Lifetime ovulatory years and risk of epithelial ovarian cancer: a multinational pooled analysis.

This study included 25 case-control studies (see Table 1 ) ( 16-42 ) from OCAC ( 43 ). Participants provided informed consent for original studies, whose protocols were approved by their respective institutional review boards. Characteristics of the 25 case-control studies from the Ovarian Cancer Association Consortium, conducted in Asia, Australia, Europe, and North America from 1989 to present and included in the lifetime ovulatory years (LOY) analyses Employed a nested-case control study design within a cohort study. OCAC’s harmonized core data provided LOY component variables: age at last menstrual period (LMP) before diagnosis (case participants) or interview (control participants); age at menarche; number of pregnancies; number of full-term births; and total durations of pregnancy, breastfeeding, and OC use. LOY was calculated with 12 algorithms ( Supplementary Table 1 , available online) ( 8 ) using the formula: where “menstrual span” was calculated from age at LMP minus age at menarche. The algorithms were divided into 4 classes based on how “years of anovulation” was defined (see Figure 1 ). Flowchart for algorithms to calculate lifetime ovulatory years. OC = oral contraceptive. Seven studies recorded age at LMP (cases = 6881 [32.4% of total]; controls = 8316 [31.7% of total]). For the remaining studies, we imputed age at LMP (see Figure 2 ) ( 44 ) and assessed the imputation algorithm by comparing actual vs imputed age at LMP for the 7 sites ( Supplementary Table 2 , available online). Sites with 50% or more missing values in any LOY component except age at LMP were excluded from algorithms using those components ( Supplementary Table 3 , available online) ( 45 , 46 ). Flowchart for imputation of age at last menstrual period (LMP). HRT = hormone replacement therapy. Variables considered a priori as potential confounders included age at diagnosis (cases) or interview (controls), self-reported race (Asian, Black, Other, Unknown, White, where Other was defined by each participating study as not Asian, Black, or White), education, body mass index (BMI) 1 year to 5 years prior, family history of ovarian or breast cancer in a first-degree relative, smoking status, history of endometriosis, and tubal ligation. We used random effects meta-analysis to assess interstudy LOY-EOC heterogeneity. Because we observed no substantive heterogeneity ( Supplementary Figure 1 , available online), we used the pooled data set adjusted for study site for all analyses. We used Pearson correlation to assess pairwise correlations of LOY calculated among algorithms limiting analyses to observations with complete data for each algorithm in the pairwise comparison. Pearson correlation was also used to assess the correlations of individual components with LOY calculated by each algorithm. Multivariable logistic regression was used to estimate odds ratios (ORs) with 95% confidence intervals (CIs) for the association between LOY and EOC overall and by histotype. Models were adjusted for study site, age at diagnosis or interview, race, education, BMI, smoking status, and family history. Inclusion of tubal ligation and endometriosis in models did not alter any findings; thus, tubal ligation and endometriosis were omitted from final models. Because OCAC only recorded total months of breastfeeding across all live births and not months per breastfeeding episode, to account for return of ovulation once food is introduced typically at 6 months, we performed sensitivity analyses replacing breastfeeding duration with either 1) number of live births times the average duration of breastfeeding per live birth if the average duration was less than 6 months or 2) number of live births times 6 months if the average duration was 6 months or greater. Similar sensitivity analyses were performed for algorithms containing a term for breastfeeding duration (algorithms I-L). Sensitivity analyses were performed with multiple imputation by chained equation to assess the effect of missing values on LOY-EOC associations ( 47 ) including the same covariates as main models. Nested imputations were done for number of pregnancies, number of full-term births, duration of breastfeeding, and duration of OC use using the binary variables of ever pregnant, ever breastfed, and OC use, respectively. Imputations were done 5 times with auxiliary variables defined as Pearson correlation larger than 0.4 ( 48 ). Sensitivity analyses also examined limiting models to population-based studies and using only observations with complete data for all variables. To assess the relationship between LOY and EOC histotypes, we present results using algorithm K because this algorithm most closely reflects lifetime ovulatory years accounting for OC use, pregnancy type, and breastfeeding. Prior studies suggest that the relationship between LOY and EOC may not be linear ( 49 ); thus, we constructed models using LOY and log(LOY). Because log(LOY) did not improve model fit when included with LOY and models with