Social Science & Medicine 110 (2014) 10e17

Contents lists available at ScienceDirect

Social Science & Medicine

Incorporating intersectionality theory into population health research

a r t i c l e i n f o a b s t r a c t

Article history: Intersectionality theory, developed to address the non-additivity of effects of sex/gender and race/
Received 31 December 2013 ethnicity but extendable to other domains, allows for the potential to study health and disease at
Received in revised form different intersections of identity, social position, processes of oppression or privilege, and policies or
18 March 2014
institutional practices. Intersectionality has the potential to enrich population health research through
Accepted 24 March 2014
Available online 25 March 2014
improved validity and greater attention to both heterogeneity of effects and causal processes producing
health inequalities. Moreover, intersectional population health research may serve to both test and
generate new theories. Nevertheless, its implementation within health research to date has been pri-
Keywords:
Population health
marily through qualitative research. In this paper, challenges to incorporation of intersectionality into
Epidemiology population health research are identified or expanded upon. These include: 1) confusion of quantitative
Intersectionality terms used metaphorically in theoretical work with similar-sounding statistical methods; 2) the question
Research methodology of whether all intersectional positions are of equal value, or even of sufficient value for study; 3) dis-
Health inequalities tinguishing between intersecting identities, social positions, processes, and policies or other structural
Health disparities factors; 4) reflecting embodiment in how processes of oppression and privilege are measured and
Social inequity analysed; 5) understanding and utilizing appropriate scale for interactions in regression models; 6)
Quantitative method
structuring interaction or risk modification to best convey effects, and; 7) avoiding assumptions of
equidistance or single level in the design of analyses. Addressing these challenges throughout the pro-
cesses of conceptualizing and planning research and in conducting analyses has the potential to improve
researchers’ ability to more specifically document inequalities at varying intersectional positions, and to
study the potential individual- and group-level causes that may drive these observed inequalities. A
greater and more thoughtful incorporation of intersectionality can promote the creation of evidence that
is directly useful in population-level interventions such as policy changes, or that is specific enough to be
applicable within the social contexts of affected communities.
Ó 2014 The Author. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-SA
license (http://creativecommons.org/licenses/by-nc-sa/3.0/).

1. Population health research and the need for explicit theory susceptibilities), population groups often experience extremely
different incidence or prevalence of disease (Rose, 1985). The
The term “population health research” can be used to refer to drivers of specific health inequalities can involve intrinsic biological
quantitative research across a range of disciplines (e.g. population factors, such as inherited differences in genetic susceptibilities
epidemiology, social epidemiology, public health, medical sociol- across populations. However, where inequalities are structured
ogy, health promotion, community medicine, community psy- across socio-demographic factors, they are often driven by social
chology) that aims to understand and impact the health and well- inequity, or social policies and practices that create the context for
being of populations. In a classic paper, Geoffrey Rose (1985) increased incidence of disease in some groups while protecting
distinguished between the causes of disease among individual others. These factors represent what Rose described as “the de-
persons and the causes of disease incidence among populations. terminants of population incidence rate”.
Even in cases where the causes of individual disease are the same Currently, a full examination of such causes remains hampered
(e.g. the same virus, the same individual genetic or environmental by a focus on measuring health inequalities and production of
research documenting corresponding social gradients (Lofters &
O’Campo, 2012; Mowat and Chambers, 2012). While documenta-
E-mail address: greta.bauer@schulich.uwo.ca. tion of inequalities is important, it too often fails to provide

http://dx.doi.org/10.1016/j.socscimed.2014.03.022
0277-9536/Ó 2014 The Author. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-SA license (http://creativecommons.org/licenses/by-nc-sa/3.0/).
 G.R. Bauer / Social Science & Medicine 110 (2014) 10e17 11

evidence that can be used to intervene (Lofters & O’Campo, 2012), one master category of social position is of primary research in-
either on a population level (e.g. through policy) to shift overall risk, terest (Hancock, 2007). For example, all analyses can focus on sex/
or at a specifically local level within the social contexts of highly gender or on race/ethnicity or on socioeconomic status. A multiple
affected communities. Moreover, repeatedly documenting health approach in which more than a single category is of interest
inequalities that apply to broad segments of a population may serve operates under an additive assumption that treats multiple mar-
to reinforce existing notions of the intractability of injustice, while ginalisations or privileges as individual categories that can be
failing to identify intervenable factors that might be candidates for layered (Hancock, 2007). While this allows for consideration of a
potential solutions. greater number of social categories, it is not in itself an intersec-
Documentation of health inequalities is often done with a focus tional approach. Using such an approach, the health status of
on one unitary category of difference, which is itself simplified. For Aboriginal women in Canada, for example, would be assumed to be
example, race-based inequalities are still sometimes theorized as sufficiently understood through adding together the independent
biological, or are followed with speculation on a range of possible health impacts of being Aboriginal with those of being female. In
causes, such as racism, family structure, diet, or even poverty; re- contrast, the intersectional approach assumes that an individual’s
searchers in race, ethnicity and health have urged other researchers experience, and their health, are not simply the sum of their parts,
to avoid using race/ethnicity as a proxy for such factors (Jones, and that, for example, what it means to be a woman and what the
2001; Muntaner et al., 1996). While “race” may be a biological fic- health implications are, may be different for Aboriginal women
tion, the social process of racialization is real. The structural and versus non-Aboriginal women. This makes sense in that gender can
interpersonal discriminatory processes of racism are themselves be constituted (and health affected) through cultural meanings and
measurable (Krieger et al., 2005). Likewise, within sex/gender processes including those that are potentially positive, such as
research, research on inequalities is often seen as confirming ex- indigenous cultures, and also through negative policies and their
pectations of “obvious” biological differences, with little attention impacts, such as through gendered aspects of historical trauma in
given to verifying biological similarities, distinguishing the effects residential schools or under policies such as the Indian Act. Sex,
of biologically sexed mechanisms from gendered social processes, gender, race, ethnicity, income, social class, education, age, sexu-
or allowing for their interaction (Springer et al., 2012a,b). Exam- ality, immigration history. each may be understood in greater
ining such unitary approaches to research surfaces the need for complexity through intercategorical approaches to intersection-
careful delineation of related constructs that are often conflated ality, which use categorization pragmatically to explore the health
under a lowest common denominator approach of documenting impacts of multiple identities or social positionalities (McCall,
socio-demographic variation. Moreover, such research studies may 2005).
