Do health literacy, physical health and past rehabilitation utilization explain educational differences in the subjective need for medical rehabilitation? Results of the lidA cohort study – BMC Public Health

The present study investigated the mediating effects of HL, past medical rehabilitation utilization and physical health in the association between education and SNR. For the analysis, data from the lidA (leben in der Arbeit) study were used. The lidA study is a prospective cohort study on work, age, health, and labour participation, representative for older workers from the German Baby Boomer generation from the 1959 and 1965 birth cohorts with respect to age, sex, nationality, education, and occupation [27, 28]. In the first study wave (t₀, 2011), the primary response rate, calculated as the ratio of completed interviews (n = 6,637) to the operational sample (n = 24,322) according to the definition RR5 of the American Association for Public Opinion Research, was 27.3% [28, 29]. Of the 6,637 completed interviews, 6,585 were valid. These respondents were eligible for the follow-up waves, which took place in 2014 (t₁, n = 4,244) and 2018 (t₂, n = 3,586). A more detailed description of the study and sampling procedures can be found elsewhere [27, 28, 30, 31]. The present analysis included persons participating in all three study waves (n = 3,232), excluding respondents without valid information on the included covariates (n = 102). Full data were available from 3,130 subjects.

Figure 2 shows the assumed causal relationship for our analysis model. With the description of the variables in the following paragraphs, the time of their measurement is indicated in parentheses. The arrows in the model in Fig. 2 describe both direct and indirect effects (mediated through HL, past rehabilitation utilization and physical health) of education on the SNR as well as the confounding influences on the exposure-mediator- and mediator-outcome associations.

Subjective need for rehabilitation (t₂)

The outcome of this analysis is the SNR. Participants were asked “Would you wish to participate in a rehabilitation programme (regardless of whether you have already had rehabilitation)?” The answer was restricted to “yes” or “no”. This item was assessed in study wave three (t₂).

Educational status (t₀)

Educational status was considered the exposure. A score [32], which combines the level of schooling and vocational training, is used to determine the educational status attainment. The results are classified into three categories: high (tertiary education), medium (upper secondary vocational education and postsecondary nontertiary education) and low (primary, lower secondary and upper secondary general education). This item was assessed at wave one (t₀).

Health literacy (t₂)

Health literacy was measured using the seven questions on the “coping with illness” dimension of the Health Literacy Questionnaire (HLQ) 16 [33]. Given the limited survey time, coping with illness was chosen, because it is the most important dimension in the context of our research question. This part of the HQL-instrument asks for the ability to find information about therapies for one’s own illness, whom to address professional medical aid, to understand the doctor’s or pharmacist’s therapeutic advice, to make one’s own health decisions on the basis of the doctor’s advice, to ask for a second medical opinion and to follow the doctor’s advice. The responses “very easy” and “fairly easy” were combined into “easy”, and “fairly difficult” and “very difficult” were combined into “difficult”. Each question answered with “easy” was scored with one point. Then, three categories were set up nearest to the categorization of the complete HLQ-instrument with 16 items defined by Röthlin et al. [33]: inadequate (0 to 3 points), problematic (4 to 5 points), and sufficient HL (6 to 7 points). This item was assessed in wave three (t₂).

Physical Health (t₂)

Physical health was assessed with the Short Form Health Survey (SF-12). Based on the items, a physical component score (PCS_Score) was created as described by Nübling et al. [34]. The PCS_score (0-100) was subsequently divided into tertiles. The lowest tertile (PCS_Score 2).

Past rehabilitation utilization (t₀ – t₂)

To assess past rehabilitation utilization (yes/no), participants were asked at each study wave whether they had previously participated in a rehabilitation measure. The answers reflect participation in the period between 2008 and 2018.

Confounders

Age (born in 1959 or 1965), sex (female/male) and migrant status (no migrant background/1st generation migrants with German citizenship/1st generation migrants with foreign citizenship/2nd generation migrants) were considered confounders of the exposure-mediator and mediator-outcome associations. Persons who were born abroad and who subsequently immigrated were defined as having a 1st generation migrant background. Persons of the 2nd generation were born in Germany but had at least one parent who was born abroad [35].

