Linear Discriminant Analysis - Sample size requirements

#1
As general question, Are 70 observations sufficient to design a reliable decision rule for discriminating between two groups, with a Fisher linear discriminant function?
 
#2
depends how many variables you intend to use for the classification. if the number of variables is much higher than 70, you will most likely go into overfitting
 
#3
thank for your help.
I have 9 variables.

And if I reduce the number of observations til 30, it will still be reliable?
I would like to understand which is the threshold criteria, I should follow, in order to trust the prediction. As I understand if I use 30 or 70 observations with 9 variables in the model to distinguish between two groups,
If the Manova test highlight a significant difference (p < 0.01) between the groups in the classification function created. The error estimation by bootstrap and Leave-One-Out is more than ok. And an external sample test set gives good prediction. Could still someone argue that the sample size is not enough, under which criteria?
 
Last edited:
#4
30 doesn't sound like a sufficient sample size to me...

be very careful of the leave one out and other validation methods, they are sensitive to overfitting. I would split the file into a training set and testing set and make a "blind test".