Breast cancer incidence in Sweden has always been approximately twice as high as in Singapore. In recent years, this difference is limited to postmenopausal women. The aim of this study was to explore the reasons behind these differences through the use of age-period-cohort modeling. This population-based study included all breast cancer cases reported to the Swedish and the Singapore cancer registries from 1968 to 1997, with a total of 135,581 Swedish and 10,716 Singaporean women. Poisson regression using age-period and age-cohort models was used to determine the effects of age at diagnosis, calendar period and birth cohort. Incidence rate ratios were used to summarize these effects. An age-cohort model provided the best fit to the data in both countries, indicating that changes over lifetime, rather than recent differences in medical surveillance, might account for the observed differences in these 2 populations. The changes over birth cohort were much greater among Singaporean women. The relative effect of age was very similar in the 2 countries. Analyses show that age and cohort effects may explain the differences in trends of female breast cancer incidence between Sweden and Singapore. The larger cohort effect seen in Singaporean women may be attributed to more rapid changes in reproduction and lifestyle patterns than that of Swedish women during the period studied. The incidence of breast cancer in postmenopausal women in Singapore will probably continue to rise in the coming decades to match the current Swedish rates.
Although breast cancer is the most common cancer among women in the world, there is wide geographic variation in incidence rates.1 While the incidence rates in several Western populations are high (for example, 104/105 in Los Angeles, CA; 81/105 in New South Wales, Australia; 74/105 Birmingham, U.K.), they are low in Asian populations such as Madras, India (24/105), Shanghai, China (27.2/105) and Osaka, Japan (27.9/105).1 Differences in health care systems and access, as well as completeness of cancer registration, probably explain some of these differences. However, the almost 2-fold higher age-standardized incidence seen in the late 1990s in Sweden (81/105) compared to Singapore (47/105),1 2 highly developed countries, could hardly be explained by these factors and alternative explanations need to be explored. Possible explanations for the differences are lifestyle factors, reproductive history and, more speculatively, genetic background factors.
Sweden has a total population of nearly 9 million people and comprises predominantly Caucasians, while Singapore has a population of approximately 4 million people consisting of 3 main ethnic groups: Chinese (77%), Malays (15%) and Asian Indians (6%). Both countries have well-established population-based cancer registries that serve to monitor cancer incidence.2, 3 This allows for the analysis of trends in incidence rather than mortality, which tends to be influenced by treatment and prognostic factors.
Approximately 25% of all breast cancers are thought to have a hereditary component.4 However, the highly penetrant BRCA1 and BRCA2 genes appear to account for only a small proportion of all inherited breast cancers, with estimates of the prevalence of BRCA1 or BRCA2 mutations in the general population ranging from 1/200 to 1/1,000.5 Ethnic variation in the polymorphism of low-penetrant susceptibility genes, most likely involved in estrogen metabolism or the transcriptional effect of the estrogen and estrogen receptor complex, could explain some of the geographic differences.
It has also been suggested that breast cancer consists of several entities and that the disease could be divided into at least 2 groups, pre- and postmenopausal breast cancer.6 Recently, as many as 6 distinct and separate subgroups were suggested based on gene expression analyses of tissue samples.7 Different subgroups of breast cancer could thus indicate different etiologic factors causally related to the disease.
Young age at menarche, advanced age at menopause and at first birth and use of hormone replacement therapy increase breast cancer risk, while early first birth has a protective effect.6, 7, 8 The Western lifestyle and reproductive pattern could thus increase breast cancer risk.
Age-period-cohort (APC) modeling is often used to examine temporal variations in disease incidence and provide clues to etiology. The predominance of period effect would suggest that changes in a particular calendar year period affecting all age groups would account for observed trends. Hence, changes in screening, diagnosis and treatment would contribute to a period effect. On the other hand, a cohort effect would suggest differences between the different generations of individuals in the population, for example, changes in the lifestyle. In this study, modeling of period and cohort effects was used to enable identification of similarities and differences between breast cancer incidence rates in the 2 countries over time to identify possible contributing factors as well as to speculate on changes in breast cancer incidence in the coming decades.
