Reproductive factors and breast cancer risk according to joint estrogen and progesterone receptor status: a meta-analysis of epidemiological studies

We identified epidemiological studies (cohort or case-control studies) in MEDLINE from the year 1966 to Dec. 1, 2005, by running searches with the key words "Breast Neoplasm/ep [Epidemiology]" and "(ER or PR).mp [mp = title, abstract, name of substance, mesh subject heading]". We identified additional studies by tracking the references in all identified articles. We noticed that the studies published before 1995 all defined their receptor subtypes according to either ER or PR status and that most of them had a hospital-based study design, whereas most of studies published since 1995 used joint ER/PR status to define receptor subtypes and had a population-based study design. Using joint ER/PR status could reduce the chance of including any tumors in which one of the receptor statuses was mislabeled. Therefore, for inclusion into this meta-analysis, the identified articles have to have estimates of relative risk (RR) for ER⁺PR⁺ and ER^-PR^- breast cancer. We did not summarise the data for the two rare subtypes (ER^-PR⁺ and ER⁺PR^- breast cancer), because few studies reported estimates of RR for them. For exposure variables, we focused on the summary of results for reproductive factors that had been more frequently tested across studies, although some studies had also examined other factors such as body mass index, hormone replacement therapy, and so on. We thoroughly reviewed two cohort [9, 10], five population-based case-control [7, 11‐14], and two hospital-based case-control studies [15, 16] that investigated these issues. We also included one population-based case-control analysis in press [8] (Table 1).

We extracted study-specific estimates of RR (odds ratios, rate ratios, and risk ratios) and their 95% confidence intervals (CIs) for highest versus lowest category of parity, age at first birth, breastfeeding, and age at menarche. For this analysis, we used RR to refer to any RR measure. Two studies presented their results by subgroups only, either by menopausal status [13] or by age at birth [10]. To obtain one summary estimate, we combined the RRs for the subgroups through weighting their log (base e) RRs by the inverse of their variances. For RRs of parity, seven studies [8‐10, 12‐15] used nulliparous women as the reference group, whereas one of our own studies [7] used never-pregnant women as the reference group. For consistency with other studies, we re-computed our results to have nulliparous women as the reference group using our original data [7] and changed the lower limit in the highest category from five to three births. For the same reason, we also changed the lower limit in the highest category from 24 months of breastfeeding to 7 months for two of our own studies [7, 8].

For parity, we also extracted or computed study-specific trend estimates. One study provided the estimates in the original publication [15]; we computed trend estimates using original data for two of our own studies [7, 8] and computed trend estimates using RRs (and 95% CIs) for categorical variables in four studies [9, 12‐14] using the method described by Greenland and Longnecker [17]. We used the midpoint of each category in these calculations; for the open-ended highest category, we used its lower limit plus one as an estimate of the mean number of births. For the study of women under age 45 years from Georgia, Washington, or New Jersey, that reported RRs for ever versus never having given birth [12], we used two as our estimate of the number of births because that was close to the mean number of births for parous women under age 45 years in two of our own studies [7, 8]. To calculate the estimate for one study, we combined the RRs given for one birth at age 20 and one birth at age 35, weighting their log RRs by the inverse of their variances [10].

To summarise the results by menopausal status, we did as follows. We accepted the definitions used in three studies [9, 11, 13]. Three studies restricted eligibility to young women (under age 40 [14], under age 45 [12], and under age 50 [8]). We combined these young women with premenopausal women to form a group representing women who were premenopausal or young.

We used fixed effects models to calculate summary RRs for all studies combined, by type of study design and menopausal status (premenopausal or young and postmenopausal status), because we did not detect statistically significant heterogeneity of effects between studies (p ≥ 0.10) [18]. To ensure that each particular study contributed the same weight to the summary log RRs for ER⁺PR⁺ and ER^-PR^- subtypes, we used the inverse of the average variances of the two subtypes as the weighting variable. This is necessary for the calculated summary values to be directly comparable. In Tables 2, 3, 4, 5, we list the study-specific RRs (95% CIs), then the summary RRs (95% CIs) by type of study design and menopausal status, and the overall summary estimates for all studies.

To test for potential heterogeneity in risk by receptor status, we first calculated the average variance weighted difference in the log RRs for the two receptor status groups. The variance for each difference in log RRs is estimated by the sum of their variances. The differences are divided by their pooled variance, these weighted values are summed across studies, and the total is divided by the sum of the inverses of these pooled variances. To create a test statistic, χ² (with 1 degree of freedom), this value is squared and then divided by its variance, which is the inverse of the sum of the inverses of the pooled variances across studies. Potential heterogeneity in effect by study design and by menopausal status was examined by using standard homogeneity tests. All p values reported for homogeneity are two-sided, and p values less than 0.05 were considered statistically significant.

