Modeling Strategy Guide

Decision Framework:
1. Start with the simplest model that could answer your question
2. Add complexity only when diagnostics reveal inadequacy
3. Prefer interpretable models unless prediction accuracy is the sole goal
4. Always have a baseline (mean, majority class, last observation)

Model Complexity Ladder:
  Level 1: Descriptive statistics, cross-tabulations
  Level 2: Linear/logistic regression
  Level 3: Regularized regression (Lasso, Ridge, Elastic Net)
  Level 4: Tree ensembles (Random Forest, Gradient Boosting)
  Level 5: Deep learning (only with sufficient data and clear justification)

import pandas as pd
import numpy as np

def engineer_features(df: pd.DataFrame, config: dict) -> pd.DataFrame:
    """
    Apply systematic feature engineering based on domain knowledge.

    config example:
    {
        'log_transform': ['income', 'citations'],
        'interactions': [('experience', 'education')],
        'polynomial': {'age': 2},
        'time_features': 'date_column',
        'lag_features': {'metric': [1, 7, 30]}
    }
    """
    df = df.copy()

    # Log transforms for right-skewed variables
    for col in config.get('log_transform', []):
        df[f'{col}_log'] = np.log1p(df[col])

    # Interaction terms
    for col_a, col_b in config.get('interactions', []):
        df[f'{col_a}_x_{col_b}'] = df[col_a] * df[col_b]

    # Polynomial features
    for col, degree in config.get('polynomial', {}).items():
        for d in range(2, degree + 1):
            df[f'{col}_pow{d}'] = df[col] ** d

    # Time-based features
    if 'time_features' in config:
        time_col = config['time_features']
        df[time_col] = pd.to_datetime(df[time_col])
        df[f'{time_col}_month'] = df[time_col].dt.month
        df[f'{time_col}_dayofweek'] = df[time_col].dt.dayofweek
        df[f'{time_col}_quarter'] = df[time_col].dt.quarter

    return df

from sklearn.linear_model import LogisticRegression
from sklearn.neighbors import NearestNeighbors

def propensity_score_match(df, treatment_col, covariates, caliper=0.05):
    """
    Match treated and control units based on propensity scores.
    """
    # Estimate propensity scores
    X = df[covariates].values
    y = df[treatment_col].values

    lr = LogisticRegression(max_iter=1000, random_state=42)
    lr.fit(X, y)
    df['pscore'] = lr.predict_proba(X)[:, 1]

    # Match using nearest neighbor within caliper
    treated = df[df[treatment_col] == 1]
    control = df[df[treatment_col] == 0]

    nn = NearestNeighbors(n_neighbors=1, metric='euclidean')
    nn.fit(control[['pscore']].values)

    distances, indices = nn.kneighbors(treated[['pscore']].values)

    # Apply caliper
    valid = distances.flatten() < caliper
    matched_treated = treated[valid].index.tolist()
    matched_control = control.iloc[indices.flatten()[valid]].index.tolist()

    return {
        'matched_treated': matched_treated,
        'matched_control': matched_control,
        'n_matched': sum(valid),
        'n_unmatched': sum(~valid),
        'balance_check': 'Run standardized mean differences on covariates'
    }

from scipy import stats
import numpy as np

def design_experiment(baseline_rate, mde, alpha=0.05, power=0.80):
    """
    Calculate required sample size for a two-proportion z-test.

    Args:
        baseline_rate: Current conversion/success rate
        mde: Minimum detectable effect (absolute change)
        alpha: Significance level
        power: Statistical power
    """
    from statsmodels.stats.power import NormalIndPower
    effect_size = mde / np.sqrt(baseline_rate * (1 - baseline_rate))
    analysis = NormalIndPower()
    n = analysis.solve_power(
        effect_size=effect_size, alpha=alpha, power=power, ratio=1.0
    )
    return {
        'sample_size_per_group': int(np.ceil(n)),
        'total_sample_size': int(np.ceil(n)) * 2,
        'baseline_rate': baseline_rate,
        'minimum_detectable_effect': mde,
        'alpha': alpha,
        'power': power
    }