scale (IPOS) for dementia patients. Using machine‑learning covariate adjustment, the standard error of the estimated QUALIDEM treatment effect shrank by roughly 30 % compared with the original mixed‑effects analysis, and the duration‑specific model revealed a cumulative improvement over successive periods. The second re‑examines a city‑wide procedural‑justice training program for Chicago police officers, an individually randomized stepped‑wedge design. QIF identified a correlation pattern that improved efficiency relative to the independence assumption, while the flexible covariate adjustment uncovered a time‑varying effect: modest immediate impact that grew larger after several months of training.

In sum, the paper delivers a comprehensive, theoretically grounded, and computationally feasible approach to maximize precision in stepped‑wedge designs. By marrying modern machine‑learning tools with quadratic inference functions, it achieves robustness to both model misspecification and correlation‑structure uncertainty, while retaining the ability to model rich treatment‑effect heterogeneity. This methodology is poised to become a standard tool for researchers across health, education, and social sciences who employ stepped‑wedge trials.