As part of a consulting project, an investigator came for assistance with a preliminary analysis of data from an observational study on change in health-related quality of life (HRQL) from pre- to post-hospitalization for an acute condition. Analysis of the data presented several issues, likely encountered by others analyzing similar data:
- multiple adjustment variables deemed clinically important by investigator, some with 7+ categories and low/no cell counts: does it make sense to adjust for these variables in the preliminary analysis?
- accounting for variable post-hospitalization response time: can/should we adjust for time to response?
- confounding by adjustment variables: relationship between measures of illness severity and clinical measures may make sense.
- Adjusting for baseline response
Current thoughts for the first issue:
Multiple adjustment variables with low/no cell counts
- Recommend that client obtain more subjects in categories with low/no cell counts, as it is impossible to estimate contrasts for these groups
- Refer to statistical “rule of thumb” of at least 10 observations per fitted variable, and note that categorical variables with k categories (k>2) require fitting k-1 dummy variables.