If I hypothesize that a gene signature will identify subjects at a lower risk of recurrence, that is decrease by 0.5 (hazard ratio of 0.5) the event rate in 20% of the population and I intend to use samples from a retrospective cohort study does the sample size need to be adjusted for unequal numbers in the two hypothesised groups?

For example using Collett, D: Modelling Survival Data in Medical Research, Second Edition - 2nd Edition 2003. The required total number of events, d, can be found using,

\begin{equation} d = \frac{(Z_{\alpha/2} + Z_{\beta/2})^2}{p_1 p_2 (\theta R)^2} \end{equation}

where $Z_{\alpha/2}$ and $Z_{\beta/2}$ are the upper $\alpha/2$ and upper $\beta/2$ points, respectively, of the standard normal distribution.

For the particular values,

  • $p_1 = 0.20$
  • $p_2 = 1 - p_1$
  • $\theta R = -0.693$
  • $\alpha = 0.05$ and so $Z_{0.025}= 1.96$
  • $\beta = 0.10$ and so $Z_{0.05} = 1.28$,

and taking $\theta R = \log \psi R = \log 0.50 = -0.693$, the number of events required (rounded up) to have a 90% chance of detecting a hazard ratio of 0.50 to be significant at the two-sided 5% level is then given by

\begin{equation} d = \frac{(1.96 + 1.28)^2}{0.20 \times 0.80\times (\log 0.5)^2}= 137 \end{equation}

  • $\begingroup$ I hope you don't mind, but I converted your question into latex. One thing, should $Z_{\alpha}$ not be $Z_{\alpha/2}$ $\endgroup$ – csgillespie Feb 22 '11 at 17:04
  • $\begingroup$ If it is clear for folks I don't mind at all. You are right should be a 2-sided alpha. $\endgroup$ – user712 Feb 22 '11 at 17:35
  • $\begingroup$ What are $\theta R$ and $\psi R$ ? Should they be $\theta_R$ and $\psi_R$? $\endgroup$ – onestop Feb 22 '11 at 19:02

Yes, your power will change based on the ratio of exposed to unexposed. For example, in a recent study I did the power calculations for, at an equal sample size, an Exposed:Unexposed ratio of 1:2 achieved power = 0.80 at a HR of ~1.3. It took until HR ~1.6 or so for a ratio of 1:10.

In your case, since the sample size will vary but your HR won't, the smaller the ratio, the larger your sample size will need to be.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy