Probabilistic Narrative Identification: Evidence from Tax Policy
This paper generalizes narrative identification by interpreting documentary classifications as probabilistic information about latent exogeneity. Traditional narrative approaches treat selected policy actions as fully exogenous, implicitly imposing a zero-contamination restriction on a chosen subset of observations. Instead, I treat exogeneity as a latent binary attribute and use narrative evidence to construct a probability that a policy action belongs to the exogenous class. Identification proceeds at the level of reduced-form moments: variation in narrative support disciplines outcome–policy comovement, allowing the orthogonal-class moment to be recovered. I construct this probability with a document-grounded language-model procedure that retrieves policy-specific passages, extracts stated motivations, and maps those motivations into exogeneity probabilities.
Empirically, I first validate the measure against the benchmark classifications of Romer and Romer (2010) and Cloyne (2013). I then construct a new dataset of routine U.S. federal tax actions. The resulting aggregate tax responses are contractionary, closely aligned with the canonical narrative tax literature, and operate mainly through investment. Finally, I construct a new narrative dataset of French VAT changes from 1960 to 2024 and combine it with UK VAT changes from Cloyne’s dataset. The pooled UK–France evidence provides the indirect-tax counterpart: VAT increases contract output mainly through private consumption and generate a short-lived rise in inflation, consistent with tax pass-through.