Evidence-Based Pathology: Interobserver Variability Can Result in Significantly Different Estimates for Prognostic Outcome in Clinicopathologic Entities Such as UIP or NSIP
AM Marchevsky, R Gupta. Cedars Sinai Medical Center, LA, CA
Background: Clinicopathologic entities are based on findings associated with significantly different prognoses. The effect of interobserver variability (IOV) on the prognostic estimates associated with overlapping entities like UIP and NSIP is unknown.
Design: UIP and NSIP studies providing prognostic estimates were systematically reviewed. A simulation was performed keeping the number of surviving patients in each cohort as a constant. The proportion of UIP cases that could have been classified as NSIP by putative other pathologists was arbitrarily changed at 10-30% IOV intervals, by study. The variable simulated survival proportion was analyzed with kappa and chi-square statistics to determine at what level of IOV and kappa value the different diagnoses by various pathologists would have significantly affected the prognostic estimates reported by each study.
Results: Seven level III studies reporting 518 UIP and 189 NSIP patients were analyzed. Two of these studies reported IOV resulting in kappa=0.59-0.42. Overall survival of UIP and NSIP patients ranged between studies from 11% to 58% and 39% to 100% respectively. Simulation showed that IOV at the 20% level and kappa=0.68 would have significantly changed the survival estimates of all studies at the p<0.02 level.
Conclusions: Prognostic estimates are dependent on IOV, effect size and sample size.IOV at agreement levels that would be classified as moderate to substantial by kappa statistics can significantly influence prognostic estimates. Definitions of overlapping entities such as UIP and NSIP need to be made more stringent to decrease IOV to lower levels in order to provide more consistent prognostic estimates.
Monday, March 9, 2009 1:00 PM
Poster Session II # 227, Monday Afternoon