Re: st: RE: gpscore for propensity score of 4 observed groups

[Austin Nichols <austinnichols@gmail.com>]

Stephen O'Connor <oconnor7@u.washington.edu>:
Why not just run three propensity score reweighting models comparing
each level of treatment to the control?
Of course, there is no magic in these methods; you are still assuming
selection on observables as you would in OLS.

On Tue, Oct 12, 2010 at 6:24 PM,  <oconnor7@u.washington.edu> wrote:
> Ok, thank you Juan!
>
>
>
> On Tue, 12 Oct 2010, Villa Lora, Juan Miguel wrote:
>
>> no, you misunderstood. A control group is what you need. But I would
>> expect a 'clean' control group without any kind of treatment or a 'placebo
>> taker'. If you want to use a control group with certain grade of treatment
>> you won't get a reliable causal inference. I recommend an article found at
>> http://ftp.iza.org/dp4451.pdf "Evaluating Nonexperimental Estimators for
>> Multiple Treatments: Evidence from Experimental Data"
>> ________________________________________
>> From: owner-statalist@hsphsun2.harvard.edu
>> [owner-statalist@hsphsun2.harvard.edu] On Behalf Of
>> oconnor7@u.washington.edu [oconnor7@u.washington.edu]
>> Sent: Tuesday, October 12, 2010 6:09 PM
>> To: statalist@hsphsun2.harvard.edu
>> Subject: RE: st: RE: gpscore for propensity score of 4 observed groups
>>
>> Thanks for your help. Yes, actually the groups consist of three levels of
>> TBI severity and a control group of arm injured patients. Sorry for the
>> confusion. I did not think that having a control would affect the propensity
>> score in this situation.
>>
>> Stephen
>>
>> On Tue, 12 Oct 2010, Villa Lora, Juan Miguel wrote:
>>
>>> Stephen,
>>> I haven't dealt with such case. Anyway, it seems to me that you don't
>>> have a control or comparison group without treatment. This situation may
>>> confound your evaluation and your results will be invalid.
>>> Juan
>>>
>>> ________________________________________
>>> From: owner-statalist@hsphsun2.harvard.edu
>>> [owner-statalist@hsphsun2.harvard.edu] On Behalf Of
>>> oconnor7@u.washington.edu [oconnor7@u.washington.edu]
>>> Sent: Tuesday, October 12, 2010 5:32 PM
>>> To: statalist@hsphsun2.harvard.edu
>>> Subject: Re: st: RE: gpscore for propensity score of 4 observed groups
>>>
>>> Hi Juan,
>>>
>>> Yes, this is true, the treatment variable is categorical. I was worried
>>> that this would be an issue. That said, what command what provide a
>>> propensity score for a categorical variable with four levels?
>>>
>>> Stephen
>>>
>>> On Tue, 12 Oct 2010, Villa Lora, Juan Miguel wrote:
>>>
>>>> Hey Stephen!
>>>> Before answering your question I'd like to understand your treatment
>>>> variable. You are stressing that you got four different groups, which makes
>>>> me think that you are dealing with a categorical variable instead of a
>>>> continuous variable as it's required for gpscore. Remember that Hirano and
>>>> Imbens (2005) spell out that this treatment variable must be a normal
>>>> distributed (one of the basic assumptions of the analysis).  Are you sure
>>>> your treatment variable is continuous?
>>>> Juan
>>>> ________________________________________
>>>> From: owner-statalist@hsphsun2.harvard.edu
>>>> [owner-statalist@hsphsun2.harvard.edu] On Behalf Of
>>>> oconnor7@u.washington.edu [oconnor7@u.washington.edu]
>>>> Sent: Tuesday, October 12, 2010 5:15 PM
>>>> To: statalist@hsphsun2.harvard.edu
>>>> Subject: st: gpscore for propensity score of 4 observed groups
>>>>
>>>> Hi,
>>>>
>>>> I have technical question regarding gpscore.
>>>>
>>>> If my treatment variable is type of injury severity, what should I put
>>>> for "cutpoints", "index", and "nq_gps"? The treatment level has four
>>>> different levels of injury severity.
>>>>
>>>> What I am seeking is to create a propensity score for the likelihood
>>>> that a person develops a certain type of injury, in this case TBI, based
>>>> upon a list of covariates.
>>>>
>>>> This is the syntax I used for a preliminary analysis:
>>>>
>>>> gpscore  age child_gender prev_anxiety prev_depn prev_adhd
>>>> prev_behavioral prev_other_mh, t(inj_severity) gpscore(mygps)
>>>> predict(hat_inj_sev) sigma(hat_sd) cutpoints (inj_severity) index(p50)
>>>> nq_gps(4)
>>>>
>>>> Again, I am unsure of cutpoints, index, and nq_qps given that I have
>>>> four established groups for the treatment variable.
>>>>
>>>>
>>>> Thank you,
>>>> Stephen
>>>>
>>>>
>>>>
>>>>

*
*   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/