Marie-Luise sent me a private email with a question about the user-written command -psmatch2-. I responded to her and urged her to post questions such as hers on Statalist in the future so that others may benefit from her question. I have posted her question and my response below.
** You can use -psmatch2- (user-written command) to get propensity score weights or matches for use in other commands, but you do need to specify an outcome variable. This allows -psmatch2- to identify the subset of observations with non-missing values to use in the creation of a propensity score matched or weighted sample. If you do not want to see the ATT that -psmatch2- calculates, you can specify that -psmatch2- runs quietly (i.e. type “qui psmatch2 … “).
** If you calculate your propensity score in a separate step from your treatment effect estimate, you need to account for uncertainty in the propensity score estimation in your treatment effect standard errors. For samples that are weighted by a propensity score, you can calculate bootstrapped standard errors. For samples that are matched by a propensity score, you need to use the Abadie-Imbens (AI) method, which is incorporated into Stata’s -teffects- command.
(See Abadie & Imbens 2012. “Matching on the Estimated Propensity Score.” National Bureau of Economic Research Working Paper No. 15301. Available at http://www.nber.org/papers/w15301)
Without knowing more about your data, I can’t comment on the use of cluster-robust SE in your later analyses.
** I have not used –cgmreg- before (user-written command from http://www.econ.ucdavis.edu/faculty/...er/statafiles/), but it appears that it is meant to deal with clustered data. Perhaps someone else on Statalist can comment on this approach.
Hope this helps!
Melissa
- “I am conducting a PSM with psmatch2 in order to enhance the comparability of my sample before running a regression of interest i.e., I am trying to do doubly-robust estimation of an ATT. For that purpose, I use panel data on all Italian municipalities for the years 2005-2010 (2005-2006 is the pre-treatment period). I am not interested in knowing the ATT after the matching (although I may consider it as an additional hint anyway). Is is right that if I do not define the option outcome(varlist) using psmatch2, I just get the same information in terms of output with the only difference (from my output I guess the answer is positive) of not having the ATT-table below the regression?”
** You can use -psmatch2- (user-written command) to get propensity score weights or matches for use in other commands, but you do need to specify an outcome variable. This allows -psmatch2- to identify the subset of observations with non-missing values to use in the creation of a propensity score matched or weighted sample. If you do not want to see the ATT that -psmatch2- calculates, you can specify that -psmatch2- runs quietly (i.e. type “qui psmatch2 … “).
- “What SE specification shall I use and on what depends this choice? In my later regression I use two-way cluster-robust SE and I am confident of doing the right thing there.”
** If you calculate your propensity score in a separate step from your treatment effect estimate, you need to account for uncertainty in the propensity score estimation in your treatment effect standard errors. For samples that are weighted by a propensity score, you can calculate bootstrapped standard errors. For samples that are matched by a propensity score, you need to use the Abadie-Imbens (AI) method, which is incorporated into Stata’s -teffects- command.
(See Abadie & Imbens 2012. “Matching on the Estimated Propensity Score.” National Bureau of Economic Research Working Paper No. 15301. Available at http://www.nber.org/papers/w15301)
Without knowing more about your data, I can’t comment on the use of cluster-robust SE in your later analyses.
- “Thirdly, after matching I would like to use the psmatch2-provided variables (above all _weights) for my regression which I conduct by cgmreg-command. Is there anything that argues against this approach?”
** I have not used –cgmreg- before (user-written command from http://www.econ.ucdavis.edu/faculty/...er/statafiles/), but it appears that it is meant to deal with clustered data. Perhaps someone else on Statalist can comment on this approach.
Hope this helps!
Melissa
