Methods for structural sexism and population health research: Introducing a novel analytic framework to capture life-course and intersectional effects

Abstract

Accumulating evidence links structural sexism to gendered health inequities, yet methodological challenges have precluded comprehensive examinations into life-course and/or intersectional effects. To help address this gap, we introduce an analytic framework that uses sequential conditional mean models (SCMMs) to jointly account for longitudinal exposure trajectories and moderation by multiple dimensions of social identity/position, which we then apply to study how early life-course exposure to U.S. state-level structural sexism shapes mental health outcomes within and between gender groups. Data came from the Growing Up Today Study, a cohort of 16,875 children aged 9–14 years in 1996 who we followed through 2016. Using a composite index of relevant public policies and societal conditions (e.g., abortion bans, wage gaps), we assigned each U.S. state a year-specific structural sexism score and calculated participants’ cumulative exposure by averaging the scores associated with states they had lived in during the study period, weighted according to duration of time spent in each. We then fit a series of SCMMs to estimate overall and group-specific associations between cumulative exposure from baseline through a given study wave and subsequent depressive symptomology; we also fit models using simplified (i.e., non-cumulative) exposure variables for comparison purposes. Analyses revealed that cumulative exposure to structural sexism: (1) was associated with significantly increased odds of experiencing depressive symptoms by the subsequent wave; (2) disproportionately impacted multiply marginalized groups (e.g., sexual minority girls/women); and (3) was more strongly associated with depressive symptomology compared to static or point-in-time exposure operationalizations (e.g., exposure in a single year). Substantively, these findings suggest that long-term exposure to structural sexism may contribute to the inequitable social patterning of mental health among young people living in the U.S. More broadly, the proposed analytic framework represents a promising approach to examining the complex links between structural sexism and health across the life course and for diverse social groups.

1. Introduction

Gendered inequities in health remain glaring, with marginalized gender groups (including cisgender girls/women and transgender, nonbinary, and gender nonconforming people) continuing to experience considerably higher rates of adverse physical and mental health outcomes relative to groups holding more gender privilege (including cisgender and gender conforming boys/men).^1,2 Most studies aiming to identify the determinants of these patterns focus narrowly on factors operating at the individual- and/or interpersonal-level of influence (e.g., psychosocial stress, perceived discrimination).¹ However, an important new stream of research is beginning to shift the analytic lens “upstream” to consider the impact of more macro-level societal factors, including structural sexism – i.e., “systematic gender inequality in power and resources manifest in a given [patriarchal] gender system”.³ Informed by the structural racism and health literature,⁴ this work is leveraging geographic variation in relevant policies (e.g., abortion bans), institutional practices (e.g., wage gaps), and other area-based indicators (e.g., social norms) to reveal how the magnitude of a health inequity depends on the degree of structural sexism enacted in a given locale, in-turn providing crucial insights into what (and who) may ultimately drive the gendered social patterning of health.⁵

Despite these significant advancements, however, population health research on structural sexism is still in its infancy and key gaps in knowledge remain. First, the vast majority of prior studies have used cross-sectional data and methods,⁵ precluding assessments of causality or examinations into any long-term effects. And yet, leading models of health (in)equity such as fundamental cause and ecosocial theories suggest that exposure to structural forms of discrimination and disadvantage accumulate and compound over the life-course to shape population health patterns and drive inequities^6–8 – a hypothesis that is supported by a robust literature from life-course epidemiology.⁹ Moreover, many individuals in the U.S. live in multiple locations over the course of their lives, and places themselves change with respect to their relevant sociopolitical conditions over time, raising significant concerns about the validity of studies that use more static conceptualizations of structural sexism exposure.^10,11

Second, very little attention has been paid to intersectionality, a critical social theory from Black feminist scholarship and practice that focuses on the mutually constitutive nature of structural discrimination and systems of oppression (e.g., sexism, racism, classism, heterosexism, transphobia) and the ways in which they differentially and simultaneously shape the lived experience of individuals in relation to their unique constellation of held identities and social positions (e.g., gender, race/ethnicity, class, sexuality).^12–15 Consistent with this perspective, a growing literature is revealing how health status differs both between and within gendered groups, with particularly pronounced inequities experienced by cisgender girls/women and transgender, nonbinary, and gender nonconforming people who are also Black, Indigenous, or other people of color; working class/low-income; lesbian, gay, bisexual, or queer; and/or disabled.^16,17 Additionally, two recent studies have found that living in places characterized by high levels of structural sexism disproportionately – and sometimes exclusively – harms the health of these and other multiply marginalized populations,^18,19 suggesting that structural sexism may drive not only gendered health inequities, but also compound the effects of other forms of structural discrimination on the health of those who bear the brunt of multiple systems of oppression, underscoring the need to consider such complexity within this line of work.

It is thus becoming increasingly apparent that research on structural sexism and health would benefit from the use of methods that are more aligned with its guiding theoretical frameworks, although strategies for doing so have yet to be put forth. To begin to address this gap, this paper proposes a novel methodological approach to exploring the life-course and intersectional effects of structural sexism. We note that this approach is just one of many possible formulations – our primary intent is to advance a framework for re-conceptualizing structural sexism as a dynamic and context-dependent exposure, rather than to prescribe a definitive methodology for doing so. The rest of this paper is structured as follows. First, we briefly review methods that have been increasingly used within life-course epidemiology and quantitative intersectionality research to situate and motivate our approach. Next, we introduce sequential conditional mean models (SCMMs) as a promising analytic strategy for structural sexism research that we argue may better accommodate life-course and intersectional effects compared to more conventional techniques. Here we provide: (1) conceptual and technical overviews of SCMMs, (2) step-by-step instructions for their implementation, and (3) an applied case example for illustrative purposes. Lastly, we discuss important considerations regarding the use of SCMMs in “real-world” research on structural sexism (as well as other forms of structural discrimination and systems of oppression) and outline future directions for work in this area.

