# Most surprising finding (partial answer) **The blue-light filter appears to do almost nothing for sleep latency or sleep duration, while phone use is tightly linked to both.** This is a correlational finding from 8,500 rows, and the investigation was cancelled before I could test it further. ## What the data show **Phone use is strongly tied to sleep outcomes** (Pearson r, n = 8,500): - Phone use vs. sleep latency: **r = 0.858** - Phone use vs. next-day fatigue: **r = 0.711** - Phone use vs. total sleep hours: **r = -0.554** **The blue-light filter barely registers:** - Filter vs. sleep latency: **r = -0.057** - Filter vs. sleep hours: **r = 0.039** - Screen brightness vs. latency: **r = 0.093** **Group comparison (filter off vs. on):** | Metric | Filter off (n = 4,524) | Filter on (n = 3,976) | |---|---|---| | Avg sleep latency | 41.6 min | 39.6 min | | Avg sleep hours | 6.21 | 6.33 | | Avg fatigue | 3.99 | 3.58 | | Avg deep sleep % | 21.4 | 22.1 | Filter users fall asleep about 2 minutes faster, sleep about 0.11 hours longer, and report about 0.4 points less fatigue. These gaps are small next to the phone-use effects. I did not run significance tests, so I can't say whether they are statistically meaningful. ## Other notable patterns - **Snoozes track fatigue closely** (r = 0.881), and **sleep hours track fatigue inversely** (r = -0.879). These are the strongest relationships in the data. - **Caffeine after 5pm** has almost no link to total sleep hours (r = -0.038). It does correlate with **less deep sleep** (r = -0.343) and **longer latency** (r = 0.199). So caffeine seems to affect sleep quality more than quantity. - **Physical activity** is essentially unrelated to sleep hours (r = 0.004) and fatigue (r = -0.029). ## Caveats - The correlations are all linear and pairwise. I did not control for confounders, so none of this establishes causation. - The filter's small effect could be real, or it could be hidden if filter use is confounded with phone use. I did not check whether filter users use their phones more or less. That would be the next step. - The phone-latency correlation of 0.858 is unusually high for behavioral data. I did not verify how the phone-use variable was measured, so it could be partly an artifact of how the data were generated or recorded. - I picked "most surprising" by judgment. The filter's near-zero effect stands out because it contradicts a common assumption, but the phone-use effect is the larger finding.