A multilevel approach to modeling health inequalities at the intersection of multiple social identities
Introduction
Intersectionality is a theoretical framework that is increasingly used to study the patterning of health inequalities because of its focus on the multidimensional, multiplicative nature of disadvantage (Bowleg, 2012, Farmer and Ferraro, 2005, Schulz and Mullings, 2006, Veenstra, 2011, Warner and Brown, 2011), which corresponds with discipline-specific theories such as fundamental causes (Link and Phelan, 1995) and ecosocial theory (Krieger, 2011). Intersectionality theorists posit that inequalities are generated by numerous interlocking systems of privilege and oppression such as racism, classism, sexism, and ageism (Collins, 1990, Crenshaw, 1989, McCall, 2005), and push back against the “additive approach,” which treats the advantages or disadvantages conferred through simultaneous occupation of multiple social positions as simply accumulated. Care must be taken in the adoption of intersectionality by public health researchers, however, to ensure that it is properly framed within the context of ongoing debates in epidemiology—namely between the so-called “risk factor” epidemiology and “eco-epidemiology” (Susser and Susser, 1996). Conventional approaches to quantitative intersectionality analysis have also presented several methodological limitations, including issues of scalability, model parsimony, small sample size, and interpretability of results.
In this study, we explore an alternative analytic approach (Evans, 2015, Green et al., 2017, Jones et al., 2016) that resolves some of the key theoretical and methodological tensions inherent to this research. This approach involves applying hierarchical, multilevel models to study large numbers of interactions and intersectional identities while partitioning the total variance between two levels—the between-strata (or between category) level and the within-strata (or within category) level. This analytic approach is a valuable tool for exploring multiple interactions simultaneously and the patterning of inequalities across society. We apply and compare this new approach with the conventional, fixed effects approach to interaction models. To demonstrate its potential application in health research, we explore an empirical example that will be familiar to many social epidemiologists—body mass index among U.S. adults.
McCall (2005) has identified three distinct orientations within the current intersectional literature—the intercategorical, the intracategorical and the anti-categorical. The anti-categorical approach involves the critique and deconstruction of analytic categories. The intracategorical approach tends to “focus on particular social groups at neglected points of intersection … in order to reveal the complexity of lived experiences within such groups” (p.1774). The approach to intersectionality most often adopted in social epidemiology is the intercategorical approach because of its natural fit with quantitative analyses of inequalities. The intercategorical approach involves the provisional adoption of analytic categories to document inequalities among groups and explore the interactions between different dimensions of identity, position, and social processes. The conventional intercategorical approach to studying interactions, which we will refer to as the fixed effects approach, involves fitting a single-level regression model with a full complement of parameters to account for all points of interaction. This can be accomplished either by using a full set of dummy variables (one representing each possible combination of social identity and position, e.g., young college educated high income black woman) or by including main effects and a saturation of interaction terms. Mathematically these approaches are equivalent, and so we refer primarily to the version with interaction parameters.
When using intercategorical intersectionality in population health research it is critical to correctly situate this framework within existing debates in epidemiology. Rose (1992) famously distinguished between causes of individual cases (i.e., why did this person get sick with this illness at this point in time) and causes of population incidence (i.e., what caused this population to have a higher disease incidence than that population). Causes at these distinct levels may or may not resemble each other. Susser and Susser (1996) expanded on this to differentiate risk factor epidemiology with its focus on identifying causes of cases from eco-epidemiology, which takes a multilevel perspective and considers causal pathways ranging from the societal level to the molecular level. As others have noted (Merlo and Wagner, 2012, Merlo, 2014), this distinction is not always appreciated in modern epidemiology.
The risk factor approach in epidemiology involves the identification of risk factors through the comparison of group averages. In practice the use of variables such as gender, race/ethnicity, and socioeconomic status (SES) in quantitative intersectionality research may make it appear that the mission of intersectionality research corresponds with the risk factor approach, and involves identifying ever-narrower and more specific “risky identities” that are particularly burdened by health inequalities. This is, however, diametrically opposed to a central tenet of intersectionality—namely, that intersectionality does not situate the problems associated with particular identities within individuals or the identities themselves, but within the structural power hierarchies, social processes, and social determinants that shape the social experiences of individuals with those intersectional identities. While categorical variables (gender, race, class) may be used in regression models, care should always be taken to recognize that these may be intended as proxies for the interactions of systems of oppression (sexism, racism, classism) and other social processes in producing population-level incidence (Bauer, 2014).
