Elevated risk of stillbirth in males: systematic review and meta-analysis of more than 30 million births

We searched ISI Web of Science and Medline for studies providing sex-specific rates of stillbirth and published between 1990 and 2012 inclusive. The techniques specified by Stroup et al. [12] for conducting and reporting meta-analysis of observational studies in epidemiology (MOOSE) were followed Additional file 1. The medical subject headings (MeSH) and search terms used were: “stillbirth”, “still-birth”, “fetal death”, “foetal death”, “pregnancy loss”, “sex”, “fetal sex”, “foetal sex”, and “gender”. The terms were combined with the Boolean operators “OR” or “AND”. The reference lists and cross-references of eligible studies were also searched, as were the bibliographies of recent reviews. Google and Google Scholar were used to search the `grey’ literature (information outside scientific journals) for relevant information. Studies not published in English and randomised trials testing interventions to reduce stillbirth and/or perinatal mortality were excluded. We first checked the abstracts of the studies identified on the basis of having relevant titles. The full text of studies regarded as potentially eligible was then assessed to determine whether they had provided data, stratified by sex, on both the numbers of stillbirths and the numbers of live births/total births. If the relevant numbers were not directly provided, the paper was included if the data could be extracted, using standard methodology, from other information provided (e.g., where un-adjusted odds ratios and total numbers of births were provided). The methodological quality of the studies was also assessed: those with incomplete definitions of, or missing data on, the population, study design, exposure, or outcome variables were excluded. Two reviewers (DM and FM) independently reviewed the methodological quality of the studies and differences were resolved by consensus and discussion.

The definition of stillbirth has changed over time, and there are also differences between countries despite the availability of international guidelines [13]. The research therefore included studies which used a variety of definitions of stillbirth, for example ≥28 weeks’ gestation or birth weight >500 g; or ≥20 weeks’ gestation or birth weight >400 g. Table 1 provides a summary of the studies and the definitions used by each.

Data were included in the meta-analysis if they came from studies that fulfilled all of the following criteria:

Original observational epidemiological study (including community-based cross-sectional studies; case-control studies and cohort studies)

Randomised controlled trials and other assessments of interventions designed to reduce stillbirth or neonatal death rates were excluded.

The national data archive of population Vital Statistics in England and Wales was used to derive information on the numbers of live births and stillbirths by sex of the fetus for the years 1990 to 2010 (see [34]; theme population). This archive records all registered live and stillbirths.

After reading each relevant article that appeared to meet the inclusion criteria, data on pregnancy outcome by sex was abstracted into 2 × 2 tables. In addition, the following data were noted for each study: year of study, publication, sample method, sample size, study design (cohort, case-control, population-based cross-sectional summary reports of vital statistics), definition of stillbirth, and any adjustments made for potential confounding variables in the original article, the adjusted RRs with confidence interval (CI), and the variables included in the final models. The data abstraction was done by PM and checked by FM. Decisions on whether to exclude studies, all of which were based on the absence of necessary data defined in the inclusion criteria, were reached following discussion between the authors.

RRs were used to quantify the relationship between sex and the prevalence of stillbirth. We combined the data using DerSimonian and Laird random effects (inverse variance) models to estimate the pooled RRs and associated 95% CIs. These models are preferred because we suspected a priori that differences between populations (particularly low-/medium- vs. high-income countries) and study designs could be important [35]. The random effects model assumes that all studies are estimating different effects resulting from variations in factors, such as study population, and samples variation within and between studies. As a result, it generally produces wider confidence intervals than fixed effect models, the pooled results of which we also present for comparative purposes. Population-attributable risk estimating the percentage of the stillbirth in the population due to male sex was calculated [36].

Funnel plots were assessed for symmetry to evaluate possible publication bias. Statistical heterogeneity across the studies was explored through subgroup analyses of the following study-level covariates on the RRs associated with sex: study setting, study population, and study design. These assessments were supplemented with the inspection of I² statistics which estimates the percentage of outcome variability that can be attributed to heterogeneity across studies [37]. An I² value of 0% denotes no observed heterogeneity, whereas 25% is “low”, 50% is “moderate”, and 75% is “high” heterogeneity. Finally, sensitivity analyses were performed to explore the effects of heterogeneity between study types. The potential influence of each individual study on the overall summary estimates was assessed by re-running the meta-analysis omitting one study at a time. Analyses based on rates adjusted for covariates increase the likelihood that any associations represent independent relationships between the exposure variable (fetal sex) and outcome (stillbirth). We therefore assessed if the effect sizes would have been materially altered by using adjusted rather than crude data, and provide the pooled adjusted RRs for comparative purposes where the data permit.