Material and methods
The Swedish Cancer Register and the Singapore Cancer Register are both well-documented nationwide registers.2, 3 Established in 1958, the Swedish Cancer Register receives reports of newly diagnosed incident cancers from clinicians, pathologists and cytologists. In the 1970s, the completeness of reporting was almost 96%, while in the late 1990s, it was close to 100%.9, 10 The Singapore Cancer Register started in 1968 and receives notifications of incident cancers from all medical practitioners and pathology laboratories as well as reviews of all hospital discharges and death certificates of all cancer patients. The completeness of reporting was 97.8% for the period 1968–1977 and 99.0% for 1993–1997.3
All breast cancer cases diagnosed in Sweden and Singapore from 1968 to 1997 were selected for analysis. The total number of cases was 135,581 in Sweden and 10,716 in Singapore. Women less than 30 years of age at diagnosis were excluded from analysis, as there were too few events, leaving 135,034 cases in Sweden and 10,494 in Singapore. The 5-year age-specific incidence rates by 5-year calendar periods were calculated using denominators derived from population censuses. In Sweden, the incidence pattern of the cases aged 45 years and older changed dramatically after 1987, most likely due to the introduction of population-based mammography screening. Because our interest was in the changes in incidence rather than the artifactual effects of screening, Swedish women aged 45 and older diagnosed after 1987 were excluded from the Poisson regression analysis, which was performed on women aged less than 75 years at diagnosis. In all, 65,662 Swedish and 9,750 Singaporean women were included in the regression analysis.
The APC model is a Poisson regression model, where the number of events in age group i, period j and birth cohort k (yijk) is modeled as a Poisson random variable with mean θijk as follows: Yijk = number of events in age group i, period j and cohort k; Nijk = number of person-years in age group i, period j and cohort k; rate rijk = Yijk/Nijk.
where μ denotes the baseline rate in the chosen reference stratum, αi denotes the effect of the ith age interval, βj denotes the effect of the jth calendar period and γk is the effect of the kth birth cohort (derived from period minus age).
Thus, there is an exact linear dependency between these age, period and cohort effects since the birth cohort k = j − i + m, where m is the total number of age classes. As a result of this collinearity, we cannot estimate all the effects from the full model; so we limit our analysis to fitting age-cohort (AC) and age-period (AP) models to the data. The deviance statistics and Z-test were used to determine the goodness of fit of the models and significance of the effects, respectively.11 The deviance statistic measures the closeness of the model predictions to the observed rates; hence, a nonsignificant value indicates a good fit. The Akaike information criterion (AIC) was also calculated as it enables the comparison of models with different complexity, with smaller values indicating better fit.12 Incidence rate ratios (IRRs) were used to summarize the effects of age groups, calendar periods and birth cohorts. These were derived from separate models for women younger and older than 45 years at diagnosis.
The age-standardized incidence rate of female breast cancer in Singapore doubled from 20/105 in the 1968–1972 period to 44/105 for the period 1993–1997. In contrast, the age-standardized incidence rate in Sweden rose from 54/105 in the 1968–1972 period to 77/105 for the period 1993–1997.
The overall breast cancer incidence rates in Sweden and Singapore begin to rise at 30 years of age and start diverging 5–10 years later (Fig. 1). While in Singapore the increase levels off after age 50, it continues to rise slowly in Sweden, reaching nearly 350/105 at age 80+ years. Figure 2 shows a similar comparison for different calendar periods 5 years apart. Regardless of calendar period, breast cancer incidence approached 350/105 (range, 329–361 per 105) for Swedish females aged 80+ years. Coinciding with the introduction of population-based mammographic screening in Sweden in the late 1980s, the incidence peak moved to younger ages and was followed by a notable decline in incidence for higher ages. A different temporal pattern was seen in Singapore. Incidence rates rose among younger women (< 45) notably from 1983 to become similar to Swedish rates. In women 50 years of age and older, the striking difference between Sweden and Singapore persisted through the 25-year period of study. A slow gradual increase took place in Singapore, but it affected all the older-age groups similarly, with the age incidence curves remaining approximately horizontal among postmenopausal women.
An age-cohort model provided a good fit in both the premenopausal and postmenopausal women in Singapore (Table I). In Sweden, the age-cohort model was the best fit for premenopausal women and both age-period and age-cohort models provided similar and reasonable fit for postmenopausal women.
