Data and population

We used data from the 1958 National Child Development Study (NCDS), one of the national British birth cohorts, which includes all people born during one week in 1958 (n = 18,555). Data on life conditions and experiences, about family, education, work, and health, were collected in twelve waves from birth to age 62 by the Centre for Longitudinal Studies. The NCDS has been described in detail elsewhere [20]. Detailed review of the ethical practices throughout NCDS is available at [https://cls.ucl.ac.uk/wp-content/uploads/2017/07/NCDS-Ethical-review-an-Consent-2014.pdf].

For this study, we used data collected during the first (1958, birth, N = 17,638), fourth (1981, 23 years, N = 12,357), fifth (1991, 33 years, N = 16,174) and biomedical waves (2002–2004, 44–45 years, N = 9,377). See the flow chart in Fig. 1. To reduce selection bias, we included all the living subjects at the time of the biomedical waves [21], when outcome variables had been collected (N = 17,272). Indeed, the total cohort is assumed to be representative of the generation, but the subjects are not missing at random at each wave. As a consequence, including only non-missing participants can leads to collider bias [22]. We therefore chose to include all living participants, to preserve the population structure, and imputed missing data. We however also performed a sensitivity analysis on participants who participated at the four used collection waves, involving more selection bias but fewer missing data (N = 7,021).

Fig. 1Gender concepts

In this study, we explored gender as (1) the level at which the social characteristics and behaviours of an individual fits the stereotypes/ norms of masculinity or femininity (gender performance); and (2) the fact that, or the process by which, social characteristics and behaviours are differently distributed according to the binary sex at birth (gender pressure) [9]. These concepts can be operationalized in three ways within the epistemological and methodological framework of epidemiology, as detailed elsewhere [9] (see Fig. 2):

(a)

Gender as an individual characteristic: gender refers to how an individual performs their gender, according to the norms of gender in the population in which they are socially active. This corresponds to the concept of gender performance. E.g., an individual can be said to have a “feminine” gender if they have mainly social characteristics considered as feminine, like having more care activities (childcare, looking after older people, nursing). This conceptualization implies understanding gender as an individually defined variable.

(b)

Gender as an effect of sex on socio-behavioural characteristics: gender refers to the fact that socio-behavioural characteristics are differently distributed according to the sex at birth. This corresponds to gender as a gender pressure [9]. E.g., gender refers to the systemic process by which women are more likely to engage in caregiving activities than men in a given population. This conceptualization involves understanding gender as an effect of sex on one or more social-behavioural characteristics.

(c)

Gender as an interaction between sex and the early-life social environment: if sex differences are not stable between social groups, we can explain these sex differences by gender mechanisms [9]. This third way of thinking about gender also refers to gender as a gender pressure as in conceptualization (b), but it takes into account the fact that the systemic process of gender varies, in its form or intensity, across social groups. E.g., if care activities are more often found in women in population A but in men in population B, we can conclude that the fact that care attitudes are associated with a sex is not “natural” but linked to systemic gender mechanisms. This is symmetrically equivalent to the fact that a given social environment does not have the same effect, through socialisation, on an individual, depending on their sex attributed at birth [9]. This conceptualisation involves understanding gender as a difference in effects, i.e. an interaction. In theory, this effect can concern the whole social environment, at any age, but to simplify the approach, we here considered only the early-life social environment, which is a priori independent of the sex at birth. This third conceptualisation can be seen as part of an intersectional approach to gender [23].

Fig. 2

Conceptual graphs for three conceptualizations of gender

Here, we did not address gender as an experience of self (gender identity) or as a given kind of psyche (gender personality).

The conceptualizations of gender imply specific analytical strategies to meet the objective of identifying gender mechanisms to explain sex differences in biomarkers (See “Analyses” section). We refer to the corresponding strategy by the letter for each corresponding conceptualization (a, b, c).

MeasuresOutcomes: biomarkers

When individuals were about 45 years old, biomedical data were collected through a survey and a home-based clinical assessment (blood, saliva samples and anthropometric measurements) [24]. We explored several of these biomarkers, representing the four most frequent systems used to construct the score of allostatic load [25]: systolic blood pressure (SBP) for the cardiovascular system; triglycerides, low density lipoprotein (LDL) cholesterol and haemoglobin A1c (HbA1c) for the metabolic system; C-reactive protein (CRP) for the inflammatory system; and cortisol for the neuroendocrine system. When the distribution was too asymmetric, the variables were log transformed (triglycerides, CRP and cortisol).

