A Mixed Model Approach wtih Control for Modeled Baseline in Longitudinal Analysis, with Applications to the Multi-Ethnic Study of Atherosclerosis

Abstract

Background Previous research has shown that longitudinal analyses that control for measured baseline can induce bias in the longitudinal association between an outcome and a risk factor where the outcome and risk factor are associated at baseline and the outcome is measured with error. An alternative to control for the measured baseline is to model the baseline outcome and follow-up outcomes jointly using a mixed model approach. Methods 6814 Participants in the Multi-Ethnic Study of Atherosclerosis (MESA) were assessed for coronary artery calcium (CAC) up to 4 times over the course of 10 years. We chose a subset of 3191 participants with complete follow-up to compare the results of several longitudinal methods to assess the rate of change in CAC over time, including difference, ANCOVA (control for measured baseline), individually-derived slopes, and a mixed model approach with control for modeled baseline (MMCMB). We also executed a simulation study, in which we compared the bias and uncertainty in the estimates of several parameters after applying different levels of iid normal errors. Results MMCMB and ANCOVA showed similar bias and uncertainty for longitudinal associations with covariates that were associated with change in the outcome but not baseline outcome. MMCMB was less biased than ANCOVA for covariates associated with both baseline outcome and change in the outcome. These results were consistent between the simulation and the data analysis. Conclusion Overall, we were satisfied with the performance of MMCMB compared to other methods assessed and are optimistic that we can communicate its form, interpretation, and advantages in a way that will lead to broader use.