In ecosocial theory, Krieger (2011) theorizes health inequalities between populations as resulting from numerous interacting pathways of embodiment across the life course, through which we come to “literally incorporate, biologically … the material and social world in which we live” (p. 214). Ecosocial theory encourages a broad vision for the determinants of health inequalities—including both the interlocking systems of oppression and privilege (sexism, racism, classism) implicated by intersectionality and other social processes. For instance, Krieger points to issues of social and economic deprivation, environmental hazards, and the targeted marketing of harmful commodities to low income populations as key pathways of embodiment, which may not readily be classified as forms of intersectional “classism” per se.
Eco-epidemiologists have argued strongly against the “tyranny” (Merlo and Wagner, 2012) of comparing group averages, both because it risks framing inequalities as individual-level issues resolvable by individuals (resulting in “blaming the victim”) and because such averages obscure the relatively low predictive power of these labels to distinguish between cases and non-cases. In other words, risk factors typically are unable to discriminate between individuals who will become sick and those who will not (Merlo, 2014), which should caution all of us to frame intercategorical intersectionality research in the health inequalities domain as explicitly eco-epidemiologic.
Paradoxically, as Merlo (2014) has noted, many existing eco-epidemiologic studies continue to utilize a framework reliant on comparing group averages—though admittedly these new risk factors are situated at higher contextual levels, such as comparing neighborhood averages. Eco-epidemiologic approaches should balance consideration of group averages with what Merlo (2014) has called a “multilevel analysis of individual heterogeneity”—or a multilevel examination of variation within and between groups. The approach presented here is explicitly framed with this intention and allows for consideration of both group averages and multilevel variation within and between groups.
As a brief aside, we will henceforth refer to these points of intersection as “social strata” rather than as categories or groups. The intersectionality literature has encouraged us all to become more skeptical of the reification of categorical labels, and therefore we feel that the term “strata”—which alludes to stratified analyses—implies provisional acceptance of labels for the purposes of studying inequalities, while remaining aware of the inherent danger in treating social labels as monolithic, unchanging, and inflexible. Similarly, we sometimes use the word “identity” as a shorthand to refer to dimensions of identity, position, and resources. We do not mean to imply that income, for instance, is best understood as a social identity.
The conventional fixed effects approach to operationalize intercategorical intersectionality is open to two related theoretical criticisms and poses additional empirical concerns. First, including interaction terms encourages us to only study the intersectionality of marginalization. For instance, in a comparative quantitative intersectional analysis where white males are the reference group, we might include main effects for “black” and “female” and an interaction term for “black and female.” Following current standards, finding this interaction term to be statistically significant would be interpreted as support for the interaction of racism and sexism. However, this setup enables us to only evaluate the interaction effect experienced by black women, while those experiencing multiple privileges (e.g., white men) or mixes of privilege and disadvantage (e.g., white women and black men) are treated as having no observable interaction effect. Theorists have voiced this concern and called for consideration of the points of intersection that mix privilege and marginalization (Bauer, 2014, Choo and Ferree, 2010, Hancock, 2007, Nash, 2008). While studying intersectionality of privilege could be accomplished by switching the reference group to “low SES black females” or by constructing alternative post hoc analyses, ideally, we would be able to determine simultaneously whether all intersectional identities exhibit evidence of an interaction (or intersectional) effect above and beyond the contributions of the additive main effects.
