Thanks for the suggestion - mafc.cauchit seems to work with p_ID and audio_name as random effects, but unfortunately not with audio_name as a fixed effect, as I'd hoped for ("cannot generate feasible simplex"). Oh well!
-----Original Message-----
From: R-sig-mixed-models [mailto:r-sig-mixed-models-bounces at r-project.org] On Behalf Of ken knoblauch
Sent: 08 May 2015 15:29
To: r-sig-mixed-models at r-project.org
Subject: Re: [R-sig-ME] Logistic modelling with guessing parameter
Peter Harrison <pharr011 at ...> writes:
Hello Ben and Ken,
Thanks very much for your useful responses!
Ben - thanks for your hint re non-finite values in PIRLS, I'll bear
that
in mind for the future. I added
non-finite warning checks to the link functions etc. and nothing
came up this time, unfortunately.
You're right, all of the groups show complete separation - I hadn't
thought too carefully about this but
realise now that it will be a problem. I think I might have to merge
the
musical tracks into larger groups,
perhaps by metre, genre, etc.
Thanks a lot for the great graph, too!
Ken - thanks for the heads-up about using the link directly from
the psyphy package, that definitely
simplifies things! Sorry, the code I gave and you ran is
an example that does work: responses ~ accuracy +
(1|p_ID). The model I was having problems with
was when I added the "audio_name" predictor, i.e.
responses ~ accuracy + audio_name + (1|p_ID).
Something that you can try is the mafc.cauchit link that would impose a less steep slope because of the heavy tails of the Cauchy. This might be less sensitive to complete separation sort of like a regularization of the psychometric function slope.
I have used this link to correct for the bias introduced by lapses at the upper asymptote instead of introducing a lapse parameter which is not obvious how to do otherwise with glmer.
Best wishes,
Peter
[[alternative HTML version deleted]]