Singh and colleagues analyzed prospective data on 164 white men and 143 black men who were interviewed while awaiting a biopsy at a Veterans Affairs hospital. The authors were interested in whether reported levels of leisure-time physical activity were predictive of biopsy results and whether this relation differed by race. White and black men had similar levels of reported exercise, and white men had a lower risk of a positive biopsy (37% vs 45%). The authors fit multivariable logistic regression models with adjustments for several covariates, including serum prostate-specific antigen, digital rectal examination (DRE) findings, and previous biopsies. They reported adjusted odds ratio (OR) estimates for the effect of reported exercise and on heterogeneity of these estimates by race, although the outcome was common, rendering the OR an overestimate of the relative risk. Singh et al concluded that white men experience a protective effect of exercise, but that black men do not.
The reported differential race effect is that the adjusted impact of continuous exercise on cancer for white men had an OR of 0.90 (95% confidence interval, 0.82-1.00; P = .041; whereas for black men, the OR was 1.02 (95% confidence interval, 0.91-1.13; P = .76). However, the conclusion of racial difference is based on a statistical fallacy, which is to view the distinction between P < .05 in 1 stratum and P > .05 in another stratum as indicative of heterogeneity. In fact, using either Cochran Q or a Wald test on the difference in log OR estimates, the heterogeneity test for this comparison yields P > .05. The 2 stratum-specific effects are not statistically distinguishable from a pooled OR value of 0.96.
The validity of this comparison is further threatened by adjustment for postexposure covariates, such as prostate cancer screening test results. If physical activity were to affect cancer incidence, then it would also affect cancer screening results, such as suspicious DRE findings. It is for this reason that epidemiologic guidelines for confounder selection specifically prohibit adjustment for factors intermediate between exposure and outcome.
Singh et al have implicated race as a significant factor in the relation between exercise and prostate cancer and have speculated that this may be explained by various biologic mechanisms, including different sex hormone profiles and genetic susceptibilities. But this is based on faulty analysis of the data. It would be helpful if they would report valid statistical inference regarding heterogeneity in this effect across strata, using appropriate covariate selection and choice of effect parameter.
Jay S. Kaufman, PhD
Department of Epidemiology, Biostatistics, and Occupational Health, McGill University, Montreal, Quebec, Canada