Misleading documentation for pymc.normal_like

[strumke]

Hi!
The documentation [1] for pymc.normal_like fails to mention that it actually returns the logarithm of the function, as the source code [2] reveals in the following lines

like = 0.0
like = like - 0.5 * tau_tmp * (x(i)-mu_tmp)**2
like = like + 0.5dlog(0.5tau_tmp/PI)

[1] https://pymc-devs.github.io/pymc/distributions.html#pymc.distributions.normal_like
[2] https://github.com/pymc-devs/pymc/blob/8733c6686787e0e98bd2445ea5408fe988adf0c9/pymc/flib.f

Thanks!

[fonnesbeck]

The documentation does say log-likelihood:

However, if you have ways of improving the docs, we always welcome pull requests!

[strumke]

Hi and thanks for the reply My suggestion is to actually write the logarithm in the function, or state explicitly that the function stated is the normal likelihood (and not the normal log-likelihood) Inga

…

On 31 January 2018 at 01:52, Chris Fonnesbeck ***@***.***> wrote: Closed #175 <pymc-devs/pymc#175>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <pymc-devs/pymc#175 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ANoQKEtpqhjl1jPjBherOKXbcTb9xqxJks5tP7legaJpZM4RyAvH> .

-- Inga Strümke Skoglien 35 5056 Bergen Norway +47 95730153 +34 619371726

[fonnesbeck]

I see what you mean. The log-likelihood is just a computational convenience for doing more stable statistical computation, whereas users generally think of distributions in terms of its PDF/PMF, and not the log-transformed scale, so we present those in the documentation. Most users will not have to worry about the fact that the distributions are log-transformed, except when more advanced computations are required. That’s a good point, though, and we should probably state that the formula of PDF/PMF is displayed in the docs.

[strumke]

Hi Thank you for the reply, and for considering my concern. Inga

…

On 31 January 2018 at 17:28, Chris Fonnesbeck ***@***.***> wrote: I see what you mean. The log-likelihood is just a computational convenience for doing more stable statistical computation, whereas users generally think of distributions in terms of its PDF/PMF, and not the log-transformed scale, so we present those in the documentation. Most users will not have to worry about the fact that the distributions are log-transformed, except when more advanced computations are required. That’s a good point, though, and we should probably state that the formula of PDF/PMF is displayed in the docs. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <pymc-devs/pymc#175 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ANoQKC6TSzJKaVBTvbTKtA3SF_13IOU-ks5tQJSdgaJpZM4RyAvH> .

-- Inga Strümke Skoglien 35 5056 Bergen Norway +47 95730153 +34 619371726