2. Quantitative methods for capturing life-course and intersectional effects

In the past several years, important methodological advancements have been made in the population health sciences to better accommodate life-course and intersectional perspectives within quantitative analyses. Many of these advancements are rooted in epidemiology’s potential outcomes framework,²⁰ an approach to causal inference based on counterfactuals (i.e., a set of hypothetical outcomes that would have been observed under alternative exposure conditions) and typically, robust statistical techniques. Regarding life-course perspectives, several methods have been proposed for examining how social and structural factors shape health across development and over time.⁹ One of the most rigorous of these methods is referred to as the “Structured Life-Course Modeling Approach”, or SLCMA, which involves fitting a series of regression-based models to describe and quantify the relationship between some longitudinally-assessed exposure and a subsequent health outcome;²¹ in such models, the operationalization of the exposure corresponds to a specific life-course hypothesis of interest (e.g., accumulation vs. critical/sensitive periods). While standard regression can be implemented within SLCMA, scholars have increasingly advocated for the use of more advanced techniques such as marginal structural models, structural nested models, and parametric g-computation to address the myriad challenges that arise in life-course analyses.^22,23 Of particular concern is time-varying confounding, which refers to confounding by factors that simultaneously confound and mediate an exposure-outcome relationship, given the fact that traditional model-based adjustment of such factors can actually lead to underestimating the association or effect of interest. Graetz et al. recently illustrated this dilemma in their study on structural racism and cardiovascular health – by employing parametric g-computation, they were able to appropriately account for the joint confounding and mediating influences of socioeconomic status, in-turn revealing the full magnitude of effect of a key driver of racialized health inequities.²⁴ Time-varying confounding and other similar biases are likely relevant to structural sexism research too, although the application of the appropriate methods to handle them within the field has been limited.

Several methods have also been recently advanced for quantitative intersectionality research that are adept at revealing within-group heterogeneity.²⁵ To-date, the majority of methods have been descriptive, in that they focus on documenting the prevalence or risk of a given outcome across subgroups defined by intersecting dimensions of social identity/position, and especially, identifying inequities experienced by subgroups that are multiply marginalized.²⁶ While an important and necessary first step, there has also been a growing emphasis on developing a comparative body of methodological literature for “analytic intersectionality” – i.e., research that seeks to identify the (potentially causal) drivers of these population health patterns.^25,27 A pertinent example comes from Bauer and Scheim,²⁸ who advanced a causal mediation approach to analytic intersectionality research that allows one to estimate the extent to which a given social or structural factor contributes to health inequities between intersectional subgroups, while allowing for the factor’s magnitude of effect to vary across these subgroups. Other examples include the work of Evans^29,30 and Homan et al.,¹⁹ both of whom have developed multilevel modeling methods for capturing high-dimensional interactions between multiple dimensions of social identity/position and their associated systems of power and oppression. Although there are pertinent limitations to these approaches (e.g., it is unclear the extent to which they can accommodate long-term/life-course exposure trajectories), they have laid important groundworks for thinking about how to capture the heterogeneous, intersectional effects of structural and social determinants of health within quantitative analyses.

To-date, these methodological advancements have largely occurred in isolation from one another and, with the exception of Homan et al.’s structural intersectionality approach,¹⁹ have yet to be applied to health research on structural sexism. In the current paper, we thus draw on the pioneering work from life-course epidemiology and quantitative intersectionality and put forth a cohesive analytic framework for studying how structural sexism shapes population health patterns over time and for diverse social groups. At its core, this framework involves re-conceptualizing structural sexism as a dynamic, longitudinal exposure and allowing for its effects on health to vary along axes of gender, race/ethnicity, class, sexuality, and other such categories. While there are several relevant methodological approaches that one could take depending on data availability and/or specific research interests, here we focus on sequential conditional mean models (SCMMs), a class of causal models that use propensity scores to estimate the effect of an exposure on a subsequent outcome.³¹ In terms of the methods mentioned above, SCMMs are most similar to marginal structural models in that they can accommodate time-varying exposures, outcomes, and confounders (thus making them well-suited to life-course analyses, which require such longitudinal data);^31,32 SCMMs also have several additional unique advantages:

1. First, as compared to marginal structural models, SCMMs are relatively simple. They are fit using more conventional statistical techniques (specifically, generalized estimating equations [GEE] vs. inverse probability weighting) and can be readily implemented using standard statistical software. This makes them accessible to a broader research audience.

2. Second, because SCMMs are estimated via GEE, they can handle and appropriately account for correlation arising from complex (e.g., non-hierarchical) clustering patterns.³³ Take, for example, a study examining how exposure to U.S. state-level structural sexism during adolescence shapes mental health in adulthood. Participants may have lived in multiple states throughout their teen years, all of which would likely exert some degree of a contextual effect and thus engender correlation that needs to be accounted for to obtain correct standard errors. While complex clustering patterns can also be addressed through the use of multiple-membership or cross-classified multilevel models, these can be computationally inefficient.