Statistical analysis

First, we conducted a descriptive analysis to display the sample characteristics and the proportion of participants with SNR. A chi-square test was used to test the associations between the covariates and the outcome (Table 1). P values were derived from Cramer’s V test.

To quantify the multiple mediating effects of HL, past medical rehabilitation and physical health in the association between education and SNR, an inverse odds weighting (IOW) approach [36] was used. This counterfactual-based approach to mediation has advanced during recent decades and has several advantages over traditional product / difference of coefficients approaches. The IOW approach allows for the decomposition of the total effect into direct and indirect effects in mediation analysis with binary outcomes and regardless of the presence or absence of an exposure-mediator interaction, which may be of particular importance when investigating a social status indicator such as education. Using the IOW, we can furthermore accommodate multiple mediators of any measurement scale simultaneously [37]. Detailed descriptions of the IOW approach of causal mediation including practical implications and examples are given by Nguyen et al. [37] and VanderWeele [38].

Using the counterfactual-based IOW approach, the total effect (TE) of the exposure (education) on the outcome (SNR) was decomposed into a natural direct effect (NDE) and a natural indirect effect (NIE). The NDE describes the effect of changing the exposure (e.g., from high to low education) but fixing the mediator. In this way, the indirect pathway through the mediator(s) is deactivated [37,38,39,40]. The NIE describes the effect of changing the mediator to the value it would naturally take had the exposure changed (e.g., from high to low education), while the exposure is actually fixed (direct pathway deactivation) [37,38,39,40].

In line with Nguyen et al. [37], the mediation was conducted in six consecutive steps. First, in an exposure model, education level was regressed on the mediator(s) and confounders using multinomial regression. To compute the inverse odds weight (IOW), the inverse of the predicted log odds for each observation in the exposed group was taken from this first model. This step included two postestimation predictions, one for the low educated and one for the medium educated. In the second step, the IOW was assigned to the exposed group (firstly low education level, then medium education level), and a weight equal to 1 was assigned to the reference group. Third, in an outcome model regressing the SNR on education level and confounders, the TE was estimated using a generalized linear model from the Poisson family and log link function. Poisson regression was preferred over logistic regression because, for nonrare outcomes (> 10%), the odds ratio is noncollapsible, leading to downward biased indirect effects [38, 41]. Fourth, the NDE was estimated using the same model but specifying the IOW. Afterwards, the NIE was obtained by subtracting the NDE from the TE. TE, NDE, NIE and their 95% confidence intervals were subsequently computed by bootstrapping with 1,000 replications. All steps were carried out for each mediator under investigation separately and combined, adding the mediators sequentially in a temporally plausible order. Importantly, to be able to interpret the obtained NDE and NIE causally, several assumptions about confounding had to be made. We assumed that there was no unmeasured confounding of the (i) exposure–outcome relationship, (ii) mediator–outcome relationship or (iii) exposure–mediator relationship; furthermore, we assumed (iv) the absence of a mediator–outcome confounder, which itself is affected by the exposure [42]. To determine the extent of effect mediation, the proportion mediated (PM) was calculated using the equation for ratio measures by VanderWeele [42], here for the relative risks (RR):

The significance level was set at p 

Sensitivity analysis

To check whether selection bias may have influenced the results, we repeated the analysis using a non-response weight. The weight accounts for selective dropout by inverse probability weighting for migrant status and education level. With this method, the data are standardized on the population of the lidA baseline in 2011. The weight is calculated for each subgroup by percentage in wave 1/percentage in wave 3. For example, for the group of non-EMB with high educational level 16.44%/18.03%=0.9116. The weighting factors for all other subgroups can be found in the supplementary table S1. In this case, the exposed group (first low education level, then medium education level) was assigned the IOW multiplied by the non-response weight, and the reference group was assigned the non-response weight only.