Comprehensive individual-user analysis on the GLOBEM dataset — a longitudinal passive-sensing + mental-health study of college students. Use this skill whenever a task involves analyzing a specific participant (e.g. "Analyze user INS-W_002") from the GLOBEM dataset, exploring behavioral patterns from smartphone sensors, correlating behavioral signals with mental health outcomes, or producing a comprehensive user profile from multimodal sensing data.
Comprehensive individual-user analysis on the GLOBEM dataset — a longitudinal passive-sensing + mental-health study of college students. Use this skill whenever a task involves analyzing a specific participant (e.g. "Analyze user INS-W_002") from the GLOBEM dataset, exploring behavioral patterns from smartphone sensors, correlating behavioral signals with mental health outcomes, or producing a comprehensive user profile from multimodal sensing data.
GLOBEM Individual User Analysis
Dataset Overview
The GLOBEM dataset tracks college students over a 92-day period (~April–July)
using passive smartphone sensing. Each participant has:
ERQ_reappraisal — cognitive reappraisal use (higher = more adaptive coping)
ERQ_suppression — emotional suppression (higher = more suppression)
CHIPS — health stressors count (higher = more stressed)
MAAS_7items — mindfulness (higher = more mindful, good)
SocialFit — social fit (higher = better fit)
2waySSS_receiving/giving_emotional — emotional social support exchange
2waySSS_receiving/giving_instrumental — instrumental social support exchange
BFI10_* — Big Five personality (extroversion, agreeableness, conscientiousness, neuroticism, openness)
Analysis Pipeline
Phase 1 — Orientation (1–2 calls)
# 1. list_files to confirm available files# 2. get_field_description on 2-3 sensor files to learn any extra column names needed# Do NOT call get_field_description for dep_weekly, ema, pre, post, dep_endterm, platform
Phase 2 — Per-modality stats (one call per modality)
Filter all sensor DFs by pid == '<user_id>'. For each modality compute:
Sleep — summary_rapids_avgdurationasleepmain (minutes), summary_rapids_avgefficiencymain (already 0–100, NOT decimal — never multiply by 100), summary_rapids_avgdurationtofallasleepmain. Convert bedtime/wake minutes-since-midnight to HH:MM for readability.
Communication — rapids_outgoing_count, rapids_incoming_count, rapids_missed_count, rapids_outgoing_meanduration, rapids_outgoing_distinctcontacts. Compute outgoing/incoming ratio (>1 = proactive).
Location — barnett_disttravelled, barnett_rog, barnett_homelabel (minutes/day at home), barnett_circdnrtn (0–1 circadian consistency), barnett_siglocsvisited. Filter GPS outliers: drop values > median × 10 for distance and rog before averaging.
Depression: weekly flag rate, end-term BDI2 + dep status
EMA: mean, std, min/max, first-half vs. second-half trend
Pre→Post survey changes for ALL key scales using exact column names above (report Pre value, Post value, and change with ↑↓ arrows). Interpret direction: improved or worsened.
2. High vs. low EMA day comparisons:
Split days by EMA median → report behavioral metric means for each group:
ema_median = user_ema['negative_affect_EMA'].median()
high_ema_dates = user_ema[user_ema['negative_affect_EMA'] > ema_median]['date']
low_ema_dates = user_ema[user_ema['negative_affect_EMA'] <= ema_median]['date']
# Compare steps, screen time, home time, etc. on high vs. low EMA days
3. Depression-flagged week behavior comparison:
Aggregate daily sensor data to weekly means (group by week). Join on dep_weekly dates (±7 days), then compare depressed vs. non-depressed week means for steps, sleep duration, phone unlocks, home time.
4. Temporal trends across modalities:
Summarize first-half → second-half changes for all key metrics in a consolidated table (steps, sleep, phone unlocks, distance, EMA).
Phase 5 — Data quality check (integrate into output)
For each modality, report valid days as n/92. Flag modalities with <20% coverage as "critically sparse — interpret with caution."
Synthesis Template
## Comprehensive Analysis of User <pid>
### Study Context
- Platform, study period (date range), data completeness per modality (n/92 days)
### Physical Activity
- Steps (mean ± std, min, max, valid days), sedentary/active balance, weekday vs. weekend
- Temporal trend: first-half → second-half step change
### Sleep
- Duration (hours), efficiency (%), timing (bedtime HH:MM, wake HH:MM), variability
- Temporal trend: sleep duration first-half → second-half
### Communication
- Call frequency, outgoing/incoming ratio, social diversity (distinct contacts)
- Temporal trend: outgoing call count first-half → second-half
### Location & Mobility
- Daily distance, home time (hours/day, %), circadian routine score, radius of gyration
- Temporal trend: distance or home time first-half → second-half
### Phone Usage
- Unlock count, screen time (hours), home vs. elsewhere split
- Temporal trend: unlock count or duration first-half → second-half
### Social Proximity (Connectivity)
- BT scan rate, unique devices per day
### Mental Health
- Depression trajectory: weekly flag rate (n/total weeks), end-term BDI2 + dep status
- EMA negative affect: mean ± std, trend (first-half → second-half mean)
- Pre→Post survey changes (all key scales with ↑↓ changes and interpretation)
### Cross-Modal Patterns
- EMA correlations with behavioral signals (list r values)
- Behavioral differences on high vs. low negative affect days
- Behavioral differences in depressed vs. non-depressed weeks
- Notable first-half → second-half behavioral shifts
### User Profile
- 3–5 sentence synthesis connecting behavioral patterns, temporal trends, and mental health
Common Pitfalls
Sleep efficiency: summary_rapids_avgefficiencymain is already percentage (e.g., 93.5). Never multiply by 100.
Survey columns must use exact names with _PRE/_POST suffix. Column discovery via df.columns.tolist() recommended on first access; get_field_description does NOT work on survey files.
Minute encoding: Bedtime/wake times are often minutes-since-midnight. Convert: f"{int(m//60):02d}:{int(m%60):02d}". Values like 1500 → 25:00 means 1:00 AM next day.
Sparse data: Some users have very few valid rows (<14/92 days for some modalities). Always check df[col].notna().sum() before computing stats and note effective n.
Location GPS outliers: barnett_disttravelled and barnett_rog can have extreme values from GPS errors. Use values[values < values.median() * 10] filtering before averaging.
Weekly vs. daily merge: dep_weekly is weekly; sensor data is daily. To compare, aggregate daily data into 7-day windows aligned with each dep_weekly date row.
EMA correlation requires merging on date: User's EMA may not overlap with all sensor dates. Use inner join and report n for each correlation.
2waySSS and ERQ are often missed but frequently asked about: Always extract emotion regulation (ERQ_reappraisal/suppression) and social support exchange (2waySSS giving/receiving) from pre/post surveys.
