When comparing 2 proportions, I've seen 2 ways the standard error can be calculated.

The first I've seen is: $$ SE = \sqrt{SE_1^2 + SE_2^2} $$

The second one is: $$ SE = \sqrt{\hat{p}(1-\hat{p})(\frac{1}{n_1} + \frac{1}{n_2})} $$

Under what circumstances would I use the first one vs the second one?

What I'm trying to accomplish is calculate the z-score and confidence that $p_2$ is greater than $p_1$.

Edit, here are examples of sites that seem to talk about the same thing but use different SE formulas:

http://www.kean.edu/~fosborne/bstat/06d2pop.html https://onlinecourses.science.psu.edu/stat100/node/57

  • $\begingroup$ Why don't you just do it the Bayesian way and model each probability as, e.g., a Beta distributed belief, and then calculation the probability that one belief is greater than the other? $\endgroup$
    – Neil G
    Feb 15, 2018 at 13:20
  • $\begingroup$ Could you explain what are the variables in your equations? $\endgroup$ Feb 15, 2018 at 13:48
  • $\begingroup$ My p_1 and p_2 are just proportions of people who did something. n_1 and n_2 are the numbers of observations. So if x_1 and x_2 are the number of people who did something out of the n_1 and n_2 observations, respectively, p-hat is (x_1 + x_2)/(n_1 + n_2) $\endgroup$
    – tehc0w
    Feb 15, 2018 at 13:54
  • 1
    $\begingroup$ Since in this context it would seem that "$SE_i^2$" must refer to $\hat p(1-\hat p)/n_i$, the step from the first expression to the second involves the simplest possible algebraic manipulation. Presumably, then, the $SE_i^2$ refer to some other formula: but what is it? Is this question really trying to get at the difference between separate and pooled estimates of $p$? $\endgroup$
    – whuber
    Feb 15, 2018 at 16:28
  • $\begingroup$ I think this may be going in the right direction. If $p_1$ = $p_2$ = $\hat p$ then yes it reduces easily. Is that what is implied? If I'm using $\hat p_1$ and $n_1$ for $SE_1$ and $\hat p_2$ and $n_2$ for $SE_2$, is it possible to prove the 2 expressions are similar? $\endgroup$
    – tehc0w
    Feb 15, 2018 at 16:54

1 Answer 1


I found my answer. It comes down to pooling or unpooling.

You would use $$SE = \sqrt{SE_1^2 + SE_2^2}$$ if you believe the variances are not similar and $$SE = \sqrt{\hat{p}(1-\hat{p})(\frac{1}{n_1} + \frac{1}{n_2})}$$ if you believe variances are similar.

Some additional literature can be found here.

Thanks @whuber for the inspiration.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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