An Intersectional Analysis Providing More Precise Information on Inequities in Self-Rated General Health

Intersectionality theory combined with an analysis of individual heterogeneity and discriminatory accuracy (AIHDA) can facilitate our understanding of health disparities. This enables a more precise application of proportionate universalism for resource allocation in public health. Analyzing self-rated general health in Sweden, we show how an intersectional perspective allows for a detailed mapping of health inequalities while avoiding simplication and stigmatization based on indiscriminate interpretations of differences between group averages. we We We intersectional


Background
Health disparities have been documented by a wealth of epidemiological research for some decades (1, 2), but remain a pressing concern. Such inequalities have often being captured in the form of socioeconomic gradients (3,4) whereby people in higher positions on the social ladder enjoy better health and quality of healthcare than people in lower positions. While socioeconomic position has received a large share of this research interest (5), disparities between groups de ned by, i.e., gender (6), race, ethnicity, immigration status or racialization (7), have also been amply documented. However, some limitations to this body of research remain.
Studies of health inequalities have typically focused on one dimension at a time, such as socioeconomic position or gender, thus paying inadequate attention to how such dimensions may interact. Meanwhile, much health disparities research has tended to construe inequalities in terms of levels of risk located in or borne by individuals or groups, rather than addressing dynamics between individual or groups (8,9), or the processes or driving forces which give rise to inequalities (10). Furthermore, health disparities research has been critiqued for insu ciently considering heterogeneity, through focusing almost exclusively on group average risk rather than on variations within and overlaps between groups. This may contribute towards simpli cation or essentialization of differences between groups, as well as to unjust stigmatization of "high-risk" groups or individuals (11,12).
Due to its potential to address these concerns, intersectionality theory has increasingly been promoted and adopted in quantitative health disparities research (e.g., 10,13,14,15). In this study, we apply an intersectional perspective combined with an analysis of individual heterogeneity and discriminatory accuracy (AIHDA) (12,16,17) to the study of disparities in self-rated health levels in Sweden. This is done in order to obtain a more detailed mapping of health inequalities while mitigating simpli cation and stigmatization based on indiscriminate interpretations of differences between group average risk.
An intersectional perspective on population health research Intersectionality theory, formulated and advanced by theorists including Crenshaw (18) and Hill Collins (19), centers on the understanding of social categorizations such as gender, class and race/racialization as interconnected rather than separate, and as creating overlapping and interacting systems of discrimination or disadvantage. The principal idea is thus that the social categorizations conditioning the distribution of resources and power, and thus health, need to be considered as interlocked rather than as unidimensional. In the context of quantitative population health research, an intersectional perspective thereby motivates the study of strata de ned by the combination of several socioeconomic dimensions (e.g., age, gender, income, racialized identity and sexual orientation), contrasting with conventional analysis of socioeconomic gradients in health often based on singular dimensions. In this manner an intersectional perspective can improve the mapping of inequities in health and therefore better illustrate patterns of disadvantage.
Such improved mapping of disparities ts well within the current movement towards precision public health (20). Relatedly, it can support the implementation of proportionate universalism in resource allocation in public health (4,21). According to the principle of proportionate universalism, as formulated by Marmot and Bell (4), interventions aiming to ameliorate health disparities should be directed at the whole population (i.e., be universal) and be combined with targeted actions of a scale and intensity proportional to the level of need in speci c population groups.
Further, applying an intersectional perspective means directing interest towards dynamics of power and wealth distribution in society, rather than to levels of risk as attributes of individuals or groups, in the interest of facilitating the amelioration of health inequalities through social change (22,23). Accordingly, the intersectional strata constructed in this study should be considered in terms of social contexts (24) rather than as characteristics of individuals. This can mitigate the risk of excessive biomedical reductionism threatening current precision-based public health (25), while reducing the likelihood of 'blaming the victim' as frequently discussed when investigating socioeconomical differences conceptualized at the individual level.
In an in uential classi cation of intersectional research, McCall (26) distinguishes between anti-, inter-and intra-categorical approaches (for further discussion see (27)). Epidemiology principally consists of the quantitative analysis of demographic, socioeconomic and biomedical population categories, and thereby per se adopts an inter-categorical (henceforth referred to as categorical) approach. However, the population categories under study should be evaluated in relation to their discriminatory performance as classi ers, i.e., their capacity to accurately classify the individuals according to the health outcome of interest (28,29). Such evaluation can serve to prevent the "tyranny of the averages" (30) through which the same average value is attributed to all the members of a group without considering the individual heterogeneity around the group average or any overlap between categories. If the discriminatory accuracy (DA) is low, the validity or relevance of the categorization can be questioned in relation to the speci c outcome at hand. In this sense, an anti-categorical stance can be adopted. This is important for the purpose of avoiding simpli cation or essentialization of differences between groups, alongside under-or overtreatment and ineffective public health interventions (31).

