Paired t-test plot 95% CI shifted about dabest-python HOT 11 CLOSED

Thanks in advance,

Ana

from dabest-python.

Hi @AnaFVicente , thanks for flagging this up!

I think this is a case of the commutative property of means, vs. the inherent weirdness of medians?

Using t5 and t6 from above:

>>> np.mean(t5) - np.mean(t6)

-0.023344683345558614

>>> np.mean(t5 - t6)

-0.023344683345558614 # Same result as above.

but

>>> np.median(t5) - np.median(t6)

-0.22693218492666745

>>> np.median(t5 - t6)

0.0528625540482075 # Not the same result...

Right now, I think the best option is for us to remove the mean lines for Gardner-Altman paired median plots....

Again, thanks for bringing this to our attention.

from dabest-python.

Thanks for your response.
If I understood well, the median of the differences (black line) is calculated with a different method than the distribution of the differences (grey histogram). That's why there's a shift between both representations. Is there a way to calculate both parameters by using the same method: np.median(t5) - np.median(t6) or np.median(t5 - t6), so I get nice plots?
Ana

from dabest-python.

The problem isn't that they are calculated in a different way. It seems to be much deeper than that. The paired median difference of t5 and t6 is positive, even though the median of t6 is lower than the median of t5....

After thinking about it for a while, there might not be a good way to depict paired median difference with the Gardner-Altman estimation plot. You might have to use the Cumming estimation plot to do so.

Simply use

 two_groups_paired.median_diff.plot(float_contrast=False)

to plot the sampling error histogram below the paired slopegraph.

from dabest-python.

Thanks for your reply. However, if I plot the cumming estimation I still have the same problem: the histogram corresponding to 95% CI distibrution (grey histogram) is not aligned with the interval (black line)

from dabest-python.

from dabest-python.

Would you suggest just to remove 95% grey histogram, corresponding to CI distibrution ?

from dabest-python.

The error curve actually is aligned with the 95CI; bootstraps derived from medians often have non-normal distributions. If you find the error curve distracting, you could remove it in a vector graphics program, but I'd advise including it as it highlights:

• the non-normality of the median difference
• the graded nature of the confidence interval.

Hope this helps!

from dabest-python.

Also, we are looking into how to properly compute and display paired median differences, taking into account all we have discussed above. Thanks for flagging this up to us!

from dabest-python.

Thanks a lot for your help!

from dabest-python.

Just a last question. I don't understand how the curve can be correctly aligned if it represents the median differences distribution while the black line represents 95%CI. Most of the curve should be inside the black line, only 5% of the date could be outside. If I understand correctly. Thanks in advance.

from dabest-python.