In educational science, surveys are commonly used to investigate students’ perspectives on a range of issues. Over the past decade, however, participation rates have declined substantially, a trend that appears to be global. While attrition has long been a challenge in longitudinal research, the combination of increasing non-participation at baseline and subsequent attrition presents a more serious methodological problem that warrants closer attention.

The Educational Longitudinal Database Evaluation Through Follow-up (UGU) collects data from approximately 10% of the total population of students born in a given year, corresponding to about 10,000 students assessed in Grades 6 (age 13), 9 (age 16), and 12 (age 19). The UGU database includes measures of multiple constructs, such as self-concept, interest, motivation, well-being, and cognitive ability.

Within the ongoing project The impact of personal, environmental, and biopsychosocial factors in and out of school on students’ mental health and well-being: A longitudinal cohort study, we examine students’ well-being and mental health from a holistic perspective using the biopsychosocial framework (Porter, 2020). A growing body of research over the past decade indicates a decline in well-being and mental health among young people (Klapp et al., 2023; Högberg et al., 2021; National Agency for Public Health, 2025). In a recent study, we found that students born in 2004 reported lower levels of social, psychological, and cognitive well-being compared with students born in 1998 (Klapp et al., 2023). We also showed that students’ cognitive well-being defined as understanding the teacher and beliefs about one’s intellectual ability to learn was a stronger predictor of final grade point average (GPA) than measured cognitive ability based on cognitive tests.

When conducting longitudinal analyses using more recent birth cohorts, it is therefore necessary to apply weighting procedures to address increasing levels of missing data. This raises an important methodological question: Are the weights and missing-data handling methods currently applied, such as full information maximum likelihood (FIML) estimation in Mplus sufficient to adequately address the growing challenges posed by non-participation and attrition?

To address missing data in the survey components of the Evaluation Through Follow-Up (UGU) research database, Statistics Sweden (SCB) has developed survey weights for each wave of data collection. These weights are primarily designed for cross-sectional analyses. However, given that UGU is a longitudinal database, there is a need for methods that more adequately handle missingness and attrition across multiple waves.

Although the UGU survey data are affected by both non-response and attrition, the database also contains complete register data covering the full population, with no missing values. This provides a unique opportunity to explore alternative approaches to handling longitudinal missingness. In this study, the wave-specific SCB weights are used as reference points while implementing and comparing three different imputation-based approaches that combine register data with incomplete survey data. Specifically, we generate imputed datasets using (1) full information maximum likelihood (FIML) as implemented in Mplus, (2) multiple imputation by chained equations (MICE) in R, and (3) random forest–based imputation in R.

To evaluate which approach most closely aligns with the cross-sectional weighting strategy while remaining suitable for longitudinal analyses, we compare results from cross-sectional regression models across the original datasets, the weighted datasets, and the imputed datasets generated using each method.

The analyses are based on nine items measuring cognitive, psychological, and social well-being, as well as one measure of cognitive ability. The study is conducted within the FEEL project, which has received ethical approval.

The analyses will be based on data from the 2004 birth cohort, who entered compulsory schooling in 2011 and completed upper secondary education in 2023. Measures include items capturing psychological well-being (e.g., “I worry about tests or homework”), cognitive well-being (e.g., “I find it difficult to keep up during lessons”), and social well-being (e.g., “How do you like your current class?”). Cognitive ability is assessed using three tests: antonyms, inductive reasoning, and spatial ability. In a first step, the original survey dataset will be compared with population-level register data on a set of background variables, including parental education and income, gender, immigration background, and academic achievement, in order to assess representativeness and patterns of non-response. Next, descriptive statistics for the well-being and cognitive ability measures will be computed using the original dataset, both with and without the application of design weights. Subsequently, six analytic models will be estimated. First, the original dataset, with and without weights, will be analysed in Mplus using the default missing data handling procedure (full information maximum likelihood; FIML). Second, the same two datasets will be analysed using multiple imputation by chained equations (MICE) implemented in R. Finally, the corresponding analyses will be conducted using random forest based imputation in R.

Preliminary analyses comparing the original dataset with datasets weighted using weights calculated by Statistics Sweden (SCB) indicate that the weights perform reasonably well. Nevertheless, there appears to be scope for further improvement with respect to the first research question. By systematically testing different methods and integrating population-level register data, we aim to examine how missing data can be handled more effectively and to develop deeper methodological insight into the use of weights and the suitability of different approaches for different types of data. The feasibility of addressing the second research question concerning the combination of multiple weights will also be examined We expect that the findings from this study will provide methodological guidance for researchers working with large-scale datasets affected by missing data. In particular, the study addresses the broader challenge of handling missingness in longitudinal designs, an issue of growing importance in educational research. A specific focus concerns how design and survey weights can be combined in longitudinal analyses. Based on our results, we will offer practical recommendations for researchers on how to handle missing information in longitudinal data.

Hogberg B, Lindgren J, Johansson K, Strandh M, Petersen S. Consequences of school grading systems on adolescent health: evidence from a Swedish school reform. Journal of Education Policy. 2021;36(1):84-106. Klapp, T., Klapp, A., & Gustafsson, J.-E. (2023). Relations between students’ well being and academic achievement: evidence from Swedish compulsory school. European Journal of Psychology of Education. Doi: 10.1007/s10212-023-00690-9. Porter, R. J. (2020). The biopsychosocial model in mental health. Australian & New Zealand Journal of Psychiatry, 54(8), 773–774. https://doi.org/10.1177/0004867420944464 Public Health Agency of Sweden. (2018). Varför har den psykiska ohälsan ökat bland barn och unga i Sverige? Utvecklingen under perioden 1985–2014. [Why has mental illness increased among children and young people in Sweden during 1985 and 2014?]. Public Health Agency of Sweden.