Fig. 4
Forest plot of the mixed linear models results on the associations between TikTok, WhatsApp, and Instagram screentime, number of activations, and number of notifications and hedonic/eudaimonic well-being. Figure shows the unstandardized coefficients and confidence interval between well-being outcomes (Y) and App-specific features (X) of the same day from mixed linear models, adjusting for well-being outcome of the previous day, age, gender, and school in each model, with time (each day) as level one, and adolescents as level two. Well-being outcomes are collected from self-reported questionnaires, and App-specific features are derived using text extraction pipeline from the cellphone screenshots of the setting pages uploaded by users, both within the same EMA study. For example, happy (y1)—screentime of Instagram (x1) is one pair of well-being outcomes—App-use feature predictors. We have 288 pairs of such outcomes (12 individual outcomes and 4 aggregated outcomes) and predictors (3 features per App for 6 Apps). Figure shows the results for TikTok, WhatsApp, and Instagram. The results for Game Apps, YouTube, and Snapchat are presented in S-Fig. 4. For each pair of well-being outcomes—App-use feature predictors, we fit a random intercept model and random slope model. We apply the likelihood ratio test to select the best model following the parsimonious rule. Bonferroni correction is applied to the confidence interval to account for multiple testing. For each well-being outcome and App-use feature, we adjust the level of significance by 0.05/# of tests (# = 6) and the confidence interval based on the adjusted level of significance. The confidence interval is adjusted by \(\:Estimate\pm\:qnorm\left(1-\frac{0.05}{2*\left(\#-1\right)}\right)*Standard\_error\), where # = 6. The interactive Plotly version of this plot is available in OSF project in path “/code_and_data_for_replication /within_level”.
