SWLDA

[ysumuyi]

Hi,
I have finished the experiment of P300 Speller. But I still have a confusion to ask.
I Know the BCI2000 uses stepwise linear discriminant analysis(SWLDA) to process the feature vectors. I have read some papers about this algorithm, but these papers all shows the step of enter/remove parameters like this:
When F is larger than Fa (F > Fa), the F will be added to the discriminant function; when F is less than Fa (F < Fa), the F will be removed from the discriminant funtion.
In BCI 2000,now the step is:
the most statistically significant input feature for predicting the target label (having a p-value < 0.1) is added to the discriminant function. After each new entry to the discriminant function, a backward stepwise analysis is performed to remove the least significant input features, having p-values > 0.15.
As it showed above, the two conclusions are contradictory. I don't know how to explain.
Looking forward your answer. Thanks very much!
From: Eddy

[boulay]

First, F can be directly transformed to p, so F > Fa is the same as p < pa.

I think what you're asking is how can a feature be removed from the model for failing to meet some criterion, if it is only included in the model in the first place after meeting a stricter criterion. The answer is that the F or p-value associated with a given feature will change whenever the model is changed.

See the section titled "Algorithms" on MATLAB's stepwisefit page. http://www.mathworks.com/help/toolbox/s ... sefit.html