OBJECTIVES: To apply innovative methods to observational data to explore variability in depression scores over 12 years and examine whether there are subsets of individuals with different trajectories of change. METHODS: Latent Growth Modeling (LGM) and Growth Mixture Modeling (GMM) analyses were applied to observational data collected over a 12-year period.  Data was obtained from the RAND version of the Health and Retirement Survey (HRS), consisting of N = 5090 individuals aged 51 - 61 years at the time of recruitment.  Scores on an 8-item version of the Center for Epidemiological Studies–Depression Scale (CES-D) were examined.  LGMs were conducted to determine the level of variability in depression scores over the 12-year period.  When considerable variability was identified, GMMs were conducted to assess whether there were subsets of individuals with differential changes. RESULTS: LGMs showed an intercept (first assessment point) score of 0.75 (“no depression”) on the CES-D, which increased to 1.3 (“sub-threshold depression”) at year 12.  Substantial variability was found around the mean intercept and slope of change.  GMMs identified three subsets of individuals with differential slopes of change.  The largest subset (83% of the sample) had a mean intercept of 0.22 and a 12-year score of 1.0 (“no depression”).  A smaller subset (13.7%) had a mean intercept of 2.5 and showed no change over the 12-year period (“stable, sub-threshold depressed”).  The smallest subset (3.3%) had a high mean intercept (6.0) and showed a decrease in depression scores (4.0 at year 12; “improved, actively depressed”).  Post hoc analyses showed that these three classes were significantly different in terms of gender, self-reported health, whether health limits their work or activities, and activities of daily living CONCLUSIONS: Examining highly variable data can yield insights about subsets of respondents who show different levels of initial depression and different trajectories of change.