Multilevel analysis of individual heterogeneity and discriminatory accuracy (MAIHDA) within an intersectional framework
Introduction
The recent study by Evans, Williams, Onnela, and Subramanian (Evans et al., 2017) (termed as ‘EWOS study’ in the remainder of my commentary) provides an innovative methodological contribution to the investigation of socioeconomic, ethnic and gender disparities in health and disease risk. This study gives an excellent answer to the call for suitable quantitative methodologies within the intersectionality framework (Bauer, 2014, Bowleg, 2012, Evans, 2015, Green et al., 2017, Wemrell, 2017, Wemrell et al., 2017a). The multilevel approach to modelling health inequalities at the intersection of multiple social dimensions provides a better understanding of the health heterogeneity existing in the population. It also gives inductive information on the potential existence of a large number of stratum-specific interactions of effects in a powerful and innovative way and, simultaneously, increases and clarifies the discriminatory accuracy of the information (Merlo, 2003, Merlo et al., 2016, Merlo et al., 2017, Merlo and Wagner, 2013a, Merlo and Mulinari, 2015, Rockhill, 2001). The methodology developed in the EWOS study should become the gold standard for investigating health disparities in (social) epidemiology.
Many readers of social epidemiology will place multilevel regression analysis within the realm of neighborhoods and health, where the context is defined by geographical administrative boundaries. The innovative aspect of the EWOS study is that, in applying an intersectional framework, the authors model the individual health outcome (i.e., BMI) using a multilevel regression analysis of individuals nested within a matrix defined by combinations of variables (i.e., gender, race/ethnicity, income, education and age). A similar multilevel analytical approach has also been described by (Jones et al., 2016) in quantitative behavioral social science.
The intersectional multilevel regression analysis does not substitute the classical fixed effects analysis (Mulinari et al., 2015, Mulinari et al., 2017, Wemrell et al., 2017b), but, compared with it, the multilevel approach has many technical and conceptual advantages when it comes to model parsimony and precision-weighted estimations of small intersectional strata averages. It also allows for decomposing the individual variance into within and between intersectional strata and, thereby, evaluates the role of those strata for understanding individual heterogeneity in the health outcome. However, the most salient advantage is that this approach allows for the improved and inductive analysis of numerous stratum-specific interactions.
Because the EWOS study is a well-written tutorial, in this commentary, I do not repeat the explanation of the statistical/epidemiological concepts. Neither will I discuss the differences between the conventional fixed effects approach and the multilevel approach, as the EWOS study does it in an outstanding way. Instead, inspired by the EWOS study, I will share with the reader a number of thoughts on the theoretical consequences of using intersectional multilevel analysis, or, more generally, multilevel analysis of individual heterogeneity and discriminatory accuracy (MAIHDA) in (social) epidemiology in general and within the intersectional framework in particular. MAIHDA is a reorganization of concepts that allows for a better understanding of the distribution and determinants of individual health and disease risk in the population. This is precisely the type of analysis the EWOS study has applied.
