Session number: 1020
Abstract title: Relative Performance of the PCS and PCD to Measure Health Status in a Longitudinal Study with Deaths
Author(s):
P Diehr - VA Puget Sound Health Care System, HSR&D and University of Washington, Department of Biostatistics
SD Fihn - VA Puget Sound Health Care System, HSR&D and University of Washington, Department of Medicine
MB McDonell - VA Puget Sound Health Care System, HSR&D
Objectives: Scores from the SF-36 health instrument have no value for death. Longitudinal studies using this measure must exclude information about persons who die. A procedure has been recently published which recodes the SF-36 components as the probability of being "healthy" in the future; deaths receive the value "0" on this new variable. The performance of the recoded variable has not been examined. We compared the performance of the Physical Component Score of the SF-36 (PCS) to the transformed score that includes death (PCD).
Methods: Using data from the Ambulatory Care Quality Improvement Project (ACQUIP), a group randomized trial of physician feedback in 7 VA General Internal Medicine clinics, we compared the distributions of the PCS and PCD and investigated whether the deaths were serious outliers or influential points in the PCD. We also compared changes over time on the two measures among persons with many chronic conditions at baseline and those who were healthier at baseline. We then analyzed the original data from the ACQUIP trial using this new approach to see if there were substantive differences in the results.
Results: In this group of veterans, who tend to be sicker than other U.S. citizens, the values for death on the PCD were not serious outliers, especially after adjusting for baseline characteristics. The PCD demonstrated a greater change over time than the PCS, in both the sicker and less sick groups. The randomized trial showed significant differences(favoring the controls) when based on the PCS. There was no difference in the two groups based on the PCD.
Conclusions: In longitudinal studies of veterans' health over time, the PCD had good statistical properties, and its use to include the people who died in the analysis may provide substantively different outcomes from the PCS. This may represent a common form of undetected bias in studies that track health over time and in which there is substantial morbidity. The PCD should be considered for secondary or primary analyses in such studies.
Impact statement: Longitudinal studies of veterans' health may provide different results depending on how the people who die are handled.