3. Third, in addition to time-varying exposures, outcomes, and confounders, SCMMs can also accommodate time-varying effect measure modification (i.e., moderation of the exposure-outcome relationship by a factor that changes over time). This is highly pertinent to life-course, intersectional scholarship on structural sexism – we might expect, for example, that the extent to which structural sexism shapes health over time varies by sexual orientation (e.g., due to the inherently linked systems of sexism/patriarchy and heterosexism). However, sexual orientation is highly fluid, especially during development, and these dynamics are likely directly relevant to the relationships of interest. SCMMs can investigate this via the inclusion of interaction terms between the time-varying moderating factor and the exposure. The same is not true for marginal structural models, which rely on weights to handle time-varying covariates.^31,32

For these reasons, we see SCMMs as a promising tool for advancing the literature on structural sexism and population health.

We note that the application of these methods to research on the social and structural determinants of health is not without criticism. While there are recognized advantages, including increased validity and resultant translational potential, scholars have also highlighted important drawbacks and incompatibilities.^34,35 For example, there are questions as to whether obtaining a so-called causal estimate is preferable to more descriptive evidence, as well as uncertainties over what constitutes a relevant exposure contrast that follows the potential outcomes framework’s “well-defined intervention” criterion (e.g., are we most interested in intervening on where people live over the course of their lives, or on the actual sociopolitical climate of these places?).³⁶ Moreover, considerable tensions exist between some of the framework’s methodological conventions and the central tenets of intersectionality and other theories of health (in)equity, with a pertinent example being the routine adjustment for any “confounding effects” of gender, race/ethnicity, and class (among other social identities/positions), which intersectional scholars would instead conceptualize as mutually constitutive, inseparable, and jointly impactful on health.^27,37 Ultimately, we see the application of the methods described above as useful in advancing research on structural sexism by providing a more systematic way of thinking through the exposure-outcome relations of interest, offering study designs and statistical techniques that can reveal previously-obscured population health patterns and health inequities, and helping to produce a body of knowledge that can inform equity-focused policies and interventions. That being said, we do not see these methods as being singularly useful, and understand other forms of knowledge/knowledge production (e.g., descriptive and qualitative research, lived experience) to be indispensable. We return to these issues in the Discussion.

2.1. Implementation

The implementation of these methods can follow these steps:

1. It will likely first be necessary to construct a multilevel and longitudinal dataset that includes time-updated measures of structural sexism, individual health status, and all relevant covariates by merging several sources together via geo-identifiers (e.g., ZIP codes). Then, a life-course structural sexism exposure variable can be defined by combining information on participants’ residential histories and the level of structural sexism present in a given locale at each relevant time point (e.g., study waves). For example, to encode cumulative exposure to U.S. state-level structural sexism, one could sum the structural sexism scores associated with each state a participant had lived in from baseline up to outcome assessment; it might also be advantageous to calculate the average of these scores rather than their sum (i.e., “duration-weighted cumulative exposure”) to account for the fact that people spend unequal times in different places.³⁸

2. Once the dataset is constructed and all key variables are defined and cleaned, basic descriptive analyses can be conducted. This could also include generating maps of participants’ geographic distribution and of the structural sexism exposure variable over the study period.

3. The propensity scores are estimated by fitting a model that predicts one’s structural sexism exposure at a given time point. For situations in which structural sexism is measured continuously, the propensity score model can take the form:

   $$E[X_t|{\stackrel{-}{X}}_{t-1},{\stackrel{-}{Y}}_{t-1},{\stackrel{-}{L}}_t,C]={\beta }_0+{\beta }_{{\stackrel{-}{X}}_{t-1}}{\stackrel{-}{X}}_{t-1}+{\beta }_{{\stackrel{-}{Y}}_{t-1}}{\stackrel{-}{Y}}_{t-1}+{\beta }_{{\stackrel{-}{L}}_t}{\stackrel{-}{L}}_t+{\beta }_{C}C$$

   where $X_t$ gives the structural sexism value associated with a participant’s place of residence at time $t,{\stackrel{-}{X}}_{t-1}$ gives the structural sexism values associated with a participant’s residential history through time $t-1,{\stackrel{-}{Y}}_{t-1}$ gives a participant’s outcome history through time $t-1,{\stackrel{-}{L}}_t$ is a vector of time-varying covariate history through time $t$, and $C$ is a vector of baseline/time-invariant covariates. From this model, the time-varying propensity scores for each participant can be derived by calculating their conditional probability density of exposure at time $t$. Importantly, the covariates included in this model should be guided by a theory- and evidence-informed directed acyclic graph (DAG) to help avoid over- or under-adjustment.³⁹ Additionally, so as to help account for selection bias stemming from item non-response and/or loss-to-follow-up, covariates predicting such missingness may be included in the propensity score model as well.³¹
4. Finally, the outcome model (i.e., the SCMM) is fit using GEE with an independent working correlation structure. Given a binary outcome, a logistic SCMM can take the form:

   $$logit(E[Y_t|{\stackrel{-}{X}}_{t-1},{\stackrel{-}{Y}}_{t-1},{\stackrel{-}{L}}_t,C])={\beta }_0+{\beta }_{{\stackrel{-}{X}}_{t-1}}{\stackrel{-}{X}}_{t-1}+{\beta }_{\widehat{{PS}_t}}\widehat{{PS}_t}$$

   where $Y_t$ gives one’s outcome at time $t,\widehat{{PS}_t}$ gives one’s estimated propensity score at time $t$, and ${\stackrel{-}{X}}_{t-1}$ is as defined above. Note that ${\stackrel{-}{Y}}_{t-1},{\stackrel{-}{L}}_t$, and $C$ can also be included in the SCMM for a doubly robust estimate. Depending on the specific research question, full-sample and/or stratified models can be fit. To assess effect measure modification of the exposure-outcome relationship (e.g., by some dimension of social identity/position $A$), interaction terms between this modifying factor and both the exposure and the propensity score should be included:

   $$logit(E[Y_t|{\stackrel{-}{X}}_{t-1},{\stackrel{-}{Y}}_{t-1},{\stackrel{-}{L}}_t,C])={\beta }_0+{\beta }_{{\stackrel{-}{X}}_{t-1}}{\stackrel{-}{X}}_{t-1}+{\beta }_{\widehat{{PS}_t}}\widehat{{PS}_t}+{{\beta }_{A}A+\beta }_{{\stackrel{-}{X}}_{t-1}\mathrm{*}A}{\stackrel{-}{X}}_{t-1}A+{\beta }_{\widehat{{PS}_t}\mathrm{*}A}\widehat{{PS}_t}A$$

