Background: Latent Transition Analysis (LTA) is an effective tool used to investigate underlying groupings in multivariate categorical data over time. Variable selection is not a primary focus in LTA, however; researchers choose data elements a priori and include all variables in the model. Objective: To expand and evaluate variable search methods developed for Latent Class Analysis for application to the longitudinal LTA setting, and to apply these techniques to an expansive data set describing transitions in medication adherence. Methods: Simulated multivariate categorical data for two time periods consisted of a mixture of variables that separate statuses and noise variables. LTA models were specified via 1) a headlong search algorithm that identifies variables that account for significant between-group variation in probability, and 2) an item elimination method that sequentially removes variables that do not affect classification. Sample size, item response probabilities, number of statuses, and transition probabilities varied in simulations. The Cohort Study of Medication Adherence in Older Adults (CoSMO), a prospective study of the barriers to antihypertensive medication adherence, served as an example of these procedures applied to real-world data. Each algorithm identified status-informative variables from a survey on mental and 6 physical quality of life, and categorical variables measuring medication adherence, psychosocial factors, healthcare utilization, and clinical variables were profiled to measure and characterize changes in CoSMO participants’ statuses over two years. Results: Both algorithms performed well with fewer starting variables, large sample size relative to the number of variables included, and sizable statuses across time. Simulations, as well as the application to CoSMO data, suggested that headlong search may select fewer variables and a simpler model whereas backwards elimination favors retention of more variables and a more complex model.