LOY alone provided a better fit than those with log(LOY) alone, we report only analyses using LOY. The association of each LOY component and EOC risk overall and separately for each histotype was estimated using multivariable logistic regression adjusted for study site, age at diagnosis (case patients) or interview (control participants), race, education, BMI 1 to 5 years prior to diagnosis (case patients) or interview (control participants), smoking status, family history, and other LOY components. To assess whether each component acts through ovulation suppression alone, we compared expected beta coefficient with actual estimates obtained from regression models ( 7 ). Based on the incessant ovulation hypothesis, 1 year of ovulation suppression should have the same effect on the log odds of EOC regardless of origin. Thus, if we assign one as the expected beta coefficient for age at LMP per year (indicating that a 1-year increase in LMP, which would increase LOY by 1, would increase the log odds by 1), then the expected beta coefficient for age at menarche per year would be −1 because each additional year increase would decrease LOY by 1 year and hence decrease the log odds by 1. Similarly, the expected beta coefficients for OC use per year, number of incomplete pregnancies (assumed to be 3 months or 0.25 years), number of full-term births (assumed to be 9 months or 0.75 years), and breastfeeding per year would be −1, −0.25, −0.75, and −1, respectively. We then computed the relative coefficients, defined as the actual coefficients from regression models divided by the actual coefficient of age at LMP. This set the relative coefficient for age at LMP to 1, just as in the expected model. This enabled us to compare the relative coefficient estimates with their expected counterparts. To assess the statistical significance of individual components, χ 2 statistics and P values were obtained from the likelihood-ratio test for the removal of each component from the full model. Sensitivity analyses examined limiting models to population-based studies and using only observations with complete data for all variables. All statistical tests were 2-sided and performed in Stata/SE version 16.1 (StataCorp, College Station, TX, USA).

Among the 25 studies, there were 26 204 control participants and 21 267 case patients ( Table 2 ). Compared with controls, cases were more likely to have a family history of breast or ovarian cancer and a history of endometriosis, be hysterectomized, and be obese or overweight. Controls were more likely to have never smoked, be premenopausal, and have had a tubal ligation. Cases reported a shorter total duration of OC use and breastfeeding and fewer total pregnancies. Characteristics of ovarian cancer cases and controls included in the lifetime ovulatory years (LOY) analyses Race was self-reported by participants and provided to the OCAC Core. The category “other” refers to lack of self-identification as Asian, Black, or White. BMI = body mass index; OCAC = Ovarian Cancer Association Consortium. Among the 12 algorithms, median LOY ranged from 31.67 (interquartile range [IQR] = 25.50-35.20) to 35.75 (IQR = 32.50-37.50) years ( Figure 3 ; Supplementary Table 4 , available online). Pairwise LOY correlations ranged from 0.75 between the algorithms in the first class (inclusive of pregnancies only) and the third class (inclusive of pregnancies, OC use, and breastfeeding) to at least 0.99 for correlations within the same class ( Supplementary Table 5 , available online). Correlations between individual components and LOY are presented in Supplementary Table 6 (available online). As algorithm complexity increased, correlations between age at LMP and LOY decreased. OC duration was moderately negatively correlated with LOY (rho range: −0.68 to −0.69); correlations between the other components and LOY were low. Distribution of lifetime ovulatory years calculated from 12 different algorithms. Odds ratios for LOY per year increase across the 12 algorithms ranged from 1.014 (95% CI = 1.009 to 1.020) to 1.044 (95% CI = 1.041 to 1.048) ( Table 3 ). Associations with LOY calculated from the third class of algorithms (inclusive of pregnancies, OC use, and breastfeeding) were not changed when months of breastfeeding were truncated at 6 for participants reporting more than 6 months per birth (data not shown). LOY associations remain unchanged when adjusting models in the first class of algorithms (which included only pregnancies) for OC and breastfeeding duration, as well as when adjusting the second class of algorithms (which included pregnancies and OC duration) for breastfeeding duration (data not shown). Sensitivity analyses with multiple imputations of missing values did not alter LOY-EOC associations ( Table 3 ). Sensitivity analyses limited to population-based studies and those limited to observations with complete data also did not alter the LOY-EOC association (data not shown). Odds ratio for ovarian cancer per lifetime ovulatory year using complete data and full data with imputation Main analyses included participants