The effect estimate of interest in the SCMM without interactions is ${\beta }_{{\stackrel{-}{X}}_{t-1}}$, which, when exponentiated, would give a conditional odds ratio; specifically, and in the current context, it would give an estimate of the change in odds of the outcome associated with a 1-unit increase in cumulative exposure to structural sexism, conditional on the covariates included in the model. In a SCMM that includes interactions, exponentiating ${\beta }_{{\stackrel{-}{X}}_{t-1}}$ would now give an estimate of the change in odds of the outcome associated with a 1-unit increase in cumulative exposure to structural sexism among those for whom $A$ equals 0 (again, conditional on covariates); exponentiating ${\beta }_{{\stackrel{-}{X}}_{t-1}\mathrm{*}A}$ would give a ratio of conditional odds ratios (specifically, the odds ratio for the exposure-outcome association among $A=0$ / the odds ratio for the exposure-outcome association among $A=1$) and quantifies the magnitude of moderation. We suggest transforming the resulting estimates into predicted probabilities for increased interpretability (e.g., the predicted probabilities of the outcome for each stratum of $A$ across a range of cumulative exposure values).

3. Applied example

To illustrate the implementation of these methods and to introduce some of the challenges researchers may face when doing so, we now provide an empirical demonstration that examines how early life-course exposure to structural sexism from childhood through young adulthood shapes subsequent mental health outcomes at the intersection of gender/sex,¹ sexual orientation, and race/ethnicity. After briefly summarizing the data and measures used, we outline our analytic plan, present the results, and discuss relevant interpretation strategies, with a particular focus on highlighting the insights gained from adopting a more theory-informed analytic approach.

3.1. Data

We drew on multiple sources of data for this example. All individual-level data came from the Growing Up Today Study (GUTS; N=16,882), a population-based cohort of U.S. youth aged 9–14 years at baseline in 1996 who have since been surveyed annually or biannually; further details, including procedures for requesting data access (https://gutsweb.org/collaborate-with-guts/), can be found elsewhere.⁴² Data on structural sexism came from the Institute for Women’s Policy Research (IWPR), which provides indicators of gender-based inequalities across four domains – political participation, employment and earnings, social autonomy, and reproductive rights – by state for most years from 1996 to 2015;⁴³ these indicators are publicly available (https://statusofwomendata.org/). Additional state-level data came from the U.S. Census (https://data.census.gov/). We merged these data sources together using year-specific geo-identifiers to create a multilevel, longitudinal dataset for our analyses that spanned from 1996 to 2016 and included 14 study waves (see eFigure 1 for a study design diagram).

To construct our analytic sample, we restricted inclusion to the 16,143 GUTS participants who reported living in the U.S. throughout the study period. Among those eligible, there was a considerable amount of missing data due to item non-response and/or loss-to-follow-up: 3,525 (22%) had missingness in baseline/time-invariant covariates, 295 (2%) had missingness in time-varying covariates, and 2,520 (16%) had outcome missingness. eTable 1 lists how many participants would have been censored due to such missingness by study wave; notably, the proportion censored tended to increase across the study period, although the pattern was non-monotonic (i.e., those who were censored at a given study wave could be observed again at a subsequent wave). While SCMMs can handle complex missing data patterns via propensity score adjustment,³¹ we used multiple imputation to retain the full eligible sample and mitigate selection bias. Further details are provided in Section 3.3.1, below.

3.2. Measures

We used the IWPR indicators to construct a time-varying composite index of U.S. state-level structural sexism following procedures outlined in our prior work.⁴⁴ Briefly, we obtained the state-specific values of each indicator for all available years (1996, 1998, 2000, 2002, 2004, 2015), using non-linear interpolation to define such values for all missing years within this time frame. Guided by a factor analysis of the indicators that revealed a single factor solution (see eTable 2 for details), we standardized their values within a given year and averaged them such that states were assigned a structural sexism score for each year that corresponded to a GUTS study wave (Figure 1), with higher scores representing higher levels of structural sexism. These scores were merged to GUTS during the data linkage process using study wave-updated participant ZIP codes; we used them to define three exposure variables: (1) static exposure, given by the structural sexism score associated with participants’ state of residence at the final study wave in the exposure window (i.e., wave 13 [2015]); (2) point-in-time exposure, given by the structural sexism score associated with participants’ state of residence at a given study wave, lagged one wave prior to outcome assessment; and (3) cumulative exposure, given by the duration-weighted mean of the structural sexism scores associated with participants’ state(s) of residence from baseline through one study wave prior to outcome assessment. Table 1 provides additional details regarding the operationalization and construction of these variables.

Figure 1.

Structural sexism score by U.S. state across the exposure window. Scores are standardized, such that a score of 0 represents the average level of structural sexism in the U.S. for a given year.

Table 1.