Table I. Results of Age-Period and Age-Cohort Analyses of Breast Cancer in Sweden and Singapore, 1968–1997
Age < 45
Age ≥ 45
Stratified by age < 45 and age ≥ 45 at diagnosis. DF, the deviance statistic's degrees of freedom. Sweden postmenopausal women (≥45) diagnosed 1988–1997 are not included.
Age + period
Age + cohort
Age + period
Age + cohort
Using women aged 45–49 in each country as the reference group, the relative effects of age, represented as the IRRs, are presented in Table II and displayed in Figure 3. These IRRs represent the relative increase in age-specific rates derived from our age-cohort models for each country. The relative effect of age on breast cancer incidence is remarkably similar in the 2 countries, and the decline and plateau in the older-age group in Singapore women seen in cross-sectional calendar periods are no longer apparent.
Table II. Estimates of the Incidence Rate Ratio for Each Age Group Relative to the Reference Group (45–49)
Sweden IRR (95% CI)
Singapore IRR (95% CI)
Separate age-cohort models were fitted for younger (age 49 and less) and older (age 45 and above) women in both countries. Swedish postmenopausal women (≥45) diagnosed 1988–1997 are not included.
The IRRs in Table III represent the relative increase in age-specific incidence in the various birth cohorts (represented by their mid year), with the 1928 birth cohort as the reference group. These are presented in Figure 4, where it can be clearly seen that the cohort effects for birth cohorts after 1930 are larger in Singapore than in Sweden. In women aged ≥ 45 years, the cohort effect is negligible in Sweden, whereas in Singapore the cohort IRR rises steadily to 2.4 (95% CI = 2.2–2.8) in the most recent birth cohort (1948). In women younger than 45 years, the cohort effect was also much stronger in Singapore than in Sweden, with an IRR of 3.5 (95% CI = 2.6–4.7) in the Singapore 1963 birth cohort compared to 1.3 (95% CI = 1.1–1.5) for the same birth cohort in Sweden.
Table III. Estimates of the Incidence Rate Ratio for Each Birth Cohort Relative to the Reference Cohort (1928)
Age < 45 IRR (95% CI)
Age ≥ 45 IRR (95% CI)
Age < 45 IRR (95% CI)
Age ≥ 45 IRR (95% CI)
Separate age-cohort models were fitted for younger (age less than 45) and older (age 45 and above) women in both countries. Swedish postmenopausal women (≥ 45) diagnosed 1988–1997 are not included.
Our cross-sectional data suggest a marked difference in the temporal pattern of age-specific breast cancer incidence between Sweden and Singapore. While period and birth cohort effects both contribute to these differences, the period effect is unlikely to explain much of the differences since the age-cohort model provided a better fit. When we adjusted for birth cohort effects, the age effects became remarkably similar in the 2 countries. This suggests that differences in the cohort effect account for the differences in cross-sectional age-specific rates. Overall, the cohort effect in Swedish women is of very small magnitude in both younger and older women. Breast cancer incidence in Singaporean women, however, showed large cohort effects increasing steadily from the 1930 birth cohort. These effects implicate increasing exposure to risk factors over time.
Breast cancer risk is influenced markedly by age at first birth, parity, age at menarche6, 7, 8 and more modestly by duration of breast-feeding.13 Thus, different temporal changes in exposure to these risk factors in Sweden and Singapore may contribute to the strikingly different birth cohort effects. Changes in lifestyle and reproductive factors are in general attributed to the modernizing and Westernizing of society. While Sweden is clearly already a Westernized society, Singapore experienced rapid transition from a developing to a developed nation during 1968–1997. For example, the GDP for Singapore increased by 325% between 1960 and 1980, whereas for Sweden the increase was 64% (International Monetary Fund). The impact of such economic changes is likely to contribute to major changes in lifestyle and to have greater impact on the more recent birth cohorts.
Total fertility rates for Singapore and Sweden have declined during the period 1964–1998, with the rate of decline for Singapore being much sharper than in Sweden (Fig. 5). Singaporean women are also delaying childbirth until a later age, with the proportion of women having their first child at the age of 30–34 years increasing from 19% in 1967 to 67% in 198914 compared to an increase from 8.5% in 1973 (the earliest year for which we have data) to 17% in 1989 in Sweden (Medical Birth Register, National Board of Health and Welfare, Sweden). These factors possibly contribute to the steep rise of the IRR for women in birth cohorts from the late 1930s.
