topic Re: Class variable and interaction in PROC MIXED in Statistical Procedures

Class variable and interaction in PROC MIXED

sadiekasso — Thu, 10 Jul 2014 20:29:13 GMT

Hey All,

I'm having some confusion running a repeated measures regression model in SAS proc mixed and hoping to get some help. I'm trying to determine whether or not I need to include the dichotomous variable in the interaction term (shown in the model below) in my Class line and why when I do so, I only get results for one level of the dichotomous variable. Here is the model :

proc mixed data = ks_adh.CPTII_sts_fixed_lbdi_cent;

class patno session sex;

model bdi = lbdi_cent sitenum F1_cpt3_Mean_cent_resid|sex /s covb CL;

repeated session / type=un subject = patno ;

run;

In the model = BDI is a continuous variable and represents patients scores on a depressive symptom measure at their 2nd,3rd,4th, and 5th sessions of treatment (or their next session BDI score). We are predicting this from : (1) The main effect of their BDI score at the current session (sessions 1,2,3,4) , the main effect of site, main effect of ratings of therapists use of cognitive methods at the current session (sessions 1,2,3,4- represented by the F1_cpt3_Mean_cent_resid variable), the main effect of patients sex, and the interaction of cognitive methods at each current session and patient sex. If I include sex in the class statement - in the output window, solutions for Fixed effects are only available for one level of SEX as shown here:

Solution for Fixed Effects

Effect sex Estimate Error DF t Value Pr > |t| Alpha

sitenum -0.4326 0.4830 54 -0.90 0.3744 0.05

F1_cpt3_Mean_cent_re -0.1276 1.5644 54 -0.08 0.9353 0.05

sex -0.5 0.2167 0.4812 54 0.45 0.6543 0.05

sex 0.5 0 . . . . .

F1_cpt3_Mean_cen*sex -0.5 -6.0599 2.1393 54 -2.83 0.0065 0.05

F1_cpt3_Mean_cen*sex 0.5 0 . . . . .

Does anyone know why solutions don't appear for the .5 level of the sex variable?

Beyond this, is it necessary that I include sex in the class statement? It's important to note, that I am going to plug the estimates derived from these analyses into the Johnson-Neyman tool for probing interactions: Two-Way Interaction Effects in MLR in order to determine the simple slope of the F1 variable at each level of sex . Given this, should the estimates for fixed effects and the variance/covariances that I plug into the JN calculator be from a model that DOES include sex in the class statement? or from a model that does NOT include sex in the class statement. Simple slopes generated differ drastically among the two approaches. Any help at all or tips would be greatly appreciated!

Re: Class variable and interaction in PROC MIXED

sadiekasso — Thu, 10 Jul 2014 22:15:16 GMT

Just to give an update= After playing around with things, I noticed that when the main fixed effects of SEX and F1_cpt3_Mean_cent_re are NOT included in the model (but the interaction term is ) I do get results for both levels of the Sex variable (see model below):

proc mixed data = ks_adh.CPTII_sts_fixed_lbdi_cent;

class patno session sex;

model bdi = lbdi_cent sitenum F1_cpt3_Mean_cent_resid*sex /s covb CL;

repeated session / type=un subject = patno ;

run;

RESULTS:

Solution for Fixed Effects

Standard

Effect sex Estimate Error DF t Value Pr > |t| Alpha

Intercept 24.0615 0.2226 55 108.12 <.0001 0.05

lbdi_cent 0.9774 0.02413 55 40.51 <.0001 0.05

sitenum -0.4730 0.4603 55 -1.03 0.3086 0.05

F1_cpt3_Mean_cen*sex -0.5 -6.1841 1.4618 55 -4.23 <.0001 0.05

F1_cpt3_Mean_cen*sex 0.5 -0.1531 1.5656 55 -0.10 0.9224 0.05

HOWEVER, when the main effects AND the interaction are included ( as they were in the model I originally posted about) I do NOT get results for both levels of sex (see model and results above) . Anyone know why this is the case? and how I could get the results for the main effects AND both levels of my Sex and Sex by F1 interaction? This would solve all of my problems! thanks

Re: Class variable and interaction in PROC MIXED

SteveDenham — Mon, 14 Jul 2014 14:19:24 GMT

When main effects and interactions are included in a model, the solution vector (by default) sets the last value to zero to avoid a singular matrix. All other values need to be thought of as deviations from the intercept or overall slope as appropriate.

I see that you are treating sitenum as a continuous variable as well, so you have a multiple regression situation, where data imbalance will make estimates differ when alternative models are fit.

In any case, the last version of the output fits separate slopes, but a common intercept, by sex for f1_cpt3_mean_cen. The first fits separate sloptes, and separate deviations from the overall intercept, by sex for this regression.

Steve Denham