Any idea how to mimic deals after propensity score matching?
Hi everyone,
So I'm replicating a paper https://drive.google.com/file/d/1HBf...x3kj5MUzK/view
The general idea of my problem is that I need to create fake deals from propensity score matches to have a control group for a Dif in Dif analysis.
I'm using propensity score matching to find the 5 closest matches for acquirers and the 5 closest for targets, for every deal.
After the matching, I end up with 5 matches per acquirer and target, but some controls are the same for multiple acquirers and targets (because I have a limited pool for the controls).
What I need to do is randomly assign one match acquirer to a random match target, sum their variables, and create new deal information to add to the base as the "control sample" to perform the rest of the analysis.
The problem is that I'm not quite sure what command can I use to make this random pair between the 2 matches, also because some of them are the same matches for multiple observations the command has to be able to consider this, so I should be able to replicate the same amount of deals and don't ignore/skip them. Thus, does anyone has any idea on which command or how can I do this??? Please any idea should be great, and if you can add the STATA commands that would be very helpful.
PD: the matching results are not the best yet, so any recommendation on how to make them better? So far, %bias is not bad (min 0.9 to max -8), all the p|t| are over 0.20 (which is ok), but the V(T)/V(C) for all of them have one * (min 0.09 to max 11.60) * if variance ratio outside [0.86; 1.17]. So I understand that one * is displayed for variables that have variance ratios that exceed the 2.5th and 97.5th percentiles of the F-distrib, meaning that are variables of concern. (Pic attached) So should I be concerned and try to make them better? any suggestions? I've read that normalizing the median of the treatment and control could help, or try to balance them for the matching (not sure how to do this?). Any feedback would be much appreciated. Should I only care for % bias or p|t| or V(T)/V(C), or do I need to care for all of them to be good?
For reference: ln TA is ln total assets, and deltas are just the last year's change on those variables (trend).
I appreciate any help, thanks in advance!!!
So I'm replicating a paper https://drive.google.com/file/d/1HBf...x3kj5MUzK/view
The general idea of my problem is that I need to create fake deals from propensity score matches to have a control group for a Dif in Dif analysis.
I'm using propensity score matching to find the 5 closest matches for acquirers and the 5 closest for targets, for every deal.
After the matching, I end up with 5 matches per acquirer and target, but some controls are the same for multiple acquirers and targets (because I have a limited pool for the controls).
What I need to do is randomly assign one match acquirer to a random match target, sum their variables, and create new deal information to add to the base as the "control sample" to perform the rest of the analysis.
The problem is that I'm not quite sure what command can I use to make this random pair between the 2 matches, also because some of them are the same matches for multiple observations the command has to be able to consider this, so I should be able to replicate the same amount of deals and don't ignore/skip them. Thus, does anyone has any idea on which command or how can I do this??? Please any idea should be great, and if you can add the STATA commands that would be very helpful.
PD: the matching results are not the best yet, so any recommendation on how to make them better? So far, %bias is not bad (min 0.9 to max -8), all the p|t| are over 0.20 (which is ok), but the V(T)/V(C) for all of them have one * (min 0.09 to max 11.60) * if variance ratio outside [0.86; 1.17]. So I understand that one * is displayed for variables that have variance ratios that exceed the 2.5th and 97.5th percentiles of the F-distrib, meaning that are variables of concern. (Pic attached) So should I be concerned and try to make them better? any suggestions? I've read that normalizing the median of the treatment and control could help, or try to balance them for the matching (not sure how to do this?). Any feedback would be much appreciated. Should I only care for % bias or p|t| or V(T)/V(C), or do I need to care for all of them to be good?
For reference: ln TA is ln total assets, and deltas are just the last year's change on those variables (trend).
I appreciate any help, thanks in advance!!!
Attached Files
