65. Transforming Self-Rated Health and the SF-36 Scales to Include Death and Improve Interpretability
P Diehr, University of Washington; DL Patrick, University of Washington; J Spertus, University of Missouri at Kansas City; CI Kiefe, University of Alabama at Birmingham; MB McDonell, VA Puget Sound Health Care System; SD Fihn, VA Puget Sound Health Care System
Objectives: Health-related interventions are often evaluated by tracking the health of participants over time. In studies of older or very ill populations, and in longer-term interventions, data are often missing because of death. As a result, longitudinal analyses are frequently limited to a healthier cohort (the survivors) that can not be identified prospectively, and that may have had little change in health. Our objective was to develop and evaluate methods to transform self-rated health and the physical component score of the SF-36 (PCS) to new variables that include a value for death.
Methods: We used longitudinal data from two large studies of older adults, the Cardiovascular Health Study (CHS) and the VA's Ambulatory Care Quality Improvement Project (ACQUIP).
We examined two measures of health-related quality of life (HRQOL). The first was self-rated health, (Is your health Excellent, Very Good, Good, Fair, or Poor? - EVGGFP). Because we examined health status over time, a sixth health state, Dead, was added. For the SF-36, we looked primarily at the PCS, but also report transformations for the Mental Component Summary score (MCS), and the eight sub-scales.
We transformed each HRQOL variable to the estimated percent probability of being "healthy" in the future (time 1, or T1), conditional on the observed value. The transformed measures are on a ratio scale, taking values of 0 to 100 where 0 is death and 100 is perfect health. For EVGGFP, we defined healthy as being in Excellent, Very Good, or Good health, and compared the new probabilities to previously published values. For the SF-36 analysis, we explored three different definitions of being healthy at T1: (1) Being alive (P-ALIVE); (2) being in Excellent, Very Good, or Good health (P-EVGG); and (3) having a PCS value in the top 75% of the reference population values for men aged 65 and over (P-TOP75%). We used logistic regression to estimate the three transformation equations. We determined how well the three different PCS transformations discriminated among groups of patients and detected change over time.
Results: The new transformation for EVGGFP was similar to that previously published, and was fairly independent of follow-up time. We recommend coding the 5 health categories as 95, 90, 80, 30, and 15, with 0 assigned to dead. The three transformations of the PCS performed at least as well as the original PCS. The P-EVGG transformation had the most favorable performance.
Conclusions: These easily interpretable transformed variables permit keeping persons who die in the analyses. People with no chance of recovery can also be assigned a value.
Impact: Excluding deaths from analyses limits longitudinal evaluations to subjects in better health (survivors) and may miss important changes over time. We recommend using the transformed variables for longitudinal analyses of HRQOL where there are deaths, either for secondary or primary analysis. This approach can potentially be applied to other measures of HRQOL.