Why is the effective sample size calculated just as a sum of p_hats and not (for example) sum of absolute values of p_hats? Shouldn't we care more about the magnitude of (anit-)correlation? If, for example, the sequence of p_hats were 1, -1, 1, -1, ..., then ESS would say it's a very low autocorrelation, but this doesn't seem right.
I am also still wondering about that bit, but here's something I found which might be helpful: mc-stan.org/docs/2_25/reference-manual/effective-sample-size-section.html#estimation-of-effective-sample-size
thank you for the well presented intuition. Keep up the good work!!!
It is clear, thanks!
Why is the effective sample size calculated just as a sum of p_hats and not (for example) sum of absolute values of p_hats? Shouldn't we care more about the magnitude of (anit-)correlation? If, for example, the sequence of p_hats were 1, -1, 1, -1, ..., then ESS would say it's a very low autocorrelation, but this doesn't seem right.
I am also still wondering about that bit, but here's something I found which might be helpful: mc-stan.org/docs/2_25/reference-manual/effective-sample-size-section.html#estimation-of-effective-sample-size