Construction of the static, point-in-time, and cumulative exposure variables

	Definition	Time-varying	Variable specification
Static exposure	The structural sexism score associated with a participant’s state of residence at a single study wave of potential theoretical or empirical interest, lagged one wave prior to outcome assessment ($w_{n-1}$)	No	${sexism}_{w_{n-1}}$= structural sexism score at $w_{n-1}$
Point-in-time exposure	The structural sexism score associated with a participant’s state of residence at one study wave prior to outcome assessment ($w_{n-1}$)	Yes	${sexism}_{w_{n-1}}$ = structural sexism score at $w_{n-1}$
Cumulative exposure	The duration-weighted mean of the structural sexism scores associated with a participant’s state(s) of residence from baseline ($w_1$) through one study wave prior to outcome assessment ($w_{n-1}$)	Yes	${sexism}_{\stackrel{-}{w_{n-1}}}$ = sum of the structural sexism scores from $w_1$ to $w_{n-1}$, divided by $n$

Our outcome, depressive symptomology, was assessed at GUTS study waves 4, 6, and 7 using the McKnight Risk Factor Survey⁴⁵ and at waves 9, 10, 11, 12, and 14 using the Center for Epidemiologic Studies 10-Item Depression Scale.⁴⁶ Each of these scales ask participants to report their frequency of experiencing relevant symptoms (e.g., negative affect, trouble concentrating) over the past week and have been shown to measure comparable underlying constructs.⁴⁵ Similar to previous studies that have used these two scales together,^e.g.,47 we calculated a standardized summary score from the relevant scale’s items for each participant at each wave and defined a binary indicator that contrasted those with high (score >1) versus low (score ≤1) depressive symptoms.

We defined gender/sex (girl/woman, boy/man), sexual orientation (completely heterosexual, mostly heterosexual, bisexual, gay/lesbian), and race/ethnicity (non-Hispanic Asian, non-Hispanic Black, Hispanic/Latine, non-Hispanic White, other/unlisted) as potential effect measure modifiers of the relationship between structural sexism and depressive symptomology; gender/sex and race/ethnicity were assessed at baseline only and thus were treated as time-invariant, whereas sexual orientation was assessed repeatedly across the study period and thus was able to be appropriately modeled as time-varying. Importantly, we acknowledge that the conflation of gender identity and sex-assigned-at-birth in GUTS precluded our ability to identify and explicate the experiences of transgender and non-binary people (who are disproportionately affected by both sexist discrimination and depression^48,49). The lack of time-updated measures of gender identity and race/ethnicity was likewise limiting, given their fluid and socially constructed nature. As such, we emphasize again that our analyses are primarily illustrative, and we return to issues relating to measurement in the Discussion.

Guided by a directed acyclic graph (DAG; eFigure 2), we included participant age and mother-reported household income (both assessed at baseline only in GUTS) and state-level median household income and the Gini income inequality ratio (both assessed over time by the U.S. Census Bureau) as potential confounders in our models, given that such factors were hypothesized to be common causes of (exposure to) U.S. state-level structural sexism (e.g., by shaping a state’s sociopolitical climate or influencing where one lives) and individual-level mental health status (e.g., by shaping one’s exposure to and experiences with relevant stressors). Gender/sex, sexual orientation, and race/ethnicity were additionally included as potential confounders in our propensity score models and our non-intersectional outcome models for similar reasons, as were participants’ exposure and outcome histories. We note that this adjustment set is relatively parsimonious, as there are ongoing debates in the field regarding which factors are better conceptualized as mediators or components of structural sexism rather than as confounders of the relationships between structural sexism and health (e.g., area religiosity).⁵⁰ Moreover, and as discussed previously, the designation and treatment of gender/sex, sexual orientation, and race/ethnicity as confounders is not fully aligned with how these axes of social identity/position would be conceptualized within intersectionality.^51,52 Our intentions in adjusting for them here were to aid in reducing bias so as to obtain more intervention-actionable results on structural sexism (e.g., by removing the influence of individual selection into states); however, as with the aforementioned issues related to measurement, we return to tensions related to confounding and directives for future research in the Discussion.

3.3. Statistical analysis

Our analysis followed the steps outlined in Section 2.1, above. First, we conducted descriptive analyses and preliminary data checks. Second, we estimated the time-varying propensity scores for the SCMMs by fitting a pooled linear model for participants’ structural sexism exposure level at a given study wave, conditional on their exposure and outcomes histories and all potential confounders. We then fit a series of outcome models that predicted depressive symptomology as a function of structural sexism, using varying model specifications for comparison purposes: (1) model 1 was a traditional logistic model that used the static exposure variable to estimate associations between structural sexism exposure at wave 13 and high depressive symptoms at wave 14 (i.e., the final exposure and outcome assessments, respectively), conditional on baseline/time-invariant and time-varying confounders measured at wave 12; (2) model 2 was a logistic SCMM that used the point-in-time exposure variable to estimate the total effect of structural sexism exposure at wave $w_{n-1}$ and high depressive symptoms at wave $w_n$, conditional on baseline/time-invariant and time-varying confounders measured at wave $w_{n-2}$; and (3) model 3 was a logistic SCMM that used the cumulative exposure variable to estimate the total effect of structural sexism exposure history from baseline through wave $w_{n-1}$ and high depressive symptoms at wave $w_n$, conditional on baseline/time-invariant and time-varying confounders measured at wave $w_{n-2}$. To be consistent with most prior research on structural sexism,^3,5 we also fit these three models stratified by gender/sex. For a second set of models, we re-fit the gender/sex-stratified SCMMs to include interaction terms between the cumulative structural sexism exposure variable and either race/ethnicity (a baseline/time-invariant effect measure modifier) or sexual orientation (a time-varying effect measure modifier) to explore within-group intersectional heterogeneity. eTable 3 provides subgroup analytic sample sizes for these models.

