Difference-in-Differences causal analysis to identify demographic drivers of behavioral changes with p-value significance testing. Use for event effects, A/B testing, or policy evaluation.
DiD framework for causal inference with automatic method selection (Multivariate vs Univariate) based on sample size.
Both methods use: Y = α + βX + γPost + δ(X × Post) + ε where δ is the DiD estimate.
Multivariate Heterogeneous DiD
n ≥ k × min_sample_ratio (default: 10k samples minimum)Univariate DiD
CRITICAL: Both use statsmodels regression to provide p-values for statistical inference.
CORRECT approach - Only aggregate actual transactions:
# Group actual transactions by entity-period
intensive_df = transactions.groupby(['entity_id', 'period'])['value'].sum().reset_index()
# Result: Only rows where events occurred (e.g., 3,404 rows for active participants)
Why this matters: Intensive margin measures "how much MORE did participants engage", not "did non-participants start engaging". Including zeros for non-participants dilutes the effect and changes the research question.
CORRECT approach - Create complete entity × group × period grid:
# Create all possible combinations
from itertools import product
all_entities = entities_df['entity_id'].unique()
all_groups = target_groups # e.g., categories, regions, products
all_periods = ['baseline', 'treatment']
complete_grid = pd.DataFrame(
list(product(all_entities, all_groups, all_periods)),
columns=['entity_id', 'period']
)
# Mark who participated (1) vs who didn't (0)
events_lookup = transactions[['entity_id', 'period']].drop_duplicates()
events_lookup['participated'] = 1
extensive_df = complete_grid.merge(events_lookup, how='left')
extensive_df['participated'] = extensive_df['participated'].fillna(0).astype(int)
# Result: All entities × groups × periods with binary outcome
Why this matters: Extensive margin measures "did the participation rate increase", which requires knowing both who participated AND who didn't participate.
DIDAnalyzer(min_sample_ratio=10, significance_level=0.05, min_group_size=1)min_sample_ratio=10: Samples-to-features ratio for multivariate DiDsignificance_level=0.05: P-value threshold 0.05-0.1 based on the context and requirementmin_group_size=1: Min observations per group for univariate DiDintensive_margin(df, features, value_col, top_n=None, sort_by='estimate', asc=False)Analyzes intensive margin: how much behavior changed among participants.
Input Data: Sparse entity-period data (only participants)
df: Sparse DataFrame - only entities with participationfeatures: List of demographic column namesvalue_col: Continuous outcome column (e.g., 'amount', 'spending')top_n: Top N features (None = all)sort_by: 'estimate' or 'p_value'asc: True for ascending, False for descendingReturns: List of {"feature": str, "did_estimate": float, "p_value": float, "method": str}
extensive_margin(df, features, value_col='purchased', top_n=None, sort_by='estimate', asc=False)Analyzes extensive margin: how many people changed behavior.
Input Data: Complete entity-period panel (all entities including non-participants)
df: Complete DataFrame - ALL entities × ALL periods (includes non-participants)features: List of demographic column namesvalue_col: Binary outcome column (0/1), default 'purchased'top_n: Top N features (None = all)sort_by: 'estimate' or 'p_value'asc: True for ascending, False for descendingReturns: List of {"feature": str, "did_estimate": float, "p_value": float, "method": str}