Reflection on modern methods: shared-parameter models for longitudinal studies with missing data

Abstract

A primary goal of longitudinal studies is to examine trends over time. Reported results from these studies often depend on strong, unverifiable assumptions about the missing data. Whereas the risk of substantial bias from missing data is widely known, analyses exploring missing-data influences are commonly done either ad hoc or not at all. This article outlines one of the three primary recognized approaches for examining missing-data effects that could be more widely used, i.e. the shared-parameter model (SPM), and explains its purpose, use, limitations and extensions. We additionally provide synthetic data and reproducible research code for running SPMs in SAS, Stata and R programming languages to facilitate their use in practice and for teaching purposes in epidemiology, biostatistics, data science and related fields. Our goals are to increase understanding and use of these methods by providing introductions to the concepts and access to helpful tools.

Keywords: Missing data; censoring; dropout; informative missingness; joint models; longitudinal data; missing not at random; reproducible research; sensitivity analyses; shared-parameter models.