The sample is a sub-sample of the Norwegian Mother, Father and Child Cohort Study (MoBa), a prospective population-based pregnancy cohort study conducted by the Norwegian Institute of Public Health [24, 25]. Participating mothers from all over Norway were recruited during routine ultrasound assessment in week 17 or 18 of their pregnancy in the period from 1999 to 2009. 41% of the invited women consented to participate. MoBa participants received questionnaires in gestational week 17 or 18, week 22 and week 30, at child’s age 6 and 18 months, 3, 5, and 8 years and onward. The study is still on-going. The reported analyses also use information from the Medical Birth Registry of Norway .
The study sample is comprised of children whose mothers indicated developmental or behaviour problems in MoBa’s age five years questionnaire, and for whom information about outcomes in the age eight years questionnaire are available. This study focuses on children with one or more of the following developmental or behavioural problems: Attention deficit hyperactivity disorder, language development, oppositional defiant or conduct disorder, autism spectrum disorder, and learning disabilities.
The current study used rating scales from MoBa questionnaires sent out at child ages five, and eight years. Exposure and inclusion criteria were based on responses in the five year questionnaire, whereas outcome measures were taken from the eight year questionnaire. The first, 1.5 and three year MoBa questionnaires and the Medical Birth Registry of Norway provided covariates.
Exposure. To measure the provision of SEA, we relied on following question: “Does your child receive, or has received any extra resources in the kindergarten?” If mothers responded “Yes” to this question, they were additionally asked about the number of hours per week. SEA is provided to individual children, both inside and outside the context of regular preschool activities.
Outcome variables. Outcome variables (PSD8 in Figure 2) were sum scores from different scales about psycho-social difficulties. Outcome dimensions were attentional, hyperactivity/impulsivity, and behavioural (ODD or CD) problems measured with the Parent Rating Scale for Disruptive Behaviour Disorders (RS-DBD, ), emotional problems measured with the Short Mood and Feelings Questionnaire (SMFQ, ) and the Screen for Child Anxiety Related Disorders (SCARED, ), and communication problems measured with the Children’s Communication Checklist-2 (CCC-2, ).
Adjustment variables. Adjustment variables and those to control for loss to follow up were chosen based on the directed acyclic graph (DAG) shown in figure Figure 2. One important set of confounders includes children’s psycho-social difficulties at baseline, because these can be seen as causes of treatment and are related to later psycho-social difficulties. A number of scales in MoBa assessed psycho-social difficulties at age five and served as baseline measures (PSD5 in Figure 2). These included the Conners’ Parent Rating Scale-Revised, Short Form (CPRS-R (S), ), Child Behaviour Checklist (CBCL, ), the Ages and Stages Questionnaire (ASQ, ), and the Children’s Communication Checklist-2. While the baseline assessment considers the same mental health and development difficulties as the outcome, MoBa used different scales for five and age year olds.
Additional variables used for adjustment or prediction of loss to follow-up included maternal age, education, ADHD symptoms measured with the Adult ADHD Self-Report Scale  at child age three and depressive symptoms measured with the SCL-5  at child age five, parity, preterm birth, birth-month, hours special education per week, number of developmental of behaviour problems, and contact with rehabilitation services, Child and Adolescent Psychiatric Units, or Educational and Psychological Counseling Service at child age five years.
Classification into groups with different developmental or behavioural problems
To classify if and in which area a child had developmental or behavioural problem (DBP), we used MoBa questions about mental health problems at age five. Mothers were asked if their child “suffered, or is currently suffering from any of the following long-term illnesses or health problems.” In addition, mothers were asked if they had been in contact with a Child and Adolescent Psychiatric Unit or the Educational Psychology Counseling Services and if the health problem was confirmed by a professional. Only children for whom mothers reported a health problem and who indicated that the problem was evaluated by a mental health professional were included in the sample.
Disorders or health problems for which MoBa’s age 5 questionnaire has questions included Epilepsy, Cerebral Palsy, impaired hearing, which were excluded from the current analysis, together with children for whom mothers indicated a chromosomal defect. MoBa also asked mothers about autism spectrum disorders (ASD), hyperactivity and attention problems (ADHD), language difficulties (Lang), and behavioural problems (Beh). Additional questions about learning disabilities (LD) were also used to identify cases of interest for this study. Each child was classified in one of the following DBP groups: 1. ASD, 2. LD, 3. ADHD & Beh & Lang, 4. ADHD & Beh, 5. ADHD & Lang, 6. ADHD, 7. Lang, 8. Beh. For some children, mothers indicated multiple DBP, in which case the child was assigned to the first group it fell into. If, for example, a mother indicated ASD, ADHD, and language problems, the child was assigned to the ASD group (details in supplementary materials and Table S1). The rational underlying this classification scheme was to use existing psychiatric diagnoses, and to classify children according to their most impairing problem because these have typically more severe and persistent effects on psycho-social development.
All analyses were performed using R . The Bayesian hierarchical regression model was implemented with the brms package . The analyses are described in more detail in the supplementary material, and analysis scripts are available at https://github.com/gbiele/SPS358.
Bias from treatment by indication and loss to follow up. Estimation of treatment effects from observational data is difficult because treatment is not assigned randomly. Instead, individuals with more psycho-social difficulties at age five, who are also more likely to have psycho-social difficulties in the future, more likely receive treatment (treatment by indication). In addition, loss to follow up makes estimation of treatment effects difficult. Therefore, we used a directed acyclic graph [DAG, 38, see Figure 2] to explicate the assumed causal structure and to determine with which approach to deal with potential biases. Given this structural model, inverse probability of continued participation weighting was needed to reduce bias from loss to follow up , whereas adjustment for predictors of SEA was sufficient to control bias from treatment by indication. This means that we effectively estimated the effect of SEA on the change of psycho-social difficulties from preschool to elementary school.
Estimation of the treatment effects. We used a Bayesian adjusted and weighted hierarchical ordinal regression to estimate effects of SEA [37, 40, 41]. A hierarchical regression induces partial pooling (shrinkage) of estimates, which reduces the variance of estimates  and controls the multiple comparison problem . Importantly, when analysing related patient groups hierarchical regression results in more accurate association estimates then independent analysis of these groups . We used an ordinal regression model, because the estimation of latent, normally distributed traits that underlie the rating-scale responses facilitates the presentation of results in terms of standardised mean differences (SMD). The reported results were obtained by pooling over the independent analyses of the 50 imputed data sets . Consistent with recent recommendations to focus on estimation of effect sizes instead of significance testing [45, 46] we generally report means and the 90% credible intervals.