expand beyond one master category of social position to consider
multiple categories, but do not consider the unique intersections 3. Intersectionality theory in health research
between the categories or intersectional positions within a
category. As an overarching concept, intersectionality has much to offer to
Population health research has been increasingly critiqued for population health in providing more precise identification of in-
its failure to explicitly acknowledge the theory (or lack of theory) equalities, in developing intervention strategies, and ensuring re-
underlying analyses, and for the failure of research teams to sults are relevant within specific communities. It was recently
deliberately consider theoretical frameworks on which their identified as an important theoretical framework for public health
research may then be built (Krieger, 2003; Bartley, 2004). It has also (Bowleg, 2012), and as well as for sex, gender and health (Springer
been critiqued for stripping away the context of people’s lives et al., 2012a).
through identifying single sets of health determinants for entire While intersectionality has been explicitly incorporated into
populations (Raphael and Bryant, 2003). Several recent books have feminist academic work for over two decades, its use in health
begun to integrate population health theory and methodology research has been primarily in the form of qualitative studies. For
(Bartley, 2004; Krieger, 2011). However, even books that incorpo- example, two recent journal special issues on intersectionality
rate a range of theoretical models and address health inequity may were devoted entirely to qualitative work (Phoenix and Pattynama,
address inequalities in only a unitary way, for example, exploring 2006; Bilge and Denis, 2010). While intersectionality scholars have
health inequalities through a master category of sex/gender, or acknowledged that such scholarship can use quantitative as well as
alternatively through race/ethnicity (Bartley, 2004). qualitative methods (Hancock, 2007; McCall, 2005), and examples
of explicitly intersectional quantitative research exist in fields such
2. Intersectionality theory as sociology of health (Veenstra, 2011; Warner and Brown, 2011;
Sen and Iyer, 2012; Seng et al., 2012; Hinze et al., 2012), epidemi-
First termed “intersectionality” by AfricaneAmerican feminist ology (Marcellin et al., 2014), psychology (Stirrat et al., 2008), and
legal scholar Kimberlé Crenshaw (1989), intersectionality theory education (Covarrubias, 2011), some have posited that qualitative
sought to complicate understandings of race- and sex/gender- research is better suited to the examination of intersectionality
based scholarship by arguing that multiple marginalisations, such (Wilkinson, 2003; Bowleg, 2008). However, it may well be that
as those experienced by AfricaneAmerican women, were mutually intersectionality theory has much to offer population health
constituted and could not be understood or ameliorated by ap- research, and even that population health research may turn out to
proaches that treated race and sex/gender as distinct subjects of have some surprising things to contribute to intersectionality the-
inquiry. Though developed as a response to second-wave feminist ory and knowledge. As intersectionality scholars acknowledge the
ideals that were implicitly white and middle-class, and to anti- potential for quantitative work, and population health researchers
racist organizing that was implicitly male in its issues and ideals, call for greater theorization of analyses, much unrealized potential
intersectionality has the potential to improve research not only on exists in building theoretical and methodological bridges between
sex/gender and race/ethnicity, but on all other domains of social intersectionality and population health research.
position, such as socio-economic status, legal Aboriginal status, Within population health research, the importance of inter-
educational background, or age cohort. sectionality may be better grasped by researchers if its relationship
Intersectional approaches differ from unitary and multiple ap- to core methodological (e.g. validity) concerns were made clear,
proaches to research (Hancock, 2007). In a unitary approach, only underscoring its importance for all researchers, and not just those
12 G.R. Bauer / Social Science & Medicine 110 (2014) 10e17

whose work has sex/gender, race/ethnicity, and/or social inequality mixed locations”. Nash (2008) has identified the question of
as primary foci. Similarly, researchers working in intersectionality whether all identities are intersectional, or only those of multiply
may have a greater appreciation for the potential of quantitative marginalised subjects, as a major tension within intersectionality
research to provide population- and intervention-relevant infor- research. Hancock’s position that all intersectional positions are of
mation if they were able to clearly understand the relationship equal interest (2007) offers the potential to represent the embodied
between social, biological or psychological theory and population positions of all research participants in large population samples,
health methodology, and understand that statistics may be esti- few of whom will experience marginalisation nearly exclusively,
mated and interpreted in ways that are neither simply positivist nor without concurrently experiencing some form of privilege (see
atheoretical. More fully incorporating intersectionality theory into challenge 4.4 for more on embodiment, oppression, and privilege).
population health research presents a range of challenges, seven of Considering all intersectional positions within the domains
which are discussed in some detail below. Some challenges are under study, and generating absolute measures of the incidence or
conceptual or linguistic, some relate to measurement and specifi- prevalence of disease or other health-related phenomena for each,
cation, and others arise from difficulties or confusion in matching has the potential to provide new and interesting observations on
the social theory to the statistical theory underlying particular the distribution of the burden of disease across social location, a
quantitative analysis methods. However, each challenge also pre- sort of socio-demographic mapping. Moreover, where economies of
sents an opportunity to improve the quality of research, particu- time and intensity of analysis in qualitative research may place
larly with regard to its potential to more accurately document limits on the breadth of analysis across intersectional positions that
health inequalities, and to identify causes of these inequalities and is feasible, the potential provided by large population data sets
their potential solutions. presents no such restrictions. Through examining a larger set of
intersectional positions, comparisons across position may also
4. Challenges in incorporating intersectionality theory in illuminate the effects of privilege as well as marginalisation and the
population health research health impacts for those at positions that are both privileged and
marginalised can be better understood, without neglecting the
4.1. Quantitative theoretical language versus quantitative methods study of those at multiply marginalised intersectional locations.
