Ok, I am totally confused, this is probably a pure math rather than a statistic question but please enlighten me.

I found the logrank statistic formula being reported in the literature in 2 ways, but I am not sure which one is correct. When I plug the same number in both formula, i get different results for the chi-square.

Which one should I use?

$O$ = total of observed events

$E$ = total of expected events

  1. $$\frac{(O-E)^2}{E} $$

  2. $$\frac{O-E}{\sqrt{E}}$$

Which one should be used? The main reason I am asking is because when I apply the weights (W) as in the formula below, it looks like the weights are being cancelled out:

one-sample logrank statistic

  • $\begingroup$ That zero in #1 ('0') is meant to be an 'O' as in observed events, right? $\endgroup$ – Matt Krause Jul 25 '15 at 18:01
  • $\begingroup$ yes, sorry my bad, edited $\endgroup$ – Paris Char Jul 26 '15 at 17:16
  • $\begingroup$ I think both are correct. $\endgroup$ – Deep North Jul 27 '15 at 1:13
  • 4
    $\begingroup$ The first thing you mention is simply the square of the second. Neither is exactly a log rank statistic, since that would involve some kind of sum of terms. Compare your second proposal with the statistic here for example and note several differences. $\endgroup$ – Glen_b -Reinstate Monica Jul 27 '15 at 1:15
  • $\begingroup$ O and E are sums. Sorry my bad. $\endgroup$ – Paris Char Jul 27 '15 at 1:59

enter image description here

Just read this paragraph from the same page of your book and remember that square of a standard normal distribution is a chi-square distribution. Therefore the two tests are equivalent. You can use any of them.

  • $\begingroup$ I understand that. However I don't understand how the weights (W) don't cancel out in either of these 2 formulae. How can we apply weights when they cancel out in both numerator and denominator?Also another thing that confuses me (from the same page of the book): "a weight function is the weight W(t) = Y(t) which yields the one-sample log-rank test". So the one-sample logrank test is using Gehan weight by design? Am I missing something here? :S $\endgroup$ – Paris Char Jul 28 '15 at 13:22
  • $\begingroup$ Do you mean you try to cancel W for this ratio $\frac{\left \{ \sum_i^DW(t_i)\frac{d_i}{Y(t_i)}-\int_{0}^{r}W(s)b_0(s)ds \right \}^2}{\int_{0}^{r}W^2(s)\frac{b_0(s)}{y(s)}ds}$ ?W is a function or a random variable, I don't understand how you cancel them out? Now, i even don't know how to show it is a $\chi^2$ distribution. $\endgroup$ – Deep North Jul 28 '15 at 14:01
  • $\begingroup$ Doesn't the above formula simplify to W * (Σ(Observed) - Σ(Expected)) / (W * Var() ) ? $\endgroup$ – Paris Char Jul 28 '15 at 15:15
  • $\begingroup$ No, I think $\int_{0}^{r}W^2(s)\frac{b_0(s)}{Y(s)}ds$ itself is the variance when the Expected value $ E(W(s))$ is zero. Also, do you remember the definition of variance is $E(X^2)-[E(X)]^2$ and $E(X^2)=\int_{-\infty}^{\infty}x^2f(x)dx$ this is exactly the same when E(X)=0. I think you cannot factor out W here. It is not a constant, it is a random variable(function). $\endgroup$ – Deep North Jul 29 '15 at 0:01
  • $\begingroup$ Ok, so if I have observed deaths per year and Expected deaths per year, how do I calculate a weighted one-sample logrank test? I don't have the h0 so I cannot use the formula in the picture. All i know is that Var = sum of Expected. $\endgroup$ – Paris Char Jul 29 '15 at 1:11

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.