Methods/Statistics
Using an Online Panel to Crosswalk Alternative Exposure Measures in Two Studies: Deriving Comparable Estimates of the Effect of Physical Activity on Self-Rated Health Anna M. Pederson* Anna M. Pederson Scott C Zimmerman Peter Buto Yingyan Wu Eleanor Hayes-Larson Paola Gilsanz Elizabeth Rose Mayeda Rachel Whitmer M. Maria Glymour
Different assessments of physical activity (PA) across studies pose a major challenge when estimating effects of PA on health. We fielded an online survey (Qualtrics and Centiment) to develop a crosswalk between PA questions as asked in 2 cohorts: the Health and Retirement Study (HRS) and the Kaiser Healthy Aging and Diverse Life Experiences (KHANDLE) study. We compared associations of each question version of heavy PA and self-rated health, and whether associations differed by cohort.
Panel participants (n=2,335) were asked about heavy PA frequency as phrased in both HRS (4-level Likert scale) and KHANDLE (5-level Likert scale). HRS (n=11,094) and KHANDLE (n=1,823) participants aged 65+ self-rated their health (5-level Likert scale). We distance matched each HRS participant to 5 panel participants based on the HRS PA question and then imputed their responses to the KHANDLE PA question from the matched panel participants’ responses (and vice versa). We pooled the imputed data and regressed self-rated health on heavy PA.
The measurement survey included panel participants aged 50+; 63% Female; 22% Hispanic, 25% Black, 26% White, and 27% Other racial/ethnic identity. In this sample, responses to HRS and KHANDLE versions of the heavy PA questions were moderately correlated (R2 = 0.47). In the pooled data, when using the HRS PA question with a 4-level PA scale, a 1 unit higher heavy PA was associated with better self-rated health (b=0.21 [0.20,0.22]), although the associations differed by cohort (interaction b=-0.089 [-0.107,-0.071]). When using the KHANDLE question version with a 5-level PA scale, heavy PA was associated with better self-rated health (b=0.13 [0.12,0.14]), and the association was similar across cohorts (interaction b=0.006 [-0.011,0.023]).
Fielding a survey with multiple question versions allows us to directly compare findings across multiple studies and pool or meta-analyze studies.
