Introduction

This project investigates how daily screen behaviors and sleep relate to self-reported mood (1–10), using a scale-aware, nonlinearity-aware analytics workflow. We analyze a daily, cross-sectional dataset with the following variables: screen_time_hours, hours_on_TikTok, sleep_hours, social_media_platforms_used, stress_level (1–10 higher=worse), and mood_score (1–10 higher=better).

Methodology

• Data hygiene: Schema standardization, type coercion (hours → float, scales → int), range/logic checks (0–24h; 1–10 scales; TikTok ≤ total screen), outlier flags (pre-registered), and a missingness audit (no meaningful missingness; MICE pipeline ready).
• EDA (scale-aware): Pearson and Spearman correlations; binned means; linear vs. spline nonlinearity checks.
• Modeling plan (overfitting-aware): Pre-registered hierarchy
  M1 sleep → M2 +screen → M3 +TikTok+platforms → M4 +interactions (screen×sleep; share×sleep) → M5 +splines (sleep & screen) + TikTok + platforms.
  Metrics: Adj-R², AIC/BIC, 5-fold CV: R², RMSE, MAE.
• Robustness: Ordinal outcome check (Ordered Probit for mood 1–10); VIF for collinearity; bootstrap CIs for effects.
• Exploratory mechanisms: Mediation (screen → sleep → mood; bootstrapped indirect effect) and moderation (stress × screen; simple slopes).
• Visualization standards: Transparent scatter + LOWESS (stratified by stress and TikTok share), PDP/ICE from a simple gradient boosting regressor; 7–8h sleep band annotated.

Models Used & Outputs

• Primary regression model: OLS with B-splines for sleep_hours and screen_time_hours (M5), plus hours_on_TikTok and social_media_platforms_used.
  • Out-of-sample performance (5-fold CV): CV R² ≈ 0.63, CV RMSE ≈ 0.78, CV MAE ≈ 0.59 on the 1–10 mood scale.
  • Info criteria: M5 had the lowest AIC/BIC among the hierarchy.
  • Robustness: An Ordered Probit yielded consistent signs/significance with the OLS/spline model.
• Classification side-task (bonus): Predict high stress (≥8) with a standard-scaled Logistic Regression (main effects + screen×sleep interaction).
  • ROC-AUC ≈ 0.936, PR-AUC ≈ 0.855, Brier ≈ 0.105 (good discrimination with reasonable calibration).

Key Findings

• Sleep ↔ Mood: Strong positive, nonlinear association; benefits rise to ~7–8 hours then plateau.
• Screen/TikTok ↔ Mood: Negative associations; TikTok time adds a distinct negative effect beyond total screen time; platform count adds little.
• Nonlinearity matters: Splines improved fit vs. linear models (CV R² ~0.63 vs. ~0.55).
• Mediation: Indirect effect via sleep ≈ 0 (bootstrap CI spans 0) in this cross-section.
• Moderation: At higher stress, the screen→mood slope is more negative (steeper decline).

Limitations & Scope

• Cross-sectional and self-reported → findings are associational, not causal; unmeasured confounding may remain (e.g., baseline mental health, workload).
• Ratios can inflate VIF when combined with their components; we avoid such combinations in final models.

Deliverables

• Reproducible Colab notebook with: data checks, EDA, model hierarchy + CV, effect sizes with 95% CIs (plus bootstrap), ordinal sensitivity, mediation/moderation, PDP/ICE visuals, and saved figures in /content/figures/.

In sum: A disciplined, scale-aware analysis shows that sleep is the strongest correlate of better mood, screen (especially TikTok) relates to worse mood, stress amplifies the negative screen effect, and nonlinearity is essential to model these relationships accurately. The best model achieves CV R² ≈ 0.63 with RMSE ≈ 0.78, adequate for group-level insights.