Results differ slightly from PMCMR about scikit-posthocs HOT 6 CLOSED

@maximtrp It is numerically better to use scipy distribution sf for upper tail pvalues
in cases like this, cdf should be replaced by sf
p_value = 2. * (1. - ss.t.cdf(np.abs(t_value), df = x_len_overall - x_len))
to
p_value = 2. * ss.t.sf(np.abs(t_value), df = x_len_overall - x_len)

1 - cdf is limited in values to floating point precision at 1, while sf in most distribution can compute much smaller tail probabilities.

from scikit-posthocs.

@dyballa
Thanks for giving it a try. What test are you using?

from scikit-posthocs.

@maximtrp With PMCMR, I'm using posthoc.kruskal.conover.test with p.adjust.method="none".

from scikit-posthocs.

@dyballa I've compared the original PMCMR algorithm for Conover's test and my port of it line by line. This is not related to posthocs package. The differences arise on calculating p-values, using SciPy's scipy.stats.t.cdf() function. Anyway, differences are seen only in extremely small values. Probably, it is related to rounding error. Needs further investigation.

from scikit-posthocs.

@maximtrp Interesting! Thanks, I appreciate your checking this.

from scikit-posthocs.

@josef-pkt Thank you!

from scikit-posthocs.