Prognostic value of continuous variables in breast cancer and head and neck cancer. Dependence on the cut-off level

Statistical aspects of prognostic factor studies in oncology

“…Altman et al (1994) showed that for this procedure the falsepositive error rate when the marker is not prognostic is close to 40% rather than the nominal 5%. This was also illustrated by Courdi et al (1988) using data for hormonal receptors and labelling index for patients with primary breast cancer. Altman (1992) showed how the statistical significance of the log-rank test comparing survival curves should be adjusted for the use of 'data-derived' optimal cut-off point.…”

Tumour growth rates in squamous carcinoma of the head and neck measured by in vivo bromodeoxyuridine incorporation and flow cytometry

“…Cell kinetic studies in man, until recently, have been limited to in vitro investigations because of the ethical constraints of administering radioactive DNA precursors to patients. Tritiated thymidine labelling of freshly excised tissue followed by autoradiography has been widely used (Gentili et al, 1981;Brandt & Olsson, 1987;Courdi et al, 1989;Meyer & Bauer, 1975). The development of flow cytometric methods for measuring DNA content, particularly in paraffin embedded archival material, has resulted in DNA ploidy and S-phase fraction analysis being performed for many tumour types (Cornelisse et al, 1987;Armitage, 1985;Feichter et al, 1987 …”

Categorising continuous variables

“…A similar approach has been used by other authors (e.g. Courdi et al, 1988;Clark et al, 1989;Tandon et al, 1989;Coiffier et al, 1991), but this does not make the procedure valid.…”

“…However, there is no recognised procedure for making this adjustment, and there is no corresponding adjustment to the estimated difference in prognosis between the two groups. Courdi et al (1988) and Sigurdsson et al (1990) observed that different authors have obtained different 'optimal' cut-off points, but did not seem to recognise the role of sampling variation in this context. We should expect different authors to find considerably different 'optimal' cutoffs for the same measurement in different samples from the same population.…”

“…Some of these points were addressed by Courdi et al (1988), in particular the need to adjust the P value for multiple testing. However, there is no recognised procedure for making this adjustment, and there is no corresponding adjustment to the estimated difference in prognosis between the two groups.…”