Dear Statalist community,
I am facing the following problem and would highly appreciate any advice. To check the robustness of my regression results I wanted to run a nearest neighbor matching by firm size and then run a panel regression on the residual subsample to check the effect of my treatment on the DV without the influence of size. Thanks to very detailed coding support (https://www.statalist.org/forums/for...-year-clusters) I was able to conduct the matching and run the respective regressions.
Now, I am facing the following puzzle. The distribution of the treated and untreated observations in my original sample is unbalanced with N(treated) > N(untreated). Furthermore, I decided to conduct a matching with replacement.
In the first step I matched/assigned treated to untreated observations (using some treated observations multiple times). This decision was rather intuitively driven than theoretically justified.
However, if I match/assing untreated to treated observation (in this case using untreated observations multiple times) not only does the sample size increase (since N(treated) > N(untreated)) but I also receive contrasting regression results (closer to the orginal unmatched sample).
Therefore, my question is whether there is an objective decision criterion that tells me in which direction to match (treated -> untreated or untreated -> treated) and whether this decision has any effect on the validity of my regression results?
I highly appreciate any help. Thanks so much for your valuable time.
Best regards
Marc
I am facing the following problem and would highly appreciate any advice. To check the robustness of my regression results I wanted to run a nearest neighbor matching by firm size and then run a panel regression on the residual subsample to check the effect of my treatment on the DV without the influence of size. Thanks to very detailed coding support (https://www.statalist.org/forums/for...-year-clusters) I was able to conduct the matching and run the respective regressions.
Now, I am facing the following puzzle. The distribution of the treated and untreated observations in my original sample is unbalanced with N(treated) > N(untreated). Furthermore, I decided to conduct a matching with replacement.
In the first step I matched/assigned treated to untreated observations (using some treated observations multiple times). This decision was rather intuitively driven than theoretically justified.
However, if I match/assing untreated to treated observation (in this case using untreated observations multiple times) not only does the sample size increase (since N(treated) > N(untreated)) but I also receive contrasting regression results (closer to the orginal unmatched sample).
Therefore, my question is whether there is an objective decision criterion that tells me in which direction to match (treated -> untreated or untreated -> treated) and whether this decision has any effect on the validity of my regression results?
I highly appreciate any help. Thanks so much for your valuable time.
Best regards
Marc