Beyond such broad mapping, intersectional analyses involving
Multiple statistical methods have been proposed or used for deeper or more locally specific research questions and theoretical
incorporating intersectionality into quantitative analysis, including formulations will need to focus on those positions or identities that
ANOVA (Warner, 2008), hierarchical class analysis (Stirrat et al., are relevant. Given that a focus on inequalities driven by social
2008), cross-tabulation (Covarrubias, 2011), dichotomous or pol- inequity necessitates an anti-oppressive or social justice approach
ytomous logistic regression (Veenstra, 2011; Hinze et al., 2012; to research, here positions of multiple marginalisation will
Seng et al., 2012; Marcellin et al., 2013), multi-level modelling continue to require prioritization in order to address potential
(Black and Veenstra, 2011), and latent class analysis (Garnett et al., remedies for those who are multiply oppressed.
2013), though it is not clear which, if any, provide a good match
between statistical theory and specific intersectional research 4.3. Intersecting identities, positions, processes, and policies
questions. This lack of clear methodology for studying inter-
sectionality constitutes one tension within intersectionality Many research studies using intersectional approaches, as well
research (Nash, 2008). Interestingly, quantitative applications of as many papers discussing intersectionality theory, have consid-
intersectionality can be obfuscated by the predominance of ered primarily intersecting identities or intersecting categories of
mathematical-like language in intersectionality theory, though its social position, whereas others have extended an intersectional
use there is conceptual rather than strictly mathematical. In the framework to processes. Drawing on earlier work (Hankivsky and
original paper in which Crenshaw coined the term intersectionality Cormier, 2009), Dhamoon and Hankivsky (2011) now characterize
(1989), she refers to “the interaction of race and gender”. Ange- intersectionality as “concerned with simultaneous intersections
Marie Hancock’s often-cited paper (2007) is simply titled “When between aspects of social difference and identity (as related to
multiplication doesn’t equal quick addition: Examining inter- meanings of race/ethnicity, indigeneity, gender, class, sexuality,
sectionality as a research paradigm” and Lisa Bowleg’s paper (2008) geography, age, disability/ability, migration status, religion) and
is titled “When black þ lesbian þ woman s black lesbian woman: forms of systemic oppression (racism, classism, sexism, ableism,
The methodological challenges of qualitative and quantitative homophobia) at macro and micro levels in ways that are complex
intersectionality research” (Bowleg, 2008). Thus, the very language and interdependent.” This makes an important distinction between
used in intersectionality can create confusion for quantitative re- social identities or social positions that are related to potential
searchers. For example, studying how gender and race interact privilege or oppression and the social processes or policies that may
multiplicatively (in an intersectional sense) does not imply that one generate, amplify or temper inequalities between groups, both of
needs to e or even should e use a multiplicative-scale statistical which can be studied intersectionally. Without an emphasis on
interaction model. If intersectionality is to be implemented in intervenable processes or policies, a quantitative intersectionality
quantitative research, then terminology will need to be dis- focused purely on intersecting identities or positions would run the
aggregated in order to allow for clear communication and to pre- risk of continuing to reinforce the intractability of inequity, albeit in
vent the conflation of identical- or similar-sounding concepts. a more detailed or nuanced way.
However, there are methodological considerations in examining
4.2. Questioning whether all intersectional identities or social intersections of identity or social position (e.g. ethnoracial group,
positions are of equal value, or of sufficient value to merit study sexual orientation), versus intersecting processes (e.g. racism, ho-
mophobia), versus analysis of policies or practices, versus combi-
Given that intersectionality research originated in a critique of nations of these (e.g. whether experiences of homophobia have a
the failure of unitary and multiple approaches to address issues for different impact on health among members of different ethnoracial
those who were multiply marginalised (Hancock, 2007), it is not groups). For example, in a descriptive intersectional analysis (e.g.
surprising that research has, as McCall (2005) notes, “tended to one that has as its aim to identify the burden of disease among
reflect multiple subordinate locations as opposed to dominant or those at different socio-demographic intersections) it would
 G.R. Bauer / Social Science & Medicine 110 (2014) 10e17 13

generally be inappropriate to statistically adjust for other variables. that “it is virtually impossible, particularly in quantitative
However, in a process-oriented analysis examining what Lofters research, to ask questions about intersectionality that are not
and O’Campo (2012) call “solution-focused variables”, those that inherently additive”. She questions whether quantitative research
drive heterogeneity across descriptive categories, careful attention is really compatible with intersectionality, contending that not
to the concept of confounding and its control is necessary in order only are approaches to individual questions typically additive, but
to identify potential interventions. Moreover, depending on theory, also unsuccessfully so, building on her experience of asking
the relationship between intersecting identities, positions, pro- research participants at a particular intersection (black lesbian
cesses or policies may be constructed as an interaction, as effect- women) to attempt to decompose their identities into single ad-
measure modification, as mediation, or as moderated mediation. To ditive layers of experience (e.g. the experience of being black). In
add to this complexity, policies and institutional practices play a applying an ecosocial approach to the study of discrimination and
structural role in discrimination that cannot simply be captured at health, Nancy Krieger (2012) maintains that, as researchers, “our
the individual level, and so a group-level or multi-level analysis research needs to integrate these conjoint social facts the same
may be necessary (see further discussion of the issue of level in ways our bodies do, each and every day”. The embodied nature of
challenge 4.8). human beings, whose selves cannot be stratified into parts rep-
In addition to making distinctions between social categories and resenting their multiple dimensions, needs to be both reflected
the processes that generate inequalities, it may be important to and respected in research.