3.3.1. Estimation

Analyses were conducted in R version 4.2.2 using the geepack and mice packages.^53,54 All models were fit with GEE; model 1 accounted for the clustering of participants within states, and models 2 and 3 additionally accounted for the clustering of repeated measures within participants.³³ Additionally, as mentioned, we used multiple imputation via the MICE (multivariate imputation by chained equations) algorithm to address missingness in our data. Twenty imputed datasets were generated under the assumption that missing values were missing-at-random, conditional on the observed values of all study variables; we fit the models described above using these imputed datasets and pooled the resulting estimates following Rubin’s rules.⁵⁵ Example code for fitting SCMMs is provided in the Appendix.

3.4. Results and interpretations

Descriptive characteristics of the analytic sample at baseline are presented in Table 2. Just over half of the participants (54.2%) were girls/women and the majority were non-Hispanic White, completely heterosexual, and had a household income greater than or equal to $100,000. Approximately 18% were experiencing high depressive symptoms at wave 4 (the first wave at which depressive symptomology was assessed), and as shown in eTable 4, girls/women had a higher prevalence than boys/men at most time-points.

Table 2.

Descriptive characteristics of the Growing Up Today Study at baseline

	Analytic sample (n = 16,143)
Individual-level covariates
Gender/sex, n (%)
Boys/men	7,401 (45.8)
Girls/women	8,742 (54.2)
Sexual orientation, n (%)
Completely heterosexual	15,112 (93.6)
Mostly heterosexual	841 (5.2)
Bisexual	117 (0.7)
Gay/lesbian	73 (0.5)
Race/ethnicity, n (%)
Non-Hispanic Asian	240 (1.5)
Non-Hispanic Black	152 (0.9)
Hispanic/Latine	244 (1.5)
Non-Hispanic White	15,062 (93.3)
Other/unlisted	445 (2.8)
Age in years, mean (SD)	11.5 (1.6)
Household income, n (%)
Less than $50,000	2,120 (13.1)
$50,000 to $74,999	3,821 (23.7)
$75,000 to $99,999	3,650 (22.6)
$100,000 or higher	6,552 (40.6)
State-level covariates
Median household income in $, mean (SD)	39,270 (2,751)
Gini income inequality ratio, mean (SD)	0.6 (0.03)

Notes. Percentages are column percentages that may not sum to 100 due to rounding.

Table 3 presents results from the first set of models comparing the static, point-in-time, and cumulative structural sexism exposure operationalizations. Here, the effect estimates from model 1 give the change in odds of experiencing high depressive symptoms at wave 14 associated with a 1-unit increase in exposure to structural sexism at wave 13, conditional on the included potential confounders; the effect estimates from models 2 and 3 give the associations between structural sexism exposure at wave $w_{n-1}$ or structural sexism exposure history from baseline through wave $w_{n-1}$, respectively, and experiencing high depressive symptoms at wave $w_{n-1}$, also conditional on the included potential confounders. Considering first the estimates from models including the full analytic sample, we found a small yet significant association between cumulative exposure to structural sexism and subsequent depressive symptomology; specifically, the odds of experiencing high depressive symptoms at wave $w_n$ was 1.3 times as high with each standard deviation increase in cumulative exposure from baseline through wave $w_{n-1}$ (odds ratio [OR]: 1.30, 95% confidence interval [CI]: 1.07–1.57, corresponding to a Cohen’s d effect size of 0.2). Conversely, associations between depressive symptomology and both the static and point-in-time structural sexism exposure operationalizations were close to the null value of 1 and non-significant (OR: 1.01, 95% CI: 0.91–1.12 and OR: 1.03, 95% CI: 0.99–1.06, respectively, corresponding to effect sizes of 0.001 and 0.02). A similar pattern of results was obtained from models examining these relationships among girls/women; among boys/men, the effect estimate for the association between cumulative exposure to structural sexism and subsequent depressive symptomology was diminished slightly and the CI included 1 (OR: 1.15, 95% CI: 0.83–1.59, corresponding to an effect size of 0.08).

Table 3.

Associations between structural sexism and depressive symptomology in the Growing Up Today Study: comparing static, point-in-time, and cumulative exposure operationalizations

	Odds ratio (95% CI)
Full sample
Static exposure at $w_{13}$ and outcome at $w_{14}$a	1.01 (0.91, 1.12)
Point-in-time exposure at $w_{n-1}$ and outcome at s $w_n$b	1.03 (0.99, 1.06)
Cumulative exposure through $w_{n-1}$ and outcome at $w_n$b	1.30 (1.07, 1.57)
Girls/women
Static exposure at $w_{13}$ and outcome at $w_{14}$a	1.04 (0.89, 1.21)
Point-in-time exposure at $w_{n-1}$ and outcome at s $w_n$b	1.04 (1.00, 1.09)
Cumulative exposure through $w_{n-1}$ and outcome at $w_n$b	1.40 (1.11, 1.76)
Boys/men
Static exposure at $w_{13}$ and outcome at $w_{14}$a	0.97 (0.78, 1.20)
Point-in-time exposure at $w_{n-1}$ and outcome at s $w_n$b	1.01 (0.94, 1.07)
Cumulative exposure through $w_{n-1}$ and outcome at $w_n$b	1.15 (0.83, 1.59)

Notes. CI = confidence interval. $w_n$ denotes study wave $n$, from $n$ = 1 (baseline) to $n$ = 14.

^aEstimated from a traditional logistic model adjusted for baseline/time invariant covariates and time-varying covariates assessed at $w_{12}$ (i.e., lagged one study wave prior to exposure assessment).

