**The most counterintuitive finding is that social media use looks almost mechanically tied to wellbeing, yet weekend extra hours are statistically unrelated to everything.** **1. Daily use tracks every outcome very tightly (n = 4,500).** - Daily usage hours correlate at **-0.85** with the Mental Health Index. Real survey data on behaviour and wellbeing rarely gets this clean. - Daily usage correlates **-0.72** with sleep hours, **-0.71** with GPA and **-0.68** with sleep quality. It correlates **+0.75** with perceived stress. - The outcomes also move together: sleep and mental health at +0.77, stress and mental health at -0.77, GPA and mental health at +0.66. **2. Weekend extra hours show no relationship with anything.** - Weekend extra hours vs. daily usage: **-0.010**. - Weekend extra hours vs. mental health: **0.004**. - Weekend extra hours vs. GPA: **0.012**. - Weekend extra hours vs. sleep: **0.017**. - Weekend extra hours vs. stress: **0.001**. - Age vs. GPA is also essentially zero, at **0.028**. You would expect people who use social media heavily on weekdays to also use it more at weekends. Here the weekend measure behaves like independent noise, while the weekday measure behaves like a near-deterministic driver. **3. The self-rated impact labels line up with usage.** The Overall_Impact groups are: | Overall_Impact | Students | Avg daily hours | Avg mental health | Avg GPA | Avg sleep (hours) | |---|---|---|---|---|---| | Beneficial | 3,681 | 4.4 | 84.3 | 3.55 | 7.0 | | Neutral | 654 | 8.5 | 63.4 | 3.00 | 5.5 | | Negative | 165 | 12.3 | 47.1 | 2.54 | 4.2 | About 82% of students (3,681 of 4,500) are in the "Beneficial" group, so most students describe social media as helping them. Within that group, usage is lower and outcomes are better, so the negative effects are concentrated in a small high-usage tail. **How to read this** - The data is observational, so none of this shows that usage causes the outcomes. Sleep loss or stress could just as well drive heavy use. - The strength and cleanliness of these relationships, together with the lack of any weekend signal, make the data look engineered or synthetic rather than organically collected. That is my inference, not something the dataset states. Treat the effect sizes as illustrative, not as real-world estimates.