# mixed model differences question

1 message
Open this post in threaded view
|

## mixed model differences question

I’d like to understand what accounts for the differences I see. I’m using v26.0.0.0 if that matters.

Consider these two models:

mixed behavior with wave txgroupn/fixed wave wave*wave txgroupn/

print solution/random intercept wave | subject(rid) covtype(un).

mixed behavior by txgroupnrev with wave/fixed wave wave*wave txgroupnrev/

print solution/random intercept wave | subject(rid) covtype(un).

The first model is the standard covariate setup. Txgroupn is a dichotomous variable coded 0, 1. Wave is the time coding for a growth curve.

The second model makes txgroupn a factor and so that the correct group is the reference, txgroupn is reversed coded as txgroupnrev.

My current understand says the results ought to be the same. But they’re not.

Type III tests fixed effect for model 1 then model 2

 Type III Tests of Fixed Effectsa Source Numerator df Denominator df F Sig. Intercept 1 165.400 1006.201 .000 wave 1 87.621 8.105 .005 wave * wave 1 74.521 4.831 .031 txgroupn 1 82.545 .042 .838 a. Dependent Variable: Behavior.

 Type III Tests of Fixed Effectsa Source Numerator df Denominator df F Sig. Intercept 1 138.703 1324.826 .000 wave 1 87.621 8.105 .005 wave * wave 1 74.521 4.831 .031 txgroupnrev 1 82.545 .042 .838 a. Dependent Variable: Behavior.

Now the estimates table for model 1 then model 2

 Estimates of Fixed Effectsa Parameter Estimate Std. Error df t Sig. 95% Confidence Interval Lower Bound Upper Bound Intercept 3.869160 .121976 165.400 31.721 .000 3.628329 4.109990 wave .059161 .020781 87.621 2.847 .005 .017861 .100460 wave * wave -.002448 .001114 74.521 -2.198 .031 -.004667 -.000229 txgroupn -.023950 .117075 82.545 -.205 .838 -.256826 .208926 a. Dependent Variable: Behavior.

 Estimates of Fixed Effectsa Parameter Estimate Std. Error df t Sig. 95% Confidence Interval Lower Bound Upper Bound Intercept 3.869160 .121976 165.400 31.721 .000 3.628329 4.109990 wave .059161 .020781 87.621 2.847 .005 .017861 .100460 wave * wave -.002448 .001114 74.521 -2.198 .031 -.004667 -.000229 [txgroupnrev=0] -.023950 .117075 82.545 -.205 .838 -.256826 .208926 [txgroupnrev=1] 0b 0 . . . . . a. Dependent Variable: Behavior. b. This parameter is set to zero because it is redundant.

The intercept df for the F and the F itself differ but the dfs and Fs for the other effects don’t. I’d always understood that F = t*t. Well, 1006.201 is not quite 31.721 squared but squaring works for the other parameters.

But on a model with another binary predictor beside txgroupn, the intercept and wave dfs and their Fs differ even though the estimates tables don’t.

This could just be something I don’t understand, if so I’d like to understand better. If not my understanding, is this fixed in a patch?

Thanks, Gene Maguin

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD