Skip to main content
         

Vol. 11 Issue 3, Summer 2006

Using "Population Attributable Fraction" to Understand: To What Extent Can Breast Cancer Incidence Be Attributed to Higher Socioeconomic Status?
The Ribbon 

Christine A. Erdmann, Ph.D., MPH, Assistant Professor
Department of Epidemiology, University of Michigan School of Public Health

Introduction

With few exceptions, higher socioeconomic status (SES) has been found to be associated with higher breast cancer risk in both developed and lesser developed countries using a variety of measures of SES, including income, education level, and occupation type (Faggiano et al., 1997). This article examines the extent to which breast cancer incidence can be attributed to higher socioeconomic status in the United States (U.S.). An oft-recited but frequently misunderstood statistic used for estimating the extent to which disease incidence can be attributed to a risk factor is the "population attributable fraction." Therefore, this article first examines the concept of population attributable fraction.

Understanding Population Attributable Fraction

Population attributable fraction is a statistic used to estimate the proportion of cases that can be attributed to one or more specified risk factors. More precisely, population attributable fraction is the "proportional reduction in average disease risk over a specified time interval that would be achieved by eliminating the exposure(s) of interest from the population while distributions of other risk factors in the population remained unchanged. This also can be interpreted as the proportion of disease cases over a specified time that would be prevented following elimination of the exposures, assuming the exposures are causal" (Rockhill et al., 1998). As illustrated by the following formula, the population attributable fraction (PAF) depends on the prevalence (pe) of the risk factor and the relative risk of the risk factor (RR).

Figure 1 illustrates the calculation and interpretation of a population attributable fraction for an imaginary breast cancer (BC) risk factor: handling two or more frogs during adolescence. In this example, 30% of the population has handled two or more frogs during adolescence. For use in this formula, however, the prevalence should be expressed as a proportion (i.e., 0.30). The relative risk (RR) estimate for this risk factor is 1.6. That is, women who handled two or more frogs during adolescence were 1.6 times more likely to develop breast cancer than women who never touched a frog during adolescence. In other words, women who handled two or more frogs during adolescence were 60% more likely to develop breast cancer than those who never touched a frog during adolescence. After plugging in Pe = 0.30 and RR = 1.6 into the formula, the calculated PAF is 0.15. This can be interpreted correctly as, 15% of breast cancer cases could be avoided if frog handling in adolescence were eliminated assuming that handling two or more frogs during adolescence is causally related to increased breast cancer risk and that the distribution of all other risk factors remains unchanged.

Figure 2 shows how the population attributable fraction increases for increasing values of RR while the prevalence (Pe) is held constant at 0.30. Given Pe = 0.30 and RR = 1.6, one can look up the population attributable fraction in using this graph. The frog icon is situated at the place on the plotted line that corresponds to RR = 1.6; this corresponds to a population attributable fraction of 0.15 (or 15% if expressed as a percentage). Again, this can be interpreted correctly as, 15% of breast cancer cases could be avoided if frog handling in adolescence were eliminated assuming that handling two or more frogs during adolescence is causally related to increased breast cancer risk and that the distribution of all other risk factors remains unchanged.

Figure 3 illustrates how the population attributable fraction increases as the prevalence (Pe) of the exposure increases. Each black line plotted on this graph corresponds to prevalences from 0 to 1.0, inclusive. The black lines plotted between the bottom and top black plotted lines correspond to prevalences of 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, and 0.9, respectively. For example, where the red and blue lines intersect in Figure 3 corresponds to Pe = 0.30, RR = 1.6, and a PAF = 0.15, the same situation illustrated in Figure 2. Back to the frog handling example: If the prevalence of frog handling were reduced to 0.10 (or 10% if expressed as a percentage) and the relative risk remained equal to 1.6, corresponding to where the red and yellow lines intersect in Figure 3, the PAF would be reduced to 0.06 (or 6% if expressed as a percentage). If the prevalence of frog handling were greater in the population, then a greater number of cases could be prevented if frog handling were eliminated. For example, if all women in the population handled two or more frogs during adolescence (Pe =1 or 100% exposure) and the RR remained equal to 1.6, corresponding to where the red and green lines intersect in Figure 3, then the population attributable fraction would be 0.38 (or 38% if expressed as a percentage).

