Table 1 Overview of quantitative analysis strategies for modeling intersectional stigma

Stratified analyses

The relationship between a measure of stigma and a health outcome is analyzed in separate samples disaggregated by an identity of interest (e.g., illness status, gender, race)

• Simple, easy to perform and interpret

• Cannot necessarily test for statistical significance [101]

• Difficult to interpret and cumbersome to perform when multiple axes are considered

• Cannot use for two continuous measures of health-related stigma

• Exploration of educational outcomes among individuals of Mexican origin only within a sample of women [102]

• Exploratory questions about how the relationship between stigma and health might vary in the presence of an additional identity-related factor that is discrete (e.g., HIV-status, gender)

Factorial design

Vignettes that describe individuals with different combinations of characteristics or identities are presented, typically randomly, to a general population sample. Individuals’ responses to a measure of stigma (e.g., social distance) across vignettes are compared [103]

• Allows for decomposition of stigma related to different identities or factors into unique and shared components [103]

• Experimental design

• Can reflect additive assumptions about the nature of intersectional stigma [52]

• Difficult to interpret and cumbersome to perform when multiple axes are considered

• Difficult to include explanatory or process variables [52]

• Disentanglement of stigma associated with HIV from stigma associated with risk practices (e.g., injection drug use) [104]

• How the level of community discrimination or stigmatizing attitudes and beliefs may vary based on the presence or absence of a small number of additional behavioral or identity-related factors

Moderation analysis

The main effects of two (or more) stigma-related variables are modeled along with the product of those variables (e.g., race × gender × HIV status)

• Simple to do, and in the case of two-way interactions, to interpret

• Flexible

• Can assess positive or negative changes in magnitude and directionality of effects [63]

• When main effects explain much of the variance in the outcome, the ability to assess interactions between those terms is limited [20]

• Three-way or higher-order interactions are difficult to depict and comprehend

• Examination of how social adversity, HIV status, and race interact to explain depression [105]

• Assessment of how stigma related to HIV and substance use interact to explain depression [106]

• Assessment of how weight discrimination interacts with race and socioeconomic status to shape mental health among women [107]

• When large sample sizes are available and variation is present within subgroups to test how the relationship between stigma and health might vary in the presence of an additional identity-related factor that is discrete (e.g., HIV-status, gender)

Latent class or latent profile analysis

Identifies subpopulations of individuals based on their endorsement of different stigma or discrimination experiences

Predictors of membership in these populations (such as identity characteristics like race or health status) can be evaluated and latent class regression can be used to assess how these different patterns of experiences differentially predict health outcomes

• A person-centered, rather than variable-centered, approach to assessing intersectionality

• Treats different patterns of stigma experiences as latent and allows these to be empirically determined

• Can require large sample sizes

• More difficult to explain to lay audiences, including policymakers and funders in some cases

• Identification of patterns of bullying and discrimination experiences related to different identities and assessed to what extent these patterns differentially predicted mental health outcomes [73]

• When large sample sizes are available and the question of interest is how the nature of stigma may vary based on the presence of different combinations of stigmatized behaviors or identities

Multilevel models

In addition to fixed effects, random effects (intercepts and slopes) at the cluster level (e.g., neighborhood, city, country) are included in regression models

Covariates can be included at both levels of analyses and cross-level interactions can be modeled

• Enables modeling of structural level influences on stigma and health

• Can be used for analysis of intensive longitudinal data by accounting for correlation of observations within person over time

• More difficult to explain to lay audiences, including policymakers and funders in some cases

• Requires data collection in multiple contexts and, in some cases, may require existing data at higher levels (e.g., state or country level data)

• Exploration of whether the relationship of gender, class, and race to self-rated health varied by neighborhood [63]

• Assessment of how country-level and individual level factors interact to influence the mental health of male sexual minority European migrants [19]

• Examination of how everyday experiences of discrimination impact internalized stigma among people living with HIV using a smartphone-based experience sampling method survey [64]

• When multiple time points are available or data is available from multiple clusters (the number necessary will vary, but 10–15 would be considered few clusters for an analysis [108]) and contextual influences on the relationship between stigma and health are of interest

Structural equation modeling

Allows for simultaneous estimation of measurement and structural components, including pathways between observed and latent variables

• Appropriately models measurement error associated with inclusion of latent variables

• Flexible strategy: can simultaneously assess the impact of multiple exposures on multiple outcomes, include group-based or multilevel modeling, and assess moderated mediation or mediated moderation [109]

• Can assess how exposures and outcomes predict each other over time

• Modeled relationships may be inappropriately interpreted as causal

• Depending on the number of parameters included, may require larger sample sizes to be estimable

• Not all models may be identifiable and sensitive to model misspecifications

• Simultaneous assessment of experiences of racial discrimination and HIV-related stigma on quality of life among African and Caribbean Black women in Canada [74]

• For estimating complex models including multiple stigma-related factors as predictors or multiple related health outcomes of interest, particularly when including psychosocial variables that are not directly observable (e.g., stress, coping)