All Times ET
Keywords: Longitudinal data analysis, GLMMs, GEEs
Patient registry data offer opportunities for rich longitudinal insight within select populations. However, the choice of a longitudinal modeling framework can affect the inferences drawn from a registry sample in unexpected ways. We analyzed a US-based registry of pulmonary arterial hypertension (PAH) patients to assess the impact of COVID-19 on various outcomes. We first used generalized linear mixed models (GLMMs) to evaluate differences in pandemic outcomes versus the pre-pandemic period, handling within-patient correlation using random intercepts. For several key outcomes including employment, GLMM results contradicted what marginal summary statistics suggested; while the proportion of employed individuals in the registry remained stable through the pandemic, GLMM suggested that the odds of employment were significantly reduced by over 45%. Upon further investigation, we found that this disconnect resulted from the facts that (1) long-time registry participants had greater weight in the estimation of odds ratios compared to newer registry entrants, and (2) long-time registry participants tended to be near retirement age. We found generalized estimating equations (GEEs), adjusted for age, mitigated this issue and yielded estimates that more-closely resembled marginal trends; our GEE model estimated a non-significant reduction in the odds of employment of less than 4%. Those in the statistical community working with patient registry data may wish to consider a GEE-based approach that complements the results of marginal summary statistics, as we did here. Compared to GLMMs, our approach offered a more coherent set of results within the registry sample.