Towards Robust Local Projections Job Market Paper
Local projections with proxy/instrumental variables have been increasingly used for causal inference in empirical macroeconomics. In this study, I propose Bayesian inference for this model under a distributed lagged representation of data. In the proposed parametrization, researchers place priors directly over the structural impulse responses and the first-stage parameter. Dispersed priors over the structural impulse responses allow for weak-instrument robust inference in finite samples and serially correlated data, but
this usually leads to poor identification. I show how additional assumptions about the instrument can be incorporated through the first-stage coefficient prior, greatly sharpening inference. To illustrate the method I perform two empirical exercises. First, I estimate US marginal income tax shock’s impulse responses on a 16-variable yearly-frequency local projection. Marginal tax shocks are contractionary but their effect on real activity and consumption lasts only two years. I attribute this quick recovery to substitution
effects. The second exercise is a replication of Gertler and Karadi (2015) monetary policy analysis using two-year Treasury rates as the policy variable. Even in the presence of a weak instrument, monetary policy shocks are unambiguously contractionary and money is non-neutral after four years.
On The Empirics of Optimal Tax policy Under Parameter Uncertainty.
In this paper, I investigate optimal income tax policy when policymakers don’t know the elasticity of taxable income, the key structural parameters which is sufficient to determine optimal tax in a stylized model. In this setting, a question of how parameter uncertainty should be represented in the optimal tax policy arises. The standard practice is to derive optimal taxes as mappings of structural parameters and plug-in point and interval estimates in this mapping. I show the policy maker prior information is not fully incorporated in this case and propose an alternative — using welfare as a loss function when summarizing the posterior distribution of structural parameters. In a simulation exercise, I show the optimal policy interval estimates using the plug-in approach are wider than the ones using my proposed approach without any gain in coverage.
Other Working Projects
The Dynamic Impact of Social Security Spending: Evidence from Brazil.
This paper provides new evidence of the effects of government spending shocks on the Brazilian economy between 1990-2020. Identification of spending shocks is carried out in Bayesian VAR and LP and by using a new instrumental variable: statutory variation in public pension spending around the four social security reforms that have taken place in Brazil during the sampled period. Early results so far show spending shocks are pro-cyclical. The two year spending multiplier is between 0.7 and 1.8, large considering Brazil is an indebted developing economy. Multipliers are zero in the long run.
Building Counterfactuals with Spatio-Temporal Data. (with Marcelino Guerra)
In this paper we propose a new identification strategy to perform causal inference in georeferenced spatial panel data. We utilize predictions drawn from pre-treatment periods as counterfactuals for the post-treatment period. The difference measures the impact of policy treatment without assuming an explicit functional form of how shocks propagate through space. The model is applied to measure the effect of the construction of football fields in poor neighborhoods of Fortaleza, Brazil, on criminal activity.