^bEstimated from a sequential conditional mean model adjusted, via propensity scores, for baseline/time invariant covariates and time-varying covariates lagged one study wave prior to exposure assessment.

Results from the second set of models exploring intersectional heterogeneity in these relationships are presented visually as predicted probabilities in Figure 2; corresponding ORs and 95% CIs are given in eTable 5. There are several interesting findings to highlight. First, the likelihood of experiencing high depressive symptoms increased as cumulative exposure to structural sexism increased for most intersectional subgroups, which is consistent with our findings from the models using the full analytic sample. However, there was considerable heterogeneity in the magnitude of these increases, with marginalized or multiply marginalized subgroups typically displaying the steepest slopes. One example of this is shown in Figure 2a: these gender/sex- and sexual orientation-specific estimates indicate that cumulative exposure to structural sexism disproportionately increased the likelihood of experiencing high depressive symptoms for sexual minority girls/women – especially those identifying as bisexual – compared to their completely heterosexual peers. Lastly, despite findings from our gender/sex-stratified models indicating that cumulative exposure to structural sexism was associated with depressive symptomology only for girls/women, these intersectional models revealed how some groups of boys/men (e.g., those who were non-Hispanic Asian or non-Hispanic Black) had slopes suggestive of a positive association as well. That being said, the CIs around many of these estimates were wide, likely due to small subgroup sample sizes.

Figure 2.

Associations between cumulative exposure to structural sexism and depressive symptomology by intersectional subgroup in the Growing Up Today Study.

Collectively, the results from this illustrative applied example help to extend the burgeoning literature on structural sexism and population health by providing new insights into the nature and course of these relationships. With respect to the first set of models comparing structural sexism exposure operationalizations, our finding of a small yet significant association between cumulative exposure and depressive symptomology that was unique to girls/women is consistent with theoretical models of health (in)equity that suggest that the impacts of structural forms of discrimination and disadvantage may not be immediate but rather develop over time as exposure accumulates, especially for targeted groups.^6,8,9 Our intersectional models add further nuance by demonstrating how cumulative exposure to structural sexism disproportionately impacted the mental health of (multiply) marginalized subgroups, a finding that is in-line with the few existing studies in this area and with intersectional theorizing and scholarship more broadly.^13,18,19 If the assumptions underlying causal modeling (e.g., consistency, exchangeability, positivity) were met, our findings would point to a role of early life-course exposure to U.S. state-level structural sexism in the production and/or maintenance of intersectional inequities in mental health. However, we reiterate that these analyses were conducted primarily for demonstrative purposes and the findings should be interpreted with caution, as several limitations (e.g., residual confounding, small subgroup sample sizes) likely preclude rigorous causal inference.

4. Discussion

Our goal in this paper was to help advance population health research on structural sexism by introducing a novel analytic approach for addressing life-course and intersectional effects. Drawing on recent methodological advancements from life-course epidemiology and quantitative intersectionality,^22,25 we outlined how SCMMs can be used to model the long-term, cumulative impacts of structural sexism on a subsequent health outcome, even in the context of time-varying confounding; we also demonstrated how SCMMs can be used to test for the potential moderating effects of social identity/position variables, including those that are accurately measured as being fluid over time (e.g., sexual orientation) so as to explore between- and within-group intersectional heterogeneity. Importantly, these models rely on relatively common statistical techniques that many quantitative researchers will be familiar with (e.g., generalized estimating equations) and they provide results that are easily interpretable and relevant to understanding and addressing social inequities in health.

The methods presented in this paper offer several strengths. By accounting for both the changing nature of places over time and participant mobility, SCMMs can better capture how structural sexism may actually be experienced and embodied – i.e., cumulatively across the life-course, rather than statically or discretely.⁷ This is important, because as the findings from our applied example illustrate, ignoring individuals’ life-course exposure dynamics may inadvertently result in under-estimating (or perhaps in some situations, over-estimating) the impacts of structural sexism on population health and health inequities. Moreover, the unique ability to incorporate time-varying effect measure modification within SCMMs means that the modeled effects of structural sexism are allowed to vary across intersectional subgroups (including when membership in these subgroups changes over time), likely improving the internal and external validity of this work.⁵⁶ Overall, we believe that the continued use of these methods will help to better align structural sexism research with its guiding theoretical principles, in turn providing more nuanced, accurate, and intervention-actionable results.

4.1. Limitations and future directions

There are several limitations to note. First, the implementation of our analytic approach requires rich, longitudinal data, which may be difficult to access or even non-existent for certain applications. While there can be potential workarounds when only cross-sectional data are available or feasible (e.g., asking participants about residential histories), prospectively collected and geocoded cohort data with time-updated measures will be the gold standard for limiting recall bias, time-varying confounding, and other forms of error. Relatedly, and as illustrated through our example analyses, the ability to explore and document intersectional inequities with rigor and nuance relies on the repeated assessment of participants’ held identities and social positions via validated, inclusive measures, which are often omitted from even the most comprehensive cohort studies. We encourage researchers who are interested in implementing these methods with secondary data to consult recently published guidelines regarding the analysis and interpretation of health inequities when less-than-ideal measures of gender, race/ethnicity, class, sexuality, etc. were used,^e.g.,57,58 and to continue to advocate for the inclusion of better measures in existing data sources.

