Here are the three questions I think are most worth asking. The first rests on the correlations I computed across all 4,500 students. The other two are proposals I haven't tested yet. ## 1. Is there a usage level where wellbeing and GPA start to fall off sharply? The correlations are strong across all 4,500 students: | Pair | Correlation | |---|---| | Daily usage vs. mental health index | -0.85 | | Daily usage vs. perceived stress | +0.754 | | Daily usage vs. sleep duration | -0.716 | | Daily usage vs. GPA | -0.71 | Average usage is 5.3 hours a day and average GPA is 3.43. A linear correlation can't say whether harm rises steadily or jumps past a threshold, for example around 6 to 8 hours. Banding `Daily_Usage_Hours` and comparing the average GPA, sleep, stress and mental health index in each band would show that. It would directly support the question of whether to recommend usage limits, and where to set them. ## 2. Does sleep explain the damage, or is it a separate effect? Sleep is linked to the outcomes too: - Sleep duration vs. GPA is +0.595. - Sleep quality vs. mental health is +0.662. This suggests sleep may carry part of the effect of heavy use. A useful test is to compare students with the same sleep duration but different usage, and see whether heavy users still have lower GPA or mental health. If they do, cutting screen time would help beyond what better sleep alone achieves. `Late_Night_Usage` (a yes/no flag) is a natural extra lever to test here. ## 3. Which student segments and platforms are most at risk? The dataset has age, gender and academic level, plus primary platform, device type and how often students compare themselves to others (`Social_Comparison_Frequency`). Comparing the stress, mental health and GPA gaps between heavy and light users within each segment would show: - where outreach should be prioritized; - whether a platform or device type is riskier once usage hours are held constant; - whether social comparison is a distinct driver of stress or just a marker of heavy use. The last point decides whether an intervention should target comparison behavior or simply screen time. ## Caveat The data is observational, so these relationships are associations, not proof that social media causes the outcomes. The correlations are also unusually strong for survey data. It would be worth checking how the dataset was generated before relying on the exact magnitudes.