The rise and fall of breast cancer rates

Around 2002, after several decades of steady increase, the incidence of breast cancer in many parts of the developed world began to decrease.1 2 Several explanations for this decline have been proposed, but a favoured theory has been a reduction in the use of hormone replacement therapy (HRT) after release of the results of the Women’s Health Initiative randomised trial.1 2 3 4 In the linked study (doi:10.1136/bmj.e299), Weedon-Fekjær and colleagues present ecological data supporting a role for mammography screening and HRT use in recent trends in the incidence of breast cancer.5

Ecological studies use aggregate data to explore correlations and time trends. No information is available on individuals, so no inferences can be drawn about cause and effect. Thus, Weedon-Fekjær and colleagues’ study provides no data on whether the women who participated in the screening programme were the same women identified with breast cancer, or whether women who used HRT developed breast cancer more often than those who did not. Instead, the correlations reported are established by time trends within the population of Norway and by statistical modelling techniques. One of the shortcomings of aggregated data is the potential for erroneous conclusions—the introduction of screening programmes and increased acceptance of HRT use may have occurred at the same time as other changes that affected the incidence of breast cancer.

Weedon-Fekjær and colleagues apply a method proposed in the mid-1980s to estimate the age group and calendar period specific incidence of breast cancer while simultaneously accounting for screening variables and HRT use on a population level.6 7 They conclude that much of the rise and fall of breast cancer rates over the past two decades can be explained by the surge in incidence of breast cancer owing to the diagnosis of preclinical disease and (over)diagnosis of breast cancer by mammography screening between 1995 and 2004, and then the decline in breast cancer detection when women left the screening programme at age 69 years. A similar proportion of these trends can be accounted for by the increasing popularity in HRT until 2002, then the sharp decline in its use after the Women’s Health Initiative was prematurely stopped.

Although it is reasonable to describe time trends in mammography screening, HRT use, and the incidence of breast cancer on the basis of data from the Norwegian population, the authors do not discuss artefacts that can arise in ecological data and age-period-cohort analyses when non-linearities are present—problems that were noticed only after the method was introduced. The basic problem is that such analyses are based on relations across group summary statistics, which are inevitably averages and cannot properly capture non-linearities and interactions correctly.8 9 Thus, the authors claimed estimates of non-linear cohort and period effects must be at least partly distorted relative to the actual individual relations of breast cancer risk to HRT and screening; the degree of distortion can be assessed only with individual data.

In addition to these subtle modelling problems, several possible sources of confounding are unaccounted for. The time trends in mammography and HRT use were paralleled by changes in other factors in the same birth cohorts that may have contributed to the observed increase in breast cancer incidence trends—earlier onset of menarche, delayed child birth, lower number of children, and increasing weight gain after menopause. This problem and problems from non-linearities have no relation to goodness of fit of the model—the author’s models could fit perfectly even if they were completely confounded or distorted by underlying non-linearities. Consequently, the estimated fraction of cases attributable to screening and HRT use is far more unreliable than the statistics make it appear (because those statistics account for random errors only, not biases). In particular, the authors’ conclusion—that trends in breast cancer incidence since the early 1990s can be “fully attributed” to the introduction of mammography screening programmes and adoption of hormone treatment—cannot be justified on the basis of their data and analyses, because the data do not allow full adjustment for potential confounding variables and cannot properly reflect non-linearities and interactions.

This does not mean that the authors’ conclusions are wrong, however, but that other evidence has to be considered too. Previous studies that have correlated time trends in breast cancer incidence with changes in screening and HRT use over time have also been based on population based estimates,1 2 3 4 so individual level data are needed to evaluate the contribution of other factors.

Between 2004 and 2008, the incidence of breast cancer was steady in the United States,10 11 consistent with a stabilisation in screening and hormone use. According to Weedon-Fekjær and colleagues’ data, the incidence in Norway continues to decrease and, after accounting for trends in mammography and HRT use, continues to increase until 2008 (see fig 2, top panel, in Weedon-Fekjær and colleagues’ paper).5 Whether the discrepant trends between these countries in the past few years are a modelling artefact or caused by factors beyond screening and hormone use will be the focus of studies in the years to come.