Using Egger's regression asymmetry test, we assessed the possibility of publication bias [19]. This analysis is based on a regression model in which the standard normal deviate is regressed against the study-specific estimate of the precision of log RR. When no publication bias is present, the points will scatter around a regression line that runs through the origin. We considered there to be publication bias if the intercept of the Egger's regression line deviated from zero with a two-sided p value of less than 0.10.

All analyses were performed using the Stata statistical software (Version 8; StataCorp LP, College Station, TX, USA).

Table 1 presents the basic characteristics of the ten studies that we reviewed. Nine of the ten studies were published between 1995 and 2005 [7, 9‐16], and the other is currently in press [8]. Seven of the ten studies are from the U.S. [7‐12, 16], and the others are from Canada [13], Australia [14], and Japan [15]. These studies include women of all ages. The main source of hormone receptor information was medical records [7‐14]; in one study, a single laboratory provided the data on receptor status [16], and one study did not specify the source of their receptor data [15]. The percentage of participating cases with available ER/PR status was at least 65% in nine studies [7‐14, 16], whereas it was only 40% in one hospital-based case-control study [15]. The number of subjects included in these analyses ranged from 104 to 2,130 for ER⁺PR⁺ and 80 to 1,081 for ER^-PR^-. On average, the number of ER^-PR^- breast cancer cases involved in the analyses was approximately 46% of that of ER⁺PR⁺ cancer cases. All studies considered confounding in their analyses although the confounders included in the models varied by study (Table 1).

Eight studies were included in the meta-analysis of parity and breast cancer risk by ER/PR status (Table 2). Both the summary RRs for the highest versus the lowest category and the summary RRs per birth indicated that the protective effect of parity was confined to ER⁺PR⁺ cancer. Each birth reduced the risk of ER⁺PR⁺ cancer by 11% (RR per birth = 0.89, 95% CI = 0.84–0.94), and the p value for homogeneity between ER⁺PR⁺ versus ER^-PR^- cancer was less than 0.001 (Figure 1).

Nine studies were included in the meta-analysis of age at first birth and breast cancer risk by ER/PR status (Table 3, Figure 2). Women in the oldest age at first birth category were on average at a 27% greater risk (summary RR = 1.27, 95% CI = 1.07–1.50) for ER⁺PR⁺ cancer than women in the youngest age category, but age at first birth was not associated with risk of ER^-PR^- cancer (summary RR = 1.01, 95% CI = 0.85–1.20). The difference in effects between ER⁺PR⁺ and ER^-PR^- was statistically significant (p for homogeneity between ER⁺PR⁺ and ER^-PR^- cancer = 0.010). The summary RR for ER⁺PR⁺ cancer appeared greater among postmenopausal than premenopausal or young women (postmenopausal women: summary RR = 1.65, 95% CI = 1.15–2.38; premenopausal or young women: summary RR = 1.24, 95% CI = 0.96–1.62), but the difference was not statistically significant (p = 0.211).

The summary RRs from the seven studies of breastfeeding show that breastfeeding was associated with reduced RRs of both ER⁺PR⁺ and ER^-PR^- cancer (summary RRs [95% CIs]: 0.95 [0.87–1.05] and 0.91 [0.83–1.00] for ER⁺PR⁺ and ER^-PR^- subtypes, respectively) (Table 4, Figure 3). The protective effect from breastfeeding was observed among population-based case-control studies, but not among hospital-based case-control studies. This difference is marginally significant (p = 0.071 for ER⁺PR⁺ and p = 0.054 for ER^-PR^- cancer). One reason for this discrepancy could be that the two hospital-based studies included a small number of cases and therefore had insufficient power to find any effect.

Nine studies were included in the meta-analysis of age at menarche and breast cancer risk by ER/PR status (Table 5, Figure 4). The overall summary RRs of the oldest versus the youngest age at menarche category show that age at menarche was negatively associated with the risk of both ER⁺PR⁺ and ER^-PR^- cancer (summary RRs [95% CIs]: 0.72 [0.64–0.80] and 0.84 [0.75–0.94] for ER⁺PR⁺ and ER^-PR^- subtypes, respectively). The protective effect of late age at menarche was statistically significantly greater for ER⁺PR⁺ than ER^-PR^- cancer (p for homogeneity between ER⁺PR⁺ and ER^-PR^- cancer = 0.006).