i'm currently planning to write my bachelor thesis but it's been a while since my last statistics seminar, i'm extremely rusty and so any guidance here would be appreciated.
I have a relatively small sample of n=40, with 20 participants each with a learning disability and 20 without as control. Each participant completes three different tasks from which a multitude of datapoints is generated.
A quick example of how my data looks like:
| Participant | Disease | Task | Var1 | Var2 | Var3 | ... | Var38 |
|---|---|---|---|---|---|---|---|
| 1 | 0 | 1 | .67555 | .5353 | .4456 | .2345 | .3812 |
| 1 | 0 | 2 | .5514 | .7753 | .2239 | .3567 | .2234 |
| 1 | 0 | 3 | .4563 | .6675 | .3345 | .5681 | .5234 |
| ... | |||||||
| 40 | 1 | 2 | .8898 | .7798 | .7887 | .9989 | .5662 |
| 40 | 1 | 3 | .7854 | .8334 | .7687 | .7878 | .6534 |
My plan was/is to build a logistic regression model with the dichotomous DV (disease/non-disease) for each task, with the question if one of them can separate the DV better than the other and which of the generated datapoints are the best predictors in this task. After brief research one should predict the DV better, as its the field standard for measuring it, but I am "methodically insecure" how to go about it.
My first question would be:
- Is this approach feasible and is it possible to compare the tasks directly (e. g. in indices of fit like AUC, RMSE and r2)?
And secondly, as Im using an analysis technique which produces a lot of possible predictors (38), which im trying to reduce to avoid overfitting. Ive heard good things about LASSO, but am stuck using SPSS for licensing reasons. As univariate and stepwise techniques seem to be advised against, which would be an advisable method to reduce them to a more manageable level.
My apologies for the broad questions and sparse understanding, its been a long time in which i have endeavored to forget everything ive learnt.
