Dear CEM List members,
I have about 45 treatment firms and I am trying to create a control
sample. Using that control
sample, we plan to collect additional information about the firm, which
would take some time and
effort. I was wondering if I could use different CEM specifications for
each treatment firm to get the
most accurate control firms.
The issue is that for some treatment firms, I get a huge strata, on the
other
hand, for other groups I have two control firms. I was wondering if I could
use a more stringent
matching criteria for those treatment firms with large number of control
firms and a less stringent
matching criteria for those that have a low number of control firms.
My concern is how would I run a regression later because the weights are
derived from different
matching specifications. Or does it matter?
Thank you for your time and consideration.
Kind Regards,
Pureum Kim
This issue is an adequate match. The extent to which they cannot be derived from a single CEM run should be irrelevant if the purpose is to generate the best match. Matching does not solve statistical problems it merely helps strengthen the arguments that the treatment is the difference between the groups.
Sent from my iPad
> On Apr 29, 2015, at 9:14 AM, pureum kim <pureum.kim(a)usc.edu> wrote:
>
> Dear CEM List members,
>
> I have about 45 treatment firms and I am trying to create a control sample. Using that control
>
> sample, we plan to collect additional information about the firm, which would take some time and
>
> effort. I was wondering if I could use different CEM specifications for each treatment firm to get the
>
> most accurate control firms.
>
> The issue is that for some treatment firms, I get a huge strata, on the other
>
> hand, for other groups I have two control firms. I was wondering if I could use a more stringent
>
> matching criteria for those treatment firms with large number of control firms and a less stringent
>
> matching criteria for those that have a low number of control firms.
>
> My concern is how would I run a regression later because the weights are derived from different
>
> matching specifications. Or does it matter?
>
> Thank you for your time and consideration.
>
> Kind Regards,
>
>
> Pureum Kim
> -
> --
> cem Mailing List, served by HUIT
> Send messages: cem(a)lists.gking.harvard.edu
> [un]subscribe Options: http://lists.gking.harvard.edu/?info=cem
> More information on cem: http://gking.harvard.edu/cem
> Cem mailing list
> Cem(a)lists.gking.harvard.edu
>
> To unsubscribe from this list or get other information:
>
> https://lists.gking.harvard.edu/mailman/listinfo/cem
I would however be cautious about why treatment firms are substantially different and attempt to control for any bias that might arise because of the apparent difference in treatment firms but this could simply be a sensitivity analysis indicating the conditions for which results continue to hold.
Sent from my iPad
> On Apr 29, 2015, at 9:14 AM, pureum kim <pureum.kim(a)usc.edu> wrote:
>
> Dear CEM List members,
>
> I have about 45 treatment firms and I am trying to create a control sample. Using that control
>
> sample, we plan to collect additional information about the firm, which would take some time and
>
> effort. I was wondering if I could use different CEM specifications for each treatment firm to get the
>
> most accurate control firms.
>
> The issue is that for some treatment firms, I get a huge strata, on the other
>
> hand, for other groups I have two control firms. I was wondering if I could use a more stringent
>
> matching criteria for those treatment firms with large number of control firms and a less stringent
>
> matching criteria for those that have a low number of control firms.
>
> My concern is how would I run a regression later because the weights are derived from different
>
> matching specifications. Or does it matter?
>
> Thank you for your time and consideration.
>
> Kind Regards,
>
>
> Pureum Kim
> -
> --
> cem Mailing List, served by HUIT
> Send messages: cem(a)lists.gking.harvard.edu
> [un]subscribe Options: http://lists.gking.harvard.edu/?info=cem
> More information on cem: http://gking.harvard.edu/cem
> Cem mailing list
> Cem(a)lists.gking.harvard.edu
>
> To unsubscribe from this list or get other information:
>
> https://lists.gking.harvard.edu/mailman/listinfo/cem
Dear CEM List,
I am attempting to run cem in Stata MP 14:
cem months (#15) if laac_none_base50p!=., treatment(laac_none_base50p) show
I get the following error:
cemStata(): 3301 subscript invalid
<istmt>: - function returned error
r(3301);
My data:
479,808 observations
months ranges from 1-156 and is a real number
laac_none_base50p is either 0 (n=1,140) or 1 (n=472,312)
Please help; I can't find any support for this error.