The analyses were performed using statistical software R-version 2.15.2 [38]. Meta analyses were conducted with the package Meta-version 2.0-2 [39], and attributable fractions were calculated using EpiR version 0.9-48 [40].

The literature search identified 795 unique citations. After detailed review, 21 of these studies met the inclusion criteria (Figure 1).

The characteristics of the studies are summarized in Table 1. All the published studies reported data from a single study population, with the exception of Xu et al. [16], who reported data from three different study populations, and Gadow et al. [15], who reported data from 11 Latin American countries. Each of these countries was treated as a separate population in the analysis. With the original dataset from England and Wales added, 34 populations were available for the meta-analysis. Only two of the populations were studied using a case-control design. Given that stillbirth is a relatively rare outcome, we interpret the various estimates of risk (i.e., odds ratio, RR) as being approximately equivalent, and here report RRs.

The risk of stillbirth was greater for male than female fetuses (Figure 2): the crude mean rate (stillbirths/1,000 total births) was 6.19 for males and 5.71 for females in the study cohorts. There was significant heterogeneity among studies (I² = 71.9% (95 CI: 60.4%-80.0%)), but the pooled risk estimates were similar whether we used a random effects model (RR =1.10, 95% CI: 1.07-1.13) or a fixed effects model (RR =1.09, 95% CI: 1.08-1.10). The attributable fraction due to sex in the population was 4.17% (95% CI: 3.70-4.63), and the attributable fraction among male fetuses was 7.82% (95% CI: 6.97-8.66).

The contour-enhanced funnel plot for this meta-analysis is shown in Figure 3. There is some asymmetry indicating the possibility of missing studies. However, the asymmetry is not marked, and inspection of the graph indicates that the apparently missing studies largely fall in the area of statistical significance (shaded area) rather than non-significance (white area). True heterogeneity, rather than a failure to publish non-significant results, is therefore the most plausible explanation of the observed patterns.

There was no evidence for a difference in RRs according to the income classification of the country (Q =0.37, df =1, P =0.543, Additional file 2). Since prenatal sex selection could bias the data on sex-specific stillbirth rates, we conducted subgroup analyses of countries identified by the World Health Organisation [41] as having biased prenatal sex selection (n =3) and those that do not (n =31) (Figure 4). This reduced the observed heterogeneity. For the former, the overall risk of stillbirth was higher (mean rate of 8.5 stillbirths per 1,000 births in affected countries compared with 6.0 elsewhere), and females had a higher relative risk (RR =0.64, 95% CI: 0.50-0.82). This switch in the direction of sex-linked risk appears to be based on higher than expected death rates for females rather than reduced fatalities in males. Given the live-birth rates in the study cohorts from India and China, if stillbirth patterns had followed those seen in other countries we would expect approximately 186 cases, 89 of them female (95% CI: 87.0-91.5) (based on the random effects model), whereas 116 male and 152 female stillbirths were actually recorded. There were only two studies from countries without known significant prenatal sex selection, where the risk of stillbirth was higher in females than males [11],[14]. In each case, overall stillbirth rates were unusually low, suggesting that prenatal sex selection is unlikely to be responsible for the observed patterns.

For the countries with no known sex-biased selection, the overall heterogeneity estimate (I²) was 69.3% (95% CI: 55.7%-78.8%). We therefore conducted sensitivity analyses, based on random-effects models, to explore the reasons for the remaining heterogeneity (Additional file 3). First, we explored the effect of the definition of stillbirth, by grouping studies according to the gestational age cut-off. Two thirds (20/31) of the studies used a definition that included relatively early gestational ages (≥20 weeks). Within this subgroup, heterogeneity was reduced (I² = 39.8%: 95% CI: 0%-64.7%) and the stillbirth RR for males remained similar to the pooled estimate derived from all studies. Where a gestational age cut-off ≥28 weeks was used, the RR of stillbirth remained similar (RR =1.12, 95% CI: 1.00-1.26 compared with RR =1.11, 95% CI: 1.08-1.13), but heterogeneity remained high (I² = 85.4%, 95% CI: 74.2-91.8). Hence, for gestational age ≥28 weeks, we explored the effect of population type (hospital- versus population-based) and study design (case-control vs. cohort/population cross sectional survey). In all subgroups, the risk was elevated for male fetuses, and there was no difference in the effect size between the subgroups. The subgrouping had no material effect on heterogeneity (Additional file 3).

We explored whether the findings were consistent for the studies (n =8) which reported adjusted RRs (Additional file 4) (note that covariates varied widely between the studies). The pooled adjusted RRs had similar effect size as the overall result, but heterogeneity was lower (Additional file 3).

Sensitivity analyses which left out individual studies in turn found no evidence that any study exerted particular influence on the pooled RR estimate (Figure 5). This suggests that findings of the meta-analysis are robust.