Probabilistic Narrative Identification: Evidence from Tax Policy
This paper develops a probabilistic approach to narrative identification. Traditional narrative studies retain policy actions classified as exogenous and exclude the remainder from identification. I instead assign each action a probability of exogeneity, allowing the broader policy record to remain informative. Variation in this probability allows policy actions to be reweighted so that the endogenous component cancels, recovering the component associated with exogenous policy. I construct these probabilities through a replicable mapping from policy documents using large language models. Applied to the canonical tax-policy sample of Romer and Romer (2010), the framework yields an estimated medium-run GDP contraction about two-thirds smaller than under binary selection. I then extend the analysis to indirect taxation by constructing two new datasets of U.S. budget and French VAT proposals. VAT illustrates how probabilistic identification makes narrative analysis feasible when binary selection leaves too little credible identifying variation. Together, the applications reveal contrasting transmission patterns: U.S. tax increases lower output and inflation, with similar consumption and investment contributions to the GDP decline, whereas VAT increases initially raise prices and contract output, with consumption contributing more than investment.