Rotemberg Weight Package

This program estimates the Rotemberg weights outlined in Goldsmith-Pinkham, Sorkin and Swift (2018). Each weight returned corresponds to the misspecification elasticity for each individual instrument when using the Bartik instrument defined by the weights. The discussion below pertains to the Stata implementation -- see the R-code subdirectory for an R implementation.

Installation

To install the Stata package, clone or download this repo, and then copy bartik_weight.ado and bartik_weight.sthlp to your personal ado folder. You can find this folder using the sysdir command.

Example

In the code folder, I provide two examples of datasets whose datasets are prepped to run through the bartik_weight function: example_ADH.do and example_BAR.do

Example 1: Canonical Bartik - example_BAR.do

There are N commuting zones with T time periods in this example, with K industries. Hence, following Goldsmith-Pinkham, Sorkin and Swift (2018), there are NT observations and KT instruments, where the industries are defined in the initial period (1980), interacted with a time fixed effect for 1980, 1990, and 2000. For this implementation, the controls are also taken in the inital period, and interacted with time fixed effects. Since we control for a commuting zone fixed effect, we exclude one time period of the controls, since it is not seperately identified.

For each of the KT instruments, bartik_weight.ado constructs a Rotemberg weight. This weight corresponds to the misspecification sensitivity for that industry-period pair.

Example 2: China Shock - example_ADH.do

There are N commuting zones with T time periods in this example, with K industries. Hence, following Goldsmith-Pinkham, Sorkin and Swift (2018), there are NT observations and KT instruments, interacted with a time fixed effect for 1990, and 2000. In this example, industry shares are defined using the lagged (previous decade) industry share for a commuting zone.

For each of the KT instruments, bartik_weight.ado constructs a Rotemberg weight. This weight corresponds to the misspecification sensisitivty for that industry-period pair.

How to construct a dataset from the bartik_weight output

The final output from the Stata packages is returned in three Stata matrices. I provide examples on how to convert these outputs into readable and labelled data in both examples above, and repeat the code here:

bartik_weight, z(t*_`ind_stub'*)    weightstub(`growth_stub'*) x(`x') \\\
	y(`y')  controls(`controls') weight_var(`weight') by(year)
mat beta = r(beta)
mat alpha = r(alpha)
mat G = r(G)
qui desc t*_`ind_stub'*, varlist
local varlist = r(varlist)

clear
svmat beta
svmat alpha
svmat G

gen ind = ""
gen year = ""
local t = 1
foreach var in `varlist' {
	if regexm("`var'", "t(.*)_`ind_stub'(.*)") {
		qui replace year = regexs(1) if _n == `t'
		qui replace ind = regexs(2) if _n == `t'
		}
	local t = `t' + 1
	}

Author

• Paul Goldsmith-Pinkham -- Contact me at @paulgp or paulgp [at] gmail [dot] com

License

This project is licensed under the MIT License - see the LICENSE.md file for details