Windows of developmental sensitivity to social media

The relationship between social media use and life satisfaction changes across adolescent development. Our analyses of two UK datasets comprising 84,011 participants (10–80 years old) find that the cross-sectional relationship between self-reported estimates of social media use and life satisfaction ratings is most negative in younger adolescents. Furthermore, sex differences in this relationship are only present during this time. Longitudinal analyses of 17,409 participants (10–21 years old) suggest distinct developmental windows of sensitivity to social media in adolescence, when higher estimated social media use predicts a decrease in life satisfaction ratings one year later (and vice-versa: lower estimated social media use predicts an increase in life satisfaction ratings). These windows occur at different ages for males (14–15 and 19 years old) and females (11–13 and 19 years old). Decreases in life satisfaction ratings also predicted subsequent increases in estimated social media use, however, these were not associated with age or sex.


Supplementary Figure 1: Estimated social media use and life satisfaction ratings across the lifespan (Extended version)
Top: Cross-sectional correlation between estimated social media use and a one-item satisfaction with life measure for 72,287 UK participants of the Understanding Society dataset between the age of 10 and 80 years. The results are split by age and sex: female = red, male = blue. The 95% confidence intervals represent the lower and upper Gaussian confidence limits around the mean based on the t-distribution. Middle: Frequency distribution of estimated social media use by age and sex. Bottom: Shading of each rectangle represents whether a model relating estimated social media use and life satisfaction ratings that takes into account a possible sex difference is more likely to represent the data than a model that does not take into account sex: darker shade = model with sex differences is more likely.
It should be noted that as high levels of social media use are very rare in the youngest and oldest age groups present in the data (e.g., ages 10, 11 and 60+, Supplementary Figure 1), we cannot evaluate functional form in these groups. Further, as most participants were measured multiple times, more than one data point per participant will appear in this graph. Source data for this figure are provided as a Source Data file .   10  11  12  13  14  15  16  17  18  19  20  21  22−25  26−29  30s  40s  50s  60s  Difference between the predicted life satisfaction of those participants who report using 7 or more hours of social media vs those who report using no social media The graph plots the difference between the model predicted life satisfaction ratings for those who report using 7+ hours of social media (score: 5) and those using no social media (score: 1) for different age groups (shown on the x axis). The model was a linear regression predicting life satisfaction from social media use, social media use 2 and the covariates of log household income, neighbourhood deprivation and year of data collection (life satisfaction ratings ~ estimated social media use + estimated social media use 2 + log household income + neighbourhood deprivation + year of data collection). No statistical significance tests were carried out, and the statistical assumptions of the regression were not tested, further no correction for multiple comparisons were made.
It shows that the link between social media use and life satisfaction ratings is most negative in early adolescence compared with older adolescence and adulthood, but there are other ages when the relationship also becomes slightly more negative (e.g., males aged 26-29). Source data for this figure are provided as a Source Data file. Results from an Akaike weights procedure where two models were fit to data of different age groups. The models had the following form: life satisfaction ~ a*social media + b*social media 2 (and including control variables for log household income, neighbourhood deprivation and year of data collection). One of the models allowed the parameters a and b to vary by sex, while the other model did not. The model fit indices (AIC) were compared using an Akaike weights procedure, and the ratio of the weights are plotted on the y-axis: the higher the bar, the more a model with sex differences is more likely to be a better model for the data than a model without sex differences at that age group. The graph gives no indication about which direction the sex difference is in. The y-axis is on a log scale to account for the rapidly increasing Akaike weights ratios. Source data for this figure are provided as a Source Data file.

Supplementary Methods 1
In the extension of the cross-sectional analyses, we analysed a range of questionnaires that were only completed by 10-15-year-olds in the Understanding Society survey and questionnaires completed by 13-and 14-year-olds in the Millennium Cohort Study.

Measures: Understanding Society
We constructed supplementary cross-sectional plots examining well-being and mental health questionnaires completed only by the younger adolescent sample (age 10-15 years). These included a longer well-being questionnaire at every wave which supplemented the single life satisfaction question used above with questions regarding satisfaction with school work, appearance, family, friends and school. Further, at every even wave adolescents completed an 8-item self-esteem scale and at every odd wave they filled in a 25-item Strengths and Difficulties Questionnaire (SDQ). The SDQ is a commonly used and widely validated measure for psychosocial functioning applied in school, home and clinical environments 1 . It encompasses five questions each about emotional symptoms, conduct problems, hyperactivity/inattention, peer relationship problems and prosocial behaviour (0 = not true, 1 = somewhat true, 2 = certainly true). We analysed conduct scores as a measure of externalising symptoms and emotional scores as an internalising symptoms measure.

Measures: Millennium Cohort Study
In the Millennium Cohort Study the SDQ measure was the same as in Understanding Society, but it was completed by the adolescent's caregiver. Furthermore the Millennium Cohort Study included a slightly different self-esteem measure: the 5-item Shortened Rosenberg Self-esteem scale 2 . Lastly the adolescents were asked to fill out a depressive symptoms scale in form of the "Mood and Feelings Questionnaire short form" 3 , asking adolescents: "For each question please select the answer which reflects how you have been feeling or acting in the past two weeks." Responses were "I felt miserable or unhappy," "I didn't enjoy anything at all," "I felt so tired I just sat around and did nothing," "I was very restless," "I felt I was no good any more," "I cried a lot," "I found it hard to think properly or concentrate," "I hated myself," "I was a bad person," "I felt lonely," "I thought nobody really loved me," "I thought I could never be as good as other kids," and "I did everything wrong" (1 = not true, 2 = sometimes, 3 = true; scale subsequently reverse scored).

