Choice of instruments for MR analyses

We chose MR instruments for age at menarche based on findings from the large genome-wide association (GWA) meta-analysis by Day et al. [16], which included 329,345 women of European ancestry from the ReproGen consortium (40 studies, N = 179,117), 23andMe (N = 76,831), and UK Biobank (N = 73,397). In all, 389 independent single nucleotide polymorphisms (SNPs) were identified at a genome-wide significant level (p < 5 × 10⁻⁸), explaining overall 7.4% of the population variance in age at menarche. Of the reported 389 SNPs, 372 were available (or had a proxy, linkage disequilibrium r² > 0.7) in UK Biobank.

The instrument strength of each SNP for both early and late menarche was assessed using the F statistic, a function of the magnitude and precision of the SNP’s effect on menarche [17]. To avoid bias associated with the use of weak instruments, we excluded SNPs with a low F statistic in UK Biobank, using the common threshold of 10 [18]. This left us with 206 and 151 SNPs (254 in total) as instruments for early and late menarche, respectively; F statistic values varied from 10.1 to 382.5 for early menarche and from 10.0 to 150.0 for late menarche (S1 Table). All SNPs had good imputation quality (info score > 0.8).

In the GWA meta-analysis by Day et al. [16], 127 of the 389 age-at-menarche SNPs were also associated in the same direction with age at voice breaking at nominal significance level; of these, only 40 were strong instruments (F statistic > 10) for either early or late puberty in males. In order to increase our number of instruments, we searched the literature for additional genetic studies on age at voice breaking, using the NHGRI GWAS Catalog [19] and HuGE Navigator [20], as well as crosschecking of references. We identified 3 other GWA studies [21–23], 2 of which added new independent SNPs, the studies by Day et al. [22] and by Pickrell et al. [23] (S2 Table). After combining all SNPs and excluding weak instruments, 24 and 37 SNPs (46 in total) remained for the MR analyses of early and late voice breaking, respectively. Their F statistic values varied from 10.2 to 52.4 for early voice breaking and from 10.8 to 101.6 for late voice breaking (S2 Table). All SNPs had good imputation quality (info score > 0.9). All 46 SNPs were independent, except for 3 pairs of correlated SNPs (linkage disequilibrium r² of 0.3, 0.6, and 0.6). As a post hoc analysis to investigate the impact of the correlation between these SNPs on our MR results, we used the method by Burgess et al., which accounts for the correlation between instruments [24].

Details on genotyping and imputation methods and quality control procedures in UK Biobank are available elsewhere [25].

MR methods

In 1-sample MR investigations (where the genetic associations with both risk factor and outcome are estimated within the same study), the standard MR approach is the 2-stage least squares (2SLS) method: the risk factor is regressed jointly on all SNPs, and the outcome is regressed on the genetically predicted values of the risk factor from the first regression [26]. As this method does not account for pleiotropy (where SNPs chosen as instruments for the risk factor affect the outcome through additional independent pathways [14]) and we cannot rule out the possibility of pleiotropic effects for some of our SNPs, we used methods developed for 2-sample MR. Here estimates of the genetic associations with risk factor and outcome are obtained separately for each SNP. If all SNPs are valid instruments (e.g., no pleiotropy) and linear models with no interactions faithfully describe the SNP–risk factor and risk factor–outcome relationships [27], the individual MR estimates of causal effect will vary only by chance, with no between-instrument heterogeneity [28]. For each SNP, the association with early or late menarche/voice breaking (G–X, expressed as log odds ratio [OR]) and the association with asthma (G–Y, log OR) were estimated using multiple logistic regression adjusted for assessment centre, first 10 ancestry principal components, and genotyping array. It should be noted that with binary variables the linearity assumption is strictly violated [29], but in practice the magnitude of the bias has been shown to be negligible [30].

An MR estimate was obtained for each SNP using a Wald estimator (ratio of G–Y over G–X), with standard error derived using the Delta method [31], and individual MR estimates were pooled using inverse-variance weighted (IVW) fixed-effect meta-analysis [32]. The pooled MR estimate is asymptotically equal to the MR estimate obtained with the 2SLS method [32], which we also performed as a sensitivity analysis. Since the IVW fixed-effect meta-analysis method assumes no pleiotropy, we performed secondary analyses using approaches based on different assumptions about pleiotropy across SNPs: (1) IVW random-effect meta-analysis, assuming random pleiotropic effects (i.e., they cancel out) [33]; (2) MR-Egger regression with penalised weights, assuming directional pleiotropic effects [34]; and (3) weighted median analysis, assuming that at least 50% of the SNPs are valid instruments (but with no assumption about direction of pleiotropic effects) [35].

