Data and participants
The student-level data was obtained through the School Health Promotion Study (SHP), a nationwide classroom survey. The SHP has monitored the health and well-being of Finnish adolescents since 1996, and it is conducted by the Finnish Institute for Health and Welfare (THL). Our data collection was approved by the THL’s ethical committee (THL/1704/6.02.0 1/2016).
We focused on students in the eighth and ninth grades (14–16 years old). The adolescents were informed of the aim and content of the survey, and they had the opportunity to decline to take part. Their parents and guardians were also informed. Written consent was not necessary, since the survey was conducted anonymously. The data was collected in 2017 during school lessons. In total, 84% (N = 730) of Finland’s lower-secondary schools participated.
The school-level data was obtained from the Benchmarking System of Health Promotion Capacity-Building’s (BSHPCB) data collection from comprehensive schools. This data too was collected in 2017. The BSHPCB is a nationwide benchmarking tool for local governments and schools to manage, plan and evaluate their own health promotion activities and resources in basic education. The data collection form is completed by the school’s principal together with a student welfare team. The BSHPCB is run by the THL and the data collection in basic education is done in collaboration with Finnish National Agency for Education. In total, 91% of Finland’s lower-secondary schools participated.
We included schools in our analyses using two variables from the BSHPCB. The first variable measured when the most recent inspection of the health and safety of the school environment and the well-being of the school community had been carried out. This inspection is required by Healthcare Act 1326/2010, which states that all schools in Finland should be checked every 3 years. The triennial official inspection is conducted in cooperation with the school health service, representatives of the school (e.g. the principal), representatives from the health authority, occupational healthcare, occupational health and safety, and the authorities responsible for the construction and maintenance of school buildings [31]. The inspection that focuses on building-related factors is reported in detail elsewhere [21]. For our analyses, we selected only schools where the inspection had been carried out in 2016 or 2017.
The second variable measured whether mould and dampness had been observed in the school (see section A building-related predictor). We included in our analyses the schools where a) mould and dampness had been identified during the check and the problems had not been remediated, and b) no mould and dampness had been identified during the check. We excluded from our analyses schools with fewer than 10 students (N = 51), students that needed special education (N = 89) and students who did not report their age or reported that their age was less than 14 (N = 340).
The final data set consisted of 25,101 students from 222 schools where both the inspection of the health and safety of the school environment and SHP were conducted.
Measures
Outcome variables
The perceived quality of teacher-student relationships was measured by three items: ‘teachers encourage me to express my opinion in class’; ‘teachers are interested in how I am doing’; ‘teachers treat us (students) fairly’. The response scale was 1 = fully agree, 2 = agree, 3 = disagree, 4 = fully disagree. A mean rating of the items was calculated. Only if the respondent had answered all three items was the score calculated. These items have also been used in many previous studies as indicators of teacher-student relationships [30, 32]. The reliability was reasonable (Cronbach’s alpha = 0.75). The data source was the SHP.
Class spirit was measured by three items: ‘it’s peaceful to work in my class’; ‘the atmosphere in our class is such that I dare to express my opinion freely’; ‘the pupils in my class get along well’. The response scale was the same as above, and the mean rating was calculated similarly. These items have also been used in many previous studies as indicators of class spirit [21, 27]. The reliability was reasonable (Cronbach’s alpha = 0.68). The data source was the SHP.
Mediator
Our mediator was the subjective assessment of IAQ (subjective IAQ). It was measured by two items: ‘have any of the following things bothered you at your school during this school year? a) ‘Stuffy air (bad indoor air)’; b) ‘unpleasant odour’. These items were measured on a three-point scale (1 = not at all, 2 = some, 3 = a lot). A mean rating of the items was calculated. If the respondent had not answered both items, the score was not calculated. Cronbach’s alpha was 0.71. The data source was the SHP.
A building-related predictor
Observed mould and dampness was measured by one item: ‘were the following issues evaluated in the most recent inspection of the health and safety of the school environment: problems with mould and dampness?’ The response options were: 1) no data available; 2) not included in the inspection; 3) inspected, no deficiencies detected; 4) inspected, deficiencies detected but not yet corrected; 5) inspected, deficiencies detected and corrected. In this study, we included in the analyses only on the third and fourth options, and they were recoded as follows: 0 = inspected, no deficiencies detected; 1 = inspected, deficiencies detected but not yet corrected. The data source was the BSHPCB.
