Introduction

Observational studies have reported a positive association between adult height and ovarian cancer risk.1,2,3,4 However, these studies were subject to the biases inherent in conventional observational studies, including selection bias, differential and non-differential reporting bias and confounding. The degree to which these factors could account for the observed association is uncertain. Mendelian randomisation (MR) uses genetic markers as proxies for environmental exposures and, due to the singular qualities of genotype data, can provide complementary evidence by overcoming many biases affecting conventional studies.5 We used MR to examine the relationship between height and ovarian cancer risk in the Ovarian Cancer Association Consortium (OCAC),6 aiming to provide more certainty about the relationship between height and ovarian cancer risk. We hypothesised that greater genetically predicted height would be associated with increased risk.

Materials and methods

Study population and outcomes

We pooled data from 16,395 genetically European women with primary ovarian/fallopian tube/peritoneal cancer and 23,003 controls from 39 OCAC studies (Table 1; Supplementary Table 1). The data set and methods have been described previously.7 Participants were genotyped via the Collaborative Oncological Gene-Environment Study.8 Twenty-two studies provided height data (16 provided parity, oral contraceptive (OC) use, education and age at menarche information) for >50% of their participants. We first considered all cases, then stratified by tumour behaviour. Secondary analyses stratified by histologic subtype/behaviour.

Table 1 Characteristics of 39 OCAC studies and 39,398 participants of European ancestry included in the Mendelian randomisation analysis
Full size table

Genetic risk score

The Genetic Investigation of ANthropometric Traits (GIANT) Consortium had previously identified 697 single-nucleotide polymorphisms (SNPs) significantly associated with height.9 In our sample, 92 of these SNPs had been genotyped and the remainder were imputed using 1000 Genome Project data.8, 10 After excluding poorly-imputed SNPs (quality r2 < 0.6), 609 remained (92 genotyped/517 imputed) (Supplementary Table 2). In controls, minor allele frequencies (MAFs) were >5% (except for 16 SNPs, MAFs 1.7–4.9%).

We constructed a weighted genetic risk score (GRS) for height by summing height-increasing alleles across the 609 SNPs (‘GRS-609’/‘the GRS’), weighting alleles by β-coefficients for their association with height reported by GIANT. The score represents predicted additional height conferred by these variants, compared to having no height-increasing alleles. We report results for 5 cm increments.

Statistical analysis

Statistical methods have been described previously.7 Briefly, we used individual-level OCAC data for two-stage predictor-substitution MR 11, 12: first, we predicted height from the weighted GRS for all participants using coefficients from linear regression in 17,649 controls with height data; second, within each study, we used logistic regression to model disease status on GRS-predicted height. Models adjusted for age and five principal components for population structure.8 We combined study-specific estimates using meta-analysis,13 generating pooled odds ratios (pOR) and 95% confidence intervals (CI) for the trend in risk per 5 cm increase in predicted height. We had 97% power to detect an OR of 1.10 (mRnd tool).14

Sensitivity analyses included removing 16 SNPs with MAFs <5%, and restricting to SNPs with imputation r2 ≥ 0.9 (‘GRS-363’), SNPs representing distinct loci9 (‘GRS-377’), and directly-genotyped SNPs (‘GRS-92’). We examined whether potential confounders of the association in observational studies were associated with the GRS. To assess robustness to pleiotropy (where SNPs may influence risk via pathways not mediated through height), we conducted MR-Egger regression15 and assessed smaller GRSs excluding SNPs with the highest probability of acting via other pathways from GRS to outcome (SNPs associated with ovarian/other hormonal cancers (breast, prostate), hormone levels and in/near tumour initiation/growth genes). We identified these potentially pleiotropic, pathway-specific SNPs via the NHGRI GWAS Catalog,16 the UCSC Genome/Table Browsers17, 18 and from lists of SNPs nominated for iCOGS genotyping by ovarian, breast and prostate cancer researchers (to capture SNPs of interest unpublished at the time of analysis).

Secondary analyses defined cases by histologic subtype/behaviour. Among 16 studies with height/confounder data, we conducted conventional analysis (adjusted for parity, OC use, education, menarche age; stratified by study, 5-year age group) and compared results with MR-estimates among the same women.

Analyses were performed using SAS 9.2 (SAS Institute Inc., Cary, NC) and Stata 13.0 (StataCorp LP, College Station, TX). This work and each contributing study was approved by the appropriate institutional review board/ethics committee. All participants provided informed consent.

Results

Population characteristics

We included 16,395 cases (14,549 invasive tumours, 1691 borderline, 155 of unknown behaviour) and 23,003 controls (Table 1). The median diagnosis year was 2003, with 74% diagnosed post-2000. Participants were aged 18–94 (median 56) years at diagnosis/interview. Mean height ranged from 159 to 167 cm across 22 studies with data, and was 163 (standard error (SE) = 0.05) cm for controls and 164 (SE = 0.06) cm for cases (p < 0.0001).

Genetic risk score characteristics

The GRS-609 was normally distributed in controls, ranging from 15.45 to 18.99 (median = 17.23; interquartile range = 16.92–17.54). It explained 13% of variance in height, 17% after adjusting for age and principal components (partial-R2 = 12%; first-stage regression partial-F-statistic = 2505.8 (df = 1), p < 0.001). A 1-unit GRS-609 increase was associated with 5.2 cm greater height. Average height was 6.2 cm greater in the highest vs. lowest GRS quartile.

