Assessment of robustness of linear mixed model under irregular longitudinal data

Résumé

Introduction Observational data collected from electronic health records (EHR) are often obtained at irregular visit time points because the visit process may depend on the patient's medical condition or disease severity, which is usually the outcome of interest. The use of traditional models, such as the linear mixed model, to analyse such outcomes has been shown to produce biased estimates. In this study, we aim to determine the extent to which irregular visits warrant the use of alternative models for analyzing extreme irregular longitudinal data. Methods We compared the performance of the linear mixed model (LMM), broken stick model (BSM), generalized estimating equation (GEE), and weighted GEE based on bias, coverage probability, standard error, and mean squared error using simulated data. The simulated data were generated under extreme irregular visit, and extreme irregular visit with missingness scenarios, with a moderate sample size ( n = 500). The performance of these models was also assessed using real data from a study of patients who underwent metabolic and bariatric surgery at Tygerberg Hospital. Results Under extreme irregular visits, the LMM showed varying performance depending on the degree of informativeness. When the degree of informativeness was set to 0.5, the LMM achieved the smallest ARB (1.3%) and an acceptable coverage probability (94%) compared with the other models. Under extreme irregular visit patterns with missingness, all models performed well at 10% and 20% missingness proportions; however, BSM and weighted GEE did not perform well at 40% missingness. For the real data set, the models produced similar estimates of the treatment effect. This may be because the real data exhibited moderate to no irregularity, as suggested by the Pearson correlation between the gap times and the longitudinal outcome. Conclusion The study findings revealed varying performance of LMM across degrees of informativeness, with the smallest ARB and an acceptable coverage probability achieved when the degree of informativeness is set to 0.5 for extreme irregular visits only. Finally, under extreme irregular visits with missingness, the LMM, weighted GEE, and GEE are preferred at missingness proportions of 10% and 20%, whereas GEE and LMM are preferred at 40%.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0