Difference needed for power in fixed population.

#1
Hi,


I am designing a study to examine the difference made by implementing a new treatment strategy for a medical condition. I have become a little stuck on some of the statistics.

I want to determine if the new strategy is effective.

I can assume that the population number is fixed as all patients that underwent the old strategy last year will be switched to the new one this year.

I have a figure for the total number of complications for last year and want to know how many fewer complications must occur this year for the difference to be statistically significant. (p<0.05).

Does anyone know what test I should use or how I calculate this?


Any help would be much appreciated.

Thank you.
 
#2
Would this be ok?

I figured that I have the have the sample size for both groups, and the proportion for group 1 and so can simply use trial and error to find the proportion for group 2 that would result in a sig difference?

Does anyone think that would be an ok method?
 
#3
That would be ok, though you could also use the formula to get an exact number. A quick Google search found this:

http://www.jerrydallal.com/LHSP/sizenotes.htm (the formula is halfway)

By the way, your data is best modelled using a binomial distribution. This won't matter much if your n is large, but it might matter a little bit if your n is small.
 
#4
Oops, seems like I misread your post. Your groups are not independent, since every patient is measured twice. So you should use a paired test.
 
#6
It depends. What is your n (sample size) and p (proportion with complications) in the first year? If n is sufficiently large and p is not too close to the extremes (0 or 1), you should be fine using a normal approximation for the binomial test. Though you could also find out some info on the paired binomial test; that would give you an even more exact answer.

But that's all from a statistical point of view. Even though I know little about medicine, I would say the way you measure the risk of complications is not very informative. From your description, I'm assuming you have the followiong data: whether each patient experienced complications during the year (yes/no). Wouldn't it make more sense to incorporate the number/severity/type of the complications somehow? A yes/no distinction on such an important variable seems really limited.

But since you can't really measure anymore for the past year, that doesn't help you much. Although maybe you could incorporate more useful variables this time around, so that you can do a more informative follow-up next year.
 
Last edited:
#8
Thank you again Junes. I have been looking at the McNemar's test, however, I don't have any of the data for test 2 yet as it has not yet occurred and so I cannot see how I can complete the equation?