Dear All,
I work with a STATA 14 on the estimation of the ATET. I have population data (N=900,000) and for the sake of transparency, I want to keep the large sample. I get estimates with psmatch2 for PSM and kelner matching (after a day), however I can never get estimates for NN (command teffects nn match). The strange thing is that if I make the sample smaller, the estimates work.
Basically, I have the following questions:
1) Is there any way how to use different command to estimate teffects nn match?
2) Do you have any idea, how to estimate this effect with teffects nn match (I have been working in this issue really a lot of days and I have no clue, how to overcome this).
Many thanks for your replies in advance.
Dear Ondrej, maybe this will help: https://www.statalist.org/forums/forum/general-stata-discussion/general/1338002-teffects-psmatch-stata-stops-running-when-using-large-samples 😊
I figured out there is basically no relevant answer so I posted my question there again and I have also contacted STATA support. Lets see.
I hope you will get your answer soon from STATA support team. I am also very interested in figuring it out 😃.
I got a reply from STATA team:
Dear Ondra,
There are some differences in the results produced by the two commands:
- One of the differences is related to the fact that the results from -psmatch2- are affected by the sort order. In fact, the output for -psmatch2- may include a warning message stating that the sort order of the data could affect the results.
The results from -teffects- are not affected by the sort order and this is one of the key for the differences in processing time in our evaluation.
- Another difference is that -teffects psmatch- produces a robust variance, which is not the case for -psmatch2-.
Let me comment that according to our records, the serial number you provided corresponds to Stata IC. However, I believe our records also indicate that your University owns Stata 15 MP license. The code for -teffects is parallelized, so by using Stata MP you could significantly reduce the processing time. Here is an example prepared by a colleague where you can see the difference in processing time between one processor (which would be the case for IC) and two processors:
clear set seed 1234 set obs 20000 gen double x1 = rnormal() gen double x2 = rnormal() gen double x3 = rnormal() gen double et = rnormal() gen double treat = (0.75*x1 + 0.75*x2 + et) > 0 gen double e0 = rnormal() gen double e1 = rnormal() gen double y0 = 3 + 0.8*x1 + 1.2*x3 + e0 gen double y1 = -1 + 0.5*x1 + 0.5*x3 + e1 gen double y = treat*y1 + (1-treat)*y0
* Running -teffects nnmatch- with 1 and 2 processors: timer clear
set processors 1 display c(processors) timer on 1 qui teffects nnmatch (y x1 x2 x3) (treat), ate timer off 1
set processors 2 display c(processors) timer on 2 qui teffects nnmatch (y x1 x2 x3) (treat), ate timer off 2
* Display results: timer list display r(t2)/r(t1)
And the output for the last two commands:
. timer list 1: 14.89 / 1 = 14.8940 2: 7.59 / 1 = 7.5860
. display r(t2)/r(t1) .50933262
Thus, you can confirm that working with MP and two processors the processing time is reduced to almost half in the example above.
Sincerely,
So finally, I could not reach any acceptable reply from STATA:
Regarding the difference between -teffects nnmatch- and -psmatch2-: my colleague already pointed out some reasons as to why the results may differ. If you think the differences are due to something else, you may consider contacting the authors of -psmatch2-. Since this is a user-written command, support is provided by its authors.
However, basically, if I run
bootstrap, reps(5): teffects nnmatchh
bootstrap r(att), reps(5): psmatch2
I get roughly similar estimates, so for large samples combining psmatch with mahal neighbor matching can be estimated (somehow).
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size