The sample was based on the Northern Swedish Cohort, a 27-year prospective study comprising all pupils in the ninth grade of compulsory school living in Luleå in 1981, when the participants where 16 years of age (N = 1083; 506 girls and 577 boys) . The questionnaires were filled in during school hours, with the PI present and able to assist the pupils. Follow-up surveys were conducted in 1983, 1986, 1995 and 2008. Participants completed a comprehensive questionnaire at all follow-ups covering health, social and socioeconomic conditions, and school/working conditions. The majority of the items originated from the Swedish Survey of Living Conditions  and the Low-Income Study . In order to allow follow-up, the questionnaires were not anonymised.
In 1981, personal interviews were conducted with all form teachers (n = 65) for each individual pupil by the project leader (AH). The interviewed teacher had been responsible for the class during the last three last years of compulsory school. The interviews were performed in a quiet room during school time. The teachers were interviewed with a previously used questionnaire consisting of 35 questions regarding the pupils' situation at school.
Of the original cohort, 1040 (96 %) subjects had enough information on mental health across the life course to be assigned to one of the trajectories described below, 951 (88 %) in addition had complete data on the studied exposures, and 940 (87 %) additionally had full information on all covariates, thus forming the analytic sample in the present paper.
Social background was assessed by the participants at age 16 using self-completion questionnaires. Social class was measured with two questions: What is/was your father’s occupation? and What is/was your mother’s occupation? Responses were coded according to the Swedish classification of social classes into working (blue-collar) class, lower white-collar class, and upper white-collar class. We defined parental social class as the highest of the father’s and the mother’s social class, with the highest value indicating upper class. Family unemployment was assessed by a yes response to the question Has anyone in your family been unemployed during the last 12 months? Parental mental health and alcohol problems were assessed with the question Is your father healthy? (If he is dead: indicate how his health was most of the time when he was alive). Tick one or more options. – and an identical question about the mother. The response options were: Yes, as far as I know; No, (s)he has some kind of somatic disease; No, (s)he has psychological complaints; (S)He has alcohol problems; Don’t know. Ticking psychological complaints and/or alcohol problems for father and/or mother was defined as parental mental health or alcohol problems.
Parental interest in their offspring’s studies was assessed with a question about Knowledge about the parents’ interest in the pupil’s studies, which was asked of the main teachers regarding each participant in 1981. There were five response options ranging from ‘probably very small’ to ‘probably very large’.
Assistance with homework assignments was based on the question Can you get help with homework assignments when you need it? with five response options from ‘no, never’ through ‘yes, always’ asked in the questionnaire to the participants themselves at age 16.
Mean grades at exit from compulsory school at age 16 were obtained from school records with the participants’ permission. Grades were given on a scale from 1 to 5 and by design normally distributed. The mean grades were thus also normally distributed and vary between 1 (the lowest possible) to 5 (the best possible) with 3.00 as the national mean. In some analyses mean grades were dichotomised as average or below vs. above average based on this theoretically defined national mean.
Internalised mental health symptoms (IMHS) between the ages of 16 and 43 years. In each of the five questionnaires the participants were asked if they had experienced any worry/anxiousness and anxiety/panic, in the past 12 months. The respondents were also asked how often, in the past 12 months, they had experienced sadness or feeling low. Based on these three questions, an ordinal scale index was created, ranging from 0 = not experiencing any symptoms to 8 = having worry/anxiousness, anxiety/panic and often or always feeling sad and low. A detailed description of the scale of IMHS has been published .
Extraction of trajectories of mental health
We used latent class growth analysis (LCGA) to retrieve trajectories of IMHS from age 16 to 43. The analysis classifies individuals in the population into subgroups following distinct developmental trajectories of the outcome. Both the shapes and the number of trajectories are detected by the analysis. LCGA assumes that the variance and covariance for the latent growth factors within each class are zero, i.e., all individual within a class follow the same trajectory .
We considered growth trajectories with intercept, slope and a quadratic term, and tested models with two to six trajectory classes. Because a large proportion of individuals was clustered at the scale minimum each measurement time we assumed the distribution of the IMHS score to be censored normal . The full information maximum likelihood method (FIML), assuming missing at random mechanisms, was used for handling missing data (default in the Mplus software). The trajectories were determined using information on IMHS only, but a sensitivity analysis adjusted for sex gave a similar solution with more than 99 % of the individuals assigned to the same latent class (data not shown).
We used several criteria to choose the optimal LCGA model for further analysis. As the Bayesian Information Criterion (BIC)  and the Lo-Mendell-Rubin (LMR) test  may favour less parsimonious models that are more difficult to interpret , we additionally considered entropy (value near 1 for a good model), size of the resulting classes (enough individuals in each to be meaningful in further analyses) and high posterior probabilities (near 1, on average, for a good model) and the interpretability of the fitted model . In order to avoid local solutions that do not represent the global maximum of the likelihood, the analyses were repeated with increased number of random sets and start iterations, once plausible models were identified .
We considered the five class model to be the most sensible representation of the data. Compared with the three and four class models it demonstrated clearer heterogeneity and at least as high entropy (Additional file 1: Table S1). Adding a sixth class did not yield trajectories that were substantially different from those identified with the simpler models (Additional file 1: Figure S1). Although the BIC continued to increase when adding a sixth class (Additional file 1: Table S1), the increase was now smaller and the LMR test was only borderline significant. Moreover, the smallest class with this model was only 1.6 % (17 individuals). Altogether, we considered the five class model as optimal.
Correlation coefficients were calculated for all pairs of variables in the study. To examine if parental involvement was related to the chance of following a certain trajectory of internalised mental health, we used multinomial regression to obtain odds ratios and 95 % confidence intervals for being assigned to the most likely trajectory as derived by LCGA. We analysed the whole sample as well as women and men separately, with parental social class (and in the former case also sex) as covariates. The individuals were weighted with the posterior probability of belonging to the assigned trajectory because ordinary statistical inference is not directly applicable for trajectory groups that are based on probabilities and thus are not fixed constructs. To test if parental involvement influences mental health independently of academic achievement, we entered mean school grades on leaving compulsory school in a final step of the analysis. The trajectories were derived using MPlus version 7.11; whereas all other analyses were done in IBM SPSS for Windows, version 20.0.0. All p-values are based on two-sided tests.
This study was conducted with the approval of the Regional Ethical Review Board in Umeå and has conformed to the principles embodied in the Declaration of Helsinki.