draw more careful distinctions between social identities and social It is precisely the impossibility of expecting individuals to
positions, in order to both more precisely communicate the do- decompose their individual experience that reveals the need for
mains under consideration and to open up possibilities to examine comparison groups, providing contrasts between participants at
additional intersections. While sometimes conflated, there is not each intersectional position under study (e.g. black gay men,
necessarily concordance between one’s personally held identity white lesbian women, straight white men) in order to render the
and a social position one occupies, as indicated either by objective health impacts at each intersectional position visible. While it
measure (e.g. income or wealth) or the way one is perceived and may indeed be difficult to impossible to ask questions that are not
treated by others (e.g. racialization). A woman migrating to the additive, it is certainly possible to conduct analyses that are not
United Kingdom may find herself racialized as black, despite additive, but rather that elucidate the effects of social position and
holding no such identity in her home country; a bisexual-identified social processes through comparisons that break free from as-
woman may be assumed by others to be heterosexual based on her sumptions of additivity and allow for intersectional multi-
male partner; and one does not have to identify as impoverished to plicativity. It may thus be that experiences of discrimination are
live in poverty. Adoption of identities is a developmental process best measured using a single discrimination scale, the effects of
(Phinney, 1989; D’Augelli, 1994), and identities themselves can be which can be compared at intersections with social position
understood as multidimensional, encompassing not just personally categories.
held identity but also degree of importance, as well as personal Since the focus of such scales is on discrimination or margin-
attachment to and social embeddedness within a group (Ashmore alisation, privilege is by default defined as the pole on the other end
et al., 2004). Moreover, identities are context-specific and may of the continuum: the absence of these experiences. While there
shift with regard to place and time, or with the need to align with remains much work to do in understanding how processes of
others around shared identity. Social position then may have an oppression impact health, there is less understanding of the im-
impact independently of identity, or may interact with it in ways pacts of privilege, social inclusion and how these may facilitate and
that impact health. protect health. It seems clear that it is not safe to assume that
Careful distinction between intersecting identities, positions, privilege functions purely through the absence of overtly negative
processes, policies and practices, as well as the methods that are experiences of discrimination. Not being fired or denied a job based
needed to analyse each, has the potential to advance health equity on who you are does not represent the same “privilege” as does
in multiple ways. It averts the problem of conflating identity with getting a job or being promoted because you know someone with
position or experience, opens up possibilities in studying in- influence, are perceived to “fit in”, or because your social in-
teractions across these different domains or for examining media- teractions are lubricated by perceptions that you are friendly,
tion models, and allows for attention to the differing smart, and decidedly not scary. Moreover, many or most individuals
methodological needs and requirements of different types of occupy social positions that include both privileged and margin-
intersectional questions. alised domains; given that they will experience processes of priv-
ilege, marginalisation, and even their complex co-constituted
4.4. Embodiment and experiences of oppression and privilege occurrence, intersectionality theory is necessary to understand
these experiences.
Measurement of oppression involves scales that may combine Moreover, the processes through which oppression and privi-
experiences of oppression within one domain, such as racism or lege function may not be those that we traditionally capture in the
ethnic discrimination (Krieger et al., 2005; Diaz et al., 2001) or “-ism” scales, which assess self-reported individual-level experi-
homophobia (Diaz et al., 2001) requiring experiences to be dis- ences of discrimination, and so a careful theorization of process is
aggregated by participants and attributed to specific domains of necessary. Both structural forms of discrimination and the sub-
discrimination. Such scales have been designed with an implicit conscious experience or internalization of marginalisation are often
additive (non-intersectional) assumption, that discrimination or unmeasured (Krieger, 2012). It is important to also note here that
marginalisation is distinct and identifiable for each type of iden- particularly in localized or sub-group studies, there may not be
tity/group. Other scales are composed of a single series of items on variation in structural factors such as policies or other group-level
discrimination experiences, but ask participants to secondarily variables within the study sample, though such factors may
attribute any experiences to one or more domain of oppression continue to play a causal role in impacting health. Schwartz and
(Krieger et al., 2005; Williams et al., 1997). However, it is clear that Carpenter (1999) have identified this as one form of conceptual
individuals who are members of multiply marginalised groups or interpretation bias that is inherent in attempting to explain
cannot simply dissect out the “types” of oppression specific to health-related inequalities through studying inter-individual vari-
each part of their identity or experience. Bowleg (2008) reasons ation within marginalised groups.
14 G.R. Bauer / Social Science & Medicine 110 (2014) 10e17

4.5. Understanding differences between types of regression models difficult, to construct measures of additive-scale interaction and
for intersectional applications their confidence intervals from multiplicative-scale models
(Skrondal, 2003; Greenland et al., 2008; Zou, 2008), and these may
It is not always clear to researchers how to produce statistical be of use to intersectionality researchers. Such measures include
measures that are most relevant to population health, or most the relative risk due to interaction (RERI), synergy index, and the
interpretable. If using regression modelling, the scale of the model attributable proportion due to the interaction; of these, the synergy
impacts both. For example, linear regression models typically used index performs best when additional covariates are included in a
with continuous outcome variables are in the additive scale, while multiple regression model (Skrondal, 2003). These measures have
logistic regression, Poisson regression, and other log-scale re- clear interpretations with regard to intersectionality. For example,
gressions typically used with dichotomous outcome variables are in in an analysis of the intersection of two social positions, the synergy
the multiplicative scale. This means that when a main effects index represents the ratio of the excess risk observed among those
analysis is conducted, multiple main effects are combined mathe- at the intersection of those positions to that expected if their out-
matically in different ways. In linear regression they are added comes were simply a function of adding together the excess risks
together, in log-scale regression they are multiplied, though neither for each of the positions.