Exposures: sex and early social environment

As our main exposure measure, we used sex attributed at birth. Relative to the effect of sex on the outcome, the early-life social environment was a competitive exposure and a confounder of the mediator-outcome relationship in strategies (a) and (b). In strategy (c), the early-life social environment was a modifier of the effect of sex on the outcome (see Fig. 2). We used two variables to characterise the early-life social environment at the time of the cohort member’s birth: educational level of the cohort member’s mother (school leaving age of 15 versus stayed at school beyond age 15, i.e., “O level”) and their other parent’s (or mother’s partner) social class (manual or non-manual social class). We used these variables to define two groups: the deprived group if the mother had a short education and the other parent a manual social class, and the non-deprived group in all other cases.

Mediators: gender scores and socio-behavioural characteristics

In strategies (a) and (b), we explore gender processes through mediator(s) (see Fig. 2). In strategy (a) gender was conceptualised as an individual characteristic which was measured by a gender score based on socio-behavioural variables. The mediator was this score. In strategy (b), gender was conceptualised as a sex effect on socio-behavioural characteristics. In this approach, mediators were the same socio-behavioural variables than those used to compute the gender score but kept separated.

Choice of socio-behavioural characteristics

In each strategy, we used the same set of socio-behavioural variables, either to compute the score (a) or separately (b). We consider that gender processes do not simply impact aspects of life classically described as gendered (like domestic load, type of occupation, etc.) but diffuses into all socio-behavioural dimensions. We had therefore chosen to use a larger set of socio-behavioural individual characteristics for which we assumed a priori to be distributed differently according to sex, because of the gender processes, and which may a priori have an impact on biomarkers and health.

Gender processes are multi-level, multidimensional and highly diffuse [7], so much so that we could say that every aspect of a human's life is impacted by the gender norms of the society in which they live. It impacts their identity (“how an individual sees themselves”), their roles (behaviours, experiences, expectations), their relations (“how individuals interact with and are treated by others”) and their relative power in different institutions (“political, educational, religious, media, medical, cultural and social institutions”) [7]. It seems impossible to capture all the aspects of life impacted by gender phenomenon [9]. We therefore sought to characterize, as broadly as possible, various dimensions of social life from the data available in the cohort.

Individual social characteristics that have an impact on health are equally multi-level, multi-dimensional and diffuse. They can be classified in two types: behaviours and social advantages/ disadvantages, i.e., resources which give the individual a varying degree of control and resilience over their environment, their experiences, and their life course. According to Bourdieu, these advantages/ disadvantages can be categorised into three dimensions: cultural capital, economic capital and social capital [26].

We therefore used several variables to characterise the three dimensions of capital and the behaviours. We made the a priori hypothesis that the distributions of these variables may vary according to sex, due to the gender processes:

Cultural capital refers to knowledge, skills and integrated attitudes that will influence the way an individual sees the world, thinks, behaves, lives and acts [26]. In this study, we have represented these resources through five measures: “educational level at 23” (more or less than O level), “literacy at 23” (declared difficulties or not), “numeracy at 23” (declared difficulties or not), “often reads books at 23” (at least once a month) and “driver’s license at 33”. The driver's license is not a classic criterion of cultural capital, but we considered that it corresponded to the definition of a skill giving an increased control on the environment, an increased capacity to act in social life.

Economic capital refers to the material and financial resources of individuals and the means to produce them [26]. In this study, we measured economic reserves with “personal savings at 23” and qualified the resources to produce them through “paid work at 33” and “social class at 33” (manual or non-manual).

Social capital refers to an individual's social network, its size, value, and the degree of usefulness of these relationships [26]. In this study, we used the frequency of friends’ visits at 23 (more or less than once a week) as a marker of social support and “being religious at 23” as a marker of belonging to a community. The three variables “child(ren) at 23”, “married at 23”, and “doing laundry at 33” are markers of an affective and family support, but also of the domestic burden, counterpart of this resource.

To characterize behaviours, we chose behavioural variables which can be considered as protective or a risk for health: “smoking (≥ 1 cigarette/day)” at 23 and 33, “everyday alcohol drinking” at 23 and 33, “frequent fried food” at 33 (more than once a week), “sport” at 23 and 33 (at least once a month). Risk taking being "a value and reality associated with masculinity" and which penalizes men by "causing them to perish" [27], we also used a proxy of risk-taking behaviours with the variable “have attended hospital or casualty department for any kind of accident or assault between 23 and 33”.

