Author reply

Abstract

Unfortunately, there are significant flaws in Mr. Promish's interpretation of our results and his Bayesian estimation procedure. First, the data contained in Table 5 of our article1 are observed (descriptive) not estimated (predicted) probabilities of lymph node metastases. To use Mr. Promish's monitory example, Table 5 shows that neither of the 2 women in our sample with ≤ 0.5-cm, nonpalpable tumors with lymphvascular invasion (LVI) had lymph node metastases at presentation. However, we should point out that whether the data contained in Table 5 are observed or estimated probabilities, the practicing clinician should never presume to be “absolutely certain” that a particular patient has any specific probability of lymph node metastases based on the results of published retrospective analyses. This is especially true for women with a relatively unusual or rare pattern of prognostic characteristics such as ≤ 0.5 cm, nonpalpable tumors with LVI. This simply is because one can never be absolutely certain that the prognostic characteristics or features of a particular patient are exactly equivalent to the characteristics or features of the sample on which the analyses were conducted. Furthermore, the practicing clinician also should keep in mind that any statistic has an associated confidence interval within which one reasonably can expect the corresponding population parameter to vary. The confidence interval on a proportion involving two patients is very wide. Mr. Promish also implies that his Bayesian estimation procedure may give a better representation of the risk of lymph node metastases than our logistic regression model. When evaluating the performance of any estimation procedure, the practicing clinician always should consider carefully the agreement between the actual, observed results and the results estimated or predicted by the estimation procedure. When disagreement is large, it is wise to question the validity of the procedure used to generate the estimates. One should be concerned about the validity of the Bayesian procedure used by Mr. Promish simply because it generated estimates that differ markedly from the actual probabilities reported in our Table 5. In contrast, our three variable logistic regression model-generated estimates that differ very little from the actual probabilities given in Table 5. Furthermore, unlike the Bayesian estimates, the monotone relation between tumor size and probability of lymph node metastases that has been observed by other authors2-5 clearly is evident in the estimates generated by our logistic regression procedure. We recently cross-validated our 3-variable model on a second, independent sample of 3108 women and found nearly identical probabilities of lymph node metastases for the risk groups shown in Table 5 in our article.1 Jeremy S. H. Jackson Ph.D.*, Ivo A. Olivotto M.D.*, Donna Mates M.A.*, * British Columbia Cancer Agency, Vancouver and Fraser Valley Clinics, Vancouver and Surrey, British Columbia, Canada