Some of the genetic variants associated with pubertal timing are also associated with BMI [16], and this may complicate the interpretation of the results given that BMI could either introduce pleiotropy (i.e., affecting asthma independently from age at puberty) or lie on the causal pathway. In secondary analyses we therefore excluded SNPs associated with BMI from our instruments. We identified SNPs associated with BMI, overweight, or obesity in previous GWA studies from PhenoScanner, a database of publicly available GWA findings (http://www.phenoscanner.medschl.cam.ac.uk/phenoscanner) [36], using a significance threshold corrected for multiple testing, p < 2.0 × 10⁻⁴ (0.05/254) for menarche SNPs and p < 1.1 × 10⁻³ (0.05/46) for voice breaking SNPs.

To provide further support for the absence of pleiotropy, we used pre-pubertal asthma as a “negative control” outcome [37]. If our instruments affect asthma solely through age at menarche/voice breaking, we expect to find no effect on asthma developed before puberty.

There was no formal pre-specified protocol for this study; all the analyses described above were decided a priori, and we performed additional post hoc analyses suggested by the reviewers. In particular, sensitivity analyses were carried out to assess the robustness of the results to the use of different AAO cutoffs for post-pubertal asthma definition in males, and to the presence of some correlation between genetic variants. Subgroup analyses were also performed to investigate the possibility that the effect of age at puberty on asthma might be modified by whether or not one is overweight.

All analyses were performed using Stata 15 (StataCorp).

MR analyses

Detailed results from all MR analyses are reported in S3 and S4 Tables, for females and males, respectively. In females, we found evidence for a detrimental causal effect of early menarche, with an OR of 1.08 (95% CI 1.04 to 1.12; p = 8.7 × 10⁻⁵), and for a protective causal effect of late menarche, with an OR of 0.92 (95% CI 0.89 to 0.97; p = 3.4 × 10⁻⁴). This suggests a continuous protective effect of increasing age at puberty, and does not support observational evidence of a U-shaped relationship. When we repeated the analysis using age at menarche as a continuous variable (332 SNPs with F > 10 as instruments), we found an OR of 1.06 (95% CI 1.03 to 1.11; p = 1.6 × 10⁻⁴) per year decrease in age at menarche. In males, we found an OR of 1.07 (95% CI 1.00 to 1.16; p = 0.064) for early and 0.93 (95% CI 0.87 to 0.99; p = 0.029) for late voice breaking. The results remained the same in sensitivity analyses allowing for the correlation between SNPs, and became slightly stronger in sensitivity analyses using higher age cutoffs for asthma AAO to define post-pubertal asthma, specifically AAO >16 and >17 years (S4 Table). These findings in males are very similar to those in females, despite the much wider confidence intervals reflecting the lower power of the analyses (fewer instruments available for age at voice breaking and possibly larger measurement error). The use of a 1-sample 2SLS approach gave similar results in both females and males (S3 and S4 Tables).

We found only modest between-instrument heterogeneity in females (I² of 16% and 20% for early and late puberty, respectively; S3 Table) and little evidence of heterogeneity in males (I² of 2% and 0%, respectively; S4 Table). The results of the analyses controlling for pleiotropy were consistent with those of the main analyses (Fig 1), with differences in the precision of the estimates in line with expected differences in statistical power [33,35]. MR-Egger regression in males suggested stronger effects compared to the main analysis, particularly for early puberty (Fig 1), but the confidence intervals were wide as the analysis had limited power. This low power is due to the low number of SNPs and their relatively similar instrument strengths (MR-Egger regression works best when there is a large spread of strengths [38]).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Estimates across different MR methods of the effects of early and late puberty on asthma risk in females (menarche) and males (voice breaking).

In red: main analysis. IVW, inverse-variance weighted; MR, Mendelian randomisation.

https://doi.org/10.1371/journal.pmed.1002634.g001

Secondary analyses with exclusion of SNPs associated with BMI (14 and 12 excluded for early and late menarche; 5 and 7 excluded for early and late voice breaking) gave similar results to those of the main analyses for both females and males (S3 and S4 Tables).

As the percentages of females in the early and late menarche categories were substantially higher than the percentages of males in the corresponding categories (Table 1), we repeated the analyses using more stringent cutoffs for age at menarche. Using a cutoff of <11 instead of <12 years for early and >15 instead of >14 years for late menarche, the percentages were very similar to those in males, 4.6% and 5.8% for early and late menarche, respectively. The results were in line with those of the main analyses, with an OR of 1.06 (95% CI 1.02 to 1.09; p = 6.4 × 10⁻⁴) for early and 0.95 (95% CI 0.91 to 0.98; p = 0.002) for late menarche.

The MR analyses on pre-pubertal asthma, used as a negative control outcome, suggested no effect for either age at menarche or age at voice breaking, further reassuring against effects of our genetic instruments on asthma independent from pubertal timing.

