This time-trend study included data from four cross-sectional surveys conducted in two Finnish cities, Rovaniemi and Salo, in 1998, 2008, 2014 and 2018. Rovaniemi is located in the northern part and Salo in the southern part of Finland. The characteristics of the populations living in Rovaniemi and Salo, with regard to population structure, mean age, gender distribution, educational structure, income distribution, ethnic background, and family composition, were comparable to the general population in Finland [20] (Supplement S1).

The Ethical Committee at the Hospital District of Turku University Hospital (1998, 2008) and The Ethics Committee of the University of Turku (2014, 2018) approved the study, and permission was obtained from school authorities. Participation was voluntary for adolescents, and their anonymity was ensured. Parental consent was requested through informing them beforehand about the study, and they had the possibility to not let their adolescent participate in the study.

The data were collected from all secondary schools in these two cities, excluding schools and classes intended for children with special needs. Due to changes to municipality structures in Finland, Rovaniemi and Salo have merged with other municipalities since the start of this study. Only the municipalities that provided data from each of the four measurement points were included in the analysis.

The study participants were adolescents in the 7th or 9th grade in Finnish secondary school, comprising adolescents aged 13–16 years (few students were as old as 17). To ensure confidentiality in the study, the adolescents filled out the questionnaires anonymously during a school lesson and returned them to the teacher in sealed envelopes. These were then placed in another envelope and returned to the research group. The teachers were informed about the confidential nature of the study, and they took the necessary precautions during the school lesson.

Teachers were instructed to ask any absent students to fill out the questionnaire at a later date in the same conditions (at school) as the respondents on the day of the survey. Despite reminders, there were non-respondents who had been absent from school that day and never completed the survey. There were also students and classes who did not want to participate. In 1998, 156 adolescents did not participate; 180 in 2008; 175 in 2014; and 241 in 2018. This resulted in a total of 6,678 returned questionnaires. Of those, 78 were excluded due to incomplete or inappropriate answering: 39 in 1998; 15 in 2008; 11 in 2014; and 13 in 2018. This left us a total of 6,600 adolescents in the final sample (Fig. 1).

Demographic details included age, school grade, city, gender, family structure and ethnic background. Family structure comprised six categories: living with two biological parents, one biological parent, remarried parents, foster parents, adoptive parents, or others. When family structure was used as a covariate in the adjusted analyses, it was divided into two categories: two biological parents and other. Ethnic background included information about the adolescents (I was born in Finland; My native language is Finnish) and their parents (My biological mother/father was born in Finland; His/Her native language is Finnish) in terms of their place of birth and native language.

Mental health was assessed with the Strengths and Difficulties Questionnaire (SDQ), assessing positive and negative behaviors, emotions and relationships. In this study, a double-translated Finnish version of the questionnaire was used [21]. There are 25 items divided between five scales: emotional symptoms, conduct problems, hyperactivity, problems with peers, and prosocial behavior [22, 23]. The possible scores for each scale of five items could range from 0 to 10. Using a three-point scale (0 = not true, 1 = somewhat true, and 2 = certainly true), respondents were asked to indicate how things had been for them over the last six months. Five items were worded positively and scored in the opposite direction. Four of the scales, not including the prosocial scale, were added together to provide a total difficulties score of 0–40, with higher scores indicating more symptoms. Of the 25 items, 10 items reflected strengths, 14 items reflect difficulties, and one item from the peer problems subscale, “I get on better with adults than with people my own age”, was neutral but scored as a difficulty item. To define any groups with the most severe problems, the cut-off points of the 90th percentile for the SDQ total scores and sub-scales were used, based on the previous European studies [19, 21,22,23,24,25,26,27] and on a Finnish normative sample from 1998 [24]. The correlations between the sub-scales were moderate, and Cronbach’s alphas varied between 0.55 and 0.70 (Supplement S2).

Substance use was assessed by asking “How often do you use alcohol [e.g., beer]?” and “How often do you use enough alcohol to get drunk?”, with the responses of “not at all”, “once a month or more often”, and “once a week or more often”. Smoking was assessed by asking “How often do you smoke cigarettes or use other nicotine products?”, with the responses being “not at all”, “not often”, “every week”, and “every day”. In the analyses, the responses “every week” and “every day” were combined as “at least once a week”.

In order to take variation between schools into account, mixed effects models with school-wise random intercepts were used to examine changes in outcomes between study years. SDQ scales were examined both as categorical and continuous variables. For total difficulties, emotional symptoms, conduct problems, hyperactivity and peer problems, mixed effects logistic regression was used to estimate the probability of a score equal to or higher than the 90th percentile cut-off point (calculated from the 1998 sample). Because a low prosocial score indicates problems, mixed effects logistic regression was used to estimate the probability of a prosocial score equal to or lower than the 10th percentile cut-off point. The means of continuous SDQ scores were compared with mixed effects linear regression. Changes in substance use were examined using mixed effects multinomial logistic regression with the option “not at all” as the reference category. Equations for the models in Supplement S3.

Association of interaction of year and sex with categorical SDQ scales and substance use was tested for with mixed logistic regression (binary or multinomial). Because differences in the effect of year were found for some outcomes, all further analyses were conducted separately for males and females. Odds ratios and differences of means were calculated for 2008 vs. 1998, 2014 vs. 2008, 2018 vs. 2014 and 2018 vs. 1998 with Bonferroni correction to control the overall probability of false positive results. This meant that a p-value lower than 0.05/4 = 0.013 was considered significant for one pairwise comparison, and estimates were calculated with a 98.75% confidence interval. Type 3 tests were performed to estimate the overall significance of the year as an explanatory variable. An overall p-value of less than 0.05 was considered significant, meaning that the average outcome was not similar in all years. All the results except background characteristics were adjusted by school grade, family structure and city.

To examine the overall yearly change in outcomes during the 20-year period, all analyses described above were repeated with the study year as a continuous explanatory variable centered on 1998. Odds ratios and differences of means between two consecutive years were estimated with 95% confidence intervals.

The statistical analyses were conducted using SAS 9.4 for Windows.