|
Abstract:
|
Micro-randomized trial (MRT) is a novel experimental design for developing and optimizing digital interventions. In MRT, each individual is repeatedly randomized among treatment options for hundreds or thousands of times. We consider estimation of the causal excursion effect using the intensive longitudinal data from MRTs. In such an estimation task, a challenge is how to utilize the high-dimensional history information in the longitudinal data with time-varying treatments to improve efficiency, while ensuring inference validty under model-misspecification. We propose a two-stage estimator to address the challenge: in the first stage, a large class of machine learning algorithms is used to fit a nuisance function; in the second stage, an estimating equation is formed with the plug-in nuisance function and the causal parameter is estimated. We show that under mild conditions the estimator is asymptotically normal, and we derive its asymptotic variance. Simulation studies are conducted to show substantial efficiency gain by using the new estimator compared to existing estimators in the literature that do not leverage the machine learning algorithms.
|