Finally, we repeated the analyses separately by BMI group, with low/high BMI defined as below/above the median (<26/≥26 kg/m² in females; <27/≥27 kg/m² in males), to investigate possible effect modification due to whether or not participants are overweight. BMI data refer to the participants’ BMI at the time of recruitment into UK Biobank (Table 1). In females, these analyses (S3 Table) showed effect estimates for early and late menarche consistent with the main analyses in the high BMI group, and no evidence for effects of early or late menarche in the low BMI group, but there was no statistical evidence of interaction (p-values of 0.23 and 0.67 for tests of the difference in estimates between BMI groups for early and late menarche, respectively). In males, there was no evidence of effect modification (S4 Table), as expected given the very low statistical power.

Strengths and weaknesses

MR is a valuable tool to assess causality in epidemiology as it is not affected by the classical confounding factors or reverse causation typical of observational studies. However, this technique is at its most reliable and transparent in the absence of pleiotropy, i.e., in this case, where the genetic instruments do not have direct effects on asthma independent of age at puberty [14]. As the risk of pleiotropy is high in MR investigations where many SNPs of uncertain biological function are included, we thoroughly investigated its potential impact on our findings. Obesity was identified a priori as a potential source of pleiotropic effects given its genetic overlap with pubertal timing, and in fact around 7% of our instruments for age at puberty were also associated with BMI. In practice, what role BMI might play in our study is unclear, as it could induce “horizontal” pleiotropy by affecting asthma independently from puberty, but it could also lie on the causal pathway either preceding puberty (high childhood BMI → early puberty → high asthma risk) or mediating its effects (early puberty → high adult BMI → high asthma risk), which is referred to as “vertical pleiotropy” [58]. Only horizontal pleiotropy would affect the validity of our MR findings. When excluding SNPs associated with BMI, we found very similar results to those of the main analyses. Pleiotropy could act through mechanisms other than obesity. We therefore investigated the likely magnitude of pleiotropy by assessing the heterogeneity among causal estimates from the individual SNPs (no heterogeneity if all valid instruments [28]), and found evidence of only modest heterogeneity. This was supported by the consistency in the findings from 3 different methods to adjust for pleiotropy (IVW random-effect, weighted median, and MR-Egger regression) with those of the main analyses. Finally, we repeated our analyses using pre-pubertal asthma as a negative control outcome, and the null results provided further reassurance against pleiotropic mechanisms independent from pubertal timing.

We used MR methods developed to investigate and adjust for pleiotropy in 2-sample MR, where estimates of the association of each SNP with risk factor and outcome are derived from separate studies, and the 2 estimates are assumed independent. The behaviour of these methods has yet to be studied in the 1-sample context, where independence of the 2 estimates is violated. However, the consistency of our findings with those derived from the 1-sample 2SLS method provides reassurance.

Our study is based on self-reported information on doctor-diagnosed asthma, age at menarche, and age at voice breaking, which introduces measurement error and potentially recall bias. However, since it is unlikely that the error in recalling age at puberty would correlate with the reporting of asthma or asthma AAO, and vice versa, the measurement error is likely to be random and only result in reduced power. Voice breaking represents the drop in the pitch of speech due to changes in vocal chords and larynx cartilage in response to androgen exposure during late puberty [59]. Despite having been suggested as a good marker of pubertal timing in males when assessed prospectively [60], age at voice breaking is much harder to recall in adult cross-sectional studies compared with the more clear-cut milestone of menarche in females, and may be subject to greater measurement error. This is a general problem and partly explains the limited knowledge about risk factors and effects of early puberty in males, for which more evidence from longitudinal birth cohort studies is needed. Another limitation is that we based the choice of MR instruments for age at menarche and, partly, age at voice breaking on findings from the GWA meta-analysis by Day et al. [16], of which UK Biobank genetic data (first release only) represented about 20% of the total sample. However, any overestimation of the SNP–age at puberty associations resulting from this would have pulled our MR estimates towards the null, therefore leading to underestimation of the true causal effects rather than to false positive results. Finally, findings from observational studies might be biased by “cohort effects” if individuals from younger generations are at different risk of both early (or late) puberty and asthma compared with older generations. Despite the secular trends of decreasing age at puberty [45] and increasing prevalence of asthma [1], this type of confounding would not bias the results of our MR study, unless the secular trends also affect the genetic variants used as instruments for age at puberty, which is unlikely.

In conclusion, we provide evidence for a causal detrimental effect of early puberty on asthma. Further studies are needed to investigate the mechanisms underlying this effect, and the similarity of our findings in females and males suggests a possible role for common biological or psychological factors accompanying early pubertal development. Taken together with evidence for other detrimental effects of early puberty on health, our study emphasises the need to investigate and address causes of the secular shift towards earlier puberty observed worldwide.