Loglinear analysis abnormality

Neuro

New Member
#1
Let me first explain the dataset, before I explain the abnormality (well according to me anyway)..

I have 4 factors: Organisation, colour, display size and accuracy. They all have two levels, resulting in 16 combinations.

There are equal numbers in every condition, except for the accuracy. There are always 60 trials in each combination of factors, of which a proportion were correct, and the others were wrong. This means, according to me, that the model can only come up with terms that interact with accuracy in order to explain the results.

The data looks like this:
Org Col Size Acc Freq
1 1 1 1 36
1 1 1 2 24
1 1 2 1 6
1 1 2 2 54
1 2 1 1 52
1 2 1 2 8
1 2 2 1 8
1 2 2 2 52
2 1 1 1 40
2 1 1 2 20
2 1 2 1 6
2 1 2 2 54
2 2 1 1 55
2 2 1 2 5
2 2 2 1 26
2 2 2 2 34



However, when I ran a hierarchical loglinear model on it in SPSS (v13.0 for windows), these are my results :

The highest order terms, according to my model, are :
size * correct
colour * size :confused:
org * correct
colour * correct


I keep getting an interaction between colour and size, which, according to me, simply cannot be there. How can I explain this? Can anyone run this through another statistical program and see if they come up with the same result?

Thanks in advance
 

JohnM

TS Contributor
#2
I use Statgraphics at work, so I am limited in capabilities, but doing something really quick using a GLM, I got what you anticipated - the only factors in the model have any effect at all are the ones interacting with Accuracy.
 

Neuro

New Member
#3
I don't think I 'm allowed to use a general linear model on this data.. This is single case data (so I'm not even allowed to do a loglinear analysis, but it's more comprehensive as a model than large amounts of chi squares, it's an acceptable way of analysing within neuropsychological single case studies)

But still, there's no explanation for the abnormal (impossible) result in the loglinear analysis. Any thoughts of somebody who 's an expert on loglinear analysis?

This is very important to me, since this is part of a paper that is to be submitted soon. Any referee who sees the colour*size interaction will raise his/her eyebrows...

Thanks anyway....