**Yes. `Overall_Impact` (Beneficial / Neutral / Negative) is the most natural way to segment these records.** The three groups separate sharply on every measure in the table above. | Segment | Students | Avg daily hours | Avg sleep (h) | Avg stress | Avg mental health | Avg GPA | Late-night share | |---|---|---|---|---|---|---|---| | Beneficial | 3,681 (~82%) | 4.42 | 7.03 | 11.16 | 84.26 | 3.55 | 0.54 | | Neutral | 654 (~15%) | 8.50 | 5.48 | 22.30 | 63.42 | 3.00 | 0.82 | | Negative | 165 (~4%) | 12.25 | 4.23 | 31.12 | 47.13 | 2.54 | 0.82 | - **The groups form a gradient.** Moving from Beneficial to Neutral to Negative, daily usage rises from about 4.4 to 8.5 to 12.3 hours. Sleep falls from 7.0 to 4.2 hours, and sleep quality falls from 2.72 to 1.04. Stress rises from 11.2 to 31.1. Mental health drops from 84 to 47, and GPA drops from 3.55 to 2.54. - **The groups are not evenly sized.** About 82% of students are Beneficial and only 165 are Negative. Any comparison involving the Negative group rests on a small sample. - **Late-night use looks like a threshold.** Its share is 0.54 in the Beneficial group and 0.82 in both the Neutral and Negative groups, so it does not keep rising. - **This is a correlation, not a cause.** `Overall_Impact` looks like a label that summarises the other measures. It is probably derived from them or self-reported, so it describes outcomes rather than explaining them. The data is also observational, so the table cannot show whether heavy usage causes the poorer outcomes. **Other ways to segment.** The dataset has other categorical columns: `Academic_Level`, `Gender`, `Age`, `Primary_Platform`, `Device_Type`, `Late_Night_Usage` and `Social_Comparison_Frequency`. These are useful for profiling who falls into each impact group. I did not test how strongly any of them separates the outcomes, so I can't say they matter as much as `Overall_Impact`. **Usage bands.** Banding `Daily_Usage_Hours` (for example under 5, 5–9 and over 9 hours) would probably recreate the same three groups from a behavioural input. That would be a more actionable segmentation for outreach, but I did not run it.