Read on Twitter
Happy to share that Stata did_multiplegt package now has an #RStats friend, thanks to the amazing work of @ShuoZhang3648 ! Available from Shuo’s github page: https://github.com/shuo-zhang-ucsb/did_multiplegt">https://github.com/shuo-zhan... and from CRAN repository. Same syntax as Stata version, but can be substantially faster. 1/n
Syntax: did_multiplegt(df, Y, G, T, D)
df = dataframe, Y=outcome, G=group (e.g. county), T=time, D= treatment. Computes DIDM estimator proposed in http://shorturl.at/lBC67 ">https://shorturl.at/lBC67&quo... 2/n
df = dataframe, Y=outcome, G=group (e.g. county), T=time, D= treatment. Computes DIDM estimator proposed in http://shorturl.at/lBC67 ">https://shorturl.at/lBC67&quo... 2/n
For all pair of periods t-1 & t, computes DID comparing Y evolution among switchers, whose treatment changes from t-1 to t, & groups whose treatment doesn’t change. DIDM = average of DID across all pairs of periods. Under // trends, consistent for switchers treatment effect. 3/n
DIDM robust to heterogeneous treatment effects, contrary to two-way FE estimator (to diagnose two-way FE: twowayfeweights Stata command, no #RStats version yet) 4/n
placebo: the number of placebo estimators to be estimated. Placebos compare switchers’ and non-switchers’ outcome evolution, before switchers switch. Under parallel trends, placebos should be 0, so placebos significantly different from 0 mean violation of parallel trends. 5/n
dynamic: number of dynamic effects to be estimated. DIDM estimates switchers’ treatment effect at the time they switch. dynamic tells you how effect evolves after. Can only be used in staggered designs with binary treatment, where groups can switch in but not out treatment. 6/n
controls: the control variables to be included in the estimation. 7/n
trends_nonparam: includes time fixed effects interacted with varlist in estimation. varlist can include one categorical variable. For instance, if one works with a county*year data set and one wants to allow for state-specific trends, write trends_nonparam= "state". 8/n
trends_lin: includes linear trends interacted with varlist in estimation. varlist can include one categorical variable. For instance, if one works with village*year data set and wants to allow for village-specific linear trends, one should write trends_lin= "village". 9/n
weight: name of variable to weight the data. For instance, if one works with a district*year data set and one wants to weight the estimation by each district*year& #39;s population, one should write weight = "population ". 10/n
breps: the number of bootstrap replications to be used in computation of estimators& #39; standard errors. cluster = " varname" computes the standard errors of the estimators using a block bootstrap at varname level. 11/n
covariances: if true, command computes covariances between instantaneous and dynamic effects requested, and between placebos requested. Useful to assess if average of instantaneous and dynamic effects is statistically significant, or test if placebos join significant. 12/n
recat_treatment and threshold_stable_treatment are two options that can be useful when treatment takes a large number of values. E.g.: you want to estimate the effect of rainfall on crops’ yield, your treatment, rainfall, is continuous. See help file for details. 13/n
Command saves the results as dataframe if specified. For instance, A<- did_multiplegt(), estimates saved as datafram A, and A$effect returns effects. Command also produces graph with all the estimators requested and their 95% confidence interval when breps option specified. 14/n
Command can be substantially faster than Stata one, with large data sets. Up to 5 times faster in some of the data sets we& #39;ve looked at. 15/n
Example: estimating effect of union membership on wages, using wooldridge data.
library("wooldridge")
Y = "lwage"
G = "nr"
T = "year"
D = "union"
controls = c("hours")
did_multiplegt(wagepan, Y, G, T, D, controls, placebo = placebo, dynamic = dynamic, brep = 5)
16/n
library("wooldridge")
Y = "lwage"
G = "nr"
T = "year"
D = "union"
controls = c("hours")
did_multiplegt(wagepan, Y, G, T, D, controls, placebo = placebo, dynamic = dynamic, brep = 5)
16/n
Low estimated premium, and potential upward bias: workers who become unionized experience higher wage growth before! 17/n
@ShuoZhang3648 and I hope you find the command useful. Let us know if any question/bug arises, and if you think of other useful options. And please share the results you obtain with the command! End of thread.