For rare exposures, the population attributable fraction will be small even when the exposure is strongly related to the disease. As is illustrated in Figure 4, regardless of the magnitude of the relative risk (RR), if the prevalence is equal to zero, then the population attributable fraction will be zero. With a very high relative risk (RR) of 10 and a low prevalence (Pe) of 0.10 (or 10% if expressed as a percentage), the population attributable fraction will be only 0.47 (or 47% if expressed as a percentage). Even with a high prevalence of 1.0 (or 100% if expressed as a percentage), an exposure with a relative risk (RR) of 10 still will not account for 100% of cases.

Warnings About Population Attributable Fractions

WARNING #1 PAF is easily misinterpreted. Often the PAF is incorrectly thought to represent the proportion of cases having any risk factors. A PAF equal to 15% does not mean that only 15% of women have any known breast cancer risk factors nor does it mean that 85% of cases do not have any risk factors. The correct interpretation of a PAF equal to 15% is that 15% of breast cancer cases would be avoided if the risk factor is causally related to breast cancer and is eliminated.

WARNING #2PAFs calculated for individual breast cancer risk factors will not sum to 1.0. Because many risk factors are correlated, population attributable fraction (PAF) estimates calculated for single risk factors should not be summed.

WARNING #3Note that the selection of exposure cutpoints used for defining "exposed" is somewhat arbitrary and can have a major impact on the population attributable fraction. For example, if the imaginary risk factor were defined as "handled 100 or more frogs during adolescence", the proportion of the population meeting this new definition would be much smaller, thus reducing the population attributable fraction. The population attributable fraction is influenced by the cutpoint used in defining the risk factor exposure because this cutpoint can affect the prevalence (Pe). A change in cutpoint (i.e., exposure definition) also may affect the relative risk (RR). Thus, it is important to pay attention to the exposure definition when considering the meaning of a particular population attributable fraction estimate.

WARNING #4 Actual reduction in disease burden after removal of the risk factor assumes that the risk factor is causally related to the disease

Back to Known Breast Cancer Risk Factors…

Consider the typical relative risks associated with the established breast cancer risk factors given in Table 1.



















PAF for Risk Factor with Strongest Association with Breast Cancer

The established breast cancer risk factor with the strongest association with breast cancer listed in Table 1 is family history of breast cancer in one or more first-degree relatives (RR = 2.6). While prevalence of family history of breast cancer will vary from population-to-population, let's consider a prevalence estimate taken from the Women's Contraceptive and Reproductive Experiences (CARE) Study (McDonald et al., 2004). The CARE Study was a large population-based case-control study of 4,575 cases and 4,682 controls between the ages of 35 and 64 years who were sampled from five metropolitan sites in the U.S. The prevalence estimate of family history among controls may be used to estimate the source population prevalence. In the CARE Study, 9.7% of controls had a family history of breast cancer in one or more first-degree relatives (in a mother, sister, or daughter, specifically). Assuming that the CARE Study population is fairly representative of the U.S. urban population, the population attributable fraction for family history of breast cancer in this population is 0.134. In other words, if family history of breast cancer were eliminated in the U.S. urban population, then approximately 13.4% of new breast cancer cases could be avoided in this population.

PAFs for Other Breast Cancer Risk Factors






















The relative risks associated with known breast cancer risk factors other than family history typically range from 1.1 to 2.0. Given this moderate relative risk range, the proportion of breast cancer cases attributed to any one of the other established breast cancer risk factors will not exceed 50% as illustrated in Figure 3. That is, for the extreme example where the prevalence is 100% (or 1.0 if expressed as a proportion) and RR = 2.0, the PAF will equal 0.50 (or 50%). Table 2 presents prevalences for some of the established breast cancer risk factors estimated from relatively large samples from various populations. While prevalences for breast cancer risk factors vary from population-to-population, the prevalences in Table 2 provide a basis for illustrating what population attributable fractions might be expected for some of the established breast cancer risk factors. Using the relative risk estimates in Table 1 and the prevalence estimates in Table 2, example population attributable fractions for some breast cancer risk factors were calculated and are given in Table 3.

