lucymcgowan/survivalpwr

survivalpwr

The goal of survivalpwr is to calculate power / sample size for survival models.

Installation

You can install the development version of survivalpwr with:

devtools::install_github("LucyMcGowan/survivalpwr")

Example

You can input parameters to compute power, for example here is a study with a hazard ratio of 1.5, an event probability of 0.8, and 200 participants:

library(survivalpwr)
pwr_coxph(
  hr = 1.5,
  eventprob = 0.8,
  n = 200)
#> 
#>      Cox Regression power calculation 
#> 
#>               n = 200
#>         nevents = 160
#>              hr = 1.5
#>       eventprob = 0.8
#>         rsquare = 0
#>          stddev = 0.5
#>       sig_level = 0.05
#>           power = 0.7272216
#>     alternative = two.sided

This indicates that this study is 72.7% powered to detect a hazard ratio of 1.5 with a two-sided test with a significance level of 0.05.

Or alternatively, you can enter a target power to determine the sample size / number of events needed to obtain that power:

pwr_coxph(
  hr = 1.5,
  eventprob = 0.8,
  power = 0.8
)
#> 
#>      Cox Regression power calculation 
#> 
#>               n = 238.7095
#>         nevents = 190.9676
#>              hr = 1.5
#>       eventprob = 0.8
#>         rsquare = 0
#>          stddev = 0.5
#>       sig_level = 0.05
#>           power = 0.8
#>     alternative = two.sided

This indicates that you would need 239 observations with an event probability of 0.8 (or alternatively 191 events total) to be 80% powered to detect a hazard ratio of 1.5 with a two-sided test with a significance level of 0.05.

This function also has the ability to incorporate r-squared if you Cox regression model is expected to have covariates that are explain some variation in the predictor of interest. For example, if you want to adjust for covariates that explain 15% of the variation in the predictor of interest:

pwr_coxph(
  hr = 1.5,
  eventprob = 0.8,
  power = 0.8,
  rsquare = 0.15
)
#> 
#>      Cox Regression power calculation 
#> 
#>               n = 280.8347
#>         nevents = 224.6677
#>              hr = 1.5
#>       eventprob = 0.8
#>         rsquare = 0.15
#>          stddev = 0.5
#>       sig_level = 0.05
#>           power = 0.8
#>     alternative = two.sided

You would now need 281 observations with an event probability of 0.8 (or 225 events).