Self-rated health
Measures of self-rated or self-assessed health, through which individuals are asked to evaluate their own health status typically on a four-or ve-point scale, are widely used in population health research. In terms of what it actually assesses and how it is linked to objective medical outcomes, this measure is not entirely understood (32). It is subjective, non-speci c and encompasses cognitive, cultural and medical, or social and biological, dimensions (33). However, its associations with mortality have been repeatedly demonstrated, for different population groups and in various countries including Sweden (33)(34)(35). Its predictive power has in fact been noted to often be stronger than that of more objective medical factors (36).
For example, the impacts of class, gender, race, migration and sexual orientation in Canada have been investigated (43,44), as has those of gender, sexual orientation and race in the United States (45,46), and of class, gender and regional context in Spain (47). Common to these studies, however, is that they adopt a categorical approach focused on between-group differences in average risk, without assessing individual heterogeneity and thus potentially allowing for a complementary anti-categorical stance.
In the present study, we use intersectional categorization combined with an analysis of individual heterogeneity and DA (AIHDA), in order to improve our understanding of inequalities in self-rated health, related to income, gender and immigration status in Sweden. Our focus lies on income, gender and immigration status partially due to the possibilities and constraints of the NPHS data, but mainly because these dimensions correspond with the categories perhaps most typically included in the intersectional study of health disparities: gender, class and race (48). While immigration status only loosely correlates with concepts of race or ethnicity, con ating issues related to racialization, migration and citizenship (13), which also concern groups in Sweden other than rst generation immigrants, immigration status is a categorization central to processes of racialization in contemporary Sweden (49).

Study population
The study is based on data from 14 consecutive National Public Health Surveys (NPHS) performed in From this sample we excluded people with missing information on income (n=922) or self-reported health (n=2,135). Thus, the nal study population consisted of 133,244 persons, i.e., 97.8% of the initial sample ( Fig 1).
The study was approved by the Swedish Ethical Review Authority (Dnr: 2019-01793) and the Data Safety Committee at the Public Health Agency of Sweden.

Variables
Our outcome variable was self-rated general health, as assessed by the NPHS participants, based on responses to the question "How do you assess your general state of health?", according to a ve-point scale (very good; good; fair; bad; vary bad). The response options were recoded, in accordance with recommendations from the Public Health Agency (51), into good (very good; good) and bad (fair; bad; very bad). The average prevalence of bad self-rated general health has been relatively stable over the survey years, although some uctuations are present (see Table 1). Due to these uctuations, our analyses adjust for survey year.
Gender was classi ed as a binary variable distinguishing between men and women, as these were the only response options enabled by the NPHS questionnaire.
Information on household income provided by the NPHS divided the income data into tertiles (low; medium; high income). The high-income group encompassed the 20% highest reported incomes, the lowincome group comprised the 20% lowest reported incomes and the medium-income group included the remaining 60% of reported incomes.
We classi ed immigration status as native, i.e., born in Sweden, or immigrant, i.e., born in another country.
As a way of operationalizing intersectional contexts, we created 12 strata by combining the two categories of gender, the three income categories and the two categories of immigration status. We used native men with high income as the reference in the comparisons, as this group was assumed to occupy the position of greatest structural privilege.
In the analysis, we adjusted for age.

