A Unified Inference Framework for Multiple Imputation Using Martingales

TL;DR

This paper introduces a unified martingale-based inference framework for multiple imputation, ensuring consistent variance estimation across various estimators and handling complex missing data scenarios.

Contribution

It develops a martingale representation for multiple imputation, enabling robust inference with a wide range of asymptotically linear estimators, including causal effect estimation.

Findings

Provides a unified inference method applicable to many estimators.

Ensures consistent variance estimation under model misspecification.

Extends to complex missing data scenarios, including non-random missingness.

Abstract

Multiple imputation is widely used to handle missing data. Although Rubin's combining rule is simple, it is not clear whether or not the standard multiple imputation inference is consistent when coupled with the commonly-used full sample estimators. This article establishes a unified martingale representation of multiple imputation for a wide class of asymptotically linear full sample estimators. This representation invokes the wild bootstrap inference to provide consistent variance estimation under the correct specification of the imputation models. As a motivating application, we illustrate the proposed method to estimate the average causal effect (ACE) with partially observed confounders in causal inference. Our framework applies to asymptotically linear ACE estimators, including the regression imputation, weighting, and matching estimators. We extend to the scenarios when both outcome and confounders are subject to missingness and when the data are missing not at random.