Comments
You can comment on the book in general on this page.
When you comment on a specific passage, please include chapter, topic and page number. If a topic increases in volume, a new page will be created.
Note: Registration is required in order to use the comment/discussion option
Discussion
chapter 15, table 15.4 95% CI for C statistic - in the example Program for testicular cancer analyses: t821 the line ”## NRI and IDI calculations with Harrell's functions improveProb(x1=as.numeric(NoLDH)-1, x2=as.numeric(YesLDH)-1, y=full$y)” did't work. how can I calculate the 95% CI for c statistic with a tree model? thank you
Calculations for the 95%CI of c are based on the rcorr.cens function in Harrell's Design library. The function gives the SE of Somer's D statistic; divide by 2, multiply by 1.96 for 95%CI.
chapter 15, table 15.4 how can i calculate 95%CI for C statistic in internal valdation. R code t821 internal validation, the line “cat(cstat[1], ”[”, cstat[1]-1.96/2*cstat[3], ” - ”, cstat[1]+1.96/2*cstat[3],”]”)” did not work. 'can not find the cstat object'. can you tell me how can i complete the R code for 95%CI for C statistic in internal valdation. cstat ← rcorr.cens(?, ?)
The issue of estimating a 95% CI after internal validation by bootstrapping is thorny. I discussed this with Frank Harrell several times. One simple approach is use the SE of the original c statistic for the optimism-corrected c statistic. Specific on the programming: The R program needs the object 'cstat' it seems; see help for the rcorr.cens function.
Question on adjusting case-control incidence back to population level
Op 2 nov. 2021, om 08:49 Niloufar Taherpour taherpour.n@gmail.com:
Dear Prof. Steyerberg, We developed a prediction model for assessing the predictors of surgical site infection after orthopaedic surgery in a shape of case-control study. As recommended in clinical prediction models book (second edition, section 3.1.7, page 41) it has been mentioned about prediction models with design of case-control study: “If a prediction model is developed, the average outcome incidence has to be adjusted for final calculation of probabilities, while the regression coefficients might be based on the case–control study ” , Would you please guide us how we should adjust for this. For example, when the incidence of a disease in a population is 10% and we have used a case-control study that 50% cases and 50% controls to model the factors that can predict the incidence of the disease, in the final model the prediction of the incidence will differ from the incidence in the population. My exact question is that whether we should run a weighted model according to the incidence of disease in the population or we should just correct the constant in the non-weighted model, and how we should do this.
Op 2 nov. 2021, om 09:52 Ewout Steyerberg e.steyerberg@erasmusmc.nl:
I think both options might work: weighting or intercept adjustment. I would prefer intercept adjustment; so if sampling was 50:50, and actual incidence 10:90, then adjust intercept with log(odds of sampling) = log(10/90 / (50/50)) = ln(0.11) = -2.20. Please verify that predictions are not around 50% anymore after correction, but around 10%? BW! Ewout