Statistical analyses
We performed ve consecutive regression analyses, modelling bad self-rated health as the dependent variable, using the explanatory variables alone and in combination and adjusting for age and survey year.
Cox proportional hazards regression with a constant follow-up time equal to one was used in order to obtain prevalence ratios (PR) (53). We calculated 99% rather than 95% con dence intervals (CI) in order to reduce the problem of multiple comparisons. As prevalence ratios do not provide information about overlaps between groups, measures of DA provide important information for the evaluation of between-group differences (31). For each model, we quanti ed the DA by means of the area under the receiver operator characteristics curve (AUC) (29). The AUC is calculated by plotting the true positive fraction (i.e., sensitivity) against the false positive fraction (i.e., 1speci city) for different binary classi cation thresholds of the predicted probability of bad self-rated health. Thus, the AUC measures the accuracy of the information provided by the variables in the model for discriminating individuals with bad self-rated health from those with good self-rated health. The AUC takes a value between 0.5 and 1, where 1 indicates perfect discrimination and 0.5 means that the studied variables have no DA at all.
There is no fully established practical guideline for the interpretation of the size of the AUC as a measure of DA when analyzing intersectional inequalities. However, based on the classi cation provided by Hosmer and Lemeshow (54), we qualify intersectional inequalities according to the DA as (i) "absent or very small" (AUC=0.5-0.6), (ii) "small" (AUC=0.6-0.7), (iii) "large" (AUC=0.7-0.8) and (iv) "very large" (AUC>0.8).
We further calculated the incremental change in the AUC value (Δ-AUC) between the models. The Δ-AUC quanti es the improvement in the DA obtained by a model, in relation to a reference model (15). The intersectional strata used in model 5 allows for the capturing of any statistical interaction of effects. If any such statistical interaction exists, the DA of model 5 will increase in comparison to that of model 4.
We used IBM SPSS (Statistical Package for the Social Sciences) version 22 for PC to perform all statistical analyses.

Results
The characteristics of the study population, and the prevalence of bad self-rated general health in the respective subgroups, are presented in Table 2. As can be seen in Table 3, the regression models 2-4 show the presence of inequalities in self-rated health in Sweden. Women had a higher risk of bad self- Among the intersectional strata, those comprising males showed lower risk of bad self-rated health than those including females, in all income and migration status categories (i.e., male natives with a high income had a lower risk than female natives with a high income, etc.). Strata including persons born in Sweden had lower risk than those encompassing immigrated persons, in all income and gender categories (i.e., native females with medium income and a lower risk than immigrant females with medium income, etc.). Similarly, strata comprising persons with high income showed a lower risk than those including persons with medium or low income, in all gender and immigrant status categories (i.e., low-income, native men had a lower risk than medium-and high-income native men, et c.).
At the same time, females were incorporated not only in the stratum with the highest risk, but also in the one with the second to lowest risk. Immigrant status was a feature of the four strata experiencing the highest risk, but also of two of the six strata experiencing lower risk. Less complexity was evident with regard to income, as all strata comprising persons with low income were among those experiencing the highest risk.
This presence of complexity and heterogeneity is con rmed by the measures of DA. The Δ-AUC shows that including gender in model 2 did not add much to the DA obtained by age and survey year alone (Δ-

