Influence of analytic methods, data sources, and repeated measurements on the population attributable fraction of...

Population attributable risk (PAR%) reflects the preventable fraction of disease. However, PAR% estimates of cancer have shown large variation across populations, methods, data sources, and timing of measurements. Three statistical methods to estimate PAR% were identified from a systematic literature review: the Levin's formula, the comparative incidence rate method, and the comparative risk assessment method. We compared the variations in PAR% of postmenopausal breast cancer in the Nurses' Health Study to evaluate the influence by method choice, source of prevalence data, use of single vs repeated exposure measurements, and potential joint effects of obesity, alcohol, physical activity, fruit and vegetable intake. Across models of the three methods, the estimated PAR% using repeated measurements were higher than that using baseline measurement; overall PAR% for the baseline, simple update, and cumulative average models were 13.8%, 21.1%, 18.6% by Levin's formula; 13.7%, 28.0%, 31.2% by comparative risk assessment; and 17.4%, 25.2%, 29.3% by comparative incidence rate method. The estimated PAR% of the combination of multiple risk factors was higher than the product of the individual PAR%: 18.9% when assuming independence and 31.2% when considering the risk factors jointly. The three methods provided similar PAR% based on the same data source, timing of measurements, and target populations. However, sizable increases in the PAR% were observed for repeated measures over a single measure and for calculations based on achieving all recommendations jointly rather than individually.