Re: st: RE: RE: RE: exact matches in propensity score matching

[Austin Nichols <austinnichols@gmail.com>]

Yang, Yong <yongyang@essex.ac.uk>:
Another point: your -psmatch2- call produces a linear regression where
"sales" is the outcome, but -glm- with a log link would be more
appropriate, assuming sales are nonnegative:
http://www.stata.com/meeting/boston10/boston10_nichols.pdf
Also, bootstrapping does not provide correct inference:
http://www.jstor.org/stable/40056514

On Fri, Jun 22, 2012 at 3:18 PM, Austin Nichols <austinnichols@gmail.com> wrote:
> Yang, Yong <yongyang@essex.ac.uk>:
> I would guess choice of NN, Kernel, LLR, etc. don't really matter in
> your application, relative to your choice of a parametric model for
> the propensity score as a logit on emp, emp2, cap, cap2.  Have you
> considered dividing emp&cap into a grid at (say) quartiles (16
> categories in each industry/year), and reweighting by the inverse of
> the fraction exporting within cell?  This is now sometimes called
> "coarsened exact matching" and has good properties relative to most
> alternatives: http://www.jstor.org/stable/1555493
> See also http://www.stata-journal.com/sjpdf.html?articlenum=st0136_1
>
> On Fri, Jun 22, 2012 at 2:42 PM, Yang, Yong <yongyang@essex.ac.uk> wrote:
>> Dear Daniel and Stata users,
>>
>> Thank you very much for these constructive suggestions.
>>
>> I have just seen a note on http://www.stata.com/statalist/archive/2010-09/msg00073.html . The idea seems like to add different constants to each industry year combination, and then add caliper (i.e. 0.5) when use propensity score matching to force exporter and domestic firm within the same industry and appear in the same year. Do you think it is fine to have a try on this way? if it is fine, how about followings codes to generate PMS estimate on exact matches from same industry and year?
>>
>>
>> egen industry_year=group(industry year)
>> logit exporter employees employees2 capital capital2 industry_* year_*
>> predict pscore if e(sample), pr
>> gen pscore2=industry_year*10+pscore
>> bootstrap r(att): psmatch2 exporter, pscore(pscore2) outcome(sales) neighbor(1) caliper(0.5)
>> bootstrap r(att): psmatch2 exporter, pscore(pscore2) outcome(sales) kernel bw(0.06) caliper(0.5)
>> bootstrap r(att): psmatch2 exporter, pscore(pscore2) outcome(sales) radius caliper(0.5)
>>
>>
>> Also, I am running stata on half million observations, and do you  know is there any way to speed up kernel matching estimation process? it is currently take hours to generate one estimate.
>>
>> Thank you very much for your time and all your help.
>>
>> Regards
>> Yong
>>
>>

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/