The present study used data from the 2002, 2006, 2010, 2014, 2018 and 2022 HBSC survey rounds. Participating countries were eligible for inclusion in the present analyses if they had collected data on both sexual intercourse and the age of first sexual intercourse at least in the last three survey rounds. The questions on sexual behaviour were only administered to 15-year-olds, therefore our analyses were limited to this age group. The analytic sample included 312,702 participants across 37 countries (51.9% girls, mean age: 15.08, SD = 0.43).

Early sexual initiation: Participants were asked if they had ever had sexual intercourse; those responding “yes” were subsequently asked how old they were when they had their first sexual intercourse. The response options were “11 years old or younger”/“12”/“13”/“14”/“15”/“16 years old or older”; in some countries, “17 years old or older” was the last option. While these are not relevant to 15-year-old participants, in some survey years and countries older adolescents, aged 16–18 were also participating in the study, hence the last two response options. We dichotomised the responses into having had early (before the age of 14) or not early (at the age of 14 or older) first sexual intercourse.

A nominal variable was created from country/region. Participants’ gender was assessed by asking whether they are a boy or a girl; 845 participants (0.3%) did not answer this item [48]. They were retained for aggregated analyses but excluded from gender-disaggregated models. Family affluence – as an indicator of SES – was assessed using the Family Affluence Scale (FAS). This is a composite score consisting of six items (e.g., whether participants have their own bedroom, or how many cars are owned by their family) [49]. The sum score was transformed into relative groups, classifying the lowest 20%, the middle 60% and the highest 20% affluent families in each country [50]. In order to prevent further reduction of the analytic sample, we created a fourth category, consisting of participants with missing responses on the FAS.

Survey year: In univariate analyses encompassing all survey rounds (Steps 1–4 and 5a–8a in Table 1), we have created five dummy variables that had two values for each two consecutive survey rounds (i.e. 2002 = 0, 2006 = 1; 2006 = 0, 2010 = 1; and so forth), to account for potential fluctuations across survey rounds. In multivariate analyses including all years (Steps 5a–10a in Table 1), survey years were centred to support examination of non-linear trends. In analyses covering only the last three survey rounds (Steps 5b–10b in Table 1), we have treated survey years as categorical variables, and used 2014 as the reference.

The quality of adolescent–parent communication was measured by four items, tapping into how easy it is for the participant to talk to their father, mother, stepfather or stepmother about the things that really bother them. For each four items, response options were “Very easy”/“Easy”/“Difficult”/“Very difficult”/“Don’t have or see this person.” We merged the four items into a variable with having at least one parental figure with whom communication is very easy or easy = 1; not having any such parental figure = 0.

In line with the stratified sampling strategy of HBSC (classrooms clustered in schools within countries), a school variable was created which accounts for the potential effects of school-level norms and other school-level sources of variations in early sex not accounted for in the models. Schools were nested within countries.

Steps of the analytic strategy are presented in Table 1. First, percentages of those who reported early sexual initiation were calculated, across countries and survey rounds. Chi-square tests were conducted to test the association between prevalence of early sexual initiation and survey years, in the aggregated (Model 1) and gender-disaggregated (Model 2) samples. Cramer’s V effect sizes for Chi-square tests were interpreted as: V ≤ 0.05 very small, V between 0.06 and 0.19 small, V between 0.20 and 0.29 medium, and V ≥ 0.30 large [51]. We used visual inspection of graphs depicting the temporal dynamics of changes across countries and genders (Figures 1, 2) to infer cross-country patterns of trends.

Next, univariate binary logistic regression models estimated participants’ odds for early sexual intercourse from the previous to the subsequent survey round, covering all survey rounds in which the given country collected data. The reference was always the previous survey round. Again, these steps were carried out for the aggregated sample (Model 3) and then separately for boys and girls (Model 4).

