topic Re: compare regression slopes in Statistical Procedures

compare regression slopes

nuitamment — Thu, 21 Jul 2016 20:50:44 GMT

Hi,

I am running a regression with proc surveyreg which looks like this: Y= x0 + X1 + X1*DD + X1*X2 + X1*X2*DD.

DD is a dummy variable and all the other ones are continuous variables.

How can I compare the effect (!) of the regression slopes between X1*DD and X1*X2*DD and test if the difference is significant? I do not want to include the intercept.

Re: compare regression slopes

ballardw — Thu, 21 Jul 2016 22:01:31 GMT

Is there some reason that X2 is not included in main effects??

Re: compare regression slopes

nuitamment — Thu, 21 Jul 2016 22:10:25 GMT

X2 and DD are additional included as main effects.

Re: compare regression slopes

Ksharp — Fri, 22 Jul 2016 00:47:26 GMT

I think both CONTRAST Statement and ESTIMATE Statement can do that .

Re: compare regression slopes

nuitamment — Fri, 22 Jul 2016 08:28:55 GMT

Thanks!
For clarification: If I use the contrast statement I can only compare the values of the dummy variable with each other under specified values of the continuous variable? Or can I use the contrast statement to something like an F-Test for groups of continuous regression slopes?
And if I use the estimate statement, I can estimate the effect for all specified variables. However I do not have the difference between effects. Am I wrong here?

Re: compare regression slopes

Ksharp — Fri, 22 Jul 2016 08:56:37 GMT

Sorry. I am not expert about it. I want leave it to @Steve .... CONTRAST can compare these two coefficience/parameter of variables are equal or not.

Re: compare regression slopes

SteveDenham — Fri, 22 Jul 2016 17:58:51 GMT

While this doesn't do a formal hypothesis test of the difference, it should tell you everything you want to know. Add CLPARM and SOLUTION to your MODEL statement. You could then match up the appropriate terms and look for overlap in the confidence intervals.

I think the only way to get a formal test will involve getting the inverse of the X'X (or X'WX) matrix and construct the proper quadratic form, because as you have identified, CONTRAST, ESTIMATE and LSMESTIMATE statements don't really lend themselves to testing between continuous covariate estimates. I suppose that you could construct an ESTIMATE statement that had everything in there, in the right place, and used the proper values, but I suspect that it would be more likely to give an estimate at the particular values of the continuous covariate than of the difference in the slope estimates.

Steve Denham

Re: compare regression slopes

nuitamment — Sun, 24 Jul 2016 18:45:20 GMT

Ok, thank you!