In the Event study, year 5 is being driven entirely by the 2005 group because that is the only group who has 5 years of data. But the coefficient in year 0 is some weighted average of all the treated groups (2005, 2006, 2007, 2008, and 2009) ?
Second, is it important to have at least a single "true" post period where nobody is being treated. Or could this regression have been run with the panel ending in 2009.
Third, what would it look like if the authors just dropped the years contributing to -8, -7, and -6. I believe this would be years 1997,1998, and 1999 in the panel.
Just trying to get a handle on these things, thanks
Hi! Thank you for this very informative post! One question: the ATT of all groups is 10.8 and significant at 5% - then why are the CI's in the eventstudy plot overlapping the zero? Because the CI is 95% there as well, right?
Hey Scott! Love the post. Helped me out quite a bit with a DD paper I have been working on. Had a question for you regarding some of the origins of the estimator and how instrumental variables might relate.
If I understand this approach correctly (big if), we are reweighting untreated units based on endogenous observables to more closely resemble treated units in the pre-treatment period. You call Abadie (2005) the grandfather of CS and I totally agree. Abadie (2003), "Semiparametric instrumental variable estimation of treatment response models," seems to be up the same alley.
I want to show the strength of my estimation of the effect of an endogenous treatment. The treatment is binary. I have what I think is a solid IV. I don't want to do 2SLS because of TWFE--basically, all the reasons I went with CS in the first place as my estimator. What is your interpretation of Abadie (2003)? My reading of it is that using the IV as the independent variable by which selection into treatment is based on (and thus, untreated units are reweighted by). Grad school econometrics beat 2SLS into my head but this seems like it would accomplish the same task in the CS DD format.
Thank you. Are there similar solutions for reversible treatments? For example, if we want understand how the number of protests in a state-year affects the number of police officers hired that year?
Thank you very much for this presentation. I was wondering how to apply a weight to this code. I have a matching weight since my treatment is endogenous and would like to use it in my analysis. Also if I want to account for fixed effects, do I have to put them in the "xformula" option? Thanks again.
Thanks for this very clear explanation. As far as I know, CS estimator has to assume "parallel trend" assumption, so how to apply CS estimator when the parallel trend assumption does not hold anymore? Roth has a paper and code to provide sensitivity analysis on relaxing the parallel trend assumption, but how to reconcile these two issues?
Thanks for this post, I've been trying to install did package in R, but I couldn't How can use did package? Is possible to use CS estimator on all type of impact evaluation analysis? i.e environmental policies topics or environmental data bases. All the best
All the applications I have seen using these new DID methods employ data collected annually. However, I am working with cross-sectional survey data collected roughly every 3 ~ 4 years over a 15 year interval. Do you know of any work using these DID methods that uses similarly spaced data? Will the gap in timing of data collection bias the estimates?
Thanks so much for writing this up! Is there any work/guidance on what (if any) of this recent TWFE work applies to settings in which treatments are *reversible*? I.e. settings in which people enter into a treated state that they then exit again. I assume the variance weighting structure still applies, but I'm not sure what else would be at work
Thanks for this. I attempted to run your code and am getting an error of:
Error in lapply(X = unname(split(e, grp)), FUN = FUN, ...) :
invalid formal argument list for "function"
Do you know why this is the case?
This is an awesome post.
Just want to ask three clarifying questions.
In the Event study, year 5 is being driven entirely by the 2005 group because that is the only group who has 5 years of data. But the coefficient in year 0 is some weighted average of all the treated groups (2005, 2006, 2007, 2008, and 2009) ?
Second, is it important to have at least a single "true" post period where nobody is being treated. Or could this regression have been run with the panel ending in 2009.
Third, what would it look like if the authors just dropped the years contributing to -8, -7, and -6. I believe this would be years 1997,1998, and 1999 in the panel.
Just trying to get a handle on these things, thanks
when I have a similar implementation example or exercise in Stata?
This is amazing and potentially life-saving for my master's thesis. Thank you Scottie
Great post, thank you.
Should I be using this method if I have small groups (say less than 5 observations in each group) and an unbalanced panel?
Thank you
Hi! Thank you for this very informative post! One question: the ATT of all groups is 10.8 and significant at 5% - then why are the CI's in the eventstudy plot overlapping the zero? Because the CI is 95% there as well, right?
All the best,
Helena
Hey Scott! Love the post. Helped me out quite a bit with a DD paper I have been working on. Had a question for you regarding some of the origins of the estimator and how instrumental variables might relate.
If I understand this approach correctly (big if), we are reweighting untreated units based on endogenous observables to more closely resemble treated units in the pre-treatment period. You call Abadie (2005) the grandfather of CS and I totally agree. Abadie (2003), "Semiparametric instrumental variable estimation of treatment response models," seems to be up the same alley.
I want to show the strength of my estimation of the effect of an endogenous treatment. The treatment is binary. I have what I think is a solid IV. I don't want to do 2SLS because of TWFE--basically, all the reasons I went with CS in the first place as my estimator. What is your interpretation of Abadie (2003)? My reading of it is that using the IV as the independent variable by which selection into treatment is based on (and thus, untreated units are reweighted by). Grad school econometrics beat 2SLS into my head but this seems like it would accomplish the same task in the CS DD format.
Any thoughts? Thanks!
Thank you. Are there similar solutions for reversible treatments? For example, if we want understand how the number of protests in a state-year affects the number of police officers hired that year?
Thank you very much for this presentation. I was wondering how to apply a weight to this code. I have a matching weight since my treatment is endogenous and would like to use it in my analysis. Also if I want to account for fixed effects, do I have to put them in the "xformula" option? Thanks again.
Thanks for this very clear explanation. As far as I know, CS estimator has to assume "parallel trend" assumption, so how to apply CS estimator when the parallel trend assumption does not hold anymore? Roth has a paper and code to provide sensitivity analysis on relaxing the parallel trend assumption, but how to reconcile these two issues?
Thanks for this post, I've been trying to install did package in R, but I couldn't How can use did package? Is possible to use CS estimator on all type of impact evaluation analysis? i.e environmental policies topics or environmental data bases. All the best
Very helpful, thank you.
All the applications I have seen using these new DID methods employ data collected annually. However, I am working with cross-sectional survey data collected roughly every 3 ~ 4 years over a 15 year interval. Do you know of any work using these DID methods that uses similarly spaced data? Will the gap in timing of data collection bias the estimates?
Thanks for your time!
Thanks so much for writing this up! Is there any work/guidance on what (if any) of this recent TWFE work applies to settings in which treatments are *reversible*? I.e. settings in which people enter into a treated state that they then exit again. I assume the variance weighting structure still applies, but I'm not sure what else would be at work
this is amazing. Thanks so much for walking us through this!