# How do I model repeated measure + random effects with SPSS mixed

TL;DR: I'm having trouble understanding what a correct model for my data should be, and also if it would even be a good idea to make it a single model.

I have a regression model with two predictors. The predictors are 'perceived ease of use' (ease) and 'Perceived Usefulness' (useful), and the dependent variable is 'Intention to Use' (intention). We measured these variables for 10 different technologies. So for each person, we have a multiple of these variables. On top of that, we have two measurements for each person (repeated measures).

If possible, I would like to fit one model (mixed model in SPSS) into this data. Of course an "easy" solution would be to do a "simple" repeated measure model for each technology separately, but then I lose a lot of information, because I cannot partial out the effect of the person, and difference between the models would be hard to explain.

So my questions are:

• Is it valid to treat all dependent variables (intention) for the different technologies as the same outcome variable, or do we have to do 10 separate analyses? As I said, the reasons to make a single model is that more information is preserved (the effect of the person), thus hopefully resulting in a more general model that is based on "more" data.

• If it is valid, how do I model this in SPSS? Right now I have this:

MIXED intention BY personID time technology WITH  ease usefulness
/FIXED = usefulness ease | SSTYPE(3)
/RANDOM=technology time|  SUBJECT(personID) COVTYPE(ID)
/REPEATED = time*technology | SUBJECT(personID)
/PRINT = SOLUTION TESTCOV.


However I'm not sure this models the dependencies in the data correctly. I also don't know what kind of covariance matrices to choose.

Does this model fits my description well? If not, what I can do to improve it?

• I vote all one dependent variable with a random intercept by person, a time independent variable possibly as a random slope (with the possibility of higher order time IVs, e.g., time^2) and any other observation or person level predictors. If you lack person level predictors, a fixed effect type of approach may be better. Commented Nov 29, 2013 at 2:17

It depends a bit on why you have two measurements per person for each technology. If this is to get a better measurement, you could get the average across the two and run a model like

lme(averageintention ~ ease + useful + technology, random = ~ 1 | person,
data=dataset, method="ML")


or with any interactions that might seem valid.

Running mixed models is much easier with R compared to SPSS, especially to compare different models and get more insights into the different effects.

If you find technology has a significant effect in the model, you could then subset the model for the different technologies and run it again. In R, you could do this by first creating a variable for the separate technologies, e.g. tech1Sub <- dataset\$technology=="tech1"and then rerunning the model as follows:

lme(intention ~ ease + useful + technology, random = ~ 1 | person,
data=dataset, method="ML", subset = tech1Sub)

• The measurement is taken once right after they were explained or showed or they tried the technology once, the second measurement is for when they have had the opportunity to try the technology in the field. So I suppose averaging the intention variable in not appropriate here? Commented Nov 29, 2013 at 18:07
• No averaging is not appropriate in that case. Think about the questions you want to ask. The full model you could run is lme(intention ~ ease + useful + technology * timing, random = ~ 1 | person, data=dataset, method="ML", subset = tech1Sub) whereby timing is either 1 or 2 based on when the 'intention' measure was tested. Commented Dec 1, 2013 at 17:10