1
$\begingroup$

I'm using SPSS to run a Mixed Model with two categorical (factor) predictors with an interaction between the two predictors. I get the following Estimates of Fixed Effects:

Estimates of Fixed Effects

In the interaction I am unclear as to why I get four redundant (reference) categories and only two estimates of fixed effects. I was expecting three reference categories (at most): where all PredictorA =1 combinations of PredictorB would be referenced against PredictorA = 2 combinations. In the output, combinations of PredictorB =10 appears, to me, to have no reference.

Some clarity on how the redundant (reference) parameters work in the interaction would be appreciated.

Here is my syntax:

MIXED DV BY PredictorA PredictorB /CRITERIA=CIN(95) MXITER(100) MXSTEP(10) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE) /FIXED=PredictorA PredictorB PredictorA*PredictorB | SSTYPE(3) /METHOD=REML /PRINT=G SOLUTION TESTCOV /RANDOM=INTERCEPT | SUBJECT(iD) COVTYPE(VC)

Thanks for your time.

Edit:

Additionally the same happens in this example, taken from UCLA

enter image description here

$\endgroup$
1
  • $\begingroup$ So I created a new variable which is categorically coded for each combination of PredictorA by PredictorB and as expected I get an estimate for every combination/category, except for the largest coded combination which becomes the default reference category as per SPSS. Happily this work around provides me with the same Akaike's Information Criterion as the above interaction does, and I can compare my results to the reference category that I am interested in.....but I am still unclear as to what is happening above. $\endgroup$
    – MutinyDK
    Commented Aug 1, 2020 at 10:59

1 Answer 1

1
$\begingroup$

Without actually inspecting the data it is not possible to be 100% sure, but from the output it appears that level 2 in predictor A has not been observed (hence there is no estimate for it's main effect) and the same for level 10 in predictor B. Consequently all of the interactions involving either of these levels of the predictors will also not be estimated.

$\endgroup$
2
  • $\begingroup$ It's a good thought, but all combinations have observations within them. In reality Predictor B does have another level not included in this example in which it creates the scenario you mentioned, but shouldn't be the case here. I have had added a comment to my question for more context. $\endgroup$
    – MutinyDK
    Commented Aug 1, 2020 at 10:55
  • $\begingroup$ Please include a cross tabulation, or a post a link to the actual data. I can't say much more without more information. $\endgroup$ Commented Aug 1, 2020 at 11:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.