08-06-2013 04:11 AM
How do you request for a Hausman Test under the PROC SYSLIN procedure? I am trying to carry out a 2 stage least squares regression and need to check for the validity of my instrument.
I tried googling but to no avail.
08-06-2013 05:00 PM
Thank you for your response. PROC MODEL does have the Hausman option. However, it gives me results of a non-linear 2SLS regression whereas what I want to test is a linear model. Is it possible then to impose linear restrictions using PROC MODEL?
08-07-2013 01:18 PM
I am probably missing something here, but say you had:
proc model data=in;
y1 = a1 + b1*y2 + c1*x1;
y2 = a2 + b2*y1 + c2*x2;
fit y1 y2 / 2sls;
(that probably needs some parms or instrument statements, but this is for illusration).
Specifying 2SLS means that derivatives are replaced with predicted values. This has meaning if the equations are non-linear, but it just means that the instrumental variables x1 and x2 are used in the first stage fit, for when all of the equations in the system are "linear in the parameters" the derivatives are just regressions on the instrumental variables.
So, PROC MODEL should be a good choice if you want to now do Hausman testing.
08-07-2013 07:20 PM
I am referring to the discussion here http://www.listserv.uga.edu/cgi-bin/wa?A2=ind0512a&L=sas-l&P=20981
"PROC SYSLIN is designed to work on simultaneous linear equations", which is what I intended.
However, "PROC MODEL is designed to work on non-linear equations".
08-08-2013 07:57 AM
Since the two methods give different results, there is obviously something different, and David gives the reason. However, that does not mean that the results obtained from PROC MODEL are incorrect, merely obtained by a different algorithm. And for me (and me only), I would trust the results from a goal seeking algorithm that accommodated correlations over an algorithm that does not accommodate the correlations (OLS and its extension to 2SLS in SYSLIN). It's the only way I see in SAS to do the Hausman testing on simultaneous equations, other than to "roll your own" through data step programming.