Latent Factors
As the questionnaires in the Understanding Society and Millennium Cohort Study surveys had more than one item, we first extracted latent factors and applied model comparison to examine measurement invariance across sex before plotting the latent factors' relation to estimated social media use. Due to the large sample size, chi-square tests for measurement invariance over sex routinely rejected the null hypothesis. We therefore claimed partial measurement invariance even if the null hypothesis was rejected, but when the completely freed model was preferred over the constrained model by the BIC (or when the BIC's are equal), which generally penalizes complexity more harshly than a likelihood ratio test.
We first analysed Understanding Society. To achieve partial measurement invariance for the well-being questionnaire, we allowed one item (satisfaction with school work) to vary across sex as it loaded more highly for males than females:  2 (4) = 15. 4 Figure 4, Panel A). We also plotted each life satisfaction question's raw scores by social media use and age in Supplementary Figure 4, Panel B to examine whether a specific aspect of life satisfaction was more negatively related to estimated social media use (see also Figure  2).

Supplementary Methods 2
The BayesFactor package was used to calculate the Bayesian regression: "The vector of observations y is assumed to be distributed as: y ~ Normal(α 1 + Xβ, σ^2 I).

Supplementary Results 1
We supplemented the analyses in the main manuscript by extracting factor scores for a wider set of well-being and mental health measures completed by 10-15-year-olds in our original dataset, Understanding Society, and 13-14-year-olds drawn from a different cohort (the UK Millennium Cohort Study, 11,724 participants). The limited age range did not allow us to compare these adolescents with other age groups, however the datasets present a similar pattern of sex differences across a wider range of psychosocial outcomes (Supplementary Figure 4, Panel A). In Understanding Society data, self-reported well-being showed the largest mean sex differences across ages 10-15 when compared to questionnaires such as selfesteem or depressive symptoms (Akaike weight of model with sex difference compared to model without sex difference: well-being 72.9%, self-esteem 55.3%, internalising symptoms 57.9%, externalising symptoms 56.2%). In the Millennium Cohort Study, the models differentiating between sex for well-being, depression and self-esteem were much more likely to be the better models of the data than those that did not differentiate for sex (100%), while those for externalising symptoms (21.3%) and internalising symptoms (35.5%) were not. For analyses of the correlations in Understanding Society over time see Supplementary Figure 5.
The well-being questionnaire (a questionnaire whose constituent questions include the satisfaction with life question analysed in the main manuscript, Figure 2) therefore seems to show the most substantial sex differences. Examining the constituent sub-questions that make up the wellbeing questionnaire (satisfaction with appearance, family, friends, school, school work and life), we found no evidence that a specific sub-component of life satisfaction was the lone driver of these sex differences (Supplementary The plot shows that the cross-sectional correlation between estimated social media use and mental health and well-being measures stays relatively stable across early adolescence, except wellbeing whose correlation becomes more negative in females during the period. Using the confidence intervals that represent standard errors around the mean, one can interpret which measures' relation to social media use is different statistically. For example, across both sexes and most ages, well-being and conduct symptoms are more negatively related to social media use than other measures. No correction for multiple comparisons was made. Source data for this figure are provided as a Source Data file. The plot shows that there are differences in how estimated social media use relates to life satisfaction measures across age and sex. Using the confidence intervals that represent standard errors around the mean, one can interpret which measures' relationship with social media use is different statistically. For example, in females the relationship becomes more negative for satisfaction with appearance, life, school and school work; while the relationship between estimated social media use and satisfaction with family and friends is less negative than other measures and stays relatively stable over age. No correction for multiple comparisons was made. Source data for this figure are provided as a Source Data file. Results from an Akaike weights procedure where four different Random-Intercept Cross Lagged Panel Models were fitted to the data: 1) a model that allowed both cross lagged paths to vary by sex and age ("Free Both"); 2) a model that allowed only the cross lagged path from life satisfaction ratings (LS) to estimated social media (SM) use to vary by sex and age ("Free LS -> SM"); 3) a model that allowed only the cross lagged path from estimated social media use to life satisfaction ratings to vary by sex and age ("Free SM -> LS"); and 4) a model that allowed none of the cross lagged paths to vary by sex and age ("Constrained"). The higher the bar the more the model is preferred over the others. Source data for this figure are provided as a Source Data file.  To examine how missingness develops over both younger and older adolescent cohorts in Understanding Society we investigated the percentage missingness in those age cohorts who remained in the adolescent survey for all seven waves of data collection (those aged 10, 11, 12, 13, 14 and 15 during the first wave). We find that missingness increases over time, but that it increases faster in later age cohorts. This shows that attrition is greater at older ages. Source data for this figure are provided as a Source Data file. In addition to the analyses in Supplementary Figure 9, we examined selective attrition further in those age cohorts who remained in the Understanding Society adolescent survey for all seven waves of data collection (those aged 10, 11, 12, 13, 14 and 15 during the first wave). To do so, we coded those adolescents as '1' who had data for wave 1 and wave 7, and those as '0' for those who had data for only wave 1 and not wave 7. We then used ordinal regression to regress this attrition variable onto the predictor variables of sex, mean income (inc), mean index of multiple deprivation (imd), social media use at wave 1 (sm) and life satisfaction at wave 1 (ls). The test was two-tailed (OR = odds ratios) We found that the only consistent predictor of attrition is sex (males showing higher levels of attrition). Further, for half of the age cohorts lower income predicts higher levels of attrition. No correction for multiple comparisons was made.