Bugs in prior plotting - selected sigmoids

[dekuenstle]

The prior plotting function should add dots for the min, max, quantiles, and point estimate. However, they appear not at the right positions in the plot.

@guillermoaguilar Could you please add the screenshots?

[guillermoaguilar]

MATLAB priors:

Python priors:

Dots positions are different. Black is the point estimate, red and blue the quartiles. Green and yellow are the same, they are the min and max of the range of possible parameter values.

[HeikoSchuett]

The black dot in the prior seems to be still off in #154 . It should be at the prior mean but is clearly not. So the placement is not right yet.

[pberkes]

@HeikoSchuett the black dot is not the mean but the MAP estimate, it looks good from the marginals.

[pberkes]

@HeikoSchuett it's the same in Matlab: https://github.com/wichmann-lab/psignifit/blob/master/plotPrior.m#L97

[guillermoaguilar]

So actually plot_prior still needs correction after I talked to Felix last week (sorry I didnt post it before, I was busy with the docs)

He told me: the points in this plot_priors are supposed to show the range of likely and not likely sigmoids.

• The black should be the estimate, so you show where you estimate lands in the prior (showing it's a flat part of the prior and not somewhere else completely crazy that would bias the estimate). I think this is correctly implemented in python.
• the red and blue should be the 25% and 75% of the sigmoid, not the prior. The logic is to show that these two points land also in a reasonable flat part of the prior.
• the green and yellow are for the min and max of the prior, showing these sigmoids are crazy off and not at all likely. This is correctly implemented.

BUT if you look at the paper, Fig. 3, it says these are the 0, 25, 50, 75 and 100% quantiles of the prior.

So now I'm confused again.

@HeikoSchuett how should it be?

[pberkes]

@guillermoaguilar "the red and blue should be the 25% and 75% of the sigmoid, not the prior" -> the 25% and 75% of the MAP sigmoid?

[pberkes]

That would work for the threshold plot, but how can we plot the 25% and 75% of the sigmoid in the width or lambda plot?

[guillermoaguilar]

That would work for the threshold plot, but how can we plot the 25% and 75% of the sigmoid in the width or lambda plot?

I don't know. Maybe I understood Felix wrong. Let's see what Heiko answers.

[pberkes]

In any case, as far as I can tell the current code is equivalent to the Matlab version

[pberkes]

@guillermoaguilar @HeikoSchuett PR #154 has been merged... I believe the implementation is equivalent to the Matlab code. Can we close this issue, or are there more changes you'd like to do?