Breast-feeding, which slightly protects against breast cancer, has also declined over the years in both countries.15, 16 A more Westernized lifestyle and diet may also result in increasing obesity,17, 18, 19, 20 which is known to exert effects on breast cancer risk.6, 7, 8 In particular, there is evidence that weight gains after 18 years of age is possibly an independent risk factor for breast cancer.21 It is thus possible that changes in the body size of women have also contributed to the observed cohort effects.
As Singapore is a multiethnic population, differential changes in risk factors among the different ethnic groups may contribute to the temporal changes in age-specific rates. However, Chinese were the main ethnic group (around 77%) over the period 1968–1997. Of the 10,494 breast cancer cases in women 30 years and older in Singapore, 85.4% were Chinese. The overall breast cancer incidence is therefore dominated by the incidence in Chinese women. Hence, even though there are ethnic differences in fertility, breast-feeding and obesity rates as well as age at first birth, ethnic differences are unlikely to account for the marked temporal trend observed.
The exclusion of Swedish women 45 years and older diagnosed after 1987 is unlikely to contribute to the differences in cohort effect in both populations. The age-specific rates for those 45 years and older were remarkably similar from 1968 to 1987. We could therefore reasonably assume that after 1987, the age-specific rates in the absence of screening are likely to be similar to those before 1987. Alternatively, the effect of screening could be adjusted for through modeling. This would, however, increase the complexity of the models and, more importantly, limit the interpretability of their parameters. Hence, we decided instead to restrict the data in our regression analyses.
Effects that manifest themselves in specific calendar periods over a wide range of age groups are known as period effects. Thus, increasing awareness through better public education, changes in diagnostic procedures or criteria, as well as screening programs are likely to manifest as period effects. Population-based mammographic screening for breast cancer started in Sweden in the mid-1980s.22 In Singapore, no such program has been implemented. The dramatic change in breast cancer incidence for postmenopausal Swedish women from 1987 onward (Figs. 1 and 2) and the contrast with Singapore at this same time is consistent with an impact of screening on the breast cancer incidence rate in this age group.
APC analyses of routinely collected data provide valuable etiologic hypotheses explaining long-term disease trends. The approach has evolved from graphic description23 to more sophisticated statistical modeling.24 The challenge in APC modeling is that the 3 terms in a full model are collinear and thus require an additional constraint, but the resulting estimates cannot be uniquely determined. The good fit of a 2-term age-cohort model to our data enables us to interpret and compare the effects, but it is important that one first checks for linear drift, which can manifest as any age, period, or cohort effect.25 In 3 of our 4 analyses, an age-drift model did not fit the data, and in the fourth (Sweden, age < 45) there was only marginal evidence (p = 0.08) of fit. Furthermore, in all analyses, age-cohort models provided adequate fit and were significantly better than simple linear (drift) models. In APC modeling, there is also a criticism when comparing age-cohort and age-period models that the age-cohort model will provide a better fit simply because it offers additional complexity due to the larger number of terms.11 The AIC provides a method for comparing models that adjusts for this unequal complexity. In 3 of our 4 analyses presented here, the AIC was smaller for the age-cohort model, indicating a better fit, while in the fourth (Sweden, age 45 and older), the AIC for the age-cohort model was only marginally higher than for the age-period model, with the p-values for the fits being similar.
Our analyses show that, excluding possible effects of screening, differences in trends of female breast cancer incidence between Sweden and Singapore are explained by age and cohort effects but not by period effects. This suggests that changes over lifetime, rather than recent differences in medical surveillance, might account for the observed differences in the trends in these populations. The stronger cohort effect occurring in Singapore is likely to be due to rapid changes in reproductive and lifestyle factors.
The absolute incidence of postmenopausal breast cancer in Sweden (excluding the screened population, 1988–1997) is about twice as high as in Singapore. Given the similarity of the age effects in the 2 countries, and the closing of the gap in the premenopausal incidence rates in recent years, the cohort effects would suggest that in the absence of population-wide screening, the incidence of breast cancer in postmenopausal women in Singapore will continue to rise in the coming decades. It is most likely that other countries around the world that are now on the threshold of an era of a more Westernized lifestyle will experience these same changes, with a major impact on the utilization of health care in the future.
The authors thank the Singapore Cancer Registry and the Swedish Cancer Registry for granting permission to use their data.