Background variables
Gender and age were used as student-level background variables only. Fathers’ level of education and student-perceived teacher-student relationships were used as both student-level and school-level background variables. Fathers’ level of education was used as an indicator of students’ socio-economic status. The response options on fathers’ education were: 1 = comprehensive school or equivalent (i.e. primary level), 2 = upper-secondary school, high school or vocational education institution (i.e. secondary level), 3 = occupational studies in addition to upper-secondary school, high school or vocational education institution (i.e. secondary level and occupational studies), 4 = university, university of applied sciences or other higher-education institution (i.e. tertiary level). All these background variables were reported by the SHP. The school size (i.e. number of students) reported by the BSHPCB was used as a school-level background variable only.
Calculation
The mediation analyses were conducted by analysing two two-level linear regression path models [33]: one where the outcome measure was student-reported teacher-student relationships, and one where the outcome measure was class spirit. Multilevel analysis is required when the data is hierarchical [33]. We built the models and then analysed them using Mplus statistical software 8.0 [34]. We used full information maximum likelihood estimation (FIML) with robust standard errors (the MLR estimator in Mplus) as an estimation method. MLR is robust to moderate violations of assumptions such as non-normality [35].
We used a latent factor approach first introduced by Jöreskog [36]. In order to estimate the student-level and school-level variance in each variable in the model, their total variance was decomposed into two latent uncorrelated components by Mplus. The first latent component (i.e. student level) represented the degree students’ answers deviated from their school mean (e.g. the cluster mean of reported symptoms). The second latent component (i.e. school level) represented the degree the school mean deviated from the grand mean [34, 37].
We started by analysing the intraclass correlations (ICC) by using a null model. In a null model only the outcome variable without any predictors is inserted in the model. It is used to estimate the variance between student and school levels and the ICC [33]. The ICC reports the proportion of the variance belonging to the school level [33]. Then we analysed design effects (DEFF) of each variable. The DEFF reports how much larger a variable’s sampling variance is from the mean than would be the case if the sample had been drawn from a simple random population [38]. When a DEFF is greater than 1.1 and the researcher is interested in estimating the effects of group-level predictors, multilevel modelling is needed [39]. The DEFF can be estimated as a function of the ICC and average cluster size [38].
Then, we estimated the total, direct and indirect effects of our two mediational models. The total effect refers to the relationship between the predictor (i.e. observed mould and dampness) and the outcome variable (i.e. teacher-student relationships or class spirit) when the mediator (i.e. subjective IAQ) is not controlled. In Fig. 2, the total effect is represented by path C. A significant total effect is not required when testing mediational models [40]. The direct effect refers to the relationship between the predictor and the outcome variables when the mediator is controlled. In Fig. 3, the direct effect is represented by path c’. The indirect effect is the product of path a multiplied by path b (see Fig. 3). In our model, the independent variable and the mediator were at school level, and the outcome variable was at student level – a so-called 2–2-1 design [41]. If one variable is a school-level variable, the indirect effect exists at the school level [42]. The analyses were conducted according to a syntax based on articles by Preacher, Zyphur and Zhang [42] and Preacher, Zhang and Zyphur [43]. The syntax is available online at http://quantpsy.org/medn.htm.
We report both unadjusted and adjusted models. In the adjusted model, we used fathers’ education, age and gender as both student-level and school-level variables. School size and observed mould and dampness were included only at the school level. All continuous predictors and background variables were centred by their grand means.
Finally, we counted the Monte Carlo confidence intervals to assess the significance of the indirect effects. These intervals accurately reflect the asymmetric nature of the sampling distribution of an indirect effect [43]. This type of analysis has been shown to be superior to the Sobel test [44]. For helping interpretation, we present the between-level standardised coefficients of direct and total effects. To report the effect size of indirect effects, we partially standardised their regression coefficients by dividing the indirect estimates by the between-level variance of the outcome variable [45]. R2 was used as an indicator of explained variance. Mplus provides separate R2 for the student and school levels [46].
Missing values
The number of missing values varied between the variables. Age, subjective IAQ and observed mould and dampness had the lowest percentages of missing values (0%), and socio-economic status had the highest (12%). Values were assumed to be missing at random [47]. In such cases FIML is a recommended method for handling missing data, because it uses all available data for estimation and produces unbiased parameter estimators [47].