without missing values in any component for LOY calculation; sensitivity analyses included all participants with imputation. CI = confidence interval; MCC = Melbourne Collaborative Cohort Study; NTC = Nijmegen Ovarian Cancer Study; OC = oral contraceptive; TBO = Tampa Bay Ovarian Cancer Study. Adjusted for study site, age, self-reported race (Asian, Black, Other [as defined by participants as not being Asian, Black, or White], Unknown, White), education (less than high school, completed high school, completed some college, completed college or university bachelor degree, completed graduate or professorial degree, unknown), body mass index 1 or 5 years prior (underweight, normal, overweight, obese, unknown), smoking status (never, former, current, unknown), and family history (yes, no, unknown). TBO was excluded from the sensitivity analyses because of limited numbers within site to impute missing values. MCC was excluded from the sensitivity analyses because of limited numbers within site to impute missing values. NTH was excluded from the sensitivity analyses because of failure to converge on observed data. NTH was excluded from the sensitivity analyses because of limited numbers within site to impute missing values. Individual components in LOY, except for age at menarche, were associated with EOC ( Table 4 ). There were substantial deviations between relative estimated coefficients and expected estimates for each component. The estimated coefficient of OC use per year was 4.45 times larger than expected, and estimates for pregnancies were 11- to 15-fold greater than expected regardless of pregnancy type. Estimated coefficient of breastfeeding per year was −13.45, instead of the expected −1. Results were similar when truncating breastfeeding at 6 months per full-term birth, when limiting analyses to population-based studies and when limiting analyses to observations with complete data (data not shown). Odds ratios, expected beta coefficients, and normalized beta coefficients for ovarian cancer by individual components of lifetime ovulatory years Adjusted for study site, age, self-reported race (Asian, Black, Other [as defined by participants as not being Asian, Black or White], Unknown, White), education (less than high school, completed high school, completed some college, completed college or university bachelor degree, completed graduate or professorial degree, unknown), body mass index 1 or 5 years prior (underweight, normal, overweight, obese, unknown), smoking status (never, former, current, unknown), family history (yes, no, unknown), and other components of lifetime ovulatory cycles in the model. CI = confidence interval; OR = odds ratio. Normalized to the beta coefficient of age at last menstrual period. LOY was associated with invasive high-grade serous (HGSOC; OR per year = 1.054, 95% CI = 1.048 to 1.061), low-grade serous (LGSOC; OR = 1.040, 95% CI = 1.019 to 1.061), endometrioid (OR = 1.065, 95% CI = 1.053 to 1.076), and clear cell (OR = 1.098, 95% CI = 1.079 to 1.117) but not mucinous EOC (OR = 1.006, 95% CI = 0.992 to 1.019) ( Table 5 ). Except for breastfeeding, estimated coefficients of LOY components were close to expected for HGSOC. In contrast, estimated coefficients of individual components, except for age at menarche, were larger than the expected for LGSOC, endometrioid, and clear cell cancers. Odds ratios, expected beta coefficients, and normalized beta coefficients for ovarian cancer histotypes by individual components of lifetime ovulatory years Adjusted for study site, age, self-reported race (Asian, Black, Other [as defined by participants as not being Asian, Black or White], Unknown, White), education (less than high school, completed high school, completed some college, completed college or university bachelor degree, completed graduate or professorial degree, unknown), body mass index 1 or 5 years prior (underweight, normal, overweight, obese, unknown), smoking status (never, former, current, unknown), family history (yes, no, unknown), and other components of lifetime ovulatory cycles in the model. β = estimated coefficient; CI = confidence interval; OC = oral contraceptive; OR = odds ratio. Normalized to the beta coefficient of age at last menstrual period. Adjusted for study site, age, self-reported race (Asian, Black, Other [as defined by participants as not being Asian, Black, or White], Unknown, White), education (less than high school, completed high school, completed some college, completed college or university bachelor degree, completed graduate or professorial degree, unknown), body mass index 1 or 5 years prior (underweight, normal, overweight, obese, unknown), smoking status (never, former, current, unknown), and family history (yes, no, unknown). Using algorithm K with complete data: (age at last menstrual period—age at menarche) – years of OC use—(0.25*number of incomplete pregnancies)—(0.75*number of full-term births) – years of breastfeeding. This algorithm was chosen because it most closely accounts for expected ovulation suppression due to pregnancies, OC use, and breastfeeding.