Discussion
Contributing to an intersectional and thus more precise epidemiological perspective based on AIHDA (15,27), our study provides a developed understanding of the socioeconomic and demographic distribution of bad self-rated general health in the Swedish population. By showing the presence of a clear intersectional social gradient, we demonstrate the relevance of an intersectional perspective in population health research. As intersectional AIHDA goes beyond conventional analysis of socioeconomic gradients based on probabilistic measurements of risk (e.g., odds ratios, prevalence ratios or relative risks) (55), we furthermore incorporated information on DA to show substantial individual variabilities or heterogeneities of links between self-rated health and income, gender and immigration status in Sweden.
The intersectional social gradient shows the salience of gender, immigration status and income as strati cation forces (13). At the same time, it reveals that the impact of these dimensions on the outcome are affected by each other (13). While all strata comprising women had a higher risk than those including males with the same income and immigration status, strata comprising women showed both relatively low and very high risk of self-rated health. This is in line with the global situation in which some primarily white and high-income females inhabit very privileged and powerful positions, while women are still overrepresented among the world's poorest and most precarious population groups (56). This, in turn, relates to the foundational insight of intersectionality scholarship that the positions of gendered subjects are fundamentally mediated by factors including racialization and class (19).
Along similar lines, all strata encompassing immigrants had a higher risk than those including natives with the same income and gender. Meanwhile, whereas the majority of strata comprising persons with immigrant status were among those experiencing very high risk, two strata -those including males and females with high income -carried a lower risk. This shows that the relevance of immigration status for self-rated health in Sweden is affected by factors including gender and income. With regard to the social gradient, the association between higher risk and lower income was quite consistent. While the Δ-AUC revealed no statistical interaction between the variables comprising the intersectional strata, these results show that the dimensions do indeed interrelate with each other with regard to the outcome.
Meanwhile, although differences between the average risk of intersectional strata are evident, the DA of the dimensions and strata is low. This means that although awareness of average differences in self-rated health is important in the interest of social change towards greater equity, average differences should not be extrapolated into generalized assumptions about speci c groups or individuals, due to substantial individual variability.
One limitation of this study lies in its observational and cross-sectional design, which allows for the study of correlations but does not enable the drawing of conclusions concerning causal relationships.
Furthermore, the variables used infer certain limitations. The outcome measure is self-assessed, and it is possible that different groups of people can have different reporting behavior (32), i.e., assess their health differently even if objective health measures are the same, and vice versa.
The act of assessing one's own health involves a cognitive process in uenced by contextual factors including an individuals' understanding of the meaning and content of the concept of health, frameworks for evaluating their own health, and norms or references to which they compare themselves to (33). Such understandings, frameworks and norms can obviously be historically and culturally contingent, and differences in reporting behavior between population groups have indeed been documented (33). For example, elderly people in countries including Sweden (57) have been found to downplay functional limitations and chronic conditions when evaluating their health. This implies that measures of self-rated health are likely to overestimate the health of older persons (32,58). For this reason, we adjusted for age in our analysis instead of including age as an explanatory variable.
Differences in reporting behavior have been found between countries or cultures (33) and between groups with different socioeconomic status. In countries including the United States, Canada and India, persons in lower socioeconomic groups have tended to rate their health more optimistically than persons in higher socioeconomic groups, as compared to "objective" health factors or diseases (32,59,60). Women have also been found to be more optimistic than men about their health (32). Such studies suggest that measures of self-rated health may overestimate the health of women and persons in lower socioeconomic groups, and thus underestimate health disparities. However, studies from Sweden found no signi cant differences in reporting of self-rated health between socioeconomic groups (35,57). While the latter studies support the comparability of the groups included in this study, the former suggest that any existing reporting bias is likely to under-rather than overestimate health inequalities.
As for further limitations, our de nition of gender as "male" or "female" was constrained by the response options offered by the NPHS, which disregads the more accurate continuum representation of both sex and gender (61). Relatedly, from an intersectional viewpoint it would have been interesting to include dimensions such as sexual orientation, cis-versus transgender and functional diversity, in order to improve the capturing of existing disparities and heterogeneities in self-rated health. However, data availability and the size of the database prevented the addition of such information.
Furthermore, our categorization of immigrated versus native-born individuals is simplistic. It con ates issues related to racialization, citizenship, migration and trauma (13), disregarding differences between groups such as refugees and immigrants from other Nordic countries (38), and excluding other groups that are affected by processes of racialization in contemporary Sweden, such as Sami populations, Jewish Swedes and second generation immigrants. Previous research has indeed critiqued the use of the category of immigrated versus native-born in the study of self-rated health in Sweden, as it encompasses large variety within and overlap between groups (62). Still, immigration status is central to processes of racialization in contemporary Sweden (49). While pointing to the "bad self-rated health" of certain population groups might be overly simplistic, essentializing and in alignment with xenophobic tendencies (63), the neglect of any existing health disparities can also prevent societal changes toward improvement.
Considering the DA of the immigration status dimension and of the intersectional strata in relation to the studied outcome aims to help resolve this dilemma (30,62).
Finally, while intersectionality theory originates in qualitatively and theoretically-oriented research, and with some researchers questioning its commensurability with quantitative analysis (64) we side with others who have pointed to the importance of applying intersectional approaches in quantitative population health research (13)(14)(15). This study could have been performed using multilevel AIHDA (i.e., MAIHDA), as described (12,35) and implemented (29)(30)(31)(32)(33)(34) elsewhere. However, our purpose was to provide a valid but simple alternative for quantitative intersectional analyses, suitable for the monitoring of health inequities in routine public health surveys.
With regard to the implications of this study, intersectional AIHDA provides information enabling public health interventions with a more accurate focus. Using this approach, we can better evaluate to which degree a universal intervention needs be proportionated to the level of the disfranchisement of speci c groups. In this way, intersectional AIHDA can contribute to precision public health within Marmot's framework of proportionate universalism (4, 21).

