This study aimed to estimate disparities in fruit and vegetable intake between groups at the intersection of educational level and gender in northern Sweden, and to assess the discriminatory accuracy of the intersectional groups. Substantial differences between intersectional groups were found in the prevalence of inadequate intake of fruits and vegetables combined, and separately. Gender made a greater contribution overall to the disparities than education did, and the excess intersectional disparities were negative and similar across outcomes. In addition, the magnitude of the joint disparities varied by outcome. The discriminatory accuracy of the covariates and gender and education was moderate and suggested an importance of these two inequality dimensions, but their cross-classification did not contribute to any additional discriminatory power.
The finding that the joint disparity for fruit intake was lower compared to the other outcomes can be explained by a higher prevalence of inadequate intake among high educated women compared to the two other outcomes (60.2% for fruit intake, vs 47.7, and 51.8% for the other two), whereas it was similar across outcomes for low educated men (83.4% for fruit intake, vs 81.4 and 84.9% for the other two).
We also found that the referent disparity for gender consistently explained a large portion of the joint disparity. This is in line with previous research indicating that women have in general a higher fruit and vegetable consumption than men [6, 19, 20], and healthier behaviors overall [15,16,17,18]. Although to a lesser extent, the referent disparity for education also explained a substantial fraction of the joint disparities. This is consistent with the literature on educational inequalities in fruit and vegetable consumption [6,7,8,9,10, 12, 13], and other types of health behaviors [11, 16, 53]. Several potential mechanisms have been proposed to explain this relationship between education level and health behaviors [14].
The finding that the excess intersectional disparity was similar across outcomes illustrates the partial overlap that exists between fruit and vegetable intake. The contribution of the excess intersectional disparities to the joint disparities observed here points to the potential usefulness of intersectional approaches in epidemiology. The intersection of multiple disadvantages for fruit and vegetable intake appeared to be associated with a unique pattern of consumption not entirely explained by the addition of gender and education: the intersection of male gender and low education appeared to be unexpectedly protective against unhealthy dietary patterns, in comparison to what would be expected from the addition of the two referent disparities.
As pointed out by Jackson et al. [33], it should also be noted that the existence and magnitude of the joint disparity should not be dismissed if the excess intersectional disparity is non-existent or negative. Even if a group does not display any excess intersectional disparity, it can still experience the sum of the two referent disparities. In this study, the joint disparity in consumption of fruits, vegetables, and fruits and vegetables between high educated women and low educated men was still great in magnitude; inadequate intake of fruits and vegetables combined was 34.6 pp. more prevalent for low educated men than for high educated women.
The discriminatory accuracy of the intersectional groups for identifying inadequate intake of fruits and vegetables as measured by the AU-ROC was close to - but below - 0.7 and could therefore be considered moderate. Most importantly, the lack of additional discriminatory power of the cross-classification found in this study is recurrent in empirical studies using intersectional approaches to study inequalities in Sweden [34, 38, 39]. Those two findings would be commonly interpreted as evidence in favor of universal rather than targeted intervention in the context of proportionate universalism. For the present study, this would suggest that the targeting of specific intersectional groups is not suitable as the sole guiding principle for health promotion planning. However, while this interpretation is straightforward when the goal is general health promotion in the population, e.g., increase fruit and vegetable intake, it is not as evident when the goal is promotion of equity in health, e.g., reducing inequalities in fruit and vegetable intake.
Moreover, there is no clear-cut threshold that would label a discrimination as high enough for intersectional approaches in epidemiology. Indeed, examinations of both novel and traditional risk factors for coronary heart disease have yielded similarly disappointing results when it comes to discriminatory accuracy [54], despite many of them being established in clinical practice. While discriminatory accuracy represents a necessary complement to means-centric examinations of health inequalities, and to traditional epidemiology more generally, a cautious interpretation of its implications specifically in the context of equity promotion is therefore warranted.
Further research and public health relevance
To our knowledge, this Swedish study is the first of its kind investigating disparities in fruit and vegetable consumption across intersectional groups defined by gender and education. The intersectional approach was helpful to shed light on the complex inequality patterns of inadequate intake of fruits and vegetables. In addition, as the estimated disparities are additive rather than relative measure of inequality, they can be used to determine the absolute benefits in population health if a disparity were to be eliminated [33].
The findings from this study will hopefully help to give a more nuanced understanding of the structural patterning of disparities in diet, and thereby improve the effectiveness of equity-promoting public health interventions and policies. Although the moderate discriminatory accuracy in this study could suggest that universal interventions and policies might be more effective for improving diet quality in Sweden, there is a risk that they would mostly benefit those already advantaged, and could therefore contribute to widening inequalities in health [55]. A combination of universal and targeted measures to improve diet quality could be a way forward, in line with proportionate universalist principles [56]. This study could then contribute to giving guidance as to where are the largest disparities, and what are the most relevant groups to prioritize.
