people-analytics（人员分析）

People Analytics

The agent operates as a senior people analytics partner, translating workforce data into actionable insights using statistical modeling, segmentation analysis, and data governance best practices.

Workflow

1. Frame the question -- Clarify the business question with the HR or business stakeholder. Examples: "Why is Sales attrition 2x the company average?" or "Are we paying equitably across gender?" Define the success metric for the analysis.
2. Assess data readiness -- Identify required data sources (HRIS, ATS, survey platform, payroll). Check for completeness, recency, and quality. Flag any gaps before proceeding.
3. Analyze -- Apply the appropriate method from the analytics toolkit (descriptive stats, regression, classification, segmentation). Document assumptions and limitations.
4. Validate findings -- Sense-check results with domain experts (HRBPs, managers). Test for statistical significance and practical significance. Check predictive models for bias across protected groups.
5. Recommend -- Translate findings into 2-3 specific, actionable recommendations with expected impact and cost.
6. Deliver and monitor -- Present insights using the dashboard framework. Set up ongoing monitoring for key metrics with alert thresholds.

Checkpoint: After step 2, confirm that all data has been anonymized or aggregated to comply with privacy policy before analysis begins.

Analytics Maturity Model

Core HR Metrics

Workforce Metrics

Performance Metrics

Engagement Metrics

Turnover Prediction Model

import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report

def build_turnover_model(employee_data: pd.DataFrame) -> dict:
    """
    Build and evaluate a turnover prediction model.

    Input: DataFrame with columns for features + 'left_company' (0/1).
    Output: dict with model, feature importance, and evaluation metrics.
    """
    features = [
        'tenure_months', 'salary_ratio_to_market', 'performance_rating',
        'months_since_last_promotion', 'manager_tenure', 'team_size',
        'engagement_score', 'training_hours_ytd', 'projects_completed'
    ]

    X = employee_data[features]
    y = employee_data['left_company']

    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=0.2, random_state=42, stratify=y
    )

    model = RandomForestClassifier(n_estimators=100, random_state=42)
    model.fit(X_train, y_train)

    y_pred = model.predict(X_test)
    report = classification_report(y_test, y_pred, output_dict=True)

    importance = (
        pd.DataFrame({'feature': features, 'importance': model.feature_importances_})
        .sort_values('importance', ascending=False)
    )

    return {'model': model, 'importance': importance, 'evaluation': report}

def score_flight_risk(model, current_employees: pd.DataFrame) -> pd.DataFrame:
    """
    Score current employees for flight risk.

    Returns DataFrame with employee_id, flight_risk_score (0-1), and risk_level.
    """
    probabilities = model.predict_proba(current_employees[model.feature_names_in_])[:, 1]

    risk_levels = pd.cut(
        probabilities,
        bins=[0, 0.25, 0.50, 0.75, 1.0],
        labels=['Low', 'Medium', 'High', 'Critical']
    )

    return pd.DataFrame({
        'employee_id': current_employees['employee_id'],
        'flight_risk_score': probabilities.round(3),
        'risk_level': risk_levels
    }).sort_values('flight_risk_score', ascending=False)

Example: Sales Attrition Root-Cause Analysis

QUESTION
  Sales voluntary turnover is 22% vs 12% company average. Why?

DATA
  Source: HRIS + engagement survey + exit interviews (n=45 exits, trailing 12 mo)

ANALYSIS
  Segmentation by tenure band:
    < 1 yr: 35% of exits (onboarding/ramp issues)
    1-2 yr: 40% of exits (comp dissatisfaction + career path)
    2+ yr: 25% of exits (manager relationship)

  Regression on exit survey scores (n=38 respondents):
    Top drivers of intent-to-leave:
      1. "I am paid fairly" (beta = -0.42, p < 0.01)
      2. "I see a career path here" (beta = -0.31, p < 0.01)
      3. "My manager supports my development" (beta = -0.28, p < 0.05)

  Compensation benchmark:
    Sales IC3 compa-ratio: 0.88 (12% below midpoint)
    Sales IC2 compa-ratio: 0.91 (9% below midpoint)
    Rest of company average: 0.98

FINDINGS
  1. Sales comp is significantly below market, especially at IC2-IC3
  2. No defined career ladder for Sales ICs beyond IC3
  3. New hires (< 1 yr) leaving due to unrealistic ramp expectations

RECOMMENDATIONS
  1. Market adjustment: Bring Sales IC2-IC3 to 95th percentile compa-ratio ($180K budget)
  2. Publish a Sales career ladder through IC5 with clear promotion criteria
  3. Redesign onboarding: extend ramp period from 30 to 90 days with milestone targets

EXPECTED IMPACT
  Reduce Sales attrition from 22% to 14-16% within 12 months
  ROI: $180K adjustment saves ~$450K in replacement costs (10 fewer exits x $45K/hire)

Pay Equity Analysis

import pandas as pd
import statsmodels.api as sm

def analyze_pay_equity(employee_data: pd.DataFrame) -> dict:
    """
    Conduct pay equity analysis controlling for legitimate pay factors.

