On the appropriateness of marginal models for repeated measurements in clinical trials

Abstract

Although models developed directly to describe marginal distributions have become widespread in the analysis of repeated measurements, some of their disadvantages are not well enough known. These include producing profile curves that correspond to no possible individual, possibly showing that a treatment is superior on average when it is poorer for each individual subject, implicitly generating complex and implausible physiological explanations, including underdispersion in subgroups, and sometimes corresponding to no possible probabilistic data generating mechanism. We conclude that such marginal models may sometimes be appropriate for descriptive observational studies, such as sample surveys in epidemiology, but should only be used with great care in causal experimental settings, such as clinical trials.

A Note on Marginalization of Regression Parameters from Mixed Models of Binary Outcomes

“…Thus, the conclusions are different depending on whether one controls for the subject effects or not. Such situations in which the conclusions vary between marginal and conditional approaches have been previously reported in the literature (Agresti, 1989; Lindsey and Lambert, 1998). …”

DNA Methylation in Benign Breast Epithelium in Relation to Age and Breast Cancer Risk

“…The prevalence of TSG methylation in benign samples was assessed in relation to age, race, menopausal status, sample cellularity, and risk level of the breast providing the sample using logistic regression in a series of univariate analyses and then by multivariate analysis that included all covariates generating a univariate P < 0.15. Although the methylation status of one breast did not predict the methylation status of the opposite breast, either for individual genes or for composite measures, each subject was treated as an independent observation for this analysis using Generalized Estimating Equations to account for multiple samples (left and right breasts) for most patients (37,38). Because at most two samples were taken from the same patient, either compound symmetry or unstructured covariance matrix gave the same results.…”

Choosing marginal or random-effects models for longitudinal binary responses: application to self-reported disability among older persons

“…The random model takes into account the underlying dependence relationship. Furthermore, the assumptions about the distributions are different, and the fact that the conditional distribution is binomial does not imply that the marginal distribution is also binomial [22]. …”