In other words, to examine whether a given social stratum shows evidence of an intersectional interaction effect, we would want to compare what is observed for that stratum with what might have been expected for it based on the additive contributions of the main effects. Establishing the magnitude and direction of this unique “interaction effect” for each social stratum is also of interest because a simplistic reading of intersectionality might imply that this interaction effect will in some way reflect the number of marginalized or privileged identities. For instance, a naïve reading of foundational works of intersectional scholarship, with their focus on race- and gender-based discrimination (Collins, 1990, Crenshaw, 1989), might lead some to conclude that intersectionality implies that possessing more marginalized identities will necessarily result in a more harmful interaction effect. The intersectionality of privilege might be assumed to work in the opposite direction—enjoying multiple interlocking privileges will result in an interaction effect that is more beneficial. This interpretation of intersectional thought has been refuted by most scholars since the early days of the approach (e.g., King, 1988). However, researchers such as Bauer (2014), who have called for greater attention to identities that mix privilege and marginalization, have highlighted our continued uncertainty about the effects of possessing a combination of marginalized and privileged identities and social positions when it comes to the social patterning of outcomes such as health inequalities.
It is important to acknowledge that our framing of quantitative, intercategorical intersectionality falls into what some intersectionality scholars have called “intersectionality as testable explanation” (Hancock, 2013), in that it involves an assessment of whether statistically significant interaction effects are detectable. Though becoming more common, this framing of intersectionality remains contentious, with some scholars arguing that intersectionality should be considered as more akin to an analytic tool to be utilized rather than a hypothesis that can be tested. The approach outlined here is also useful as a tool for exploratory analysis of inequalities, and does not necessarily require this hypothesis testing of interaction terms.
The second theoretical criticism of the fixed effects approach is that consistently comparing the multiply marginalized to the multiply privileged runs the risk of reinforcing the notion that the dominant, privileged group often used as the reference category (e.g., high SES white males) is the standard against which all other groups ought to be compared (Choo and Ferree, 2010). In other words, using the multiply privileged as the yardstick against which marginalized groups are measured reinforces the social primacy of the privileged as the “default” category. While to some extent comparisons to more privileged groups are inherent to the project of documenting inequalities, we would ideally be able to make multiple comparisons simultaneously—rather than relying only on comparisons to one privileged identity.
The fixed effects approach has yielded powerful insights into the patterning of health inequalities across society. However, the ever-increasing demands to examine interactions between dimensions of identity beyond just race and gender (McCall, 2005, Nash, 2008) has pushed researchers up against some of the methodological limitations of this approach. Namely, the fixed effects approach to interactions struggles with issues of scalability, model parsimony, reduced sample size in some intersectional strata, and occasionally, issues of interpretability.
From an empirical standpoint, including many dimensions of social identity in interaction models creates a set of additional modeling and interpretation challenges. As the number of dimensions of social identity considered increases, the number of parameters required in a fixed effects model increases geometrically to allow for all combinations of first-order, second-order, and higher-order interaction terms. While this poses minute problems when fewer dimensions of identity are interacted, model parsimony and fit does become a concern at higher dimensions. The geometric increase of parameters to interpret can also complicate the examination of results. This is especially true given the non-ideal comparisons built into the model setup described above.
Additionally, as a given sample size is parsed across more intersectional identities the issue of insufficient sample size in many social strata becomes a concern. To illustrate most clearly, imagine a model using the dummy variable fixed effects approach wherein each intersectional identity has its own indicator variable. In a linear regression with a model fully saturated in this way, we might obtain the mean expected outcome for each identity, although this mean is calculated for some groups based on only a handful of respondents. It is then up to the researcher to weed out (often by hand) those intersectional groups of insufficient sample size.
In the recent, seminal work by Jones et al. (2016) they demonstrate the novel use of multilevel models to study high-dimensional interactions in multivariate models. Applying a hierarchical random effects model, individuals are nested within what they term multivariate “contingency tables” but what we might call intersectional social strata, and apply this to study voting (and abstaining) patterns in the 2015 UK general election. A similar and explicitly intersectional approach was proposed by Evans (2015) and Green et al. (2017). These early examples of this approach build on a growing interest in using multilevel models to study variation within and between social groups (Merlo, 2014, Merlo et al., 2016). We propose applying this novel approach to interactions to study intersectional social identities in the domain of health inequalities, and offer an illustration of this alternative method and the ways in which it addresses the criticisms and resolves the tensions we mentioned above.
