[Limdep Nlogit List] PSM estimation defaults

[Jill Caviglia-Harris]

Bill:
 
Thanks for your response.  Based on this I am a little confused by my
output.  I may be interpreting the off-support incorrectly...   The
following are estimations of the ATT based on the same probit
estimations of the PSM (with the same data) and use the 'nearest
neighbor,' 'range' (caliper) and 'kernel,' in that order, - each
includes a different number of treated observations. The results read as
follows:
 
For the nearest neighbor (we have a total of 24 treated - all
included)
+----------------------------------------------------------------------+
| Number of Treated observations =     24  Number of controls =     15
|
| Estimated Average Treatment Effect   =       4.270833               
|
| Estimated Asymptotic Standard Error  =        .810050               
|
| t statistic (ATT/Est.S.E.)           =       5.272308               
|
| Confidence Interval for ATT = (     2.683135  to       5.858531) 95%
|
| Average Bootstrap estimate of ATT    =       4.003685               
|
| ATT - Average bootstrap estimate     =        .267148               
|
+----------------------------------------------------------------------+
 
For the range (caliper) - again we have a total of 24 treated - 12 are
included
+----------------------------------------------------------------------+
| Number of Treated observations =     12  Number of controls =     95
|
| Estimated Average Treatment Effect   =       1.898789               
|
| Estimated Asymptotic Standard Error  =        .701460               
|
| t statistic (ATT/Est.S.E.)           =       2.706909               
|
| Confidence Interval for ATT = (      .523927  to       3.273651) 95%
|
| Average Bootstrap estimate of ATT    =       1.876323               
|
| ATT - Average bootstrap estimate     =        .022466               
|
+----------------------------------------------------------------------+
 
for the kernel - - we have a total of 24 treated - only 20 are used (on
support) - is that what this means?
+----------------------------------------------------------------------+
| Number of Treated observations =     20  Number of controls =    169
|
| Estimated Average Treatment Effect   =       3.565033               
|
| Estimated Asymptotic Standard Error  =        .900702               
|
| t statistic (ATT/Est.S.E.)           =       3.958062               
|
| Confidence Interval for ATT = (     1.799658  to       5.330408) 95%
|
| Average Bootstrap estimate of ATT    =       3.173718               
|
| ATT - Average bootstrap estimate     =        .391314               
|
+----------------------------------------------------------------------+
If the kernel methods uses all of the treated observations to estimate
the ATT should the above read 24 and not 20? 
 
Thanks.  -Jill
 
 
Jill Caviglia-Harris
Associate Professor
Economics and Finance Department
Salisbury University
Salisbury, MD 21801
(410) 548-5591

>>> William Greene <wgreene at stern.nyu.edu> 9/11/2009 6:35 PM >>>
Jill.
There are three candidates for deciding who is a nearby observation,
'nearest neighbor,' 'kernel,' or 'range' (caliper). But, all of the
treated observations are used. There is no "trimming" built into the
program. You are right about the bandwidth for the kernel approach;
the default is .06.  If you (the user) supply a value between .001
and .250 using ;Smooth=value, then that value is used instead.
Regards,
Bill Greene

----- Original Message -----
From: "Jill Caviglia-Harris" <jlcaviglia-harris at salisbury.edu>
To: limdep at limdep.itls.usyd.edu.au 
Cc: "Erin Sills" <erin_sills at ncsu.edu>
Sent: Friday, September 11, 2009 2:36:52 PM GMT -05:00 Columbia
Subject: [Limdep Nlogit List] PSM estimation defaults


LimDep Users Group:

I recently posted a question about the kernel PSM methodology used in
LimDep but have not received a response, perhaps with further detail
this can be addressed….  My question is about the default setting in
LimDep.  It’s my understanding that there are many different
variations
on kernel, including different ways to set bandwidth and to "trim"
treated observations that are off-support  (Frolich 2004; Calienddo
and
Kopeinig 2005).  Based on recent estimation results it appears that
LimDep does trim the results – in other words does not use every one
of
the treated observations in estimating the ATT.  I assume the default
was no trimming would be done – but this appears to be a false
assumption.  The LimDep manual states that the default is a bandwidth
of
0.06 (E32-3; E32-20) but no information is provided on trimming. 
Anyone
have insight on this?

Thanks.  -Jill
Jill Caviglia-Harris
Associate Professor
Economics and Finance Department
Salisbury University
Salisbury, MD 21801
(410) 548-5591
_______________________________________________
Limdep site list
Limdep at limdep.itls.usyd.edu.au 
http://limdep.itls.usyd.edu.au 
_______________________________________________
Limdep site list
Limdep at limdep.itls.usyd.edu.au 
http://limdep.itls.usyd.edu.au