We developed a new absolute risk prediction model for invasive breast cancer for Italian women. The model includes non-modifiable risk factors and three potentially modifiable factors, BMI, leisure-time physical activity, and alcohol consumption. The model was reasonably well calibrated overall in independent data from the Florence-EPIC cohort study, but overestimated absolute risk in some subgroups such as women aged 60 years or older and women whose first live birth occurred at age 30 years or later. The discriminatory accuracy (concordance) in the Florence-EPIC cohort data was 0.62 at age less than 50 years and 0.57 for older women, and is comparable with that of other absolute risk models for breast cancer (4
A novel aspect of this work is the evaluation of the potential effects of reducing exposures from modifiable risk factors on absolute breast cancer risk, not only for the individual counselee but also for the entire population and high-risk subgroups. We developed methods based on the Lorenz curve of population absolute risk to identify high-risk subgroups. Assessment of the reduction in average absolute risk gave a different perspective than assessment of the fractional risk reduction, which is analogous to attributable risk. Indeed, in the entire population, 20-year fractional risk reductions are 20%−24%, whereas absolute risk reductions are 1.4%−1.6%. Fractional risk reductions are less useful for clinical and public health decisions than absolute risks and absolute risk reductions (30
). Our methods are also useful for designing intervention trials, because the power of such trials depends on the average absolute risk with and without intervention (31
). In our study, the absolute reduction is nearly proportional to absolute risk before modification of risk factors, reflecting proportional hazards assumptions. Thus the fractional risk reduction was nearly constant across categories of risk. However, the fractional risk reductions are greater in older women for whom lower BMI was associated with reduced risk.
Estimates of the potential effects of interventions on absolute risk can provide perspective on whether to pursue prevention research or implement interventions. For example, our estimates for risk factor modifications indicate an approximate 1.6% absolute risk reduction during 20 years in the general postmenopausal population, and an approximate 3.2% absolute risk reduction for women with a positive family history of breast cancer. For women at the highest absolute risk who account for 10% of total population absolute risk, the absolute risk reduction is approximately 4.4%. In a population of 1 million women, even a 1.6% absolute risk reduction amounts to 16 000 fewer cancers. Because programs to encourage less alcohol consumption, increase leisure activity, and encourage some weight control are likely to be safe, they can be widely administered. As emphasized by Rose (33
), broadly applicable interventions can be more effective than interventions focused on high-risk subgroups. If these interventions were restricted to the 8% of postmenopausal women with a positive family history of breast cancer, then in a population of 1 million women, only 2560 breast cancers would be prevented.
A strength of the study was the quality of the data used for developing the model and for validation. Selection bias was limited in the case–control data because the participation rate was high, and the catchment areas were comparable for case patients and control subjects. The comparability of recall between case patients and control subjects was improved by interviewing all study participants in a hospital setting (15
A number of limitations need to be considered. Although our model was reasonably well calibrated, there was evidence of overestimation in the highest quintile of absolute risk. Recalibration (35
) led to smaller odds ratios, improved fit to Florence-EPIC data in this quintile, and smaller estimates of the effects of modifying risk factors (unreported data). This analysis and other unreported numerical studies indicate that the estimated absolute risk reductions are sensitive to the estimated odds ratios for modifiable risk factors.
Another limitation was that the estimated absolute risk reductions are imprecise for the individual counselee. Population level estimates are more precise, but both individual- and population-level estimates are subject to systematic errors, which are not reflected in the confidence limits. An ideal study to estimate the effects of interventions would be a randomized intervention trial, such as the Woman’s Health Initiative (36
) or the Breast Cancer Prevention Trial (31
). Such trials yield unbiased estimates of treatment effect and information on compliance. Although trial results may not generalize quantitatively to the general population, they provide a good guide to preventative strategy. Estimates from case–control data may yield associations that are confounded by other factors or biased by differential recall. The associations observed in such data may therefore not predict actual preventative effects. Without empiric data from intervention studies, we cannot test the assumptions underlying our model. A key assumption is the proportional hazards assumption, whereby a risk factor modification acts immediately and indefinitely to multiply the breast cancer hazard by a constant factor, unless the model includes an interaction with time. In fact, it is unknown how long it will take before an intervention affects breast cancer hazard rates or how long the effect will last. Under the proportional hazards assumption, our calculations give an idea of the largest reductions in absolute risk that might be achieved.
A further limitation is that the interventions are not specified. Knowing that a woman with elevated BMI is at increased risk does not define the intervention. Yet an estimation of the causal effect of the intervention on absolute risk is desired (37
). We have made optimistic assumptions that interventions could reduce modifiable exposures to their lowest risk levels; thus our calculations would give an upper bound on the reductions in absolute risk.
With hospital-based controls, associations can be distorted by correlations between the risk factors and the control diseases; however, we chose control diseases to avoid this bias. The Florence-EPIC cohort was not a random sample of the population of Italian women, and the cohort may have had a more favorable distribution of lifestyle risk factors than the general population. Moreover, results may not generalize to other countries, where the prevalence of obesity may be larger or the frequency of mammographic screening greater.
Cummings et al. (24
) reviewed the literature on possible interventions to reduce breast cancer incidence, and concluded that lifestyle interventions such as exercise, weight reduction, low-fat diet, and reduced alcohol intake should be included in programs to prevent breast cancer. Despite their limitations, calculations of reductions in absolute risk using our model potentially provide additional perspective on the possible benefits of such prevention strategies. If combined with operational definitions of interventions and their effect sizes, our methods can provide information needed to compute sample sizes for trials to evaluate such interventions.
This research was supported by the Intramural Research Program of the Division of Cancer Epidemiology and Genetics, National Cancer Institute, National Institutes of Health and the Associazione Italiana per la Ricerca sul Cancro (AIRC IG 10415 to A.D.).