[R] Computing Confidence Intervals for AUC in ROCR Package

[David Winsemius]

On Jan 22, 2010, at 12:01 PM, Na'im R. Tyson wrote:

> Thank you for your very prompt response.  The authors of the ROCR  
> package informed me the package works as stated in the documentation  
> as long as you use R version 2.9.0--and indeed, it does!  I do not  
> mind using a slightly older version of R to get the results I need.
>
> It is useful to have the 'spread.estimate' feature for plotting, but  
> I wanted a numerical confidence interval.  In short, I am comparing  
> a few binary classifiers, and I want to show that the confidence  
> intervals for the average AUC overlap.  You can see this graphically  
> with the 'spread.estimate' option, but my dissertation committee  
> prefers numbers.

Ouch. I'm pretty sure you can find discussion on R-help offered by Dr  
Harrell regarding why overlap of CI's for ROC curves would have low  
power to detect important differences. And there is also the fact that  
overlap of 95% CI's is not an appropriate test even in the more  
straightforward situation of comparing group means.

-- 
David.


>
> Again, thank you for all of your help.  It has led me in the right  
> direction.
>
> Regards,
>
> Na'im
>
> On Jan 22, 2010, at 8:51 AM, David Winsemius wrote:
>
>>
>> On Jan 22, 2010, at 8:31 AM, Frank E Harrell Jr wrote:
>>
>>> David Winsemius wrote:
>>>> On Jan 22, 2010, at 3:53 AM, Na'im R. Tyson wrote:
>>>>> Dear R-philes,
>>>>>
>>>>> I am plotting ROC curves for several cross-validation runs of a  
>>>>> classifier (using the function below).  In addition to the  
>>>>> average AUC, I am interested in obtaining a confidence interval  
>>>>> for the average AUC.  Is there a straightforward way to do this  
>>>>> via the ROCR package?
>>>> You should probably contact the authors. When I tried using that  
>>>> package a few weeks ago, several of the annotation features were  
>>>> broken. I contacted the author who said there had been problems  
>>>> after converting to S4 method. He also said there would be a fix  
>>>> but not immediately. There has been a release since that time and  
>>>> I tried it,  but it did not appear to fix the problems I  
>>>> encountered. All I was able to get were very simple ROC curves  
>>>> without any confidence intervals or marking of levels. I ended up  
>>>> turning to the Epi package for what I needed ( but I did not need  
>>>> confidence intervals so cannot comment on that aspect.)
>>>
>>> I'm wondering what was broken with the S3 implementation that made  
>>> them change to S4.
>>
>> I was typing from memory and may not have conveyed accurately what  
>> was in the message. He mention changing versions but my attribution  
>> of that problem as switching from S3 to S4 methods seems to have  
>> been a manufactured memory. Furthermore, on loading the package in  
>> its current form, I am no longer having the problems I earlier  
>> experienced.
>>
>> So now my question to Tyson would be, what you were hoping to see  
>> with your request for confidence intervals? The "spread estimate"  
>> feature seems to have been fixed in version 1.0-4.
>>
>>>
>>> Frank
>>>
>>> -- 
>>> Frank E Harrell Jr   Professor and Chairman        School of  
>>> Medicine
>>>                   Department of Biostatistics   Vanderbilt  
>>> University
>>
>> David Winsemius, MD
>> Heritage Laboratories
>> West Hartford, CT
>>
>

David Winsemius, MD
Heritage Laboratories
West Hartford, CT