A survey of elderly Jewish women in Charlotte, North Carolina, was designed to investigate the end of life decisions that these women would make under different conditions. The conditions, presence or absence of financial resources, presence or absence of unbearable pain, and terminal versus chronic state, generated eight different hypothetical scenarios. Each woman was given the same three ordinal choices for her response to all scenarios: aggressive treatments and life supports to stay alive; non-aggressive treatment and no life supports to stay alive; and choosing to terminate her life. In addition, other covariates were of interest in the analysis. A multiple repeated measurement factor model using marginal responses that generated weighted least squares parameter estimates (from SAS PROC CATMOD) was not effective because of the large number of possible response profiles (3**8=6561). Estimating a mean response model in PROC CATMOD and estimating a cumulative logit model using generalized estimating equations (GEE) with an exchangeable working correlation structure implementing SAS/IML macro language yielded similar results. Both methods found the main effects to be significant, with pain as the most important predictor. These two methods will be compared and their respective interpretations, assumptions, strengths, and weaknesses will be discussed.
Learning Objectives: Describe the problems associated with modeling correlated ordinal data and two methods of analysis that can be applied
Keywords: Statistics,
Presenting author's disclosure statement:
Organization/institution whose products or services will be discussed: Data from University of South Carolina, Department of Health Promotion and Education
I do not have any significant financial interest/arrangement or affiliation with any organization/institution whose products or services are being discussed in this session.