Trajectory analyses in insurance medicine studies: Examples and key methodological aspects and pitfalls

Abstract

Background Trajectory analyses are being increasingly used in efforts to increase understanding about the heterogeneity in the development of different longitudinal outcomes such as sickness absence, use of medication, income, or other time varying outcomes. However, several methodological and interpretational challenges are related to using trajectory analyses. This methodological study aimed to compare results using two different types of software to identify trajectories and to discuss methodological aspects related to them and the interpretation of the results. Methods Group-based trajectory models (GBTM) and latent class growth models (LCGM) were fitted, using SAS and Mplus, respectively. The data for the examples were derived from a representative sample of Spanish workers in Catalonia, covered by the social security system (n = 166,192). Repeatedly measured sickness absence spells per trimester (n = 96,453) were from the Catalan Institute of Medical Evaluations. The analyses were stratified by sex and two birth cohorts (1949–1969 and 1970–1990). Results Neither of the software were superior to the other. Four groups were the optimal number of groups in both software, however, we detected differences in the starting values and shapes of the trajectories between the two software used, which allow for different conclusions when they are applied. We cover questions related to model fit, selecting the optimal number of trajectory groups, investigating covariates, how to interpret the results, and what are the key pitfalls and strengths of using these person-oriented methods. Conclusions Future studies could address further methodological aspects around these statistical techniques, to facilitate epidemiological and other research dealing with longitudinal study designs.