Purpose: To conduct a computer simulation to assess the effects of measurement error on prospective epidemiological studies that attribute mortality outcomes to apparent changes in the independent variable (e.g., physical fitness or activity) at baseline.
Methods: As an example, we evaluated the design of the Aerobics Center Longitudinal Study (ACLS). This study compared apparent changes in fitness between two baseline visits to mortality during a subsequent 5-yr follow-up period. Unfit men who were reclassified as fit at the second baseline examination (6.6% of sample) and fit men who were reclassified as unfit (2.3%) had follow-up mortality rates that were between those of men who were consistently classified as fit or unfit. This study design was simulated assuming that differences between baseline treadmill test durations were due to measurement error alone. Based on our own data, we estimated that repeat measurements of treadmill test duration have correlation of r = 0.89 in the absence of any real fitness change.
Results: There is excellent agreement between the published ACLS risk reductions and our simulated reductions for both cardiovascular disease (CVD) and total mortality. Compared with the "Unfit-->Unfit" (the referent group), the estimated relative risks from the simulations for men who were reclassified as fit (i.e., "Unfit-->Fit") were 0.57 for total mortality and 0.52 for CVD mortality, and for men who remained classified as fit ("Fit-->Fit"), they were 0.33 for total mortality and 0.20 for CVD mortality.
Conclusion: The imprecision of the fitness measurement alone (i.e., measurement error) is sufficient to produce the reported ACLS risk reductions in initially unfit men who get reclassified as fit in a subsequent clinic visit. This statistical artifact will apply to other studies that use this design.