Cochran’s I2 and p-values for heterogeneity19 showed no evidence of inter-study heterogeneity in the relationship between either the GRS-609 (I2 = 34%, p-heterogeneity = 0.07) or the simplified GRS-363 (I2 = 32%, p-heterogeneity = 0.08) and height among controls (Supplementary Figure 1a, b). The GRS-609 was not associated with most potential confounders of the height-ovarian cancer association in observational studies, including age, parity, OC use and education (Supplementary Table 3). The GRS was marginally positively associated with age at menarche (p = 0.03), consistent with known genetic overlap between these traits.20

Primary outcomes

Women with greater genetically predicted height had a modestly increased risk of developing ovarian cancer (pOR = 1.06, 95% CI: 1.01–1.11 per 5 cm) (Fig. 1a; Table 2) with a greater magnitude of association for borderline (pOR = 1.15; 95% CI: 1.02–1.29) than invasive tumours (pOR = 1.06; 95% CI: 1.01–1.11; Fig. 1b, c; Table 2). No significant inter-study heterogeneity was noted (Fig. 1a–c). GRS-363 (pOR = 1.06, 95% CI: 1.00–1.11, all tumours) and GRS-377 (OR = 1.07; 95% CI: 1.01–1.12) results were similar to the GRS-609. The association was stronger when we restricted to 92 genotyped SNPs (pOR = 1.14; 95% CI: 1.04–1.25). Estimates from analyses excluding low-MAF SNPs, excluding case-only studies, or adjusting for age at menarche, were similar to primary analyses. When we sequentially excluded SNPs associated with ovarian or other hormonal cancers, hormone levels and tumour development, estimates were similar or stronger than GRS-609 results. MR-Egger suggested minimal bias from pleiotropy (p = 0.1; MR-Egger beta = 0.163 corresponded to an OR per 5 cm of 1.13 (95% CI: 1.02–1.25), confirming a significant positive association).

Fig. 1
figure 1

Association between increasing genetically predicted height and risks of all, invasive and borderline ovarian tumours. Increasing height per 5 cm predicted by weighted 609-locus genetic risk score among 39 studies. Risk of a all, b invasive and c borderline ovarian tumours. The UK grouping includes RMH, SOC, SRO, UKR, SEA and UKO for a and b, and RMH, SOC and SEA for c

Full size image
Table 2 Association between increasing height (per 5 cm)—predicted by a weighted 609-locus genetic risk score—and risk of ovarian cancer, stratified by study
Full size table

In contrast, for women with height and confounder data (16 studies), the conventional analysis suggested no association (adjusted-OR = 1.01, 95% CI: 0.99–1.04 per 5 cm). Conducting MR within the same 16 studies yielded results similar to overall analyses (OR = 1.06, 95% CI: 1.00–1.13) (Supplementary Table 4).

Secondary outcomes

After stratifying by subtype/behaviour, the strongest associations were seen for clear cell (OR = 1.20, 95% CI: 1.04–1.38) and low-grade/borderline serous cancers (OR = 1.15, 95% CI: 1.01–1.30) (Table 2). However, CIs were wide and overlapping due to lower statistical power in these subgroup analyses. The estimate for clear cell cancers was also significantly elevated in our conventional analyses (Supplementary Table 4).

Discussion

We used a 609-SNP GRS to examine the relationship between height and ovarian cancer risk for women of European ancestry. Our data indicate a modest positive association between genetically predicted height and ovarian cancer risk, which may be stronger for borderline cancers. Height may be relevant to cancer risk as a marker for lifetime growth-factor levels (e.g. IGF-1) and/or early-life exposures (socio-economic/environmental/nutritional).3, 21, 22

Observational studies are subject to biases (reverse causality, selection bias, differential/non-differential reporting, confounding) which cannot be ruled out as possible explanations for observed associations. By using genotype, the MR technique can overcome some of these biases, given three assumptions. We confirmed the two verifiable assumptions: the GRS was associated with height, and not with most known confounders. The GRS-menarche age association is unlikely to explain the observed association, because age at menarche is only weakly associated with ovarian cancer, and women with later menarche have if anything lower ovarian cancer risk, so if this affected our results, we would expect the true effect to be at least as strong as the reported association. Also, removing hormone-related SNPs, or adjusting for menarche age, did not attenuate estimates. Owing to the limited current biological understanding of all 609 SNPs, we could not conclusively exclude the presence of alternate pathways from height genes to ovarian cancer (assumption three). However, MR-Egger and sensitivity analyses excluding pathway-specific SNPs provided some evidence for their absence, minimising the likelihood that our observed association is explained by pathways separate from height/growth. Although height data were not available for the entire population, this is unlikely to have affected our results as we used these data only to refine the height predictions from the GRS, and there is no reason to believe the GRS-height relationship would be different for women with and without height data. Further strengths of our analysis include the large number of SNPs and power to detect modest differences.

Despite potential limitations of conventional observational studies, our MR-estimate is almost identical to previously reported associations, suggesting previous estimates were not appreciably biased. The World Cancer Research Fund/American Institute for Cancer Research meta-analysis of 24 prospective studies, and a study pooling 47 prospective/case–control studies, both reported a significant 7–8% increase in risk (combining invasive/borderline cancers) per 5 cm height increase.3, 4 The lack of association seen in the OCAC conventional height analysis reflects the greater potential for bias in case–control studies and illustrates the value of MR in overcoming these biases. Few previous studies have examined borderline cancers separately, a strength of our analysis. Previous observational studies have not reported consistent patterns by histologic subtype2, 4, 23; our secondary analyses were under-powered to resolve this question.

Using MR, we have established that the previously observed association between height and ovarian cancer risk is unlikely to have been explained by bias, and that genetic factors influencing height play roles in ovarian cancer development. Height could therefore be used, with other risk factors, to identify women at elevated risk. Further research should continue to explore mechanisms underpinning this association.