Network General Equilibrium as a Threat to Identification: The Failure of Time Fixed Effects

Time fixed effects are not a general-equilibrium control. In network economies, aggregate shocks do not enter firm outcomes uniformly: they propagate through equilibrium prices, costs, and demand, leaving a residual network gradient after time demeaning. When an instrument, treatment, or exposure design even partially aligns with this gradient, standard panel estimators recover the direct effect plus an endogenous network-leakage component. We develop a tractable framework that signs and quantifies this Network-SUTVA bias in closed form, showing how network topology governs its magnitude. Weak instruments magnify the leakage. Difference-in-differences designs are similarly contaminated because controls can be untreated by assignment but treated in equilibrium. Calibrated to the U.S. production network, the distortion can exceed 130% of the true direct effect. We propose observed- and hidden-network diagnostics and remedies, and test them on canonical difference-in-differences designs, showing that standard estimators can confound direct policy responses with systemic network leakage.