Proper Variance Estimation for Event-Level Data Files Representing Subpopulations of Health Care Users in a National SurveySadeq Chowdhury, Agency for Healthcare Reserach and Quality
*Steven Machlin, Agency for Healthcare Research and Quality
Keywords: MEPS, domain estimation, health events
Public use data files from the Medical Expenditure Panel Survey Household Component (MEPS-HC) are released at various levels in which each record may represent, for example, a sample person, a medical condition or a medical event. The annual full-year consolidated file is the main survey file and is comprised of person level records containing individual characteristics and annual survey data (along with household and family identifiers). In addition to this person-level file, a medical conditions file and eight event-level files are also released every year. For each person in the MEPS sample, a particular health care event file may include no record, a single record or multiple records depending on the number of events of that type (e.g. hospital emergency room visit) that the person had during the year. Consequently, an event file only includes records related to a subset of health care users in the full year person file. Since the number of persons in an event file is not known before conducting the survey, estimation from event files should be treated as a domain analysis which requires the entire sample to take all variability into account in estimating the variance of domain estimates. That is, analysts should ideally include all persons with and without events by merging sampling information on persons with no events from the full person-level file to the event file. However, in practice it is convenient when producing event level estimates to analyze data directly from the events files without including sample persons having no events. This poster presents results of an analysis demonstrating that the impact of not including all sample persons when producing event-level estimates is generally negligible but can be problematic for small domain estimates. The poster also includes an overview of MEPS event files and, as an example of the use of event files in practice, shows results of an analysis of physician visits in different ambulatory care settings in 2008.