[R] General help on sample size based on recurrent events
therneau at mayo.edu
Wed Mar 18 12:25:49 CET 2009
--- begin included message -----
I would like to design a study (a two group comparison) based on reduction on
events(say hospital admissions). In a previous study hospital admission rate of
140 admissions per 72 patients (over a 4 month period) has been observed. That
is rate is about 1.9. In order to see the admission rate reduction of 50% over a
4 month period, i.e. 0.95, in a two group comparison (alpha 0.05) with
appropriate power (say 80%) what sort of group sizes are needed.
How would I go about this?
------ end inclusion --------------------
If you fit a model that has all of the relevant covariates describing the
baseline risk of each subject, then in theory the sample size depends only on
the total number of event. That is big IF -- almost certainly untrue.
For independent events, 2 sided alpha =.05, power=.8, you need 191 independent
events for sample size. I have found that a recurrent event is usually worth
1/4 to 1/3 of a "new" event, in terms of power, for a range of studies. This
value is hard to predict; I have one study where the gain was essentially 0.
My ususal strategy is to plan the study size using only the first events, and
then use all events in the analysis. The hard part, as in any survival or event
study, is figuring out how many to enroll so as to expect to see 191 of the
subjects have an event.
More information about the R-help