Conclusion
By mapping interrelating socioeconomic contexts and assessing the presence of individual heterogeneity through measures of DA, intersectional AIHDA offers a fruitful analytical approach for the investigation of health inequalities. Analyzing average risks, we found a clear social gradient of self-rated general health in Sweden. However, improving (self-rated) health is obviously a universal endeavor relevant to all intersectional strata in society. The low DA speaks against focusing only on speci c strata as it would be ineffective and could potentially lead to unnecessary stigmatization. An intersectional AIHDA approach could guide public health interventions according to the principle of proportionate universalism.

Consent for publication
Not applicable.

Availability of data and materials
The dataset analyzed during the current study is available from the Public Health Agency of Sweden (52) on request.

Competing interests
The authors declare that they have no competing interests.

Funding
This work was supported by grants to Juan Merlo (PI) from the Swedish Research Council (Vetenskapsrådet) for the project "Multilevel Analyses of Individual Heterogeneity: innovative concepts and methodological approaches in Public Health and Social Epidemiology" (https://www.swecris.se/betasearch/details/project/201701321VR?lang=en. Project id: 2017-01321. It was further supported by a research scholarship from the Research Studies Board, dedicated to students at Lund University's Faculty of Medical Sciences, awarded to Nadja Karlsson. The funders had no role in the study design, data collection and analysis, decision to publish or preparation of the manuscript.
We acknowledge the staff and participants of all the national health surveys who have made this study possible.
Authors' contributions MW contributed to the design of the study and the direction of the analysis, interpreted the results, wrote the manuscript and critically revised the work for important intellectual content.
NK contributed to the analysis, to the writing of the manuscript and to critical revision for important intellectual content.
RPV contributed to the preparation of the database and the analysis, as well as with its interpretation and the drafting and critical revision of the manuscript. JM took the initiative for the study, designed the study, acquired the data, directed the analysis and contributed to the interpretation, the drafting of the manuscript and its critical revision for important intellectual content.
All authors have approved the nal version of the manuscript and agree on all aspects of the work and ensure that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.