Giuseppe Cardillo

Dear Hannes, I took care of Michael's comment: in fact in the previous version of the function there was a bug on true positives and false positives computation that I fixed. Moreover, I erased the convention that higher values of a test are abnormal. The only convention is: 0 for the group that has not the tested character (i.e. healthy subjects); 1 for the group that has the teste character (i.e. subjects with a disease). I think everybody uses this convention.

Looking to the code:

bar0=median(x(x(:,2)==0),1); %median value of group 0
bar1=median(x(x(:,2)==1),1); %median value of group 1

for K=1:ll %For each unique value
    if bar0<=bar1 %when the values of group 0 (controls) are generally lower than values of group 1 (patients)...
        TN=length(x(x(:,2)==0 & x(:,1)<=labels(K))); %True Negative are the number of subjects of group 0 with a test value <= current unique value
        FP=length(x(x(:,2)==0 & x(:,1)>labels(K))); %False Positive are the number of subjects of group 0 with a test value > current unique value
        TP=length(x(x(:,2)==1 & x(:,1)>labels(K))); %True Positive are the number of subjects of group 1 with a test value > current unique value
        FN=length(x(x(:,2)==1 & x(:,1)<=labels(K))); %False Negative are the number of subjects of group 1 with a test value <= current unique value
    else %when the values of group 0 (controls) are generally higher than values of group 1 (patients)...
        TN=length(x(x(:,2)==0 & x(:,1)>=labels(K))); %True Negative are the number of subjects of group 0 with a test value >= current unique value
        FP=length(x(x(:,2)==0 & x(:,1)<labels(K))); %False Positive are the number of subjects of group 0 with a test value < current unique value
        TP=length(x(x(:,2)==1 & x(:,1)<labels(K))); %True Positive are the number of subjects of group 1 with a test value < current unique value
        FN=length(x(x(:,2)==1 & x(:,1)>=labels(K))); %False Negative are the number of subjects of group 1 with a test value <= current unique value
    end
    a(K,:)=[TP/(TP+FN) TN/(TN+FP)]; %Sensitivity and Specificity
end

In the first versions, I didn't take in account that group 0 could have higher values than group 1: when this condition occurred, the ROC curve was below the bisector and it was mirrored.
In the next versions, this condition is taken in account and so, I think that Sensitivity and Specificity are correctly computed.

When bar0<=bar1, the ROC points flow from right to left (or, in other words, from the point [1,1] to the point [0,0]). If you use the trapz function the result is wrong because to use trapz x data must be monotonically increasing.
In this case, I flip up-down xroc and yroc (and the labels, of course) to respect the condition to use trapz and the data are not warped. I think that is indifferent for the user to know this information.

Finally, if the bisector is the result of a random classifier, I think that there is not a classifier worse than chance and the AROC is always >=0.5

At the moment, there is not on FEX the new uploaded file (this takes 1-2 days). If you contact me in private, I'll send you the file as an attach

Hi, the code is very good. However, I encounter an error where the cut-off point is set, on line 186,

??? Attempted to access labels(7397); index out of bounds because numel(labels)=7396.

Error in ==> roc at 186
        co=labels(J); %Set the cut-off point

Is there a way to fix this?
Kind regards!

Thank you for your comment. If you type on matlab "help logrank" you'll find the reference.