Intended for healthcare professionals

Research

Value of sentinel node status as a prognostic factor in melanoma: prospective observational study

BMJ 2006; 332 doi: https://doi.org/10.1136/bmj.38849.680509.AE (Published 15 June 2006) Cite this as: BMJ 2006;332:1423

Kaplan-Meier plots paint an unnecessarily gloomy picture for patients

The article by Kettlewell and colleagues demonstrates two potentially

misleading aspects of Kaplan-Meier plots. Take, for example, the 144
patients with
tumor thickness of 1-1.9mm and a negative sentinel node biopsy. The
Kaplan-Meier curve estimates that
20% of patients in this subgroup had died by 10 years. However, table 1
shows
us that
only 2 had died of melanoma by the end of follow-up, giving a crude
overall
mortality of only 1.4%. How can there be such an enourmous
discrepancy?

Firstly, counting the ticks on the Kaplan-Meier plots shows that they
include
not only deaths due to melanoma, but also deaths from other causes, which
doubles the number of deaths from 2 to 4 in the subgroup in question.

Secondly, selection bias may be operating through differential loss
to follow-
up. In a Kaplan-Meier plot, the size of the step representing each
death is proportional to the number of survivors remaining in follow-up at

that
instant. For the last two deaths in the aforementioned subgroup, the size
of
the
steps is about 7-8%,
indicating that each death represented around 7-8% of the total survivors
remaining in follow-up at that time; only around 15 patients out of the
original 144. There has therefore
been around 90% loss to follow-up by around 6 years - presumably because
most patients had been discharged from follow-up. It seems likely that
the
15 or so
patients
who remained in follow-up at 6 years were at greater risk of dying than
those
who did not continue in follow-up. In particular, any patients with
a
recurrence of melanoma would be much more likely to remain in follow-up
than those who remained free of disease. However, a crucial
assumption of the Kaplan-Meier function is that any loss to follow-up is
non-
differential with respect to outcome. The effect of this selection bias
is that
the
Kaplan-Meier curves will overestimate mortality, and this bias will
increase
with longer follow-up times.

These considerations are particularly important in a journal that is
freely
available online to melanoma patients and their families. Many will have
looked at the right-hand tails of the Kaplan-Meier plots to try to
estimate
their long-term mortality. If my analysis is correct, these all of these
plots
(not just the one in my example) paint an
unrealistically gloomy picture of long-term prognosis and will cause a
great
deal of unnecessary
anxiety for patients.

Competing interests:
I had a 1.3mm melanoma excised
4 years ago with a negative
sentinel node and would like an
unbiased estimate of my long-
term survival.

Competing interests: No competing interests

18 June 2006
James H MacCabe
MRC Fellow in Health of the Public Research
Section of Epidemiology, Division of Psychological Medicine, Institute of Psychiatry, London SE5 8AF