2nd/10_Wiki/Topics/Architecture/Experience-Sampling-Method.md

---
id: wiki-2026-0508-experience-sampling-method
title: Experience Sampling Method
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [ESM, EMA, Ecological Momentary Assessment, Diary Studies]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [research-methodology, psychology, ux-research, mobile]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: typescript
  framework: react-native
---

# Experience Sampling Method

## 매 한 줄
> **"매 retrospection bias 의 in-the-moment self-report 의 replace"**. Experience Sampling Method (ESM, Csikszentmihalyi & Larson 1987) 매 participants 의 day 의 multiple times 의 random/scheduled prompt, 매 current activity/affect/context 의 record. Mobile-era 매 EMA (Ecological Momentary Assessment) 의 generalize — 매 mental health, UX, productivity research 의 gold standard.

## 매 핵심

### 매 Why ESM
- **Retrospection bias**: 매 "지난주 어땠나" 매 peak-end bias, mood-congruent recall 의 distort.
- **Ecological validity**: 매 in-context 매 lab 의 not-replicate.
- **Within-subject variance**: 매 person × situation 의 interaction 의 capture.

### 매 Sampling schedules
- **Signal-contingent**: 매 random beep 매 day 매 6-8 prompts.
- **Interval-contingent**: 매 fixed times (9am/12pm/3pm/6pm).
- **Event-contingent**: 매 specific event (meal, exercise) 의 trigger.
- **Hybrid**: 매 baseline random + event triggers.

### 매 응용
1. Mood / affect tracking (depression, bipolar).
2. Pain studies (chronic pain).
3. UX product research — feature use in-context.
4. Flow state research (Csikszentmihalyi original).
5. LLM agent behavior tracking — analog 매 process.

## 💻 패턴

### Mobile prompt scheduler (React Native)
```typescript
import * as Notifications from 'expo-notifications';

interface ESMConfig {
  startHour: number; endHour: number;
  promptsPerDay: number;
  minIntervalMinutes: number;
}

async function schedulePrompts(cfg: ESMConfig, days = 7) {
  const slots = generateRandomSlots(cfg, days);
  for (const slot of slots) {
    await Notifications.scheduleNotificationAsync({
      content: {
        title: 'Quick check-in (30s)',
        body: 'How are you feeling right now?',
        data: { promptId: slot.id, scheduledFor: slot.time.toISOString() },
      },
      trigger: { date: slot.time },
    });
  }
}

function generateRandomSlots(cfg: ESMConfig, days: number) {
  const slots = [];
  for (let d = 0; d < days; d++) {
    const dayStart = new Date();
    dayStart.setDate(dayStart.getDate() + d);
    const windowMs = (cfg.endHour - cfg.startHour) * 3600_000;
    const minGap = cfg.minIntervalMinutes * 60_000;
    const times: number[] = [];
    while (times.length < cfg.promptsPerDay) {
      const candidate = Math.random() * windowMs;
      if (times.every(t => Math.abs(t - candidate) >= minGap)) {
        times.push(candidate);
      }
    }
    times.sort((a, b) => a - b).forEach((offset, i) => {
      const t = new Date(dayStart);
      t.setHours(cfg.startHour, 0, 0, 0);
      t.setTime(t.getTime() + offset);
      slots.push({ id: `${d}-${i}`, time: t });
    });
  }
  return slots;
}
```

### Brief response form (PANAS-short, 30s budget)
```typescript
interface ESMResponse {
  promptId: string;
  respondedAt: Date;
  latencyMs: number;
  affect: {
    valence: number;    // -3..+3
    arousal: number;    // -3..+3
  };
  activity: string;     // dropdown: work | social | rest | exercise | other
  social: 'alone' | 'with_others';
  freeText?: string;
}
```

### Compliance tracking
```typescript
function complianceMetrics(responses: ESMResponse[], scheduled: number) {
  const completed = responses.length;
  const onTime = responses.filter(r => r.latencyMs < 15 * 60_000).length;
  const meanLatency = responses.reduce((s, r) => s + r.latencyMs, 0) / completed;
  return {
    completionRate: completed / scheduled,    // target > 0.7
    onTimeRate: onTime / scheduled,            // target > 0.5
    meanLatencyMin: meanLatency / 60_000,
  };
}
```

### Multilevel analysis (within vs between)
```python
import statsmodels.formula.api as smf

# 매 each row 매 prompt response, 매 participant_id 매 grouping
model = smf.mixedlm(
    'valence ~ activity + social + time_of_day',
    data=df,
    groups=df['participant_id'],
    re_formula='~time_of_day',
).fit()
print(model.summary())
# 매 within-person variance (situation) 매 between-person (trait) 의 separate
```

### Sliding-window mood detection
```typescript
function detectMoodEpisode(responses: ESMResponse[], windowDays = 7, threshold = -1.5) {
  const sorted = [...responses].sort((a, b) =>
    a.respondedAt.getTime() - b.respondedAt.getTime());
  const episodes = [];
  for (let i = 0; i < sorted.length; i++) {
    const start = sorted[i].respondedAt;
    const end = new Date(start.getTime() + windowDays * 86400_000);
    const window = sorted.filter(r =>
      r.respondedAt >= start && r.respondedAt <= end);
    if (window.length < 5) continue;
    const meanV = window.reduce((s, r) => s + r.affect.valence, 0) / window.length;
    if (meanV < threshold) episodes.push({ start, end, meanV, n: window.length });
  }
  return mergeOverlapping(episodes);
}
```

### Privacy: 매 on-device aggregation
```typescript
// 매 raw responses 매 device 의 stay, 매 weekly summary 만 의 server 의 send
async function uploadWeeklySummary(responses: ESMResponse[]) {
  const summary = {
    week: getCurrentWeek(),
    n: responses.length,
    valenceMean: mean(responses.map(r => r.affect.valence)),
    valenceStd: std(responses.map(r => r.affect.valence)),
    activityHistogram: histogram(responses.map(r => r.activity)),
  };
  await api.post('/esm/summary', summary);
}
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Trait measurement (depression baseline) | 매 ESM unnecessary — 매 single questionnaire 매 fine |
| Within-day variation 의 question | 매 ESM signal-contingent |
| Specific event 매 rare | 매 event-contingent |
| Compliance fragile | 매 prompt count 의 reduce, 매 incentive |
| Privacy-sensitive (clinical) | 매 on-device aggregation 또는 federated |

**기본값**: 매 6-8 prompts/day, 매 14 days, 매 30s response — 매 compliance > 70% 의 target.

## 🔗 Graph
- 부모: [[Research-Methodology]]
- 변형: [[Ecological-Momentary-Assessment]]
- 응용: [[Flow_State|Flow-State]]

## 🤖 LLM 활용
**언제**: 매 in-the-moment subjective state 의 measure. Within-person variance 의 study. Retrospective bias 의 likely.
**언제 X**: 매 stable trait. 매 single-shot decision study. 매 intrusive sampling 매 acceptable 의 X.

## ❌ 안티패턴
- **Too many prompts**: 매 12+/day 매 fatigue → compliance crash.
- **Long forms**: 매 5min response 매 ecological 의 break.
- **Ignoring missing-not-at-random**: 매 prompts during depressive episode 매 skipped — 매 selection bias.
- **Cross-sectional analysis 의 hierarchical data**: 매 multilevel model 의 use, 매 OLS 의 std error 의 underestimate.

## 🧪 검증 / 중복
- Verified (Csikszentmihalyi & Larson 1987 JNMD; Shiffman et al. 2008 Ann Rev Clin Psych).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — ESM scheduler + analysis + privacy patterns |