[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
Re: st: psmatch2/xtlogit fe/xtreg fe
"Austin Nichols" <email@example.com>
Re: st: psmatch2/xtlogit fe/xtreg fe
Thu, 11 Oct 2007 12:16:08 -0400
Claire Kamp Dush <firstname.lastname@example.org>:
Instead of matching, try using the _pscore variable (or make your own
with a logit of married on all X vars, then predict probability of
being in the "treatment" group) and make a new weight
w=_pscore/(1-_pscore) along the lines of Dinardo Fortin and Lemieux
(see -findit dfl- and references therein). Now use that weight as a
[pweight] in -clogit- instead of -xtlogit, fe- and in -areg- instead
of -xtreg, fe-. Email me or call me if you want more specifics.
On 10/11/07, Claire Kamp Dush <email@example.com> wrote:
> Dear Statalisters,
> I am having a problem attempting to conduct fixed effects regression with a matched sample obtained from psmatch2. Below is the detailed information about my problem.
> Sample: Large sample of mothers in married or cohabiting unions measured over three time points. They must be married or cohabiting at the second time point.
> Research question: Is there a difference in the impact of marital vs. cohabiting dissolution on mental health?
> I run several models that all have the same issues. Below is an example.
> First I use probit to obtain the predicted probability that an individual is in the group that dissolved their cohabiting union at Wave 3. The sample here is all those who were cohabiting at wave 2. Predictors include several Wave 1 observed variables that are related to union stability and Wave 2 depression.
> Next, I use the predicted probability to do propensity score matching using psmatch2 (nearest neighbor with replacement, caliper 0.03) where the treated group is those that dissolve a cohabiting union by Wave 3 and the control group is those that are intact by Wave 3. The outcome is depression at Wave 3. I then create a dataset with just the treated and matched controls, the weight variable produced by psmatch2, and an indicator of union status at Waves 2 and 3. I put into this dataset also a variety of observed variables measured at Wave 2 and Wave 3 that also change over time, including their child's age, mother and father's employment status (dummies), mothers school status and whether she completed her education (dummies), mother's welfare receipt (dummy), total family income, total num of adults working in the household, whether the dad has ever been in jail, days the dad saw the child, whether the mother obtains a new partner (dummy), the length of time since separ
> tion, and my outcome variable at Wave 2 and 3. I then convert this dataset to long.
> Note that I do this for each dependent variable separately - I have four measures of mental health- here I would like to discuss a continuous measure of depressive symptoms and a dichotomous measure of clinical depressed.
> Finally, I am attempting to run a fixed effects regression on the matched sample, so that my final estimate accounts for observed characteristics that do not change over time but distinguish those in cohabiting unions that dissolve and do not dissolve, observed characteristics that do change over time and may also account for a decline in mental health following a cohabiting dissolution, and unobserved characteristics that do not change over time.
> Here is where I run into the problems.
> Problem 1:
> When I attempt to run a fe xtlogit on the dummy of clinical depression, here is what I get an error "Multiple positive outcomes within groups encountered". This is for the baseline model with just the dichotomous measure of depressive symptoms (deplib) and the indicator of union status at each wave (cohdis) entered in the equation. I am getting this error because several groups do not experience a cohabitation dissolution? My code and output is listed below.
> Problem 2:
> When I attempt to run a fe xtreg on the continuous measure of depressive symptoms, I am not allowed to have a weight, so I cannot weight my controls by how many treated they matched to. I could try to re-run the analysis removing the "with replacement" option, or I could expand the data to have as many groups for each control as treated they match to. But, my question is, would my standard errors then be wrong? Is there a way to adjust my standard errors if I did this? Code for this is also below.
* For searches and help try: