|
Abstract:
|
Measures computed as the sum of a set of binary variables are often ordinal. Such is the case for the 10-item Epidemiologic Studies Depression Scale (CES-D), scored as the number out of 10 depressive symptoms endorsed. In community-based studies, the distribution of CES-D is skewed and inflated at zero because many respondents report no symptoms. For example, in the Chicago Health and Aging Project, 41% of the participants reported no depressive symptoms (see Wilson et al 2004). There are many longitudinal models available, including mixed effects models based on Poisson, inflated-zero Poisson, inflated-zero Gamma, and zero-altered Poisson distributions, and longitudinal forms of constrained ordinal models such as Proportional Odds model. Here, we introduce the longitudinal Trend Odds model. We fit all of these models to longitudinal CES-D data (2 to 22 observations per participant) collected annually on 1005 elderly participants (mean age 79.96) in the Religious Orders Study and the Rush Memory and Aging Project. All models indicate a significant change in depression over time, with the constrained ordinal model having the lowest AIC values. No trend in odds over time is found.
|