represents a multiplicative approach in an intersectionality sense. Producing clear intersectional results using regression models
Since the additive assumption is thus defined differently in can be further complicated by the use of odds ratios. Incidence risk
different models, a departure from additivity (e.g. intersectional ratios are commonly produced from longitudinal study designs
multiplicativity) has different meanings depending on model scale. with dichotomous outcomes. If designed properly, even casee
Including a cross-product (interaction) term in a model, and thus control studies can accurately estimate an incidence risk ratio
opening up potential to reflect Hancock’s intersectional approach (Pearce, 1993). However, the most commonly used regression-
and McCall’s intercategorical complexity, will produce results with based measure for dichotomous or polytomous outcomes in
differing interpretations and potentially different implications. cross-sectional population health research remains an odds ratio
Nevertheless, published studies using interaction terms explicitly produced using logistic regression, despite the existence of easy-to-
for the purpose of intersectionality analysis have been sometimes use alternatives that produce prevalence risk ratios rather than
conducted in the multiplicative scale (e.g. Veenstra, 2011; Hinze odds ratios for dichotomous outcomes (Barros and Hirakata, 2003;
et al., 2011) and sometimes in the additive scale (Seng et al., Zou, 2004; Bieler et al., 2010; Lin and Wei, 1989). Odds ratios are
2012), without an explicit rationale. difficult to interpret in main effects analysis e as often the only
With linear regression, including a cross-product term results in intuitive interpretation is to erroneously interpret them as risk
an additive-scale assessment of interaction. However, with log- ratios e resulting in sometimes dramatic overestimates of effects,
scale regression models, it results in an assessment of particularly when outcomes are not rare, for example reporting one
multiplicative-scale interaction, unless additional steps are taken. group as “40% less likely” when in fact they were 7% less likely
This can be problematic, as true absence of multiplicative-scale (Schwartz et al., 1999). Odds ratios can similarly lead to errant in-
interaction almost always indicates precisely the presence of an terpretations of multiplicative-scale interactions when they are
additive-scale interaction (Greenland et al., 2008). It is additive- assumed to stand in for risk ratios (Morabia et al., 1997), even in
scale interaction that is both more consistent with biological or cases where the rare disease assumption is met (Campbell et al.,
social causation (if studying potentially causal processes) and of 2005). Moreover, false results can also result if odds ratios are
greater relevance to population health and disease prevention used as a basis for assessing additive-scale interactions (Kalilani
(Szklo and Nieto, 2012). and Atashili, 2006).
For these reasons, assessment of additive-scale interaction is While risk ratio-generating options for regressions of polyto-
more relevant for intersectionality research. For example, an mous outcomes are less developed, in the case of dichotomous
additive-scale analysis can provide estimates of proportions or outcomes existing methodology supports the need for some addi-
numbers of people affected, to illustrate the health or disease tional effort by intersectionality researchers to familiarize them-
burden or benefit within the sub-population at each specific selves with methods that produce prevalence risk ratios with cross-
intersectional location. In an intersectional analysis with a dichot- sectional data. Some alternatives include modified Poisson
omous outcome, the presence of an additive-scale interaction in- regression with robust variance estimatation (Zou, 2004), or Cox
dicates that the number of outcome cases (e.g. the prevalence of a proportional hazards regression with a constant risk period
health-related condition) for those at an intersection differs from assigned to remove the time function (Thompson et al., 1998; Skov
what would be expected based on adding together the individual et al., 1998), and with robust variance estimation (Lin and Wei,
effects from the separate identities, positions, or processes. In other 1989). Both methods perform well (Barros and Hirakata, 2003),
words, it represents a departure from what Hancock (2007) terms can be used with clustered data, and can be conducted using
the additive assumption of the multiple approach. It may be worth commercial statistical packages such as Stata and SAS. For complex
reiterating, given the linguistic similarities, that additive-scale survey data, SUDAAN software includes an option to produce
interaction breaks free of the additivity assumption inherent in prevalence risk ratios from logistic regressions using average
Hancock’s multiple approach, and represents an analysis strategy marginal predictions (Bieler et al., 2010).
that is intersectionally multiplicative rather than additive. This
reinforces the need to distinguish these terms linguistically (e.g. 4.6. Structuring models to make effects visible: interaction and
intersectional additivity or additive assumptions versus additive effect-measure modification
scale).
In simple descriptive analyses with no adjustment for other Use of cross-product/interaction terms in regression can
covariates, assessment of additive-scale interactions can be done constitute an assessment of interaction or effect-measure modifi-
through comparisons of excess risks, structured as per textbook cation. If there is no bias, these will be the same (VanderWeele,
examples (e.g. Szklo and Nieto, 2012). When linear regression is 2009), though interpretation will obscure or highlight different
used, interactions will be in the additive scale. However, when lo- aspects of the interaction. Each assessment may be relevant for
gistic and other multiplicative-scale regression models are used, certain questions in intersectionality research. An interaction
additional steps must be taken. It is possible, and in fact not analysis can consider whether the risk of an outcome differs at
 G.R. Bauer / Social Science & Medicine 110 (2014) 10e17 15

different intersectional positions, or it can examine the risk of an modelling. Chunkwise models separate variables into conceptually
outcome at different levels of interacting processes. Such in- related sets, ordered to reflect plausible causeeeffect relationships
teractions can be identified as synergistic or antagonistic, i.e. (Cohen and Cohen, 1983; Kleinbaum et al., 2008). Within inter-
respectively as greater than or less than the sums of their parts. sectionality research, this could involve structuring models through
Importantly, while the language of statistical interaction (e.g. in- separating social position variables from the social, biological or
teractions between two or more variables) implies a dissection of psychological processes that can create inequalities. Division into
individuals into their stratified component identities, positions, or “chunks” could be undertaken on the basis of whether factors are
experiences, in actuality it allows for an embodied approach, where precede or proceed from each other in a spatiotemporal (and thus
absolute measures of disease/health frequency can be described for potentially causal) sequence, which may also represent divisions of
groups of individuals at each cross-stratified intersection. social position versus social process, or non-modifiable versus
In contrast, effect-measure modification addresses whether the modifiable factors. Separating these may help ensure that effects
impact of one factor differs across strata or level of another. For are visible when all variables are not, in fact, equidistant from the
example, does the impact of racialized discrimination on mental health outcome of interest.