Andrew
Dear CEM List members,
My research question examines whether firms adopting fair value accounting
for their
pensions change their pension plan assets. I was wondering if you could
help me with the following
quick questions about Coarsened Exact Matching.
1. What is the best practice for selecting the cut points? (I have read
that Scott and Freedman-Diaconnis are the preferred methods.)
2. My treatment sample is very small and using Scott or Freedman-Diaconnis
leads me to lose many treated firms. So I use deciles for certain matching
variables instead of the Scott or FD. Would this still be a concern if the
imbalance is low?
3. Is there a reasonable value for L1? I know that the lower value means
less imbalance but is there a rule of thumb?
4. After creating a matched sample, are there any practical guides on how
to run the regression? Some treated firms are matched with multiple control
firms, and I believe that it is not a problem to run the regression even if
some firms are matched with more control firms. Is there an issue if we
don't have exact number of paired treated and controlled firms?
Thank you so much for your time and consideration.
Kind Regards,
Pureum Kim
While I am not an expert on CEM I will provide some responses.
1. You are attempting to match so consider what provides your best match. there is no magic that provides the specific cutoffs if your goal is matching (which I presume it is).
2. If the imbalance or bias is low then you have a reasonable match.
3. I would hope for L1 less than .5 to make myself feel reasonable about the matching.
4. This is an issue concerning matching with replacement versus one to one matching without replacement. I prefer one to one matching unless I have few controls and there are arguments for both possibilities. I would check out some papers by Elizabeth Stuart that provide discussion of the costs and benefits of one-to-one w/o replacement versus with replacement. In observing studies in my discipline, I find that authors fail to acknowledge or inform the extent to which there are duplicate matches in the sample and by default the control that is used multiple times induces correlation in the control sample that should be acknowledged for at least controlled for but few take those extra steps. Using a control multiple times can induce bias that has an unobservable effect on testing between treatment and control.
On Apr 23, 2015, at 12:53 AM, pureum kim <pureum.kim(a)usc.edu<mailto:pureum.kim@usc.edu>> wrote:
Dear CEM List members,
My research question examines whether firms adopting fair value accounting for their
pensions change their pension plan assets. I was wondering if you could help me with the following
quick questions about Coarsened Exact Matching.
1. What is the best practice for selecting the cut points? (I have read that Scott and Freedman-Diaconnis are the preferred methods.)
2. My treatment sample is very small and using Scott or Freedman-Diaconnis leads me to lose many treated firms. So I use deciles for certain matching variables instead of the Scott or FD. Would this still be a concern if the imbalance is low?
3. Is there a reasonable value for L1? I know that the lower value means less imbalance but is there a rule of thumb?
4. After creating a matched sample, are there any practical guides on how to run the regression? Some treated firms are matched with multiple control firms, and I believe that it is not a problem to run the regression even if some firms are matched with more control firms. Is there an issue if we don't have exact number of paired treated and controlled firms?
Thank you so much for your time and consideration.
Kind Regards,
Pureum Kim
-
--
cem Mailing List, served by HUIT
Send messages: cem(a)lists.gking.harvard.edu<mailto:cem@lists.gking.harvard.edu>
[un]subscribe Options: http://lists.gking.harvard.edu/?info=cem
More information on cem: http://gking.harvard.edu/cem
Cem mailing list
Cem(a)lists.gking.harvard.edu<mailto:Cem@lists.gking.harvard.edu>
To unsubscribe from this list or get other information:
https://lists.gking.harvard.edu/mailman/listinfo/cem