Section snippets
Considering dimensions of social identity and position as contexts rather than as individual characteristics
Frequently gender, race/ethnicity, income, education and age are analyzed as “individual” characteristics in simple studies of socioeconomic gradients in health based on probabilistic associations. Accordingly, those variables tend to be interpreted as types of private, individual “risk factors”. However, the EWOS study makes it very clear that the quantitative intersectional analysis does not involve identifying ever-narrower and more specific “risky identities”. Instead, the EWOS study
Obtaining information on General Contextual Effects to evaluate the discriminatory accuracy of the intersectional strata
The use of intersectional multilevel regression analysis indispensably assumes that the individuals in each intersectional strata share a similar context of oppression or privilege, which conditions their health status (e.g., BMI) over and above their own individual heterogeneities. This intra-stratum dependence or similarity can be neatly expressed by means of the intra-class correlation coefficient (ICC) measure (Formula (1)), which indicates the share of the total individual variance in
The concept of discriminatory accuracy in social epidemiology
The ICC is actually a measure of discriminatory accuracy, analogous to the area under the receiver operating characteristic curve (AUC). These measures are traditionally used in Epidemiology for assessing the predictive performance of an exposure categorization (e.g., a risk factor, a screening marker) (Pepe et al., 2004, Wald et al., 1999). Measures of DA have already been applied in intersectional quantitative analyses (Mulinari et al., 2015, Mulinari et al., 2017, Wemrell et al., 2017b) and
The tyranny of the averages
The imposition of the average group value on the individual is the rule in (probabilistic) risk factors Epidemiology and it has been denominated the “Tyranny of the means” or the “Mean-centric approach” (Downs and Roche, 1979). The key concept is that common measures of average association correspond to abstractions that do not represent the heterogeneity of individual effects. This idea points to the study of inter-individual heterogeneity around group averages as fundamental to understanding
Individuals” and “groups” are not dislocated study objects: we need a new concept of population health
All social epidemiologists I know (including myself) accept the ideas of Rose's distinction between causes of individual cases and causes of population incidence. The same holds true when it comes to the acceptance of similar eco-social frameworks (Krieger, 2001), including (Susser and Susser, 1996) eco-epidemiology and its critique of an individualistic risk factor epidemiology. The problem is that current eco-epidemiology has not only failed to reduce the use of risk factors in epidemiology,
Applying MAIHDA on intersectional analysis in social epidemiology
Most empirical research framed within intersectional theory has been qualitative, and quantitative analyses are still seen with some hesitancy in this area (Bauer, 2014, Bowleg, 2008). Hancock (2013) identifies intersectionality as a normative theory that takes the insights as logical priors to research questions and, as such, does not need to meet the standard of falsifiability. It is assumed that intersectionality can be empirically operationalized but it is not at all clear that the only way
In conclusion
Public Health and Social Epidemiology need to develop innovative research methodologies bringing together a multidisciplinary perspective including, among others, expertise in the fields of medicine, epidemiology, statistics, sociology, psychology and anthropology. Social epidemiology has enriched epidemiology and public health medicine by incorporating theories and methods from humanities and social sciences and this work will continue (Kawachi and Subramanian, 2017). Epidemiology is political
Acknowledgements
I will express my sincere gratitude to my colleagues at the Unit for Social Epidemiology, Lund University Sofia Zettermark, Sten Axelsson Fisk and especially to Maria Wemrell and Shai Mulinari for everyday discussions and for their feedback on the last version of this commentary. This study was supported by a grant from the Swedish Research Council (# 2017-01321), The Faculty of Medicine, Lund University and The Region Skåne County Council, Sweden.
References (60)
Incorporating intersectionality theory into population health research methodology: challenges and the potential to advance health equity
Soc. Sci. Med.
(2014)- et al.
Causal inference challenges in social epidemiology: bias, specificity, and imagination
Soc. Sci. Med.
(2016) - et al.
Can intersectionality theory enrich population health research?
Soc. Sci. Med.
(2017) - et al.
The tyranny of the averages and the indiscriminate use of risk factors in Public Health: the case of coronary heart disease
SSM - Popul. Health
(2017) - et al.
Marginalized identities, discrimination burden, and mental health: empirical exploration of an interpersonal-level approach to modeling intersectionality
Soc. Sci. Med.
(2012) - et al.
An intersectional approach to multilevel analysis of individual heterogeneity (MAIH) and discriminatory accuracy
Soc. Sci. Med.
(2017) - et al.
Intersectionality and risk for ischemic heart disease in Sweden: categorical and anti-categorical approaches
Soc. Sci. Med.
(2017) - et al.
Intermediate and advanced topics in multilevel logistic regression analysis
Stat. Med.
(2017) - et al.
Measures of clustering and heterogeneity in multilevel Poisson regression analyses of rates/count data
Stat. Med.
(2017 Nov 8) - et al.
The median hazard ratio: a useful measure of variance and general contextual effects in multilevel survival analysis
Stat. Med.
(2017)