Pooling data from 25 case-control studies, we show a positive association between LOY and EOC, with each year of ovulation associated with a 4% increase in risk. We also found a positive association between LOY and HGSOC, LGSOC, endometrioid, and clear cell EOC but not with mucinous tumors. These LOY-EOC associations were not altered when using different algorithms to compute LOY or when imputing missing data. We further found that LOY components, except age at menarche, were associated with EOC, with the magnitude of these associations varying substantially from expectation if their mechanism of action was solely ovulation suppression. There was also notable heterogeneity in these component-specific findings among EOC histotypes. Together, these data suggest that reproductive factors comprising LOY exert their effects through means beyond ovulation suppression, and those relationships vary by EOC subtype. Most prior studies report a positive relationship between LOY and EOC ( 2 , 7-14 , 50-63 ). Differences in LOY definitions among studies make it challenging to compare specific findings across studies. In the present study, we defined LOY from available harmonized data using 12 algorithms. Like the Polish Cancer study ( 8 ) (1 of the 25 studies in this analysis), we found a high correlation for LOY among algorithms, although point estimates varied depending on the algorithm. When assessing overall EOC per 1-year increase in LOY, estimates ranged from 1.01 to 1.04, which is similar to estimates reported by the US Nurses’ Health Study (1976-2006) and Nurses’ Health Study II (1989-2005) (OR = 1.07, 95% CI = 1.05 to 1.08) ( 10 ). Although it is reassuring that our results are similar to previous work, because each study used different LOY algorithms and units of presentation (eg, quartiles, ovulatory cycles) ( 15 ), a direct comparison of estimated magnitudes is not possible. A standardized definition of LOY would facilitate cross-study comparisons and allow for more robust interstudy analyses. Our findings confirm that among algorithms that account for menstrual span, number of pregnancies, total duration of OC use, and total duration of breastfeeding, point estimates for the LOY-EOC relationship are similar. Defining LOY using these factors would facilitate interstudy analyses. We report differences in the association of LOY with EOC subtypes. We report a positive association between LOY and HGSOC and LGSOC. Whereas previous studies have reported a positive association between LOY and risk of serous tumors ( 2 , 10-15 ), only the Ovarian Cancer Cohort Consortium (OC3) ( 14 ) reported results separately for HGSOC, also finding a positive association. Separating serous EOC analyses is important because HGSOC and LGSOC are distinct diseases ( 64 , 65 ). Also consistent with most ( 10-14 ) but not all previous studies ( 2 , 15 ), we found positive associations between LOY and clear cell and endometrioid but not mucinous tumors. These results are consistent with epidemiologic evidence that suggests a different risk-factor profile for mucinous EOC ( 3 , 66 ). Results regarding the associations between LOY components and EOC appeared consistent with previous studies ( 7 , 8 , 10 , 12 , 58 , 62 ). Beyond considering statistical significance, our study also compared the magnitudes of each component’s effect on EOC risk and found the actual magnitudes varied substantially from expectation ( 7 ). Based on the incessant ovulation hypothesis ( 6 ), women with the same LOY should have the same estimated risk if ovulation is the only etiologic mechanism underlying the relationship between the components of LOY and EOC. However, consistent with 2 case-control studies ( 7 , 62 ), we show that pregnancy, OC use, and breastfeeding are associated with stronger protective effects than would be expected based on ovulation suppression alone. Moreover, the protection from 1 year of pregnancy, whether complete or incomplete, was substantially greater than that of 1 year of OC use ( 7 ). Together, these data imply that mechanisms beyond ovulation suppression, such as hormonal alterations ( 67 , 68 ) or inflammation ( 69 ), contribute to the LOY-EOC association. They further imply differences in the mechanisms whereby individual LOY components impact EOC risk, especially for non-HGSOC subtypes, suggesting that a model of EOC risk incorporating just LOY and not its component parts would be insufficient in fully capturing the effects of