    Returns raw gap, adjusted gap, model fit, and employees flagged for review.
    """
    # Raw gap
    avg_by_gender = employee_data.groupby('gender')['salary'].mean()
    raw_gap = (avg_by_gender['Female'] - avg_by_gender['Male']) / avg_by_gender['Male']

    # Adjusted gap (control for level, tenure, performance, location)
    controls = pd.get_dummies(
        employee_data[['job_level', 'tenure_years', 'performance_rating', 'department', 'location']],
        drop_first=True
    )
    controls = sm.add_constant(controls)
    controls['is_female'] = (employee_data['gender'] == 'Female').astype(int)

    model = sm.OLS(employee_data['salary'], controls).fit()
    adjusted_gap = model.params['is_female']

    # Flag outliers (residual > 2 std dev)
    employee_data['predicted'] = model.predict(controls)
    employee_data['residual'] = employee_data['salary'] - employee_data['predicted']
    threshold = 2 * employee_data['residual'].std()
    flagged = employee_data[abs(employee_data['residual']) > threshold]

    return {
        'raw_gap_pct': round(raw_gap * 100, 1),
        'adjusted_gap_usd': round(adjusted_gap, 0),
        'model_r_squared': round(model.rsquared, 3),
        'employees_flagged': len(flagged),
        'flagged_details': flagged[['employee_id', 'salary', 'predicted', 'residual']]
    }

Engagement Survey Analysis

1. Calculate response rate -- Target 80%+ for statistical validity. Flag departments below 60%.
2. Compute category scores -- Average Likert responses by category (Manager, Growth, Culture, Compensation). Compare to prior period.
3. Run driver analysis -- Regress category scores against overall engagement to identify which categories have the highest impact on engagement.
4. Segment -- Break results by department, level, tenure band, and location. Identify where scores diverge most from company average.
5. Prioritize -- Plot categories on a 2x2 matrix (Impact vs Score). "High impact, low score" quadrant = priority action areas.

Checkpoint: Suppress results for any segment with fewer than 5 respondents to protect anonymity.

DEI Metrics Framework

Data Governance Checklist

Before starting any people analytics project:

• Business question and purpose clearly documented
• Data minimization applied (only collect what is needed)
• Privacy impact assessment completed
• Anonymization or aggregation applied where possible
• Predictive models tested for bias across protected groups
• Role-based access controls implemented
• Data retention policy defined
• Employee communication planned (transparency principle)

Reference Materials

• references/hr_metrics.md - Complete HR metrics guide
• references/predictive_models.md - Predictive modeling approaches
• references/survey_design.md - Survey methodology
• references/data_ethics.md - Ethical analytics practices

Scripts

# Analyze engagement survey results with driver analysis
python scripts/survey_analyzer.py --file survey_results.csv
python scripts/survey_analyzer.py --file survey_results.csv --prior prior_survey.csv --json

# Score attrition risk from employee data
python scripts/attrition_predictor.py --file employees.csv
python scripts/attrition_predictor.py --file employees.csv --threshold 0.7 --json

# Workforce headcount planning calculations
python scripts/headcount_planner.py --file workforce.csv --growth 0.15 --attrition 0.12
python scripts/headcount_planner.py --file workforce.csv --growth 0.15 --attrition 0.12 --json

Troubleshooting

Success Criteria

Scope & Limitations

In Scope:

• Workforce descriptive analytics: headcount, turnover, retention, demographics, tenure distribution
• Engagement survey design, analysis, driver identification, and benchmarking
• Attrition risk scoring using rule-based and statistical methods (standard library only)
• Pay equity analysis: raw gap, controlled gap, outlier flagging
• DEI metrics: representation, progression rates, hiring funnel equity
• Workforce planning: headcount forecasting, scenario modeling, gap analysis
• Dashboard design and KPI framework recommendations

Out of Scope:

• Real-time predictive models requiring ML frameworks (scikit-learn, TensorFlow) -- scripts use rule-based scoring for portability
• Sentiment analysis of free-text survey responses (requires NLP libraries)
• Individual employee profiling or surveillance -- all analysis uses aggregated or anonymized data
• HRIS system administration, data pipeline engineering, or ETL development
• Legal interpretation of pay equity findings (requires Employment Law counsel)
• Organizational network analysis requiring email/calendar metadata

Known Limitations:

• Attrition risk scoring in scripts uses weighted heuristics, not trained ML models; accuracy depends on feature quality and weight calibration
• Pay equity analysis in the SKILL.md examples requires statsmodels (external dependency); scripts use standard-library approximations
• Survey analysis assumes Likert scale (1-5) responses; other formats require preprocessing
• Small population segments (< 30) produce unreliable statistical results; flag these in reporting
• Historical data biases (e.g., biased performance ratings) propagate into predictive models if not addressed

Integration Points