Statalist The Stata Listserver


[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: inverse probability of treatment weights


From   [email protected]
To   [email protected]
Subject   Re: st: inverse probability of treatment weights
Date   Fri, 06 Apr 2007 12:12:07 +0200

If pw must be constant within panel, I will work around by applying a cross-sectional command to my data. 
Thanks to the suggestions by Felix, I am now moving to -ologit-, in order to have no treatment, and several "doses" of the treatment. 
ologit treat $xvars
predict pi*, p
gen ipw=1
replace ipw=1/(1-pi0) if treat==0
replace ipw=1/pi1 if treat==1
replace ipw=1/pi2 if treat==2
...
replace ipw=1/pin if treat==n
which I am trying to re-shape as a loop, e.g. something like 
forvalues i = 1/e(k_aux) {
replace ipw=1/pi`i' if treat==`i'
}


At 02.33 06/04/2007 -0400, "Stas Kolenikov" wrote:
>On 3/27/07, [email protected] <[email protected]> wrote:
>> I heard of inverse probability of treatment weights (IPTW) and would like
>> to know if I am implementing them correctly on Stata (my data are PANEL).
>
>Is the reference for that:
>
>Little Bird (2007), On the Inverse Probability of Treatment Weights,
>Personal Communication.
>
>Or is there a more solid one? Sebastian and Maarten answered the
>question of how to code this in Stata, but as for the substantial
>questions like whether it is OK that the probabilities are strictly
>below than one, you would need to explain why those weights matter and
>how they are used in the model. (And yes, there are still chances that
>somebody who is observed to have been married might not have been --
>say a marriable guy on uninhabitted island :)).
>
>To my understanding (and I am far from being an expert on survival)
>- -logit- models are effectively implementing something like Cox model
>where you don't want to model the baseline parametrically -- which in
>the logit world transfers  to using dummies for semiparametric
>modeling of the time effects.
>
>Note also that Stata wants to have pweights constant within panel, if
>you want to use the panel data. So you need to estimate the overall
>probability of treatment for the whole panel/individual, rather than
>on the observation-by-observation basis. This may or may not
>correspond to your vision of how likely one is to get married over
>time, but that's a technical restriction with good enough underlying
>statistical reasons (the proper use of panel level estimating
>equations). 

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/



© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index