Impact of the choice of variables

The choice of variables to explore the gender phenomenon is largely based on their availability in the database and would have varied widely if another cohort had been used. To explore the impact of the variables availability on results, we constituted three different sets of variables and performed the analyses, for the strategies (a) and (b), with these different sets, "as if", in each case, we had only these variables to characterize the same phenomenon of interest. The sets were:

Complete set: all the listed-above variables had been used. It was the main analysis.

Behavioural set: only the behavioural variables had been used, as if only these variables were available.

Small set: only 4 social characteristics (educational level at 23, social class at 33, frequency of friends’ visits at 23, and marital status at 23) and 4 variables for behaviours (sport, diet, smoking and alcohol at 33), as if only these variables were available.

Scores computation

In strategy (a), we wanted to capture gender as an individual characteristic, corresponding to the level at which the social characteristics and behaviours of an individual meets the standards of masculinity or femininity. To measure the gender corresponding to this definition, we used the "gender diagnosis" method [28,29,30]. The gender score produced by this method corresponds to a measure of the level at which an individual complies with a set of elements constituting femininity or masculinity in a given population, place and time, i.e., as the probability of being "predicted male or female" from social dimensions [9].

To construct the score, we modelled sex at birth by socio-behavioural characteristics using logistic regression, for each set of variables defined above. The gender score corresponded to the predicted probability by the model of sex at birth, a continuum from 0 “predict female by their socio-behavioural characteristics”, proxy of “gendered in a feminine way”, to 1 “predict male by their socio-behavioural characteristics”, proxy of “gendered in a masculine way”.

Analyses

All the analyses were performed with R release 4.1.3, with the Tidyverse packages. To deal with missing data, we performed a single stochastic imputation using the MICE package in R [31] on each of 1,000 bootstrapped databases, also used to computed 95% confidence intervals [32, 33].

Descriptive analyses

We first described the exposures, mediators, and outcomes in excluded participants (dead at the biomedical waves) and included subjects (living at the biomedical wave), with number of missing data, and number and percentage of the variable categories, or mean and standard for quantitative variables. We also described these variables in the imputed bootstrapped databases, with mean of percentages or means and confidence intervals (2.5 and 97.5 percentiles), computed on 1,000 bootstrapped imputed datasets. The results of this description are given in Additional file 1. We then described socio-behavioural characteristics and gender scores by sex, with mean value and confidence intervals (2.5 and 97.5 percentiles) of percentages or means, computed on 1,000 bootstrapped imputed datasets.

Causal analyses

The mean value of each biomarker at 44–45 years of age was estimated under several counterfactual scenarios that differed by exposure and/ or mediator assignment. The notation \(}\left(_\right)\) represents the expected potential outcome (mean of Y) in the counterfactual scenario in which the exposure is set to \(A = a\). Under the randomization assumption (no residual confounding) and the consistency assumption (effect of \(A\) is the same whether observed or given by intervention), \(}(_\)) was estimated using g-computation. Linear regressions were used to estimate conditional expectations of the outcome, denoted \(\overline\left(A,L\right)= }(Y|A,L)\), with \(L\), the confounders. From the estimated \(\overline\left(A,L\right)\) functions, we predicted the value of biomarker \(Y\) for each member \(i\) under the counterfactual scenarios. Target causal parameters (estimands), described below, were defined in an additive scale as the difference between the mean of potential outcomes in two scenarios.

Total effect of sex

We first aimed to measure the size of sex-differences in biomarker. The estimands were the total effect \(TE\) of sex at birth \(S\) on each biomarker \(Y\), defined as the difference between the mean outcome had all the population been born male and the mean outcome had all the population been born female, denoted:

In the \(\overline\left(S,E\right)\) functions used to estimate the potential outcomes, \(E\) contained the early-life social environment variable. We included an interaction term between \(S\) and \(E\). The outcomes were first used in their original scale, the \(TE\) is therefore expressed in these units of measure, e.g., in mmHg for systolic blood pressure. They were then standardized as: \(z=\frac\) where \(\mu\) and \(\sigma\) are the mean and standard deviation of the outcome in each imputed bootstrapped data set. The \(TE\) is therefore expressed in standard deviation, e.g., a total effect of 1 corresponds to a mean difference of 1 standard deviation between men and women.

