Classification tree analysis for an intersectionality-informed identification of population groups with non-daily vegetable intake

The present study is based on the National Health Telephone Interview Survey ‘GEDA - German Health Update’ [GEDA 2009]. This cross-sectional, representative survey is repeatedly conducted by the Robert Koch Institute (RKI) as part of a nationwide public health monitoring system for diseases and health behaviour. The random sample of telephone numbers for the computer-assisted telephone interviews of the GEDA survey were drawn according to the Gabler-Häder method [19]. The participants of GEDA 2009 were randomly selected, German speaking adults (age range: 18–100 years), who were registered in Germany (n = 21,262). The cooperation rate for participants, measured as the proportion of realised interviews with individuals that have been contacted was 51.2% [19]. Further details about design, methods and nationwide representativeness of the study population of the GEDA 2009 survey have been described elsewhere [20]. The final study sample of the present study comprised 19,512 participants overall.

Frequency of vegetable intake was assessed by asking study participants the following question: How often do you eat vegetables? The following answer categories were offered: every day, at least once a week, less than once a week, never/ I do not know. The variable for vegetable intake in the present study was dichotomized into daily (every day) vs non-daily (at least once a week, less than once a week, never/ I do not know).

We selected the following nine variables of GEDA 2009, capturing different sociocultural, socio-demographic and socio-economic dimensions: Sex/gender [forced choice: female, male], Age [in years: 18–29, 30–39, 40–49, 50–59, 60–69, 70–79, 89+]; Education [categorized according to ISCED 1997 EU-classification: high, medium, low]; Employment status [full-time, part-time, occasionally, not working], Professional status [blue-collar worker, white-collar worker, civil servant, freelancer, helping family, no profession, else]; Marital status [married, married - living separately, unmarried, divorced, widowed]; Disability status [yes, no]; Migration background [21] [two-sided (non-German citizenship, respondent immigrated to Germany after birth or both parents not born in Germany), one-sided (one parent not born in Germany), no (without migration background)]; Urbanity/rurality [big city, city, rural, very rural].

In reference to a gender concept which has been developed at the Canadian Institutes of Health Research and has also served as a basis for the development of a composite measure of gender [22, 23], we defined six available variables of GEDA 2009 as sex/gender related aspects indicating possible mechanisms underlying sex/gender differences in health [24, 25]: Family constellation [with partner and child(ren), with partner and no child(ren), no partner and with child(ren); no partner and no child(ren)]; Main earner status [main wage earner in the household: one person household, respondent herself/himself, partner, other, there is no main wage earner]; Perceived social support measured by the 3-item Oslo Scale (low, medium, high) [26, 27]; Burden due to household responsibilities [5-point Likert-scale: 1 (not at all), 2, 3, 4, 5 (a lot)]; Burden due to childrearing responsibilities [5-point Likert-scale: 1 (not at all), 2, 3, 4, 5 (a lot)]; Burden due to informal care responsibilities [5-point Likert-scale: 1 (not at all), 2, 3, 4, 5 (a lot)].