Data were not available from all survey years in each country. Therefore we have carried out the next steps separately for two “arms”: (a) those 24 countries where data was available for all survey rounds; (b) for all countries, covering the last three survey rounds. First, for both arms we created null models (no prediction) accounting for potential clustering effect schools nested in countries (Models 5a/5b). Following that, for arm (a), we centred years and added three separate regression terms. In addition to a linear regression term, we tested if adding a quadratic term (modelling a parabolic curve) and a cubic term (modelling trends that may increase or decrease at varying rates) results in increased model fit (Model 6a). Next, we added gender and family affluence as covariates (Model 7a); and added having a supportive parental figure as a covariate (Model 8a). For arm (b), we added year as categorical variables (comparing change from 2014 to 2018 and 2014 to 2022, respectively) (Model 6b); added gender and family affluence (Model 7b); added having a supportive parental figure (Model 8b), then tested the potential interactions of the second and third survey year and having a supportive parental figure, to account for the potential changes in parental support from 2014 to 2018 and 2014 to 2022 (Model 9b). Since adding these interaction effects reduced model fit, we stepped back and re-ran Model 8b separately for boys and girls (Model 10b). Multicollinearity was tested and the improvement of the models were compared using Akaike (AIC) and Bayesian Information Criterion (BIC).

Where countries provided a weight variable (correcting for gender, regional or socio-economic imbalances in the sample), we applied it in Steps 3–10. In the absence of actual weights, weight was set at one. Steps 1–4 were conducted in SPSS Version 28; Steps 5–10 in Stata.

Table 2 presents the characteristics of the analytic sample are presented. Figure 1 shows the rates (in unweighted percentages) of early sexual intercourse for 24 countries between 2002 and 2022; Figure 2 shows the rates for all 37 countries between 2014 and 2022. Respective percentages and Chi-square tests showing how the rates of early sexual intercourse are associated with survey year are presented in Supplementary Table S1 for the overall sample; Supplementary Table S2 for boys and Supplementary Table S3 for girls. The prevalence of early sexual intercourse was highest in 2006 (7.3%); and lowest in 2018 (3.7%) and 2022 (3.8%). The figures and the tables show that there was a variation across countries (and sometimes gender) in the temporal dynamics in early intercourse, albeit the effects were mostly small or very small.

Weighted binary logistic regression models showing the changes in early intercourse from one survey year to another are presented in Supplementary Table S4 for the overall sample, Supplementary Table S5 for boys and Supplementary Table S6 for girls. The temporal pattern of change is summarised in Supplementary Table S7. In 14 countries, the odds of early initiation did not change, or decreased between two survey rounds. Only plateaus or increases between two rounds were seen in two countries. Odds of early initiation did not change across survey rounds, i.e., there was a stagnation, in five countries. The most prominent pattern, however, was a fluctuation over time. In other words, significant decreases, increases or no change between two survey rounds, varied in 16 countries. While in some countries the temporal pattern was similar for boys and girls, it was not uniform across all countries. We saw no clear geographical differences between the patterns.

The sequence of the multivariate models is outlined in Table 1, Column (a). The results of models featuring those 24 countries that provided data for all six survey years are presented in Table 3. Survey year had a significant effect on early sexual intercourse; AIC and BIC indices reduced as the quadratic and cubic terms were introduced, indicating that the association between survey years and early sexual initiation can be best described by a cubic model. This result rhymes with the pattern in the Supplementary Tables that the change in the rates of early sexual intercourse in many countries did not happen in a linear fashion: overall, there was an increase from 2002 to 2006, a plateau from 2006 to 2010, and then continuous decrease from 2010 to 2022. Despite being significant, the odds ratios were very close to 1, indicating no uniform temporal change in early sexual initiation between 2002 and 2022. Introducing gender and family affluence further improved model fit: girls had significantly lower odds of reporting early sexual initiation compared to boys, and so did children from medium-affluent families compared to their peers from high-affluent families. Low family affluence also meant significantly lower odds, but that odds ratio (OR) was again very close to 1. Those who did not provide information on their family’s affluence had similar odds to early sexual initiation as those from high-affluent families.

The addition of having a supportive parental figure further reduced AIC (while practically did not change BIC), implying that it increased the model fit. The respective odds ratio was significant, its value was very close to 1. In this model, 2.2% of the variance is explained by the difference between countries, and a much larger proportion (14.2%) by differences between schools.

