Article first published online: 27 FEB 2013
© 2013 BJU International
Volume 111, Issue 3, pages E20–E21, March 2013
How to Cite
Bedke, J., Chun, F. K.-H., Merseburger, A. S., Scharpf, M., Kasprzyk, K., Schilling, D., Sievert, K.-D., Stenzl, A. and Kruck, S. (2013), Reply. BJU International, 111: E20–E21. doi: 10.1111/bju.12020_13
- Issue published online: 27 FEB 2013
- Article first published online: 27 FEB 2013
We thank McMillan for sharing his thoughts on our work, because it gives us the opportunity to clarify his concerns. As indicated in his comment, the prognostic value of C-reactive protein (CRP) in RCC has been widely explored by the use of different arbitrary thresholds. Unfortunately, these analyses were not based on valid statistical calculations to identify the best suited threshold. For example, receiver operator curves have often been used with cancer-specific survival (CSS) as the endpoint [1, 2]. In fact, this approach neglects the existence of different survival times between individual patients; the endpoint of CSS is a time-dependent variable. To address this methodological shortcoming, we used Cox proportional hazard models to stratify patients according to different CRP thresholds; therefore, our analyses clearly account for the effect of time. Within our study, a CRP value of 0.25mg/dL was found to have the highest multivariable hazard ratio (2.8) and was consequently chosen for further analyses.
Prognostic models that include CRP, such as the Glasgow prognostic scale (GPS), use an arbitrary threshold of 0.1mg/dL, which is obviously not derived from valid statistical analyses . Nevertheless, GPS analysis in RCC proved CRP to be of independent prognostic value in multivariable analysis. Predictive accuracy using Harrell's concordance (c)-index, which considers the effect of time in CSS, has however not been elucidated . We therefore determined the prognostic accuracy of a simple model including the well-known prognostic variables of T-, N- and M-stage, nuclear grading and Karnofsky performance status, with and without the addition of our obtained threshold of 0.25. In this basic model, CRP did not further improve predictive accuracy quantified by Harrell's c-index . Furthermore, accuracy estimates of our model were internally validated using 200 patient resamples.
McMillan hypothesizes that Karnofsky performance status may have biased our findings. To control for the potential subjective effect of Karnofsky performance status, we recalculated our model without Karnofsky performance status using T-, N- and M-stage and nuclear grading with and without CRP (threshold of 0.25) as variables. Interestingly, the c-index remained unchanged when CRP as a variable was considered (c-index 0.867 vs. 0.867).
Taken together, our internally validated analyses clearly demonstrate that CRP represents an independent variable for CSS in patients with RCC; however, in the setting of prognostic models, discriminative ability (i.e. statistical significance) does not represent the ‘gold standard’. More importantly, potential variables must be assessed regarding their informative impact within multivariable models. In fact, only those variables that exhibit both discriminative and informative ability (i.e. accuracy or concordance when time-dependent endpoints are investigated) are helpful. Futhermore, we are able to discard McMillan's hypothesis as Karnofsky performance status does not represent a potential confounder of our results.
- 1Development and external validation of a new outcome prediction model for patients with clear cell renal cell carcinoma treated with nephrectomy based on preoperative serum C-reactive protein and TNM classification: the TNM-C score. J Urol 2009; 181: 1004–1012; discussion 12, , et al.