Appendix 3.1: handling of missing data

The common approach to handling missing data is to delete all cases having missing data. However, in an attempt to adopt statisticians' (e.g., Peeters et al., 2015; Van Buuren, 2018) recommendation to upgrade our work, we tried to employ known methods and software for handling incomplete data problems. As specified in this chapter, our study included three data sets: (1) the teacher data set for eight countries, (2) the principal data set for six countries, and (3) the combined data set of teachers and principals, also for six countries. Within the teacher data, missing items ranged from 4.8% to

0.1%. Of all 2,554 cases, 18.7% had no missing values. Overall, 95.5% of the data was included within the teacher data set. Similar to the teacher data set, all variables in the principal data set had a small percentage of missing data ranging from 7.6% to 0.8%. For principals, 84 cases (63.6%) had no missing values and of all the data (95.8%) was included in the data set.

The combined data set included all principals, but only teachers from schools of which the respective principal(s) completed the questionnaire, amounting to 1,894 teachers (78.4% of all teachers). Missing data in the combined data set ranged from 4.8% to 0.1%. Of all cases, 64.3% had one or more missing values within the questionnaire. Overall, 97.0% of the teacher data was included within the combined data set.

Multiple imputation is broadly recoimnended as the best general method to deal with incomplete data (Van Buuren, 2018). We used IBM SPSS Statistics for Windows, Version 25.0 (IBM Corp., 2017) to impute the incomplete cases based on all variables included within the regression analyses. With a small amount of missing data, substantive conclusions are unlikely to change with more than five imputations (Van Buuren, 2018) and, therefore, we chose to calculate five imputations for each missing value. Furthermore, we chose to impute the question items over the scale variables to retain as much of the original data as possible. The amounts of missing data we imputed ranged from 4.8% to 0.1%, which we considered a small amount.

Appendix 3.2