Comparison of two survival curves can be done using a statistical hypothesis test called the log rank test. It is used to test the null hypothesis that there is no difference between the population survival curves (i.e. the probability of an event occurring at any time point is the same for each population). This function use the Kaplan-Meier procedure to estimate the survival function, so it is mandatory to download
KMPLOT (http://www.mathworks.com/matlabcentral/fileexchange/22293).
Thank you Jan.
1) You are right. There is a bug in kmplot not in logrank. Now I have just upload the new version of kmplot on FEX. If you need, write me in private and I'll send you the file.
2) they are both correct! I explain. If you are doing a test of a drug against placebo, you will use a 1-tailed test (because, before to do the experiments you are thinking that drug is better of placebo). But if you are testing two different drugs or two different protocols you can only ask to logrank test: Are they different? And so you have to use a 2-tailed test. Of course, 1-tailed test are more powerful but the conditions to apply them are not always respected. Anyway, I'll clarify this in the logrank help and output
Very useful, but I had two questions. (1) LogRank gives an error at line 116, at [table1 table12 t1 T1 xcg1 ycg1 lambda1]=kmplot(x1,0.05,cflag,0);
This is fixed by deleting the last arg to kmplot, giving kmplot(x1,0.05,flag);
2) The p-value for LogRank is twice what it should be? 0.03058 vs. 0.01529 in the file header. Perhaps, mismatch between old documentation and changed function, or extra factor of 2 in p = 2*(1-0.5*erfc(-z/realsqrt(2))); %p-value
Removing Yates' gives 0.019250, which is correct. I tend to think 0.03058 is correct, with the Yates' correction. Thanks for the super-useful code!
The logrank calculation here is correct. This code applies the Yates correction in the calculation of the z-score. (Line 207 of my version of the code.) I get 0.0193 from this code when I remove the Yates correction.
I'm sorry, but using your software I do not get the p-value for the log-rank test that is consistent with the results of either R or Stata (both of which match). For the test data supplied with the function, I get a p-value of 0.0193, whereas you're getting a p-value of 0.01529. The KM curves are the same, but there is apparently something wrong in your logrank computation.
Yes it is. The 0.05 put in that lines is a dummy variable. Infact, in that lines logrank invokes kmplot to compute needed tables (i didnt want to duplicate code). If you look inside kmplot it is commented that a piece of code id jumped whan invoked by logrank and in that piece of code alpha is used to compute the confidence interval of survival curve (that is useless in logrank).
I notice that in the call to kmplot [lines 96-97] in current version, the hard-coded value 0.05 is used, even if alpha is specified differently as an input. Is that correct?
Dear Sven, thank you for your email.
T1 and T2 are cleared because they don't appear in the output. As everybody can see in the figure (without downloading the function) logrank always prints all needed output parameters. Logrank recalls another function that is kmplot: this function can give back T1 (if you want). Anyway this possibility is hidden because usually you are only interested to the plot.
Excellent submission - thank you. One note: the outputs to the function [T1,T2] are actually cleared during computation, so an error is thrown if output is requested. To conform a little bit to the MATLAB stats package, I suggest the following changes:
Line one should change to:
function [h,p] = logrank(varargin)
and the end of the function should set h as follows:
h = p<alpha;
The help section could then say:
% Displays:
% Kaplan-Meier plot
% Log-rank statistics
% Outputs:
% H : statistical significance (true if P<ALPHA)
% P : P-value for log-rank test
Thanks for the really quick reply. I appreciate it.
However, the program still crashes for me, with the same error, when I try the syntax you suggest. Maybe the version on the File Exchange is not quite the same as yours?
You are partially right. If you use only the first column, the routine crashes because the informations in the second column are mandatory. If you want no censored data:
>> x1(:,2)=0; x2(:,2)=0; logrank(x1,x2);
Anyway, a bug was present but I fixed it and upload the new version.
It seems to me that this function does not work for the case when none of the data are censored. (That shouldn't be a problem, should it?) For example, using only the first columns of your sample x1 and x2: