Hey sorry to bother again.
I'm trying to implement leads and lags in my basic diff in diff. This is basically the normal diff-in-diff augemented so that we include leads and lags of the treatment. This is the same model in my post #28, although I believe I made a silly error coding it. The new code is:
My official treatment year is 2008.
gen treatment=year>=2008
gen treatment1=year >= 2007
gen treatment2=year >= 2006
gen treatment3=year >= 2005
gen treatment4=year >= 2004
gen posttreatment1=year >= 2009
gen posttreatment2=year >= 2010
gen posttreatment3=year >= 2011
gen posttreatment4=year >= 2012
gen posttreatment5=year >= 2013
gen posttreatment6=year >= 2014
USA receives the treatment
gen lead1=treatment1*USA
gen lead2=treatment2*USA
gen lead3=treatment3*USA
gen lead4=treatment4*USA
gen lag1=posttreatment1*USA
gen lag2=posttreatment1*USA
gen lag3=posttreatment1*USA
gen lag4=posttreatment1*USA
gen lag5=posttreatment1*USA
gen lag6=posttreatment1*USA
Then throw all these in my regression with a time trend and time dummies. I think the point here is testing whether there are any anticipatory effects (i.e. leads are insignificant), which I need to test quite thoroughly in my research design. I understand that the parallel trend assumption is really just an argument and it is up to the modeler to justify the use of diff-in-diff. For example, my graphs don't necessarily show a common pre-treatment trend, but as Clyde says if I have controlled for time dummies/trends etc, I can argue that I have controlled for these trends, and hence even if my pre-treatment trends don't seem parallel, I can sort of say well my model capture so many dynamics you can't really get from the graph. I think basically my question is: In a model with time trends and time dummies and covariates, how important is visual inspection of the graph? That is, how important is it to show the classic post-treatment divergence with a common pre-treatment trend. thank you again
I'm trying to implement leads and lags in my basic diff in diff. This is basically the normal diff-in-diff augemented so that we include leads and lags of the treatment. This is the same model in my post #28, although I believe I made a silly error coding it. The new code is:
My official treatment year is 2008.
gen treatment=year>=2008
gen treatment1=year >= 2007
gen treatment2=year >= 2006
gen treatment3=year >= 2005
gen treatment4=year >= 2004
gen posttreatment1=year >= 2009
gen posttreatment2=year >= 2010
gen posttreatment3=year >= 2011
gen posttreatment4=year >= 2012
gen posttreatment5=year >= 2013
gen posttreatment6=year >= 2014
USA receives the treatment
gen lead1=treatment1*USA
gen lead2=treatment2*USA
gen lead3=treatment3*USA
gen lead4=treatment4*USA
gen lag1=posttreatment1*USA
gen lag2=posttreatment1*USA
gen lag3=posttreatment1*USA
gen lag4=posttreatment1*USA
gen lag5=posttreatment1*USA
gen lag6=posttreatment1*USA
Then throw all these in my regression with a time trend and time dummies. I think the point here is testing whether there are any anticipatory effects (i.e. leads are insignificant), which I need to test quite thoroughly in my research design. I understand that the parallel trend assumption is really just an argument and it is up to the modeler to justify the use of diff-in-diff. For example, my graphs don't necessarily show a common pre-treatment trend, but as Clyde says if I have controlled for time dummies/trends etc, I can argue that I have controlled for these trends, and hence even if my pre-treatment trends don't seem parallel, I can sort of say well my model capture so many dynamics you can't really get from the graph. I think basically my question is: In a model with time trends and time dummies and covariates, how important is visual inspection of the graph? That is, how important is it to show the classic post-treatment divergence with a common pre-treatment trend. thank you again
Comment