Network Causal Inference Laboratory: SUTVA Breakdown & Spillover Estimator

Treated ($D_i = 1$)

Control ($D_i = 0$)

Spillover Dosage Exposure ($G_i > 0$)

Potential Outcomes Decomposition SUTVA Violation Active

True Direct Effect (DE)
2.50
Direct treatment impact $\beta$

Indirect Effect (IE)
0.90
Peer network spillover $\gamma \cdot \bar{G}$

True Total Effect (TE)
3.40
Combined DE + IE

Naive SUTVA ATE

1.42

Standard A/B Test Estimate

Spillover Bias ($\text{Naive} - \text{True DE}$): -1.08

Ignoring peer interference underestimates true treatment efficacy because Control outcomes are inflated by spillover!

Causal Model Parameters

Direct Effect Coefficient ($\beta$) 2.50

Peer Spillover Effect ($\gamma$) 1.80

Unit Potential Outcomes $Y_i(D_i, G_i)$

Node	Treat ($D_i$)	Dosage ($G_i$)	Outcome ($Y_i$)