Hi Lijun,
why would you want to compare OLS regression coefficients with a matched
sample average?
I think, your question might be explained by your data:
The OLS coefficient for college does not only measure the true/caual
influence of college on income, but also the influence of other variables
(such as gender, age and race) on college attendance. But the coefficient
has not necessarily to be larger than the causal influence you obtain after
matching, as there might be compensating effects.
Example:
Suppose that men generally have a higher income than women and that also
more men get a college degree. Then, after matching, your matched sample
includes mostly men (as your treatment group
mostly consists of men, your control group also mostly consists of men).
Averaging over this subset of your observations leads to a higher average
effect of education on income because you omit the (lower) income bracket,
consisting mostly of women. OLS estimation on the other hand estimates the
effect (causal or not) of college on income for _all_ observations, thus
including the lower income bracket and thus leading to a lower estimate of
the 'influence' of education on income.
Jan
-----Urspr�ngliche Nachricht-----
Von: [email protected]
[mailto:[email protected]]Im Auftrag von Lijun Song
Gesendet: Sonntag, 28. M�rz 2004 03:07
An: [email protected]
Cc: [email protected]
Betreff: Re: st: RE: confusion about psmatch2
Hi, All,
I provide an example in detail here.
1) I first got the effect of college on income using OLS:
reg income college female age white;
2) Then, I use propensity score matching as follows:
psmatch college age sibs masei pasei,out(income) ate cal (.05) n (1);
However, the estimated coefficient of college using OLS is smaller than
the ATT and ATE derived from Propensity score matching.
Usually estimates from Propensity score matching should be less
conservative or smaller than estimates of OLS, right? It seems that OLS
ignores matching the pre-treatment characteristics, so estimates of OLS
maybe be overstated, right?
Lijun
---------- Original message ----------
From: [email protected]
To: [email protected]
Sent: Saturday, March 27, 2004 5:24:09 PM
Subject: st: RE: confusion about psmatch2
There's stuff here I've never heard about.
Can you explain how OLS can estimate "causal effects" please?
Nick
[email protected]
Lijun Song
> As a beginner at Stata, I am confused about the psmatch2.
>
> Suppose I am interested in the causal effect of college (a 1-0
> treatment indicator) on income. I also assume that race, sex, and
> family background (such as Mother and Father's SEI) will
> influence the
> treatment assignment. I also could not deny that race,sex, and family
> background also influence the outcome of interests, income directly.
>
> Then, after regressing income on college, I arrive at an estimated
> coffecient. After that, I use "psmatch2 college race sex masei pasei".
> I think the causal effect estimated by OLS estimator should be higher
> than those by Propensity Score Matching right?
>
> But my results show that the causal effect estimated by OLS
> is smaller
> than these by Propensity Score Matching. Why?
>
> In addition, the causal effect estimated by OLS is ATT or ATE?
*
* 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/
*
* 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/
*
* 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/