Future research on diet quality could consider operationalizing groups at the intersection of gender, education, and other inequality dimensions such as age, sexual orientation, migration and ethnicity, or geographical area of residence. This could help identify intersectional groups with higher discriminatory accuracy and facilitate targeted interventions to improve health equity. In addition, qualitative or mixed methods could be considered, as they could shed light on the perceptions and motivations of individuals located at specific intersections of social positions and identities that may explain the disparities observed in this study.
Strengths and limitations
Since random sampling procedures were used to draw a sample from the general population, selection bias is limited to some extent. However, the response rate varied between 42.4% for Västerbotten [46] and 45.8% for Jämtland/Härjedalen [44]. Since most people contacted to answer the survey did not answer it, it is unknown whether the final sample is truly representative of the target population.
The year when the Health on Equal Terms survey used in this study was conducted, 43% of the population aged 25–64 had some form of post-secondary education [57]. In our study, 37% are in that same category (Table 1), which is slightly lower but can be explained by the presence of survey participants older than 64 years old and by the increase in education level in the last decades [57].
It is not known to which degree the results can be generalized outside of Sweden. Indeed, the implications of the intersection of gender and education can be expected to differ between contexts, depending on the specific configuration of the gender (in)equality structure, of the educational system, and of equity-promoting policies. In addition, the results cannot readily be generalized to other countries and settings without consideration of the different patterns of health and nutrition [58,59,60].
The use of gender as one of the two variables for the operationalization of the intersectional groups can be questioned. Since this study was interested in investigating social inequalities in health behaviors, and not potential biological inequalities in health, gender seemed more appropriate as designating a social construct rather than a biological construct. However, what is measured by available register data is biological sex recorded at birth, and thus might not always represent gender identity accurately, especially since gender identities cannot be fully captured by a binary variable. The adequateness of using one term or the other in epidemiology has been discussed elsewhere [61].
In addition, dichotomizing educational attainment may have led to loss of information, since individuals with different educational characteristics were categorized into the same group. For example, individuals who did not complete high school were grouped together with individuals who did. The patterns of fruit and vegetable consumption might differ between those two groups, but this cannot be inferred from the present study.
The three outcome variables in this study were dichotomized as well, also potentially leading to loss of information. However, we dichotomized the outcome for combined intake of fruit and vegetable based on recommendations from the Public Health Agency of Sweden [43]. In addition, because one serving reported in the survey is usually estimated to be close to 100 g [43], fruit and vegetable intake resembled amounts recommended in Sweden and worldwide. Defining adequate intake as “twice a day or more” for fruit and vegetable (each) theoretically yields 400 g or more in total. This is close to the 500 g recommended in Sweden [5] and to the 400 g or 5 servings recommended by the World Health Organization [3]. Moreover, the three outcomes showed a similar prevalence of inadequate intake (Table 1). The choice of dichotomization for the outcomes is therefore intended to correspond to policy recommendations.
However, although it was assumed that the frequency of fruit and vegetable intake reported was, on average, proportional to the amount consumed, we could not assess whether this was in fact the case. Reporting bias could be a concern and over-reporting may have, for example, occurred more in those with higher education since they might be more conscious of the existing recommendations for fruit and vegetable intake. Over-reporting may also have occurred due to social desirability bias [62]. Under-reporting may have occurred among those with relatively lower education due to potentially weaker health and nutrition knowledge. Misreporting could also have occurred for other reasons, such as having a high body mass index or obesity [63], or cognitive decline in older adults in the sample, although those older adults may have not answered the survey to start with. This could therefore contribute to selection bias rather than information bias. In contrast, the exposure variables and covariates came from high-quality register data and therefore were not affected by self-report bias.
Since this study is cross-sectional in design, it cannot be used to determine causality, even though reverse causality is unlikely. In addition, despite adjusting for additional covariates, it is not impossible that some residual confounding may subsist. However, the associations found between intersectional group and prevalence of inadequate fruit and vegetable intake are not easily affected by confounders. Indeed, most third variables such as income or occupation are likely to be determined in part by gender and education, and if included would risk over-adjustment and result in under-estimation of the disparities.
Finally, the method described by Jackson et al. [33] allows to quantitatively investigate disparities at the intersection of multiple social positions and identities, in a way that corresponds to the different possibilities for an intersectional group to experience disadvantage [24]. However, this approach limits the number of social positions that can be considered simultaneously [33]. In addition, it measures additive interactions, whereas the AU-ROC is based on multiplicative interactions. This may partly explain why measures of discriminatory accuracy did not identify any intersectional interaction from Model 2 to Model 3, whereas substantial excess intersectional disparities were estimated for the three outcomes.