Dear Lien,

In SPSS the MLG procedure offers univariate and multivariate analysis repeated measures data. You can contrast both balanced or unbalanced models. A design is considered to be balanced when each model box has the same number of cases. Otherwise in a multivariate model the sum of squares due to the effects of the model and the sum of squares due to error are in the form of a matrix instead of being in the scale form of the univariate analysis. These matrixes are called SCPC ( sum of squares and vector products). MLG contrast hypothesis and also generates parametric estimations.

I hope you find it useful

Estela

Dear Lien,

before going explaining the interaction effect, I suggest that you redefine your model. Instead of using the pre and post measurements as outcome, I would calculate the difference (post-pre) of the AAAQ score to investigate the increase/decrease in crafing after stimulation. The pre value should then be considered as a covariate and the time factor is skipped from the analysis. Additional factor would be the method (sham vs real). Post-hoc tests could then be conducted for the comparison between the methods. The status of dependence from alcohol could be another factor, although this may correlate with the pre-evaluation - at least from my perspective the AAAQ score may be slightly inflated for alcoholics even without stimulation, therefore the correlation between the AAQ and AUDIT score should be investigated. Included factors and other covariates should not be correlated, otherwise you will have to consider Ridge-Regression or similar due to the collinearity. Furthermore my proposed analysis only makes sense if the number of investigated patients is larger enough (approximately 10 patients per included covariate and category of each factor). If you have multiple observations per patients, this can be either considered in the model using a random effect per patient or by applying GEE models and evaluate the patient mean difference (Only if the independent variables - covariate/factors - are the same for each observation within a patient).

Now considering interaction effects (e.g. A*B): It should be investigated if they are of qualitative value, this can be achieved with the help of the corresponding mean plots if two factors are analysed. Furthermore, the analysis of the mean plots gives you an insight whether the main effects (A, or B) can still be interpreted.

This video may help you understand the different types of interactions https://www.youtube.com/watch?v=1o4CO2mBmgU

Note that interaction terms should only be included in the model, if they are of interest since reasonably likely, or if the improve the model fit (e.g. via AIC). In general these interactions will be investigated for its type, significance and model fit.

Note that your eta squared value is almost 0 for all factors, implaying a bad model fit (the observed power corresponds to the size of the p-value - to me this holds no information, as the correlation between p-value and observed power is obvious). As you can see the eta value is much better for the stimulation and its interaction with AUDIT score. Therefore I suggest you apply the model I mentioned above (model the difference in AAAQ) including only the pre-value, the stimulation and the AUDIT value and test their 2-way interaction terms.

I hope this helps,

Best regards,

Ann