health differ between Asian women and men? Is the effect in the Krieger makes the compelling point that concepts of “proximal”
same direction, or of the same magnitude? Researchers who un- or “distal” should not be assumed to relate to concepts of level, as
derstand the different ways to structure such analyses can make group-level causal factors are not necessarily upstream from
conscious decisions that match up to underlying theoretical or individual-level causes (Krieger, 2008). However, the existence of
pragmatic concerns. For example, assessing as effect-measure causes across multiple levels, some of which exist only at a group
modification with stratification by age (youth and non-youth) level (e.g. policies) and others only at an individual level (e.g. per-
may be useful for work in fields where interventions are often sonal identity), implies a greater need for attention to level of
targeted directly at youth. measurement and analysis. Lofters and O’Campo (2012) have
Knol and VanderWeele (2012) provide recommendations, and argued that a “misguided focus on individual-level factors” has
even templates, for presentation of effect-measure modification resulted in a series of recommendations and interventions that are
and interaction results, though they assume both a causal research imaginable only as individual-level alterations in health-related
question and a reader’s familiarity with epidemiological termi- behaviour, without regard to structural factors that shape or limit
nology. Their concerns are that researchers do not always provide behaviours. This failure of imagination represents a form of con-
sufficient detail in their publications for readers to understand the ceptual bias that can be remedied through consideration of inter-
size and significance (both statistical and real-world) of their sectional effects at multiple levels during the planning phase of
findings. Examples certainly exist where even researchers who research, and the greater use of multi-level regression in the
published intersectional findings did not interpret them correctly. analysis phase.
One well-documented example resulted in author and press
statements on the lack of intensive cardiac diagnostic follow-up 5. Potential to advance health equity
among women and black people, though the results clearly
showed equal outcomes among white men, black men, and white The explicit theorization and greater application of inter-
women, with only black women receiving different levels of health sectionality within population health research has the potential to
care referral (Schulman et al., 1999; Schwartz et al., 1999). improve researchers’ collective ability to more specifically docu-
ment inequalities within intersectional groups, and to study the
4.7. Structuring models to make effects visible: avoiding potential individual- and group-level causes of observed in-
equidistance and unilevel assumptions equalities. It opens up the potential for examination of interesting
questions regarding interactions between dimensions of oppres-
Analytic strategies often fall into a grey area between a sion or privilege, including across levels. Moreover, it can serve to
completely data-driven approach to exploratory analysis (e.g. improve the validity of health research. For example, careful
backward elimination logistic regression) and a more fully theo- attention to intersectional issues has the potential to reduce mea-
rized analysis, such as with a structural equation model, mediation surement bias and improve construct validity, by identifying
model, or an analysis guided by a directed acyclic graph. It is often whether identity, position, process or policy variables are relevant,
reasonable to expect that dimensions of identity or social position and thus avoiding inadvertent use of proxy variables. It can also
may play a role in a particular health outcome, and may interact help avert conceptual and interpretation biases in preventing
with other categories of social position, even where previous misspecified levels and assumptions of equidistance from out-
research has not yet elucidated the issue. A purely data-driven comes. Lastly, it can allow for the identification of heterogeneity of
approach to multiple regression modelling assumes that all effects and the corresponding production of research results for
candidate factors that play a role in producing an outcome are more specific population sub-groups, and avoid the production of
equidistant from the outcome (Weitkunat and Wildner, 2002). results (where such heterogeneity is ignored) that may truly apply
Since this is, in general, not true, this assumption may prioritize to no one.
retention of the most spatiotemporally proximate factors, though What then, can quantitative analysis of intersectional questions
these may in fact mediate the effects of other factors such as sex/ contribute to existing qualitatively-derived knowledge on health?
gender, race/ethnicity and other categories of social position. This Quantitative intersectional research is compatible with ecosocial
type of implicit value assumption within supposedly atheoretical theory (Krieger, 2012) and other biopsychosocial approaches that
analyses suggests the need for semi-theoretical approaches that cross disciplinary divides to reflect complex realities of how social
allow for identification of both “upstream” and “downstream” context, behaviours, and life processes are biologically embodied.
factors that may play important roles in producing health and As with mixed-methods research, there is the possibility for
disease. triangulation across quantitative and qualitative results to identify
Chunkwise (also referred to as blockwise, hierarchical or set- sites of concordance or convergence and areas of potentially
wise) regression is one method that may remediate the assumption interesting divergence. This has the potential not only to provide a
of equidistance, when there is insufficient theorization or data for more rich set of observations through which to understand
more fully theorized methods such as structural equation health, but to accelerate theory generation. Well-specified and
16 G.R. Bauer / Social Science & Medicine 110 (2014) 10e17

thoughtfully designed quantitative intersectional studies that Garnett, B.R., Masyn, K.E., Austin, S.B., Miller, M., Williams, D.R., Viswanath, K.,
December 6 2013. The intersectionality of discrimination attributes and
examine how privilege and marginalisation may function together
bullying among youth: an applied latent class analysis. Journal of Youth and
to impact health have the potential to provide evidence in support Adolescence (online in advance of print).
of e or refuting e existing work in intersectionality and health. Greenland, S., Lash, T.L., Rothman, K.J., 2008. Concepts of interaction. In:
Moreover, the ability to test out intersectional hypotheses Rothman, K., Greenland, S., Nash, T. (Eds.), Modern Epidemiology, third ed.
Wolters Kluwer, Philadelphia, pp. 71e83.
regarding synergistic or antagonistic effects at intersectional posi- Hancock, A.-M., 2007. When multiplication doesn’t equal quick addition: examining
tions across a wide range of intersections in large data sets provides intersectionality as a research paradigm. Perspective on Politics 5 (1), 63e79.
an opportunity to add a higher resolution level to existing maps of http://dx.doi.org/10.1017/Si537592.
Hankivsky, O., Cormier, R., 2009. Intersectionality: Moving Women’s Health
social inequalities within populations. Building on this by exam- Research and Policy Forward. Women’s Health Research Network, Vancouver.
ining intervenable factors that may be drivers of such inequalities Hinze, S.W., Lin, J., Andersson, T.E., 2012. Can we capture the intersections? older
could allow for the identification of greater numbers of potential black women, education, and health. Women’s Health Issues 22, e91ee98.