The sequence of multivariate results featuring data from all 37 countries and the last three survey years are presented in Table 1, Column (b). Results are summarised in Table 4. In both 2018 and 2022, participants’ odds for early sexual initiation were significantly lower than in 2014, albeit they were not substantially lower than 0.9–0.8. Still, this was a robust pattern across the different models, adjusted for an increasing number of predictors. Adding gender and family affluence improved model fit: girls (contrasted to boys) and those from a medium- or low-affluent family (contrasted to those from high-affluent families) were significantly less likely to report early initiation, but again, the odds were not much different from 1. In the overall sample and among boys, those with a missing FAS response had significantly higher odds of reporting early initiation. Having a supportive parental figure did not make a practical difference in the odds of early intercourse, except an interaction between having a supportive parental figure × survey year (2022). However, this model (9b) had a higher BIC value than the one not including the interaction of parental support and survey years. Therefore, we returned to the previous model, 8(b) and applied it separately to boys and girls. The gender-disaggregated models are under column 10(b) in Table 4. Compared to 2014, there was already a small but significant decrease in the likelihood of early initiation (ORs between 0.7 and 0.9) for both boys and girls. From 2014 to 2022, the drop was significant in both genders. Family affluence was a significant predictor in both genders, but it only made a practical difference in girls’ early initiation (OR = 0.71 for medium and 0.78 for low affluence, respectively). For girls, a somewhat larger proportion of variation in early sexual intercourse was explained by school-related factors than for boys (22.2% and 14.8%, respectively); country differences had a lesser contribution, but again, they contributed more to the variation in girls’ early initiation than among boys (9.9% versus 3.0%, respectively).

This study employed the rigorous HBSC methodology, including using representative samples of adolescents across all countries, standardised protocol, comparability of data, and including those participants who did not respond to the gender or FAS items. The study, however, has limitations. In some countries, for ethical reasons, it was not allowed to use certain responses (e.g., sexual intercourse under 13) or response categories were used outside of the standard range of those in the question (i.e., sexual intercourse at the age of 17 or older). The data were weighted in some, but not all, countries and survey years.

There were large differences in the subsample sizes. In some countries and genders, only a small proportion of the total sample reported sexual initiation. These factors resulted in an analytic sample with large differences in the subsample sizes across genders, countries and survey years, which may have reduced the statistical power and the reliability of some estimates.

We used a repeated cross-sectional design over a 20-year period. This means that the participants examined over time were not the same individuals. Therefore, time and cohort effects could not be separated. In the HBSC survey rounds before 2022, no standardised questions were used to assess birth-registered sex, gender identity and sexual orientation of adolescents, and such questions are still administered at the discretion of the national HBSC research teams.

There are other factors than family SES and parental support that may influence early sexual initiation, including onset of puberty, engagement in risk behaviours, exposure to pornography and relationships with peers, which were not included in the present analysis. These also show large cross-cultural variations. Future studies need to consider these factors as potential determinants of early sexual initiation.

Early sexual initiation is associated with negative sexual outcomes in later life, and it is often non-consensual [61]. Adolescents should be empowered to make an informed and autonomous decision on when (and with whom) they want to have their first sexual intercourse. Comprehensive sexuality education, covering assertiveness and efficient communication and teaching the importance of consent, is paramount in this regard. Meta-analytic evidence supports the efficacy of such educational programmes [62, 63]. Decision-makers should ensure that these standards of comprehensive sexuality education are implemented in national primary and secondary education internationally. Furthermore, considerations of gender, socio-economic status and other potential reasons for marginalisation (e.g., immigrant background, LGBTQ+ identity) must also be included in the development of such educational programmes. Our results also underline the protective role of having supportive parental figures, whom adolescents can trust. Empowering parents, teachers and youth workers with whom children in their care can comfortably and effectively discuss sexuality-related issues should also be prioritised [60, 64–66]. Given the large proportion of the variation in early sexual initiation explained by school differences, we believe that it should be ensured that comprehensive sexuality education is made mandatory on the national level, and teachers and schools get adequate support to integrate these classes or modules in their local curricula.

Over the past two decades, there has been a slight but notable decline in the proportion of adolescents who have experienced sexual initiation before the age of 14. The rate of early initiation decreased from 6% in 2002 to 4% (3% among girls; 5% among boys) in 2022. Gender, family affluence and having a supportive parental figure also had an impact on the trends. While the proportion of youth engaging in early sex has decreased, given that early sexual initiation is associated with negative sexual health outcomes and is often non-consensual, concerted efforts are still needed from decision-makers, parents, teachers, and health and social care workers to help young people make informed decisions regarding their sexual initiation.