[R] How to calculate confidence interval of C statistic by rcorr.cens

Frank Harrell f.harrell at vanderbilt.edu
Sun May 22 21:22:37 CEST 2011


Hi Kohkichi,
What we really need to figure out is how to make validate give you
confidence intervals for Dxy or C while it is penalizing for overfitting. 
Some people have ad hoc solutions for that but nothing is nailed down yet.
Frank
 
khosoda wrote:
> 
> Thank you for your comment, Prof Harrell.
> 
> I changed the function;
> 
> CstatisticCI <- function(x)   # x is object of rcorr.cens.
>    {
>      se <- x["S.D."]/2
>      Low95 <- x["C Index"] - 1.96*se
>      Upper95 <- x["C Index"] + 1.96*se
> 
>      cbind(x["C Index"], Low95, Upper95)
>    }
> 
>  > CstatisticCI(MyModel.lrm.penalized.rcorr)
>                        Low95   Upper95
> C Index 0.8222785 0.7195828 0.9249742
> 
> I obtained wider CI than the previous incorrect one.
> Regarding your comments on overfitting, this is a sample used in model 
> development. However, I performed penalization by pentrace and lrm in 
> rms package. The CI above is CI of penalized model. Results of 
> validation of each model are followings;
> 
> First model
>  > validate(MyModel.lrm, bw=F, B=1000)
>            index.orig training    test optimism index.corrected    n
> Dxy           0.6385   0.6859  0.6198   0.0661          0.5724 1000
> R2            0.3745   0.4222  0.3388   0.0834          0.2912 1000
> Intercept     0.0000   0.0000 -0.1446   0.1446         -0.1446 1000
> Slope         1.0000   1.0000  0.8266   0.1734          0.8266 1000
> Emax          0.0000   0.0000  0.0688   0.0688          0.0688 1000
> D             0.2784   0.3248  0.2474   0.0774          0.2010 1000
> U            -0.0192  -0.0192  0.0200  -0.0392          0.0200 1000
> Q             0.2976   0.3440  0.2274   0.1166          0.1810 1000
> B             0.1265   0.1180  0.1346  -0.0167          0.1431 1000
> g             1.7010   2.0247  1.5763   0.4484          1.2526 1000
> gp            0.2414   0.2512  0.2287   0.0225          0.2189 1000
> 
> penalized model
>  > validate(MyModel.lrm.penalized, bw=F, B=1000)
>            index.orig training    test optimism index.corrected    n
> Dxy           0.6446   0.6898  0.6256   0.0642          0.5804 1000
> R2            0.3335   0.3691  0.3428   0.0264          0.3072 1000
> Intercept     0.0000   0.0000  0.0752  -0.0752          0.0752 1000
> Slope         1.0000   1.0000  1.0547  -0.0547          1.0547 1000
> Emax          0.0000   0.0000  0.0249   0.0249          0.0249 1000
> D             0.2718   0.2744  0.2507   0.0236          0.2481 1000
> U            -0.0192  -0.0192 -0.0027  -0.0165         -0.0027 1000
> Q             0.2910   0.2936  0.2534   0.0402          0.2508 1000
> B             0.1279   0.1192  0.1336  -0.0144          0.1423 1000
> g             1.3942   1.5259  1.5799  -0.0540          1.4482 1000
> gp            0.2141   0.2188  0.2298  -0.0110          0.2251 1000
> 
> Optimism of slope and intercept were improved from 0.1446 and 0.1734 to 
> -0.0752 and -0.0547, respectively. Emax was improved from 0.0688 to 
> 0.0249. Therefore, I thought overfitting was improved at least to some 
> extent. Well, I'm not sure whether this is enough improvement though.
> 
> --
> Kohkichi
> 
> (11/05/22 23:27), Frank Harrell wrote:
>> S.D. is the standard deviation (standard error) of Dxy.  It already
>> includes
>> the effective sample size in its computation so the sqrt(n) terms is not
>> needed.  The help file for rcorr.cens has an example where the confidence
>> interval for C is computed.  Note that you are making the strong
>> assumption
>> that there is no overfitting in the model or that you are evaluating C on
>> a
>> sample not used in model development.
>> Frank
>>
>>
>> Kohkichi wrote:
>>>
>>> Hi,
>>>
>>> I'm trying to calculate 95% confidence interval of C statistic of
>>> logistic regression model using rcorr.cens in rms package. I wrote a
>>> brief function for this purpose as the followings;
>>>
>>> CstatisticCI<- function(x)   # x is object of rcorr.cens.
>>>    {
>>>      se<- x["S.D."]/sqrt(x["n"])
>>>      Low95<- x["C Index"] - 1.96*se
>>>      Upper95<- x["C Index"] + 1.96*se
>>>      cbind(x["C Index"], Low95, Upper95)
>>>    }
>>>
>>> Then,
>>>
>>>> MyModel.lrm.rcorr<- rcorr.cens(x=predict(MyModel.lrm), S=df$outcome)
>>>> MyModel.lrm.rcorr
>>>         C Index            Dxy           S.D.              n
>>> missing     uncensored
>>>       0.8222785      0.6445570      0.1047916    104.0000000
>>> 0.0000000    104.0000000
>>> Relevant Pairs     Concordant      Uncertain
>>>    3950.0000000   3248.0000000      0.0000000
>>>
>>>> CstatisticCI(x5factor_final.lrm.pen.rcorr)
>>>                        Low95   Upper95
>>> C Index 0.8222785 0.8021382 0.8424188
>>>
>>> I'm not sure what "S.D." in object of rcorr.cens means. Is this standard
>>> deviation of "C Index" or standard deviation of "Dxy"?
>>> I thought it is standard deviation of "C Index". Therefore, I wrote the
>>> code above. Am I right?
>>>
>>> I would appreciate any help in advance.
>>>
>>> --
>>> Kohkichi Hosoda M.D.
>>>
>>>      Department of Neurosurgery,
>>>      Kobe University Graduate School of Medicine,
>>>
>>> ______________________________________________
>>> R-help at r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-help
>>> PLEASE do read the posting guide
>>> http://www.R-project.org/posting-guide.html
>>> and provide commented, minimal, self-contained, reproducible code.
>>>
>>
>>
>> -----
>> Frank Harrell
>> Department of Biostatistics, Vanderbilt University
>> --
>> View this message in context:
>> http://r.789695.n4.nabble.com/How-to-calculate-confidence-interval-of-C-statistic-by-rcorr-cens-tp3541709p3542163.html
>> Sent from the R help mailing list archive at Nabble.com.
>>
>> ______________________________________________
>> R-help at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-help
>> PLEASE do read the posting guide
>> http://www.R-project.org/posting-guide.html
>> and provide commented, minimal, self-contained, reproducible code.
> 
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
> http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
> 


-----
Frank Harrell
Department of Biostatistics, Vanderbilt University
--
View this message in context: http://r.789695.n4.nabble.com/How-to-calculate-confidence-interval-of-C-statistic-by-rcorr-cens-tp3541709p3542654.html
Sent from the R help mailing list archive at Nabble.com.



More information about the R-help mailing list