[ANOVA - Repeated Measures] When to include subject as a random effect?

[ProfCorrow]

I have an eye tracking study in which two groups of subjects (patient and control) are completing a face recognition task in which they learn a set of faces and then are shown faces one at a time and have to decide if each face is one that they learned earlier. During the task, we are recording their eye movement patterns to see if they look more to the top half or bottom half of the face on learning faces in the learning phase, target faces in the test phase, and distractor faces in the test phase. Thus, we have calculated the average duration of looking time to each region of interest (upper face v lower face) for each type of face (learning v target v distractor).

In one analysis we are examining the control group only. Thus, we are conducting a 2 (face half: upper v lower) x 3 (face type: learning v target v distractor) repeated measures ANOVA.

We have been told that we should always include subject as a random effect. My understanding is that, in SPSS terms, this means we would need to use SPSS MIXED. What I don't understand is WHY we should include subject as a random effect and how we would go about doing so. Should subject be included as a random effect by default simply because it is a repeated measures design?

I did find this, which seems to be on the topic. But I am having a hard time connecting the instructions here with what I am supposed to do with this particular data set. http://www.spss.ch/upload/1126184451_Linear Mixed Effects Modeling in SPSS.pdf

Thank you in advance for your help!

 

[atamalu]

It depends on the situation. A lot of the time you won't need to include subject as a random effect since SPSS uses the "within-subjects" factors to account for subject dependency with the sphericity assumption in a standard repeated measures ANOVA.

Including the subject as a "random" effect dips into multilevel models, which are used with non-independent observations. However, if you're not very familiar with multilevel models (personally I'm not yet), I wouldn't suggest using one due to interpretation issues.

So I'd suggest 1 of 2 things:
1. Run the test as Analyze -> General Linear Model -> Repeated Measures and apply the Greenhouse-Geisser correction if sphericity is violated
2. Avoid the sphericity assumption while using a standard mixed ANOVA model with single degree of freedom tests using contrast codes

Although contrast codes for repeated measures are a massive pain to deal with in SPSS so I would suggest the first option.

https://statistics.laerd.com/spss-t...a-repeated-measures-using-spss-statistics.php
Laerd is by far the best site I've found to explain SPSS methods.

 

[ProfCorrow]

Thank you Atamalu. I have since spoken with a few other researchers about this and the consensus is that the GLM is becoming an outdated way to run a repeated measures ANOVA because it uses up so many degrees of freedom. But, I'm with you! I'm not yet very familiar with multilevel models.