exposure to LOY components. Our results indicate heterogeneity in the associations between LOY components and histotype-specific risk. Notably, except for breastfeeding, the estimated coefficients for HGSOC were close to expected if only ovulation suppression underlies the component-HGSOC relationship. This suggests that ovulation may be the primary etiologic mechanism for HGSOC; however, because HGSOC is believed to arise in the fimbriated end of the fallopian tube and not the ovary ( 70-72 ), ovulation effects must extend beyond ovarian surface epithelium trauma, as originally proposed by Fathalla ( 6 ). Notably, during ovulation, fallopian tube fimbria come in close proximity to the site of ovulation, directly exposing the fimbria to ovarian follicular fluid. In vitro studies show that normal fallopian tube epithelia exposed to follicular fluid aspirates develop TP53 mutations, a hallmark of HGSOC ( 73 ). Moreover, follicular fluid has both mutagenic and tumorigenic effects facilitating the full transformation process for developing HGSOC from the fallopian tube ( 74-77 ). Thus, follicular fluid may be the link between greater number of ovulations and HGSOC. In contrast to HGSOC, factors beyond ovulation suppression underlie the link between LOY and other histotypes. For LGSOC, endometrioid and clear cell histotypes, we found that actual coefficient estimates were substantially larger than expected for OC use, pregnancies, and breastfeeding. This suggests that other mechanisms, such as increased progestin exposure ( 78 ), may play a role in the protective effects of these factors. Although we did not find any association between LOY and mucinous EOC, we report associations for several LOY components. Thus, factors other than ovulation may be driving mucinous carcinogenesis. Moreover, the relationship between LOY components and mucinous disease varied from that of other histotypes. Together, these observations suggest that factors underlying the relationship between exposures and EOC vary based on histotype and confirm the unique origin of mucinous cancers ( 79 , 80 ). The major strength of our work was pooling 25 case-control studies, allowing us to estimate more precisely the LOY-EOC association overall and by histotype. The large data set also enabled comparison of different LOY definitions and their impact on the LOY-EOC relationship. For LOY components, the sample size enabled us to separate the effects of ovulation suppression from other potential etiologic mechanisms. The range of studies from 4 continents and 9 countries supports the generalizability of our findings. Despite these strengths, there are several limitations. Because all but 2 studies ( 25 , 42 ) employed a retrospective case-control design, recall and selection bias are always a concern. Regardless of study design limitations, our estimates were consistent with previous prospective studies, including the US Nurses’ Health Study and US Nurses’ Health Study II studies ( 10 ) and the OC3 pooled analysis of prospective studies ( 14 ). We made some assumptions about LOY components that may impact results. If age at LMP was unknown, we imputed it using an algorithm based on average age at menopause by country, age at first hormone replacement therapy use, or age at hysterectomy. We compared the observed and imputed age at LMP from 7 sites, conducted sensitivity analyses using LOY calculated from the imputed value for those sites, and noted no differences in observed associations. To prevent overestimating the duration of anovulation from breastfeeding, we repeated analyses capping women at 6 months of breastfeeding per live birth. Results were unchanged. In conclusion, increasing LOY is associated with increased EOC risk, as well as the risk of HGSOC, LGSOC, endometrioid, and clear cell histotypes. Although point estimates varied slightly, the association between LOY and EOC was not altered when LOY was calculated in different ways using core components. Our study also indicated heterogeneity in the expected estimated coefficients of each LOY component on histotype-specific EOC. Together, our findings suggest that ovulation suppression is not the sole mechanism whereby reproductive factors affect EOC overall and for non-HGSOC histotypes. Identifying these mechanisms and understanding their individual and joint roles can provide deeper insight into disease etiology and potential risk-reducing approaches.

Click here for additional data file.