The complete case analyses were based on data from study participants of GEDA 2009 providing information about the frequency of vegetable intake and relevant covariables (total sample n = 19,512, 8381 men and 11,156 women). We pursued two analytical strategies with different combinations of the binary sex/gender variable, socio-cultural, socio-demographic and socio-economic variables as well as sex/gender related aspects. In strategy 1 we included socio-cultural, socio-demographic and socio-economic variables and the binary sex/gender variable. In strategy 2 we included socio-cultural, socio-demographic and socio-economic and sex/gender related aspects and subsequently calculated proportions of men and women within the identified subgroups of the model. First, descriptive statistics of all covariables and the prevalence of non-DVI within each category were computed. Second, in order to detect subgroups with differences in prevalence of non-DVI, the classification task was performed with CART [11] and CIT [14] as decision tree building algorithms using the rpart package [28], partykit package [14, 29] and R 3.6.1 [30]. The CART algorithm represents a popular decision tree technique, which has been criticised for not having a concept of statistical significance. The newer CIT-approach follows formal statistical procedures in each splitting step [14, 31]. Since no straightforward recommendation is available on which algorithm and specifications for complexity parameter (cp) in CART or α-level in CIT should be preferred in light of public health monitoring data, we compared different specifications in our analyses by calculating in total 8 classification trees, 4 for each of the aforementioned strategies. The 2 CART-analyses were conducted using either cp = 0.01 (default specification in R) or cp = 0.005. The selection of the final tree model was based on the cross-validation estimates of the error of the sub-trees of the initial tree, along with the standard errors of the respective estimates. Correspondingly, fitted trees were ‘pruned’ according to the 1-SE rule to develop a tree with the best size and lowest misclassification rate by selecting the least complex tree whose error was within one standard error above the tree with the smallest cross-validated error [11]. The 2 CIT-analyses were conducted using either α-level = 1% or α-level = 0.1% both with Bonferroni correction. In order to keep the size of the tree and the results interpretable, we did not use the default specification in R of α-level = 5%. In all models, cost weights were assigned to equally distribute sums of weights for cases and non-cases, thereby assigning equal importance to sensitivity and specificity [32]. The minimum node size allowed in a terminal node was restricted to contain at least 1% of the overall analysis population. No other survey-specific weighting factors were applied. Unweighted percentage of population as well as prevalence of non-DVI were calculated for each node separately. In strategy 2, which does not include the sex/gender binary as a covariable, we calculated proportions of men and women within the identified subgroups of the model. Subsequently we computed prevalence of non-DVI for men and women separately as well as the resulting absolute difference in prevalence.

There are some limitations to consider, when interpreting the results of the present study. Respondents in health surveys show more positive health-related behaviours compared to non-respondents [45] and therefore the prevalence of non-DVI might have been underestimated. As there were only low numbers of individuals never consuming vegetables or less than once a week, we only compared daily vegetable intake to less than daily intake. Due to data availability, we were not able to include information about quantity of vegetable intake. Furthermore, it might be criticized that the data of GEDA 2009 used in this study are quite outdated. However, given that the aim of this study was to test a theory-based methodological approach, the GEDA 2009 survey provided a higher number of variables that describe sex/gender relevant mechanisms such as gender roles compared to other more recent surveys of the RKI in Germany. Overall, men showed higher prevalence of non-DVI than women within all identified subgroups of both strategies, which points to the binary sex/gender variable as the most strongly classifying variable. It was not possible to include further sex/gender related aspects than the ones studied, which might have unmasked more of the assumed heterogeneity in prevalence within and between the group of men and women. To name a few, consideration of factors such as drive for thinness [46], favourable attitudes, perceived behavioural control over frequent vegetable intake [33], being the primary meal planner or feeling responsible for the family’s overall health [47] might have further lowered the remaining difference in prevalence of non-DVI. Furthermore, by not including the binary sex/gender as an analytical category into the multivariable analysis in strategy 2, we were able to move beyond a traditional binary male/female approach only when viewed from a methodological perspective. Unfortunately, due to data availability, we did not have the opportunity to explore other non-binary concepts of gender identity, which embrace further identities such as trans or gender queer identities. Nevertheless, strategy 2 opens up the possibility to push public health monitoring and reporting past a traditional binary sex/gender approach to gender identity, since the strategy allows the description of proportions and prevalence within the identified subgroups by several gender groups if data on gender identities are available.

A strength of our analysis is the use of classification trees as an exploratory method to identify subgroups, which substantially differ in prevalence of non-DVI [16]. To this end, we compared the results of different algorithms and specifications of CART and CIT. We found that the main findings were quite robust, since the same subgroups were identified by both algorithms and all specifications alike. Finally, another strength of our study is the inclusion of a wide range of socio-cultural, socio-demographic and socio-economic variables as well as sex/gender relevant aspects based on data of a national survey on health of adults in Germany.