I have two groups (36 patients and 30 controls) and I want to check five or six normally distributed (as I checked using skewness and kurtosis values) variables (numeric) that discriminate between the two groups. Which is the best method to analyze my data? Should I perform a logistic regression or a discriminant analysis?
$\begingroup$
$\endgroup$
2
-
$\begingroup$ Are you trying to classify such that, say, new observations can be placed in either group in the future or are you trying to infer something about the group from your variables? So: Why have you collected the data? What's the goal with the discrimination/analysis? Are those variables all the variables you measured in the experiment? If not: How did you choose those 5 or 6 variables? With feature selection, expert knowledge and/or 'pick-the-ones-that-look-promising'? $\endgroup$– BeyerOct 13, 2016 at 8:02
-
$\begingroup$ I m trying to check whether there are some cognitive variables that distinguise between patients and controls and which one is more powerful in doing so. I have primarily chosen 6 variables according to a theoritical background but, after performing an Oneway ANOVA in order to check if the two froups differ with respect to these 6 variables, I found that the two groups differ on the five of the six variables. So I guess that I should better perform my analysis with the five of the six variables? Thank you. $\endgroup$– user134402Oct 13, 2016 at 9:59
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
Do a subject-wise cross validation using both discriminant analysis and logistic regression. Whichever is giving better performance (accuracy, sensitivity, specificity and AUC) in cross-validation, use that.
Refer to this link, for more information on why to use subject-wise cross validation https://stats.stackexchange.com/a/240033/86202
answered Oct 14, 2016 at 13:01
-
$\begingroup$ The 4 measures you listed are all improper accuracy scoring rules. Tomorrow I'll post an article about that at fharrell.com $\endgroup$ Feb 28, 2017 at 13:03