NMMIPW: Inverse Probability Weighting under Non-Monotone Missing

Andrew YingOpenalex Unknown 2021 · DOI · OpenAlex

Abstract

We fit inverse probability weighting estimator and the augmented inverse probability weighting for non-monotone missing at random data.

Introduction

In statistical analysis, missing data is a pervasive challenge. Traditional methods often rely on the assumption of "monotone" missingness—a pattern where once a data point is missing for a subject, all subsequent observations are also missing. However, real-world datasets frequently exhibit non-monotone missingness, where variables are missing sporadically without a sequential pattern. The NMMIPW R package addresses this by implementing Inverse Probability Weighting (IPW) and Augmented Inverse Probability Weighting (AIPW) estimators specifically tailored for data that is Missing at Random (MAR) under non-monotone conditions.

The nmm_fit Framework

The core of the package is the nmm_fit function. This function allows researchers to estimate model parameters while accounting for the complex probabilities associated with missing data patterns. Unlike simpler tools, nmm_fit can handle a matrix of missingness indicators, enabling the tracking of overlapping and non-sequential gaps in the data.

Arguments for the nmm_fit function, including data, missing indicators, and fitting options.

Estimation Methods

The package supports two primary estimation strategies:

  1. Inverse Probability Weighting (IPW): This method reweights observed cases by the inverse of their observation probability. This creates a pseudo-population that corrects for the bias introduced by missingness, provided the missingness model is correctly specified.
  2. Augmented IPW (AIPW): By incorporating an outcome model alongside the missingness model, AIPW offers "double robustness." The estimator remains unbiased if either the propensity score model or the outcome model is correctly specified, often leading to improved efficiency.

The package integrates seamlessly with standard R modeling workflows, supporting both linear models (lm) and generalized linear models (glm).

Statistical Inference and Outputs

A key strength of NMMIPW is its focus on rigorous statistical inference. Beyond point estimates, the package provides the necessary components for hypothesis testing and uncertainty quantification.

Return values of the nmm_fit function, detailing the fitted coefficients and influence functions.

As shown in the output structure, the function returns not only the fitted coefficients but also their standard deviations (coef_sd) and influence functions (coef_IF). These influence functions are particularly valuable for advanced users performing sensitivity analyses or meta-analyses.

Discussion and Limitations

While NMMIPW provides a robust solution for non-monotone MAR data, it has certain limitations. It currently relies on lm and glm for modeling; future iterations could benefit from integrating machine learning techniques like Random Forests for more flexible propensity score estimation. Additionally, the method assumes the MAR mechanism; if data is Missing Not at Random (MNAR), results may remain biased. Finally, the computational cost of numerical derivatives and optimization may increase significantly with very large datasets or highly complex missingness patterns.

In conclusion, NMMIPW offers a specialized and accessible toolkit for researchers in fields like epidemiology and clinical trials, where "patchy" data is the norm rather than the exception.

Summary updated

Abstract

We fit inverse probability weighting estimator and the augmented inverse probability weighting for non-monotone missing at random data.
BibTeX
@inproceedings{ying2021nmmipw,
  author = {Andrew Ying},
  title = {{NMMIPW: Inverse Probability Weighting under Non-Monotone Missing}},
  booktitle = {Openalex Unknown},
  year = {2021},
  doi = {10.32614/cran.package.nmmipw},
}
Figure