2DkcTXceO

N.E. Breslow 343TABLE 30.1Regression coefficients (± SEs) in multivariate nuclear morphometric discriminantfunctions fitted to three data sets † .ProspectiveCase-Control SampleSampleRisk NWTS + JHH NWTS Alone NWTSFactor (n = 108) ∗ (n = 95) (n = 218)Age (yr) .02 .013 ± .008 .017 ± .005SNRF 1.17 1.23 ± .52 −.02 ± .26LEFD 90.6 121.6 ± 48.4 .05 ± 47.5† From Breslow et al. (1999). Reproduced with permission. c○1999American Society of Clinical Oncology. All rights reserved.∗ From Partin et al. (1994)I was convinced that most of this apparent success was due to the failureto account for the multiplicity of comparisons inherent in the selection ofthe best 2 out of 242 measurements for the discriminant function. With goodcooperation from JHH, I designed a prospective study to validate the ability oftheir nuclear morphometric score to predict relapse in Wilms tumor (Breslowet al., 1999). I identified 218 NWTS-4 patients who had not been included inthe case-control study, each of whom had an archived slide showing a diagnosisby our pathologist of a Wilms tumor having the same “favorable” histologicsubtype as considered earlier. The slides were sent to the JHH investigators,who had no knowledge of the treatment outcomes, and were processed in thesame manner as for the earlier case-control study. We then contrasted resultsobtained by re-analysis of data for the 95 NWTS patients in the case-controlstudy, excluding 13 patients from JHH who also had figured in the earlierreport, with those obtained by analysis of data for the 218 patients in theprospective study.The results, reproduced in Table 30.1, were instructive. Regression coefficientsobtained using a Cox regression model fitted to data for the 95 NWTSpatients in the original study are shown in the third column. They were comparableto those reported by the JHH group based on logistic regression analysisof data for the 95 NWTS plus 13 JHH patients. These latter coefficients, shownin the second column of the table, were used to construct the nuclear morphometricscore. Results obtained using Cox regression fitted to the 218 patientsin the prospective study, of whom 21 had relapsed and one had died of toxicity,were very different. As I had anticipated, the only variable that wasstatistically significant was the known prognostic factor age. Coefficients forthe two nuclear morphometric variables were near zero. When the originalnuclear morphometric score was applied to the prospective data, using thesame cutoff value as in the original report, the sensitivity was reduced from