topic Re: Interpreting variable significance in proc mixed in SAS Programming

Interpreting variable significance in proc mixed

Nikrenzia — Fri, 15 Dec 2017 11:09:48 GMT

Hi,

I am running a mixed model (procedure mixed) in which a 3-factor variable is highly significat (PR>F : <.0001). However, when I run lsmeans (pdiff) and check the difference between the three factors, there is no difference between them.

Does anyone know what the reason for this result can be? The dataset is large (over 10000 observations).

Regards

/Stina

Re: Interpreting variable significance in proc mixed

PaigeMiller — Fri, 15 Dec 2017 13:17:18 GMT

Naturally, it is hard to give an explanation without seeing your code and the results (hint, hint)

Re: Interpreting variable significance in proc mixed

Nikrenzia — Fri, 15 Dec 2017 14:11:51 GMT

Ok. Here is the code and the results. I'm checking if there are differences in production between plants grown in 3 different locations. Length and Age are also important parameters.

proc mixed data = mydata;
class Location Plot Plant;
model Production = Location Length Age
/ddfm = Sat noint solution outpred = OriginalScale htype = 1,3;
random Int / subject = Plot type = un ;
random Int / subject = Plant type = un ;
lsmeans Location/ pdiff;
run;title;

Solution for Fixed Effects
Effect	Location	Estimate	Standard	DF	t Value	Pr > \|t\|
Error
Location	EK	0.147	0.1205	39.9	1.22	0.2299
Location	PG	0.1119	0.1354	190	0.83	0.4095
Location	TO	-0.167	0.1422	162	-1.17	0.2421
Length		15.084	0.1213	493	12.44	<.0001
Age		0.1678	0.02495	580	6.72	<.0001
LengthFromStadiu*Age		-0.07746	0.01455	601	-5.33	<.0001
Age*Location	EK	-0.1021	0.03355	523	-3.04	0.0025
Age*Location	PG	-0.08182	0.02484	516	-3.29	0.0011
Age*Location	TO	0	.	.	.	.

Type 1 Tests of Fixed Effects
Effect	Num DF	Den DF	F Value	Pr > F
Location	3	14.4	348.94	<.0001
Length	1	424	474.32	<.0001
Age	1	454	16.52	<.0001
Length*Age	1	597	20.6	<.0001
Age*Location	2	492	6.77	0.0013

Type 3 Tests of Fixed Effects
Effect	Num DF	Den DF	F Value	Pr > F
Location	3	67.5	1.57	0.2055
Length	1	493	154.73	<.0001
Age	1	597	44.3	<.0001
Length*Age	1	601	28.36	<.0001
Age*Location	2	515	6.77	0.0012

Least Squares Means
Effect	Location	Estimate	Standard	DF	t Value	Pr > \|t\|
Error
Location	EK	12.416	0.09533	24	13.02	<.0001
Location	PG	13.022	0.07318	26.8	17.8	<.0001
Location	TO	14.102	0.07072	12.8	19.94	<.0001

Differences of Least Squares Means
Effect	Location	Location	Estimate	Standard	DF	t Value	Pr > \|t\|
Error
Location	EK	PG	-0.06065	0.1199	24.6	-0.51	0.6175
Location	EK	TO	-0.1686	0.1186	18.9	-1.42	0.1716
Location	PG	TO	-0.108	0.1019	18.3	-1.06	0.3032

Re: Interpreting variable significance in proc mixed

PaigeMiller — Fri, 15 Dec 2017 19:21:56 GMT

I am running a mixed model (procedure mixed) in which a 3-factor variable is highly significat (PR>F : <.0001).

The Type III Test shows a p-value of 0.2055, which indicates no significant differences. You probably should be using the Type III Test here.

Re: Interpreting variable significance in proc mixed

Nikrenzia — Mon, 18 Dec 2017 09:31:47 GMT

Ok, do you have any idea why the III method is better here? I normally use Type I.

Or maybe this question is more suited for a statistical forum than a programming forum?

/Stina

Re: Interpreting variable significance in proc mixed

PaigeMiller — Mon, 18 Dec 2017 13:38:37 GMT

Type I assumes that the variable has been entered into the model first, and that the sequence of terms in the model is meaningful. The sequence of your model terms is arbitrary.

Type III assumes that the variable has been entered into the model last. This is more appropriate, it measures the significance of a model term assuming the effects of the other terms has already been accounted for.

More reading: http://documentation.sas.com/?cdcId=pgmmvacdc&cdcVersion=9.4&docsetId=statug&docsetTarget=statug_introglmest_toc.htm&locale=en

Re: Interpreting variable significance in proc mixed

Nikrenzia — Mon, 18 Dec 2017 15:50:17 GMT

Thanks!
/Stina