Study design and setting

The observational study BECOSS 2 collected data at 24 Berlin primary and secondary schools at three time points in 2021 [28]: June 21–23 (T1), August 9–11 (T2), and September 27–29 (T3). The present analysis focused on T1 and T3. At T1, Berlin schools had just returned to their original class sizes in early June after a model of split classes (half the original size alternating weekly between distance learning and school attendance) had been in effect since mid-March. School attendance was not yet mandatory. T1 took place directly before the summer school vacation (June 24—August 6, 2021). At T3, students had attended school for six weeks in full classroom sizes under almost normal schooling conditions, wearing masks in classrooms and testing twice weekly for SARS-CoV-2 [29].

For random school selection, the twelve Berlin city districts were first divided into three socioeconomic strata, “low”, “middle”, and “high” [30]. Within each stratum, two districts were randomly selected. All public schools from the selected districts were then put in random order and approached in that sequence until two primary and two secondary schools from each SES stratum agreed to participate [23]. Three of the targeted schools declined to participate and were replaced by substitutes of the same stratum. The resulting 24 schools chose one or two classes depending on factors like teachers’ willingness and disposability regarding exams or school trips during the study period. We aimed to examine 20 students per class. However, the actual number of participating students per class varied from 1 to 28.

As described elsewhere [28], study information and consent forms were sent to the participating schools and distributed to the selected classes before the start of data collection. After the children and their guardians consented to study participation, paper-based questionnaires with instructions for completion were mailed to the home address of the enrolled participants at each study time point. Parents were asked to answer questions on the socioeconomic status of the household in a designated section on the first page of the questionnaire. For the participants in primary schools, parents were also encouraged to help their children with the questionnaire.

Measures

The questionnaire had a total of 54 items. Questions about the socioeconomic background were directed to parents, children were asked about their sociodemographic data, school environment, leisure time activities, and COVID-19 related circumstances like symptoms, infections, vaccinations, fear and risk of infection as well as health behavior such as wearing of masks, washing hands, and social distancing. As the numbers of infections in the sample were very low, the present study does not include an analysis of these variables. Besides HRQoL, the BECOSS 2 study assessed general anxiety disorder symptoms using the GAD-7 instrument which was the subject of another study [28]. The variables relevant to the present study were assessed as follows.

Household income (four categories: up to 1000€, 1000–2000€, 2000–5000€, more than 5000€) was defined as the average monthly net household income reported by the parents. Due to the small number of observations, we combined the three lowest categories into a “lower income” category. The “higher income” category of more than 5000€ corresponds to at least double the average monthly net income of a Berlin household with one working adult in 2021 [25]. Household education (four categories: elementary school or comparable; secondary school or comparable; higher education entrance qualification; university degree) was assessed as the highest level of completed education in the household reported by parents. The lowest two and the highest two levels were merged, respectively, resulting in a dichotomous categorization (“lower” versus “higher education”).

Additionally, we measured exact age (in days) calculated from birth dates and defined two age categories for comparison with other studies, following the reference norm data from the KIDSCREEN-10 instrument with a cutoff age of 12 years (younger = 8-11 years vs. older = 12-18 years). We documented sex, offering the three options female, male, or diverse. Family migration background (yes, no), answered by the parents, was defined as at least one parent born outside Germany, and household size (four categories: one, two, three or four, five and more persons) as the number of persons living in the same household in addition to the participant.

We assessed HRQoL using the KIDSCREEN-10 self-report Index questionnaire. Inquired fields include physical activity and fitness, moods and emotions, participation in social activities, relationships with parents and peers, as well as the perception of own school performance, referring to the last week prior to filling in the questionnaire [31]. The tool consists of 10 items (e.g., “Thinking about the last week, did you feel fit and well?”), each answered on a 5-point Likert scale from 1 (not at all) to 5 (extremely) or 1 (never) to 5 (always). The scores from 1 to 5 were summed up to raw scores ranging from 10 to 50 and transformed into Rasch person parameter estimates. Subsequently, T-values were computed based on a pre-pandemic European sample with a mean score of 50 and a standard deviation (SD) of 10. Higher values indicate a higher HRQoL. The classification of T-values into categories of low, medium, and high HRQoL was carried out in reference to pre-pandemic European norm data for children and adolescents aged 8–18 years (without accounting for age and sex). Scores half a standard deviation below the population mean were categorized as “low HRQoL”, and half a standard deviation above as “high HRQoL” [32]. Scores were also calculated for responses with one missing value based on alternative norm data. For participants younger than eight and older than 18, no HRQoL category could be computed, and they were therefore excluded from the descriptive analyses. Furthermore, given that the instrument includes two questions directly aimed at last week’s school experiences, these items were left unanswered by many participants at time point T2 directly after school vacation. We decided to exclude T2 data from analyses, because the KIDSCREEN-10 Index could not be evaluated for this time point.

Data analysis

The paper-based questionnaires were digitized by FormPro Software (OCR Systems, Version 3.1) and transformed into an Excel database. Data were analyzed using R (version 4.2.0) and RStudio (version 2022.7.2.576).