Jones, C.P., 2001. “Race”, racism, and the practice of epidemiology. American Journal
interventions, or provide stronger evidence in favour of an inter- of Epidemiology 154, 299e304.
vention. Altogether, a greater and more thoughtful incorporation of Kalilani, L., Atashili, J., 2006. Measuring additive interaction using odds ratios.
intersectionality can promote the creation of evidence that is useful Epidemiological Perspectives & Innovations 3, 5.
Kleinbaum, D.G., Kupper, L.L., Nizam, A., Muller, K.E., 2008. Applied Regression
in population-level interventions such as policy changes, or that is Analysis and Other Multivariable Methods, fourth ed. Thomson Brooks/Cole,
specific enough to be applicable within the social contexts of Belmont, CA.
affected communities. Knol, M.J., VanderWeele, T.J., 2012. Recommendations for presenting analyses of
effect modification and interaction. International Journal of Epidemiology 41,
514e520.
Acknowledgements Krieger, N., 2003. Theories for social epidemiology in the twenty-first century: an
ecosocial perspective. In: Hofricher, R. (Ed.), Health and Social Justice: Politics,
Ideology, and Inequity in the Distribution of Disease. John Wiley & Sons, San
This work is supported by an operating grant from the Canadian Francisco, pp. 428e450.
Institutes of Health Research (FRN# MOP-130489). The author Krieger, N., 2008. Proximal, distal, and the politics of causation: what’s level got to
wishes to thank Soraya Blot, Ruth Cameron, Warren Michelow, do with it? American Journal of Public Health 98, 221e230 doi:0.2105/
AJPH.2007.111278.
Ayden Scheim and Zena Sharman for valuable comments on earlier Krieger, N., 2011. Epidemiology and the People’s Health: Theory and context. Oxford
drafts. University Press, New York, NY.
Krieger, N., 2012. Methods for the scientific study of discrimination and health: an
ecosocial approach. American Journal of Public Health 102, 936e945. http://
References dx.doi.org/10.2105/AJPH.2011.300544.
Krieger, N., Smith, K., Naishadham, D., Hartman, C., Barbeau, E.M., 2005. Experiences
Ashmore, R.D., Deaux, K., McLaughlin-Volpe, T., 2004. An organizing framework of discrimination: validity and reliability of a self-report measure for population
for collective identity: articulation and significance of multidimensionality. health research on racism and health. Social Science & Medicine 61, 1576e1596.
Psychological Bulletin 130, 80e114. http://dx.doi.org/10.1016/j.socscimed.2005.03.006.
Barros, A.J.D., Hirakata, V.N., 2003. Alternatives for logistic regression in cross- Lin, D.Y., Wei, L.F., 1989. The robust inference for the Cox proportional hazards
sectional studies: an empirical comparison of models that directly estimate model. Journal of the American Statistical Association 84, 1074e1078.
the prevalence ratio. BMC Medical Research Methodology 3, 21. Lofters, A., O’Campo, P., 2012. Differences that matter. In: O’Campo, P., Dunn, J.R.
Bartley, M., 2004. Health Inequality: an Introduction to Theories, Concepts and (Eds.), Rethinking Social Epidemiology: Towards a Science of Change. Springer,
Methods. Blackwell Publishing, Cambridge, UK. New York, pp. 93e108.
Bieler, G.S., Brown, G.G., Williams, R.L., Brogan, D.J., 2010. Estimating model- Marcellin, R.L., Bauer, G.R., Scheim, A.I., 2013. Intersecting impacts of transphobia
adjusted risks, risk differences, and risk ratios from complex survey data. and racism on HIV risk among trans persons of colour in Ontario, Canada.
American Journal of Epidemiology 171, 618e623. Ethnicity and Inequaliies in Health and Social Care 6 (4), 1e11.
Bilge, S., Denis, A., 2010. Introduction: women, intersectionality and diasporas. McCall, L., 2005. The complexity of intersectionality. Signs 30 (3), 1771e1800.
Journal of Intercultural Studies 31, 1e8. Morabia, A., Ten Have, T., Landis, J.R., 1997. Interaction fallacy. Journal of Clinical
Black, J., Veenstra, G., 2011. A cross-cultural quantitative approach to intersection- Epidemiology 50, 809e812.
ality and health: using interactions between gender, race, class, and neigh- Mowat, D., Chambers, C., 2012. Producing more relevant evidence: applying a social
bourhood to predict self-rated health in Toronto and New York city. In: epidemiology research agenda to public health practice. In: O’Campo, P.,
Hankivsky, O. (Ed.), Health Inequities in Canada: Intersectional Frameworks and Dunn, J.R. (Eds.), Rethinking Social Epidemiology: Towards a Science of Change.
Practices. UBC Press, Vancouver, pp. 71e91. Springer, New York, pp. 305e325.
Bowleg, L., 2008. When black þ lesbian þ woman s black lesbian woman: the Muntaner, C., Nieto, F.J., O’Campo, P., 1996. The bell curve: on race, social class, and
methodological challenges of qualitative and quantitative intersectionality epidemiologic research. American Journal of Epidemiology 144, 531e536.
research. Sex Roles 59, 312e325. Nash, J.C., 2008. Re-thinking intersectionality. Feminist Review 89, 1e15.
Bowleg, L., 2012. The problem with the phrase women and minorities: inter- Pearce, N., 1993. What does the odds ratio estimate in a case-control study? In-
sectionality e an important theoretical framework for public health. American ternational Journal of Epidemiology 22, 1189e1192.
Journal of Public Health 102, 1267e1273. Phinney, J.S., 1989. Stages of ethnic identity development in minority group ado-
Campbell, U.B., Gatto, N.M., Schwartz, S., 2005. Distributional interaction: inter- lescents. Journal of Early Adolescence 9, 34e49.
pretational problems when using incidence odds ratios to assess interaction. Phoenix, A., Pattynama, P., 2006. Intersectionality. European Journal of Women’s
Epidemiological Perspectives & Innovations 2, 1. http://dx.doi.org/10.1186/1742- Studies 13, 187e192.