A second limitation is that logistic SCMMs use multiplicative-scale interaction terms to operationalize intersectional heterogeneity, an approach that has received valid critique for not truly capturing the framework’s “departure from additivity” principle and for introducing challenges related to scalability (e.g., convergence issues when including higher-order interactions).²⁵ Indeed, we were only able to explore the intersectionality between two dimensions of social identity/position simultaneously in our applied example due to the disproportionately small subgroup sample sizes of boys/men, marginalized racial/ethnic groups, and sexual minorities in GUTS; such cohort demographics meant that we were unable to consider any three-way or higher interactions and limits the generalizability of our findings, in turn precluding a more nuanced understanding of how structural sexism shapes health outcomes across diverse social locations. Techniques for quantifying intersectionality-relevant additive-scale interactions from multiplicative-scale models are available, although they too are limited by subgroup sample sizes and data dimensionality.

Lastly, and as discussed previously, SCMMs are a tool for causal inference rooted in epidemiology’s potential outcomes framework,³¹ which may not be the most appropriate framework for evaluating how structural sexism shapes population health patterns and intersectional health inequities.^34,35 At a conceptual level, scholars have long argued that the epistemology of epidemiology, which stems from a positivist paradigm, is fundamentally at odds with the epistemology of intersectionality, which stems from a critical interpretivist paradigm.⁵⁹ It is also worth noting that many intersectional theorists reject the need to empirically demonstrate the causal effects of systems of oppression (on health), and instead understand lived experience, community knowledge, and activism as being paramount to enacting change.⁶⁰ Moreover, the ability to draw causal conclusions from observational data is of course contingent upon the quality of these data and the degree of confounding control achieved, as well as a set of strong and largely untestable assumptions. We argue therefore that the goal in using these methods within structural sexism research is not so much to estimate as precise a causal effect as possible as it is to build an evidence base demonstrating the myriad ways by which it and other intersecting forms of structural discrimination and systems of oppression harm health – an effort that will certainly be served by making our (often implicit) beliefs more transparent and paying greater attention to exposure operationalizations and model specifications.⁶¹ That being said, we see great value in adopting a pluralistic approach to causal inference in this field, specifically by triangulating evidence across studies that use a variety of designs, methods, and guiding theoretical frameworks.^23,36

In light of these challenges, we now outline a series of adaptations and alternative strategies for researchers interested in using and/or further developing these methods within their own work. As mentioned, there are potential substitutes that could be used as or in place of prospectively collected and geocoded cohort data, including cross-sectional data with residential history measures, electronic health record (EHR) data, and location-bound data in which the duration of time spent in that location is known (e.g., data collected on college campuses that ask participants to report their year in school, transfer student status, etc.). It would of course also be possible to simplify how life-course exposure to structural sexism is operationalized, such as by analyzing only a few key exposure time points (e.g., childhood and adolescence) in relation to an outcome assessed once at the end of follow-up, which would still be consistent with most prior studies in life-course epidemiology. Regarding the potentially problematic reliance on interaction terms to quantify intersectionality, it may be advantageous to consider the use of other novel modeling approaches. For example, Evans’ MAIHDA approach to quantitative intersectionality research uses random effects to represent intersecting social identities/positions and is thus well-suited to handle high-dimensional stratification, including in the context of small subgroup sample sizes.²⁹ While MAIHDA has largely been used in descriptive intersectionality research thus far, as previously mentioned, recent adaptations now facilitate the inclusion of a contextual-level exposure variable;³⁰ this formulation may be further adapted to incorporate life-course exposure trajectories, although techniques to account for time-varying confounding need developing. There is also a growing interest in using complex systems methods (e.g., agent-based models) within population health and health equity research,⁶² which we see as a promising alternative to traditional regression-based methods that may better capture structural sexism’s dynamic, multilevel, and context-specific effects.

Finally, we note that there are a number of interesting future directions to pursue with these methods and within research on structural sexism more broadly. On a technical-level, we made a potentially over-simplified assumption that cumulative exposure was linearly related to the outcome; it is equally (or even more) likely that non-linear relationships exist, which can and should be tested. As outlined by Homan,¹¹ the particularities by which structural sexism impacts health over the life course also remain unknown – in other words, does exposure accumulate over time, or are other patterns (e.g., critical/sensitive periods, social mobility) more relevant?⁹ While SCMMs were developed to estimate so-called “short-term effects” (i.e., the effect of an exposure on a subsequent outcome), extensions for estimating “long-term effects” (i.e., the effect of a series of lagged exposures on a subsequent outcome) have been recently proposed, which would allow SCMMs to be used in exploring these alternative life-course exposure patterns.^31,32 Other topics for future inquiry include evaluating whether our available structural sexism measures are relevant across ages and social groups, assessing whether certain domains (e.g., economic, political) and/or levels (e.g., county, neighborhood) of structural sexism are more or less salient at different developmental periods, and identifying downstream mediating pathways. Regarding the assessment of different spatial scales in particular, there is a recognized need for the advancement of statistical techniques for life-course and quantitative intersectional research that are compatible with small-area estimation, given the influence that more proximal social contexts exert on health and health inequities.^63,64 Integrating all of these pursuits with methods recently advanced for capturing the intersectionality of systems (e.g., between structural sexism, racism, and cisheterosexism) will also be essential.¹⁹

4.2. Conclusions

We hope that the methods introduced in this paper will motivate further research into how structural sexism shapes population health outcomes over time for diverse social groups. In the wake of U.S. Supreme Court’s Dobbs v. Jackson Women’s Health Organization decision, the continued onslaught of simultaneously sexist, racist, and cisheterosexist policies being introduced and implemented across numerous U.S. states,^65,66 and the accelerating rise of socioeconomic inequities within and between populations,⁶⁷ this work is increasingly relevant and urgent. We believe that it is thus of utmost importance that our research reflects the dynamic and complex ways by which multiple forms of structural discrimination and systems of oppression impact population health and health inequities to more effectively reveal enacted harms and identify strategies for intervention and resistance.