RE: st: RE: Propensity score matching -balancing property

["Millimet, Daniel" <millimet@mail.smu.edu>]

If you are changing the entire level of analysis - household to individual - then it strikes me that p-score model will differ as presumably you have some individual Xs in the model as well.  In any event, as I said before, the literature is clear that it is best to over-specify the model as there is little consequence, whereas there is a huge bias from under-specifying. Beyond that, I don't think anyone can offer you firm advice without looking at the data.

Double robust estimators augment an OLS equation with an additional regressor that is a function of the p-score.  As long as either the outcome or the p-score eqtn is correctly specified, the estimator is consistent.

See

Bang, H. and J.M. Robins (2005), Doubly Robust Estimation in Missing Data and Causal Inference Models,Biometrics, 61, 962-972.

Scharfstein, D.O., A. Rotnitzky, and J.M. Robins (1999), Adjusting for Nonignorable Dorp-Out Using Semiparametric Nonresponse Models, Journal of the American Statistical Association, 94, 1096-1120

****************************************************
Daniel L. Millimet, Professor
Department of Economics
Box 0496
SMU
Dallas, TX 75275-0496
phone: 214.768.3269
fax: 214.768.1821
web: http://faculty.smu.edu/millimet
****************************************************

-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Nyasha Tirivayi
Sent: Saturday, January 15, 2011 12:10 PM
To: statalist@hsphsun2.harvard.edu
Subject: Re: st: RE: Propensity score matching -balancing property

Dear Daniel

Thank you for your response.

For PSM I am assessing outcomes for different units of observation
within the same dataset i.e. about 400 observations for household
outcomes and around 1600 for individual outcomes. Would you consider
this to be small changes in sample size? A

The double robust estimator you refer to, is it for matching or regression?

Kindly reply

Regards

N.Tirivayi
Maastricht University



On Sat, Jan 15, 2011 at 5:57 PM, Millimet, Daniel <millimet@mail.smu.edu> wrote:
> The outcome has nothing to do with balancing since it does not factor into the balancing tests (only the p-score and Xs matter).  The difference, as you note, is that the sample changes across outcomes, and these explains your changing balancing results.
>
> I would be skeptical about why the balancing test is that sensitive to (presumably small) changes in sample size.  This suggests that one should be cautious claiming that the Xs are balanced in the first chapter.
>
> In light of this, as well as based on work I (with Tchernis in J Bus & Eco Stats) and other have done on the benefits of over-specifying the p-score eqtn, I would err on the side of using the least parsimonious specification for all outcomes.
>
> You might also try other estimators in addition to matching to assess robustness, such as a doubly robust estimator.
>
> ****************************************************
> Daniel L. Millimet, Professor
> Department of Economics
> Box 0496
> SMU
> Dallas, TX 75275-0496
> phone: 214.768.3269
> fax: 214.768.1821
> web: http://faculty.smu.edu/millimet
> ****************************************************
> -----Original Message-----
> From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Nyasha Tirivayi
> Sent: Friday, January 14, 2011 9:46 PM
> To: statalist@hsphsun2.harvard.edu
> Subject: st: Propensity score matching -balancing property
>
> Hello
>
> I have a question concerning psmatch2 and general propensity score matching:
>
> The propensity score model I have used to analyse the first outcome
> for my first chapter of research does not satisfy the balancing
> property when I apply it to the other outcomes to be presented in
> later chapters. Should I use the propensity score model I have used in
> the first paper throughout the next chapters for all the outcomes,
> even if it does not balance all the time? Or each outcome might
> require a separate propensity score model with maybe different
> covariates?
>
> Each outcome also has different number of observations. Does this
> support the use of separate PSM models?
>
> Kindly respond
>
> N.Tirivayi
> Maastricht university
> *
> *   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/
>
> *
> *   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/
>

*
*   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/

*
*   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/