To account for non-response bias and non-random attrition, inverse probability weighting (IPW) was used [33]. Participants were inversely weighted separately at each of the two time points based on their age, sex, and for T3 by participation at T1. These weights were then multiplied across both time points so that children with complete observations were up-weighted to count for children of the same age and sex with missing data. Children who only responded at T1 or T3 were downweighted to 0. The result is an IPW-population, referred to as “pseudo-population”, simulating a situation in which all individuals responded at both time points. All analyses were conducted within the pseudo-population, based on the weighted responses of children and adolescents with complete observations in T1 and T3 to account for missing data from non-respondents of the same age and sex. Additionally, we conducted a sensitivity analysis for the IPW-population including age, sex, school city district, and participation at both time points and provide the results in the (S1 and S2 Files and S1 Table).

To depict the main characteristics of the study population, means and standard deviations were calculated for continuous variables, and categorical variables were presented as proportions. In addition, means and standard deviations of HRQoL (T-values), as well as frequencies and percentages of HRQoL categories (low, medium, high) were presented stratified by time point, sex, and age group.

For the analysis of the causal research questions, we followed the approach as proposed by van Zwieten et al. [34]. Assumptions on the underlying relations between the two exposures (income, education) and the outcome (HRQoL) with all potentially relevant variables from the BECOSS 2 dataset were depicted graphically using directed acyclic graphs (DAGs) separately for each effect [35, 36] (S1 and S2 Figs). Variable selection and construction of the causal diagrams to establish potential confounding and mediating variables were based on subject matter knowledge and previous research on causal relationships of SES (or income and education separately) and health outcomes [5, 9, 17, 20, 37–40]. Minimal adjustment sets for variables potentially leading to confounding according to the causal diagram were determined using the software DAGitty [41]. Given the aim of estimating the average total causal effects, mediators were not adjusted for and therefore omitted from the DAG [34].

Two generalized linear mixed models were set up to estimate the average total causal effects of household income and education on HRQoL separately. To increase the precision of the estimate, both waves of observations were included per participant. The model, therefore, provides an average estimate over the two time points which makes it more independent of pandemic-related events like lockdowns, split classes, and school holidays. The first model used T-values of HRQoL as the dependent variable and household income as well as the potential confounding variables household size, education, and family migration background as independent variables. The second model used T-values of HRQoL as the dependent variable and household education plus the potential confounding variable family migration background as independent variables. Child age was included in both models to improve precision of the effect estimate after ensuring that it was not a collider, i.e., a common consequence of exposure and outcome [42]. To account for clustered data, random intercepts for schools, districts and time point were included. The model assumptions were checked graphically (S3 and S4 Figs). Regression coefficients and 95% confidence intervals were calculated, and the effects were interpreted according to Cohen’s d effect size measure for small (0.2), medium (0.5), and large (0.8) effects [42].

Strengths and limitations

The present study assessed HRQoL in children and adolescents at a later COVID-19 pandemic stage in the summer of 2021 in Berlin, Germany. The data from two time points were combined for assessment of the causal research questions to obtain a more accurate estimate that is less impacted by pandemic-related variations. Other strengths are the use of an internationally applied and validated instrument (KIDSCREEN-10 self-report Index), IPW to partially account for non-response bias and non-random attrition based on age, sex, and participation at baseline, and DAGs to formalize the causal framework underlying our research questions. Moreover, the study included the perspective of children and adolescents themselves.

This study has several limitations. We decided to use the shortened version of the KIDSCREEN instrument instead of the KIDSCREEN-52, as the comprehensive questionnaire captures a wide range of variables and reliability measures are comparable (Cronbach’s alpha of 0.82 for KIDSCREEN-10) [53]. Our sample is unbalanced, containing more households with higher income and education. In addition, there might be selection bias concerning differing student populations in the participating schools compared to those that rejected participation, and within schools in selecting participating classes. Our sample is therefore most likely not representative of the school student population in Berlin, Germany. Specifically, only 22% of participants in our sample reported a family migration background, while more than 51% of children aged between 6 and 18 years had a migration background in Berlin in 2021 [54, 55]. Due to the study’s limits in representativeness and specific setting, generalizability of the results is limited. Migration background, socioeconomic factors, and lower HRQoL might affect non-response and non-random attrition which could not be accounted for in IPW due to missingness in some of these variables. Therefore, and also since a small effect of lower household education was shown, the prevalence of low HRQoL could be underestimated in our study due to an overrepresentation of higher SES households. Potential barriers to participation in our study include competing priorities and demands of daily life, lack of trust in research, fear of stigma, and lack of time or interest [54, 55]. Further, we asked parents of primary school students to help their children fill in the questionnaire, which may have influenced the children’s responses.

It is not possible to allocate the results of this study as direct causes of the pandemic itself or pandemic responses that may have resulted in job losses and income reduction. Because the time points in this study were within three months of each other while the pandemic was ongoing, possible changes in income due to pandemic-related circumstances cannot be determined.

Due to our study design, our causal analyses should be interpreted with caution. Several other individual and societal factors may have influenced the students’ HRQoL, some of which were not measured like single parenting, religious affiliation, parents’ professions, and number of adults and children in the household to calculate an equalized income measure adjusted for different household compositions. Nonetheless, we refrain from fully avoiding causal language, as this may lead to the adoption of ambiguous terms and could inadvertently obscure the clarity of study objectives and the plausibility of causal assumptions, as pointed out by Haber et al. [56]. Our paper strives for transparency in clearly articulating our assumptions.