Proportion of Breast Cancer Cases "Attributed" to Higher SES

Using the first National Health and Nutrition Examination Survey (NHANES I) Epidemiologic Follow-up Study, Madigan et al. estimated the proportion of breast cancer cases attributed to higher SES in the U.S. female population (Madigan et al., 1995). With higher SES exposure defined as "income in the upper two thirds of the U.S. population," the PAF estimated for the U.S. female population was 18.9%. If higher SES were eliminated, the 18.9% of breast cancer cases could be avoided in the U.S. female population, assuming that higher SES is causally related to increased breast cancer risk and that the distribution of all other risk factors remains unchanged. Of course, higher SES is not a direct cause of breast cancer. While the association between SES and breast cancer risk is strikingly consistent, it provides neither a basis for a breast cancer prevention target nor a biologic clue. Instead, higher SES acts as a proxy for a combination of breast cancer risk factors that do have a biologically plausible association with breast cancer risk. In other words, the association between SES and breast cancer risk can be explained by variation in the distributions of biologically plausible breast cancer risk factors (e.g., later age at first birth, earlier age at menarche, and hormone therapy use) (Heck and Pamuk, 1997; Braaten et al., 2005). Women in higher socioeconomic groups tend to have a higher prevalence of exposure to known breast cancer risk factors that are plausibly biologically linked to breast cancer than women in lower socioeconomic groups. So, although it is possible to calculate a population attributable fraction for higher SES, of greater interest and utility are population attributable fractions for breast cancer risk factors for which SES is a proxy and which have a biologic basis. An article by Dr. Suzanne Snedeker in this issue of The Ribbon further discusses the relationship between SES, breast cancer risk, and survivorship.

REFERENCES

Braaten, T., Weiderpass, E., Kumle, M., and Lund, E. (2005). Explaining the socioeconomic variation in cancer risk in the Norwegian Women and Cancer Study. Cancer Epidemiology, Biomarkers, and Prevention 14(11), 2591-2597.

Colditz, G.A., Atwood, K.A., Emmons, K., Monson, R.R., Willett, W.C., Trichopoulos, D., Hunter, D.J. (2000). Harvard report on cancer prevention. Volume 4: Harvard cancer risk index. Cancer Causes and Control 11, 477-488.

Erdmann C.A., Satariano, W.A., Chen, Y.Q. (2003). Traditional Risk Factor Study (TRiFS): Preliminary analysis of risk factor prevalence and population attributable fraction for breast cancer in Marin County, California. LBNL Report No. 51905. (Berkeley, California: Lawrence Berkeley National Laboratory).

Faggiano, F., Partanen, T., Kogevinas, M., Boffetta, P. (1997). Socioeconomic differences in cancer incidence and mortality. IARC Scientific Publication 138, 65-206.

Gail, M.H., Brinton, L.A., Byar, D.P., Corle, D.K., Green, S.B., Schairer, C., Mulvihill, J.J. (1989). Projecting individualized probabilities of developing breast cancer for white females who are being examined annually. Journal of the National Cancer Institute 81(24), 1879-1886.

Heck, K.E., Pamuk, E.R. (1997). Explaining the relation between education and postmenopausal breast cancer. American Journal of Epidemiology 145(4), 366-372.

Madigan, M.P., Ziegler, R.G., Benichou, J., Byrne, C., Hoover, R.N. (1995). Proportion of breast cancer cases in the United States explained by well-established risk factors. Journal of the National Cancer Institute 87(22), 1681-1685.

McDonald, J.A., Mandel, M.G., Marchbanks, P.A., Folger, S.G., Daling, J.R., Ursin, G., Simon, M.S., Bernstein, L., Strom, B.L., Norman, S.A. et al. (2004). Alcohol exposure and breast cancer: Results of the Women's Contraceptive and Reproductive Experiences Study 13(12), 2106-2116.

National Cancer Institute (NCI). (2003). Summary Report: Early Reproductive Events and Breast Cancer Workshop. http://www.nci.nih.gov/cancerinfo/ere-workshop-report (last accessed on 8/22/2006).

National Center for Health Statistics (NCHS). (2005). Health, United States, 2005 with Chartbook on Trends in the Health of Americans. DHHS Publication No. 2005-1232. (Hyattsville, Maryland).

Rockhill, B., Newman, B., Weinberg, C. (1998). Use and misuse of population attributable fractions. American Journal of Public Health 88(1), 15-19.

Back to the top

Last Update 11.07.06