Section snippets
Data
The National Epidemiologic Survey on Alcohol and Related Conditions (NESARC) was a longitudinal study launched in 2001 by the National Institute on Alcohol Abuse and Alcoholism. It was designed to include a representative sample of the U.S. non-incarcerated civilian population, including citizens and non-citizens, aged 18 years and older who are residing in the U.S. In this study, we used data from Wave 2, collected between 2004 and 2005 (Grant and Kaplan, 2005). The large sample size and
Results
From the initial sample size of N = 34,653 in Wave 2, 330 respondents were excluded because of missing responses to the height and weight items, and a further 1535 were excluded because they were not classified in the three race/ethnicity categories considered. Thus, the final sample includes 32,788 respondents. The demographic profile of the sample appears in Table 1. For details on the sample size in each intersectional social stratum, see Supplemental Table 1.
The number of observations per
Discussion
In this study, we have demonstrated a new approach for modeling health inequalities between multiple intersectional social identities. The multilevel approach has several advantages over the conventional fixed effects approach when the number of interactions becomes large, summarized in Table 3. First, multilevel models present a more parsimonious approach because they grow linearly as opposed to geometrically as new intersectional identities are added. The scalability of the multilevel
Conclusion
The present study contributes to our methodological repertoire by broadening our vision of the potential applications of multilevel models in social epidemiology. Multilevel modeling has already demonstrated its value in the public health and social science literature by enabling the study of ecological effects. The approach we have outlined illustrates that the statistical dependencies created by abstract hierarchical social structures within which respondents are embedded can also be modeled.
Funding
No direct funding supported this study.
Acknowledgements and Credits
We wish to thank the reference librarians of the Harvard University Countway Library of Medicine for their assistance in performing a systematic review of the literature.
References (37)
Incorporating intersectionality theory into population health research methodology: challenges and the potential to advance health equity
Soc. Sci. Med.
(2014)- et al.
Multiple contexts and adolescent body mass index: schools, neighborhoods, and social networks
Soc. Sci. Med.
(2016) - et al.
Are racial disparities in health conditional on socioeconomic status?
Soc. Sci. Med.
(2005) - et al.
Can intersectionality theory enrich population health research?
Soc. Sci. Med.
(2017 Apr) - et al.
Understanding how race/ethnicity and gender define age-trajectories of disability: an intersectionality approach
Soc. Sci. Med.
(2011) - et al.
Intersectionality and risk for ischemic heart disease in Sweden: categorical and anti-categorical approaches
Soc. Sci. Med.
(2017) The problem with the phrase women and minorities: intersectionality— an important theoretical framework for public health
Am. J. Public Health
(2012)MCMC Estimation in MLwiN v2.1: Centre for Multilevel Modelling
(2009)- et al.
Practicing intersectionality in sociological research: a critical analysis of inclusions, interactions, and institutions in the study of inequalities
Sociol. Theory
(2010)
Social disparities in BMI trajectories across adulthood by gender, race/ethnicity and lifetime socio-economic position: 1986-2004
Int. J. Epidemiol.
Black Feminist Thought: Knowledge, Consciousness, and the Politics of Empowerment
Demarginalizing the Intersection of Race and Sex: a Black Feminist Critique of Antidiscrimination Doctrine, Feminist Theory, and Antiracist Politics
Innovative Approaches to Investigating Social Determinants of Health - Social Networks, Environmental Effects and Intersectionality
"Source and Accuracy Statement for Wave 1 of the 2001-2002 National Epidemiologic Survey on Alcohol and Related Conditions."
Source and Accuracy Statement for the Wave 2 National Epidemiologic Survey on Alcohol and Related Conditions (NESARC)
When multiplication Doesn't equal quick addition: examining intersectionality as a research paradigm
Perspect. Polit.
Empirical intersectionality: a tale of two approaches
UC Irvine Law Rev.
Cited by (175)
What does the MAIHDA method explain?
2024, Social Science and MedicineSex and age differences in the patient-reported outcome measures and adherence to an osteoarthritis digital self-management intervention
2024, Osteoarthritis and Cartilage OpenDevelopmental psychopathology as a meta-paradigm: From zero-sum science to epistemological pluralism in theory and research
2024, Development and Psychopathology