Globem User Analysis

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:

Sensor CSVs (daily rows, columns: Unnamed: 0, pid, date, + features):

Mental-health / survey files (no _fields.json — only readable via execute_code):

Tool Usage Rules

get_field_description only works for the six sensor CSVs above. Do NOT call it for dep_weekly, dep_endterm, ema, pre, post, platform — those have no _fields.json and will always fail. Read them directly with execute_code.

Recommended Analysis Pipeline

Phase 1 — Orientation (2–3 calls)

# 1. list_files to confirm available files
# 2. get_field_description on 2-3 sensor files to learn key column names
# 3. Optional: read dataset_summary.json once for participant count & date range

Phase 2 — Per-modality stats for the target user (6 calls, one per modality)

Filter all sensor DFs by pid == '<user_id>' and compute:

Activity — key columns:

• intraday_rapids_sumsteps — daily step count
• intraday_rapids_countepisodesedentarybout / activebout — sedentary vs. active episodes
• Compute: mean ± std, min/max, weekday vs. weekend difference

Sleep — key columns:

• summary_rapids_avgdurationasleepmain — sleep duration (minutes)
• summary_rapids_avgefficiencymain — sleep efficiency (already 0–100, NOT decimal)
• summary_rapids_avgdurationtofallasleepmain — sleep onset latency
• Time of going to sleep / waking up can be inferred from related columns

Communication — key columns:

• rapids_outgoing_count, rapids_incoming_count, rapids_missed_count
• rapids_outgoing_meanduration, rapids_outgoing_distinctcontacts
• Compute: outgoing/incoming ratio (>1 = proactive communicator)

Location — key columns:

• barnett_avgflightdur, barnett_avgflightlen — mobility metrics
• barnett_homelabel — minutes at home
• barnett_circdnrtn — circadian routine consistency (0–1)
• barnett_rog — radius of gyration (meters)
• barnett_siglocsvisited — significant locations visited

Phone Usage — key columns:

• rapids_countepisodeunlock — unlock frequency
• rapids_sumdurationunlock — total screen time (minutes)
• rapids_sumdurationunlockhome — home screen time

Connectivity — key columns:

• rapids_countscans — total BT scans/day
• rapids_uniquedevices — unique BT devices detected
• rapids_countscanshighestpowerlevel — social proximity signal

Phase 3 — Mental health profile (1–2 calls)

dep_weekly = pd.read_csv('dep_weekly.csv')
dep_endterm = pd.read_csv('dep_endterm.csv')
ema = pd.read_csv('ema.csv')
pre = pd.read_csv('pre.csv')
post = pd.read_csv('post.csv')
platform = pd.read_csv('platform.csv')

user_dep = dep_weekly[dep_weekly['pid'] == uid]
user_endterm = dep_endterm[dep_endterm['pid'] == uid]
user_ema = ema[ema['pid'] == uid]
user_pre = pre[pre['pid'] == uid]
user_post = post[post['pid'] == uid]
user_platform = platform[platform['pid'] == uid]

Extract and report:

• Platform (iOS/Android)
• Pre/post survey changes: STAIS (anxiety), PSS (stress), CESD (depression), BFI10 personality, UCLA (loneliness), BRS (resilience)
• Weekly depression flag rate (dep=True weeks / total weeks)
• EMA negative affect: mean, std, trend over time
• End-term BDI2 score and dep status

Phase 4 — Cross-modal correlation & synthesis (1–2 calls)

Merge sensor dataframes on pid + date, then compute:

• Pearson correlations: steps ↔ phone usage, steps ↔ sleep duration, EMA ↔ activity/sleep
• Compare behavioral metrics between depression-flagged vs non-flagged weeks
• Temporal trend: compare first-half vs second-half metrics for activity, sleep, phone usage

Phase 5 — Data quality check (optional, 1 call)

Count non-NaN rows per modality. If a modality has fewer valid rows than the 92-day maximum, note this in the summary (e.g., activity data available only 14/92 days).

Synthesis Template

Structure the final output as:

## Comprehensive Analysis of User <pid>

### Study Context
- Platform, study period, data completeness per modality

### Physical Activity
- Daily steps (mean ± std, min, max), sedentary/active balance, weekday vs. weekend

### Sleep
- Duration (hours), efficiency (%), timing (bedtime, wake time), variability

### Communication
- Call frequency, outgoing/incoming ratio, social diversity

### Location & Mobility
- Daily distance, home time (%), circadian routine score, radius of gyration

### Phone Usage
- Unlock count, screen time, home vs. elsewhere split

### Social Proximity (Connectivity)
- BT scan rate, unique devices per day

### Mental Health
- Depression trajectory (weekly flag rate, BDI2 endpoint)
- EMA negative affect trend
- Pre→post survey changes (highlight ↑↓ on key scales)

### Cross-Modal Patterns
- Notable correlations between behavioral and mental health signals
- Behavioral differences in depressed vs. non-depressed weeks

### User Profile
- 3–5 sentence synthesis connecting lifestyle, behavior, and mental health

Common Pitfalls

1. Sleep efficiency: summary_rapids_avgefficiencymain is already in percentage (e.g., 93.5, not 0.935). Never multiply by 100 — values of 9000%+ are a sign that you double-counted.

2. Missing field descriptions: get_field_description fails for dep_weekly, ema, pre, post, dep_endterm, platform. Read their structure directly by printing df.columns.tolist() and df.head(2) inside execute_code.

3. Sparse data: Some participants have very few valid rows for activity or sleep (e.g., 14/92 days). Always check df.notna().sum() before computing averages and note the effective sample size.

4. Merging sensor + weekly data: Sensor data is daily; dep_weekly is weekly. When comparing, aggregate sensor data to weekly means before joining on pid and approximate date window.

5. Avoid re-running dataset_summary.json: It gives population-level metadata. One read is enough — don't repeat it across multiple turns.

6. Timezone / minute encoding: Bedtime and wake time are often stored as minutes-since-midnight. Convert to HH:MM for readability (e.g., 150 min → 2:30 AM).