5573-2-1. Raphael, D., Bryant, T., 2003. The limitations of population health as a model for a
Crenshaw, K., 1989. Demarginalizing the Intersection of Race and Sex: a Black new public health. In: Hofricher, R. (Ed.), Health and Social Justice: Politics,
Feminist Critique of Antidiscrimination Doctrine, Feminist Theory and Anti- Ideology, and Inequity in the Distribution of Disease. John Wiley & Sons, San
racist Politics. University of Chicago Legal Forum, pp. 139e167. Francisco, pp. 410e427.
Cohen, J., Cohen, P., 1983. Applied Multiple Regression/Correlation Analysis for the Rose, G., 1985. Sick individuals and sick populations. International Journal of
Behavioral Sciences, second ed. Lawrence Erlbaum, Hillsdale, NJ. Epidemiology 14, 32e38.
Covarrubias, A., 2011. Quantitative intersectionality: a critical race analysis of the Schulman, K.A., Berlin, J.A., Harless, W., Kerner, J.F., Sistrunk, S., Gersh, B.J.,
Chicana/o educational pipeline. Journal of Latinos and Education 10 (2), 86e Escarce, J.J., 1999. The effect of race and sex on physicians’ recommenda-
105. tions for cardiac catheterization. New England Journal of Medicine 340,
D’Augelli, A.R., 1994. Identity development and sexual orientation: toward a model 618e626.
of lesbian, gay, and bisexual development. In: Trickett, E.J., Watts, R.J., Schwartz, L.M., Woloshin, S., Welch, H.G., 1999. Misunderstanding about the effects
Birman, D. (Eds.), Human Diversity: Perspectives on People in Context. Jossey- of race and sex on physicians’ referrals for cardiac catheterization. New England
Bass, San Francisco, CA, pp. 312e333. Journal of Medicine 341, 279e283.
Dhamoon, K.R., Hankivsky, O., 2011. Why the theory and practice of intersectionality Schwartz, S., Carpenter, K.M., 1999. The right answer for the wrong question:
matter to health research and policy. In: Hankivsky, O. (Ed.), Health Inequities in consequences of type III error for public health research. American Journal of
Canada: Intersectional Frameworks and Practices. UBC Press, Vancouver, Public Health 89, 1175e1180.
pp. 16e50. Sen, G., Iyer, A., 2012. Who gains, who loses and how: leveraging gender and class
Diaz, R.M., Ayala, G., Bein, E., Henne, J., Marin, B.V., 2001. The impact of ho- intersections to secure health entitlements. Social Science & Medicine 74,
mophobia, poverty, and racism on the mental health of gay and bisexual 1802e1816.
Latino men: findings from 3 US cities. American Journal of Public Health Seng, J.S., Lopez, W.D., Sperlich, M., Hamama, L., Reed Meldrum, C.D., 2012.
91, 927e932. Marginalized identities, discrimination burden, and mental health: empirical
 G.R. Bauer / Social Science & Medicine 110 (2014) 10e17 17

exploration of an interpersonal-level approach to modelling intersectionality. VanderWeele, T.J., 2009. On the distinction between interaction and effect modi-
Social Science & Medicine 75, 2437e2445. fication. Epidemiology 20, 863e871.
Skov, T., Deddens, J., Petersen, M.R., Endahl, L., 1998. Prevalence proportion ratios: Veenstra, G., 2011. Race, gender, class and sexual orientation: intersecting axes of
estimation and hypothesis testing. International Journal of Epidemiology 27, inequality and self-rated health in Canada. International Journal of Equity and
91e95. Health 10, 3.
Skrondal, A., 2003. Interaction as departure from additivity in case-control studies: Warner, D.F., Brown, T.H., 2011. Understanding how race/ethnicity and gender
a cautionary note. American Journal of Epidemiology 158 (3), 251e258. define age-trajectories of disability: an intersectionality approach. Social Sci-
Springer, K.W., Hankivsky, O., Bates, L.M., 2012a. Gender and health: relational, ence & Medicine 72, 1236e1248.
intersectional, and biosocial approaches. Social Science & Medicine 74, 1661e Warner, L.R., 2008. A best practices guide to intersectional approaches in psycho-
1666. logical research. Sex Roles 59, 454e463.
Springer, K.W., Stellman, J.M., Jordan-Young, R.M., 2012b. Beyond a catalogue of Weitkunat, R., Wildner, M., 2002. Exploratory causal modeling in epidemiology: are
differences: a theoretical frame and good practice guidelines for researching all factors created equal? Journal of Clinical Epidemiology 55, 436e444.
sex/gender in human health. Social Science & Medicine 74, 1817e1824. http:// Wilkinson, L., 2003. Advancing a perspective on the intersections of diversity:
dx.doi.org/10.1016/j.socscimed.2011.05.033. challenges for research and social policy. Canadian Ethnic Studies Journal 35 (3),
Stirrat, M.J., Meyer, I.H., Ouellette, S.C., Gara, M.A., 2008. Measuring identity mul- 3e23.
tiplicity and intersectionality: hierarchical classes analysis (HICLAS) of sexual, Williams, D.R., Yu, Y., Jackson, J.S., Anderson, N.B., 1997. Racial differences in physical
racial, and gender identities. Self and Identity 7, 89e111. and mental health: socioeconomic status, stress, and discrimination.” Journal of
Szklo, M., Nieto, F.J., 2012. Epidemiology: Beyond the Basics, third ed. Jones & Health Psychology 2, 335e351.
Bartlett, Burlington, MA (See Section 6.6: Which is the relevant model? Additive Zou, G., 2004. A modified Poisson regression approach to prospective studies with
versus multiplicative interaction, pp. 205e207). binary data. American Journal of Epidemiology 159, 702e706.
Thompson, M.L., Myers, J.E., Kriebel, D., 1998. Prevalence odds ratio or prevalence Zou, G., 2008. On the estimation of additive interaction by use of the four-by-two
ratio in the analysis of cross sectional data: what is to be done? Occupational table and beyond. American Journal of Epidemiology 168, 212e224.
and Environmental Medicine 55, 272e277.