Statistical considerations when using a composite endpoint for comparing treatment groups