Background It is difficult to estimate the effect of exercise on future health from observational data because exercising may be both a cause and an effect of health status. Unadjusted analyses suffer from selection bias (healthier persons more likely to exercise), while adjusted analyses may adjust away some of the benefits of exercise.
Objective To obtain a "low-bias" interpretable estimate of the effect of exercise on future health.
Methods We used data from the Cardiovascular Health Study, a longitudinal study of 5,888 older adults. The number of blocks walked in the previous week, collected annually, were classified as Sedentary (less than 7 blocks per week), Moderate, or Active (28 or more blocks per week). The primary "low bias" analysis was restricted to persons who were both Sedentary and Healthy (in Excellent, Very Good, or Good self-reported health) in the two years before baseline. Self-reported health status (Healthy versus Sick or Dead) at follow-up was regressed on the level of exercise at baseline, variously including or excluding demographics, health prior to baseline, and health at baseline.
Findings Exercise trends were associated as expected with age, sex, and race. Healthy persons were more likely than Sick to start to exercise, and Sick Active persons were more likely to become Healthy than Sick Sedentary persons. In the total sample, 77% of persons who were Active at baseline were Healthy at follow-up, as compared with 49% of Sedentary persons, a difference of 28 percentage points that is difficult to interpret. In the subset who were both Sedentary and Healthy in the two years before baseline, the difference was only 14 percentage points. That difference declined to 12 points after adjustment for demographics, and to 9 points after adjusting for other health variables measured prior to baseline. After adjustment for health variables measured at baseline (possibly in the causal pathway) the difference dropped to 7 points and was no longer significantly different from zero. Similar findings occurred when survival was the outcome. The apparent effect of exercise on health was substantially smaller if persons who were Dead at follow-up were excluded.
Conclusion At least a third of the apparent benefit of exercise could be explained by selection bias. Where possible, observational studies of the effects of exercise should measure exercise at every period instead of just at enrollment. This permits incorporating exercise and health data prior to baseline. Analysis should also allow for the benefits of exercise on survival. The "low-bias" estimate of the benefit to a Healthy Sedentary older adult of becoming Active (walking 28 or more blocks per week, median = 48) was 7 percentage points for being alive 2 years later, and 9 percentage points for being alive and healthy. A modest program of walking may confer modest health benefits.
Diehr, Paula and Hirsch, Calvin, "Using Longitudinal Data to Estimate the Effect of Starting to Exercise on the Health of Sedentary Older Adults" (August 2008). UW Biostatistics Working Paper Series. Working Paper 334.