Strategy (a): mediation by a gender score

The principal objective of the study was to identify gender mechanisms that explain sex-differences in biomarkers. With gender conceptualized as an individual characteristic (a), the estimand corresponded to the proportion of the total effect of sex on biomarkers which disappear when all the individuals are gendered in the same way, i.e., the eliminated proportion by gender score \(G\), denoted \(^\). The eliminated proportion \(^\) was measured as the difference between the total effect of sex and the remaining effect of sex when all the population is gendered in the same way (gender score fixed at 0.5), divided by the total effect of sex [34]. The remaining effect of sex when all the population is gendered in the same way corresponds to the controlled direct effect \(^\), which was defined here as the difference between the mean outcome had all the population been born male and the mediator (gender score) set at a given value (here 0.5) and the mean outcome had all the population been born female and the mediator (gender score) set at the same value (0.5) [35]. In the \(\overline\left(S,E\right)\) and \(\overline\left(S,G,E\right)\) functions used to estimate the potential outcomes, \(E\) contained the early-life social environment variable (mediator-outcome confounder). We included an interaction term between \(S\) and \(E\), but not with \(G\). Finally, we had:

\(^= }[_-_]\) and \(^=\frac^}\)

Strategy (b): mediation by social characteristics

With gender conceptualized as an effect of sex on socio-behavioural characteristics (b), the estimand corresponded to the proportion of the total effect of sex on biomarkers which disappears when all the individuals have the same socio-behavioural characteristics, i.e., the eliminated proportion by the set of socio-behavioural characteristics \(\Sigma\), denoted \(^\). The eliminated proportion \(^\) was measured as the difference between the total effect of sex and the remaining effect of sex when all the population has the same socio-behavioural characteristics (see Additional file 1 for detailed fixed values \(^\) for each variables), divided by the total effect of sex [34]. The remaining effect when all the population has the same socio-behavioural characteristics corresponds to the controlled direct effect \(^\), which was defined here as the difference between the mean outcome had all the population been born male and all the socio-behavioural characteristics set at a given value (see Additional file 1) and the mean outcome had all the population been born female and all the socio-behavioural characteristics set at the same value [35]. In the \(\overline\left(S,E\right)\) and \(\overline\left(S,\Sigma ,E\right)\) functions used to estimate the potential outcomes, \(E\) contained the early-life social environment variable. We included an interaction term between \(S\) and \(E\), but not with \(\Sigma\). We therefore had:

\(^= }[_^}-_^}]\) and \(^=\frac^}\)

Strategy (c): considering an interaction between sex and the early-life social environment

With gender conceptualized as an interaction between the sex at birth and the social environment (c), the estimand corresponded to the proportion of the total effect of sex on biomarkers which disappears when all the individuals have a non-gendered social environment. We considered that an observed non-gendered social environment was not realistic, so we rather considered a “less-gendered environment”. Here, to define the social environment, we considered only the early-life social environment \(E\). We made the a priori hypothesis that the non-deprived group was less gendered than the deprived group. Therefore, the eliminated proportion \(^}\) was measured as the difference between the total effect of sex and the remaining effect of sex when all the population is exposed to a non-deprived early-life social environment \(E = 0\), divided by the total effect of sex. The remaining effect when all the population is exposed to a non-deprived early-life social environment corresponds to the total effect of sex when \(E=0\), denoted \(^\) and defined as the difference between the mean outcome had all the population been born male and the early-life social environment been non-deprived, and the mean outcome had all the population been born female and the early-life social environment been non-deprived. The model \(\overline\left(S,E\right)\) used to estimate potential outcomes under these scenarios considered an interaction term between \(S\) and \(E\). We finally had:

\(^ = }[_-_]\) and \(^}=\frac^ }\)

Complementary analyses regarding interactions

We also present in the results section a more detailed description of the interaction effects, with mean value and confidence intervals (2.5 and 97.5 percentiles) of means, computed on 1,000 bootstrapped imputed datasets, for each biomarker in each category of sex and early-life social environment. We also computed the total effect of sex in each stratum of early social environment and total effect of early social environment in each stratum of sex. We finally estimated the interaction effect \(IE\) of sex and early social environment, defined as: