We have performed a survival analysis on about 330 cancer patients and
found that a known prognostic factor, with 3 levels, good, medium and poor
prognosis, is highly significant, with 160 events in total after about 10
yrs follow-up, p<0.0001.
The clinicians now want to look at whether the same factor determines how
quickly the patient has an event - ie for those who have an event, does the
poor prognosis group have shorter time to get the event?
IF you do this and restrict the analysis to just the 160 patients who had
the event and then perform a kaplan-meier analysis with log-rank test on
this factor then this factor is no longer significant, p=0.68.
The clinician's interpretation is that depending which level of the factor
you have determines whether or not you have an event, but does not
determine how quickly the event happens if it does (and so different
interpretation about the microbiology of the tumour rather than helping to
predict anything about survival time from diagnosis for these prognostic
groups)
I'm concerned, however, that we may have seen this effect solely because of
selection bias, i.e. we have selected a group of more homogeneous women
because they have all gone on to have an event and so less likely to see
the separation in survival curves and inequality because they all end up at
zero survival time.
Incidentally, if you repeat with a different known prognostic factor then
in the subset analysis this factor's significance is still there (p=0.01)
but much reduced from p<0.001
I would be really grateful for your thoughts on this - is the subset
analysis valid? - Would you query it if you were refereeing a paper which
had done this?
Many thanks
Catherine
**************************************************************
Catherine Thomson
Information & Statistics Manager and Honorary Research Fellow
Trent Cancer Registry
5 Old Fulwood Road
Sheffield, S10 3TG
Tel: 0114-226-3573 Fax: 0114-226-3561
Email: [log in to unmask] or [log in to unmask]
web: www.trentcancer.nhs.uk
|