PROC GLM with Contrast statement

[lizzy28]

Hi all,

I am using PROC GLM Contrast statement to see if my data displays any trend like linear or quadratic. Somehow when I use two CONTRAST statements at the same time, the CONTRAST output does not come out.

Below is my code:

PROC GLM DATA=mydata;

CLASS RESP_TIME;

MODEL score= RESP_TIME ;

CONTRAST 'Linear' RESP_TIME -2 -1 0 1 2;

CONTRAST 'Quadratic' RESP_TIME 2 -1 -2 -1 2;

run;

Another question is: is it possible to do the trend plot along with CONTRAST statement?

Thanks a lot!

Lizi

[Reeza]

Please post the log from that code. I think GLM may need a QUIT instead of RUN as well, which is why it may not be completing.

[lizzy28]

Thanks Reeza.

I added QUIT, and the log remains as below:

WARNING: More coefficients than levels specified for effect resp_time. Some coefficients will be ignored.

NOTE: CONTRAST Linear is not estimable.

WARNING: More coefficients than levels specified for effect resp_time. Some coefficients will be ignored.

NOTE: CONTRAST Quadratic is not estimable.

[Reeza]

Your contrast statements are not correct then. I don't know enough on how to fix those unfortunately. I would expect to run a quadratic model to have to test the quadratic, but I could be totally off base here. I'll move this to the Stat's forum as well.

[PaigeMiller]

---

@lizzy28 wrote:

Thanks Reeza.

I added QUIT, and the log remains as below:

WARNING: More coefficients than levels specified for effect resp_time. Some coefficients will be ignored.

NOTE: CONTRAST Linear is not estimable.

---

How many levels of the variable RESP_TIME are there?

Are there levels of RESP_TIME that have no Y values (due to missing data)?

[lizzy28]

Thank you Reeza!

Paige, RESP_TIME has three levels with numeric values 1, 2 and 3, and it has not missing values. Thank you for looking into my question!

[PaigeMiller]

Then you can't specify a CONTRAST statement with 5 levels. You can only specify a CONTRAST statement with 3 levels.

My other question was about your Y variable, in this case SCORE, does it have a lot of missing values such that there is a level of RESP_TIME for which all the Y variables are missing?

[lizzy28]

I see that I should only set three levels. One further question is how I should decide the L matrix for linear vs quadratic.

Yes, there are about 8% missing values in SCORE.

[sld · Accepted solution]

Although it is possible to fit a quadratic model to only 3 time points, so few points provide very little support for the fit. 

SAS Usage Note 22912 illustrates how to acquire coefficients for polynomial contrasts.

I hope this helps.

[lizzy28]

That makes sense. Thanks for the link. It provides me a quick understanding of polynomial coefficients.