0

I want to calculate the sample size for a survival study of patients with acute myeloid leukemia, with two cohorts, one receiving chemotherapy and the other not. I want an event-free survival (EFS) at 1 year of 40% with a margin of 5%. The difference between the two cohorts has to be 5-10%, with a power of 0.8 and alpha of 0.05.

I have considered a non-inferiority design, as the new treatment may not be as efficient but has lower toxicity, and I want to prove that the new treatment is not significantly worse than the reference treatment, which has already been tested. I want a maximum difference in efficiency of 10%. Does this make sense? Would it be better to do a superiority design? I have calculated the sample size for each cohort, which must be the same, using the epi.ssninfb function from the epiR package, but I have seen that there are many R packages (that take into account different types of distribution, constant or non-constant hazard ratios, among other assumptions) and many programs to perform the calculation, and I don't know which is the best method.

The hypotheses planned are:

Ho: prob. standard treatment - prob.new treatment >= 0.1

H1: prob. standard treatment - prob.new treatment < 0.1

And the R code:

install.packages("epiR")
library(epiR)

epi.ssninfb(treat = 0.4, control = 0.4, delta = 0.10, n = NA,  r = 1, 
            power = 0.8, nfractional = TRUE, alpha = 0.05)

$n.total
[1] 593.5255

$n.treat
[1] 296.7627

$n.control
[1] 296.7627

$delta
[1] 0.1

$power
[1] 0.8

Has the sample size been correctly determined? I have also found other R functions to perform the calculation, such as power.t.test (survival package), sample.size.NI (dani package), power.prop.test (stats package), nSurvival (gsDesign package, which requires accrual and follow-up times that I don't have), etc.

I'm very confused. Can anyone help me?

Thank you very much! :)

EI_Stats

EI_Stats
  • 21
  • 2
  • Multiple methods questions. More suitable for stats.stackexchange.com. I will say that the power analyses in survival analysis can be easily simplified to a binomial test that ignore the timing and only considers the proportion of subjects with events versus the number at risk. So make reasonable assumptions about accrual rate, censoring rates and event rates at an suitable interval and use a test for differences in proportions. (I don't think any AML types that have currently effective treatments can accurately be modeled as having constant hazard rates. Typically very high early mortality.) – IRTFM May 12 '23 at 17:41

0 Answers0