I am taking part in a study that aims to enlighten the different impacts of group psychotherapy on alexithymia and emotional regulation for all comers patients in a day hospital.
The design appears simple. A patient enters in the hospital and some demographic information are taken in order to construct his/her personal, family and medical history. Then, he/she completes a battery of tests (MMPI-II; TAS-20 for alexithymia; Perceived social support; and DOE-36 for emotional openness). After this, he/she takes part in the proposed group activities. When the support end, the patient completes the same battery of tests, without the MMPI-II.
The main purpose of this work is to find signicative predictors of progression on TAS and DOE measurements. More precisely, we would like to be able in the future to obtain a base line of quantitative data (according to scores on MMPI-II, age, sex and scores on alexithymia and emotional openness) that would provide information for a better support (ideal lenght of stay; a recommended number of activities; medication; and perceived social support).
I would like to perform multivariated regressions but I haven't found an effective way to take into account the repeated measures. If I apply a delta on these repeated measures, I lose the position of the subject on the scales.
Does someone have a suggestion of statistical analyses to perform ?
Many thanks in advance.
Linear mixed effects will handle the repeated measures (as a random effect of Patient).
As I understand correctly, one patient belongs in either one investiaged group.
We have the baseline data for each patient (demography, medical history,..., MMPI-II).
Each patient has two measurement of TAS and DOE: One at the baseline (taken at the same time with demography data), and one after the intervention (the activities in each group that they are assigned).
So I suggest to use a regression model where the TAS and DOE (after intervention) are the responses (you can consider a multivariate regeression), and the group, demographic data, medical history and TAS, DOE at the baseline as covariates.
If I understand the study design correctly, each patient has only two measurements for TAS and DOE, so a linear mixed effects model is not needed here.
Yes, I think Thao is right; it sounds like a regular OLS regression would be fine.
It would also be possible to use mixed effects (instead of treating e.g. post-treatment TAS as the outcome variable and pre-treatement TAS as a predictor variable, instead you would have two datapoints for every subject, TAS would always be the outcome variable, and you would have an additional binary predictor variable ['pre' vs. 'post'] which you could use to see if there is a significant effect of treatment on TAS and if it interacts with the other predictors). It might not be necessary, though, since the analysis Thao suggested would also work.
Thank you Stephen, thank you Thao,
You have understood correctly the design of this study and I've taken careful note of both of your suggestions.
I have another little question. MMPI and activities (screened in 10 activity types) give me a big amount of data (around 60 independent continuous variables for the MMPI and 10 more for the proposed activities). According to your suggested analyses, does it sound correct to you if I firstly perform stepwise analyses in order to target the most relevant variables that might play a significative role in the predictive model ? Knowing that I have none a priori hypotheses on these variables.
Testing so many variables without any predictions is dangerous, both because many of them are likely correlated (which will cause multicollinearity, making it difficult to get an accurate estimate for any one predictor) and because it just is conceptually kind of like fishing (see e.g. Simmons, Nelson, & Simonsohn, 2011, http://pss.sagepub.com/content/early/2011/10/17/0956797611417632). You should test your data to see if there is multicollinearity, and if so, consider doing something to correct it, such as removing redundant variables, orthogonalizing variables, or doing PCA or factor analysis (something that does dimensionality reduction) to get uncorrelated 'abstract' variables out of your dataset. See e.g. Baayen et al 2006 [http://www.sfs.uni-tuebingen.de/~hbaayen/publications/BaayenFeldmanSchreuderJML2006.pdf] for discussion of these three options for reducing multicollinearity. (I remember that recently there has been some debate over whether factor analysis for this purpose is a good idea, but I can't remember the citation right now.)
Again, thank you very much for your help. I will follow your advices.
Best regards Stephen !
Good advice from Stephen.
Other options is that trying variable selection using the Lasso if you have a lot of features (variables) and no prior information is available.
If you are worrying about the multicollinearity, use the elastic net instead of the Lasso. It can take care of this problem. If you want to read more about it, I recommend to go to the book "An introduction to statistical learning with application in R" which is freely available online. The authors also have online course on Coursera which is very nice.
Hope it help.
P/s: I also used the Lasso, the elastic net for variable selection in my problem. You can work well in R. Many online document about this is available.
Many thanks to both of you for all these information, it helps me a lot.
- Badges
- Science topic