[R-sig-ME] R-sig-mixed-models Digest, Vol 155, Issue 8

[Voeten, C.C.]

Dear John,

Random slopes by listener for gender and AQ score are unnecessary/don't make sense, as these reflect listener-specific characteristics and hence are already taken into account by the random intercept. However, you may want to consider a random slope (stimulus_char|listener) as well -- that would give you the 'maximal model' sensu Barr et al 2013. I recommend using a likelihood-ratio test (with its p-value divided by 2, see Pinheiro & Bates 2000) to see if the random-slope model is a significant improvement over the random-intercept model.

By the way, for a model with only one random factor like yours, function mixed_model from package GLMMadaptive will give slightly more accurate results than glmer.

Greetings from a fellow phonologist/phonetician,
Cesko

> -----Oorspronkelijk bericht-----
> Van: R-sig-mixed-models <r-sig-mixed-models-bounces using r-project.org>
> Namens John Kingston
> Verzonden: maandag 11 november 2019 18:28
> Aan: r-sig-mixed-models using r-project.org
> Onderwerp: Re: [R-sig-ME] R-sig-mixed-models Digest, Vol 155, Issue 8
> 
> Responding to John Haman:
> 
> Thank you for your comments.
> 
> I've modeled these results using a logistic regression because the sounds at
> the two ends of the series are an unambiguous [s] and an unambiguous [sh],
> respectively. In English, the native language of all the listeners, these two
> sounds can be analyzed as contrasting in their value for just one distinctive
> feature, [s] is [+anterior] and [sh] is [-anterior]. This feature refers to the
> position of the tongue tip and blade. The intermediate steps between the
> endpoints were made by mixing the original [s] and [sh] waveforms in
> complementary proportions. So with respect to both linguistic and physical
> characteristics of these sounds, it's possible to treat responses as binomially
> distributed.
> 
> I have tried a model like the one suggested. My question is whether it's
> possible to interpret the interactions between listener gender and AQ scores
> with stimulus characteristics when other uncontrolled listener characteristics
> are accounted for in the random effect. A model like the one proposed, in
> which there's a random effect of listener on the intercept, doesn't strike me
> as hard to interpret, but what if I wanted to examine random effects of
> listener on the slopes of the fixed effects and their interactions? For
> example, is it still possible to treat a random effect of listeners on the
> interactions of listener gender by a stimulus characteristic as accounting for
> uncontrolled idiosyncrasies of those listeners on that interaction?
> 
> Best,
> John
> 
> On 2019-11-11 06:00, r-sig-mixed-models-request using r-project.org wrote:
> > Send R-sig-mixed-models mailing list submissions to
> > 	r-sig-mixed-models using r-project.org
> >
> > To subscribe or unsubscribe via the World Wide Web, visit
> > 	https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> > or, via email, send a message with subject or body 'help' to
> > 	r-sig-mixed-models-request using r-project.org
> >
> > You can reach the person managing the list at
> > 	r-sig-mixed-models-owner using r-project.org
> >
> > When replying, please edit your Subject line so it is more specific
> > than "Re: Contents of R-sig-mixed-models digest..."
> >
> >
> > Today's Topics:
> >
> >    1. a question about partitioning effects (John Kingston)
> >    2. Re: a question about partitioning effects (John Haman)
> >
> > ----------------------------------------------------------------------
> >
> > Message: 1
> > Date: Sun, 10 Nov 2019 13:15:48 -0500
> > From: John Kingston <jkingston using linguist.umass.edu>
> > To: r-sig-mixed-models <r-sig-mixed-models using r-project.org>
> > Subject: [R-sig-ME] a question about partitioning effects
> > Message-ID: <fca30459879acee26dc6b904983d1c70 using umail.it.umass.edu>
> > Content-Type: text/plain; charset="us-ascii"; Format="flowed"
> >
> > I have a data set in which a group of listeners were asked to
> > categorize a series of sounds that stepped incrementally from an [s]
> > to a [sh], as in "see" and "she". Members of this series were followed
> > by one of three vowels, each one spoken by one of four speakers. Call
> > these the "stimulus" characteristics. Past work predicts how
> > differences between the vowels and speakers' voices would influence
> > listeners' choice of [s] or [sh] as the category to which a particular
> > step in the series belongs, and those predictions are clearly
> > confirmed by the results.
> >
> > In addition to these stimulus characteristics and their predicted
> > effects, I also have "listener" characteristics, namely, their gender
> > --
> > there are 23 women and 24 men -- and their scores on Simon
> > Baron-Cohen's Autism Spectrum Questionnaire -- it provides a total AQ
> > score and scores on five subsets of questions for each listener.
> >
> > The interest in this study is how a listener's gender and their total
> > AQ or subset scores influence the effects of stimulus characteristics
> > on their categorization performance. In a mixed effects logistic
> > regression model like this, I would ordinarily represent differences
> > between listeners as a random effect, but here I also want to include
> > the gender and total and subset AQ scores as fixed effects, in
> > interactions with the fixed effects that represent the stimulus
> > characteristics.
> >
> > I should add that I have an analogous data set collected from a
> > different group of listeners, evenly divided between women and men, in
> > which the stimulus characteristics are the same but the other listener
> > characteristics are scores from the five subsets of the Big Five
> > personality questionnaire. As the subsets measure what are supposed to
> > be independent personality traits, there is no total Big Five score.
> > So my question is more general; namely, how do I model the effects of
> > these measures of listeners' personality traits while also capturing
> > uncontrolled differences between listeners?
> >
> > --
> > John Kingston
> > Professor
> > Linguistics Department
> > University of Massachusetts
> > Integrative Learning Center N434
> > 650 N. Pleasant St.
> > Amherst, MA 01003
> > 1-413-545-6833, fax -2792
> > jkingston using linguist.umass.edu
> > http://blogs.umass.edu/jkingstn/
> >
> >
> >
> >
> > ------------------------------
> >
> > Message: 2
> > Date: Sun, 10 Nov 2019 18:00:59 -0500
> > From: John Haman <mail using johnhaman.org>
> > To: r-sig-mixed-models using r-project.org
> > Subject: Re: [R-sig-ME] a question about partitioning effects
> > Message-ID: <0c253e50-af59-369a-d4fd-0442bd3528b0 using johnhaman.org>
> > Content-Type: text/plain; charset="utf-8"; Format="flowed"
> >
> > I'm curious about the basis for the logistic regression model here. A
> > correct response seems to imply that there is a true categorization
> > (or
> > coarsening) of sounds on the gradient from [s] to [sh]. Is there? Or
> > is the correct classification in the eye of the researcher?
> >
> > To try to answer your actual question, have you thought about modeling
> > the data as
> >
> > glmer(category ~ AQ*gender*stimulus_char + (1 | listener))
> >
> > or some variation thereof? This would treat AQ, gender, and stimulus
> > as fixed effects and supply a random intercept for each listener.
> >
> > -John
> >
> > On 11/10/19 1:15 PM, John Kingston wrote:
> >> I have a data set in which a group of listeners were asked to
> >> categorize a series of sounds that stepped incrementally from an [s]
> >> to a [sh], as in "see" and "she". Members of this series were
> >> followed by one of three vowels, each one spoken by one of four
> >> speakers. Call these the "stimulus" characteristics. Past work
> >> predicts how differences between the vowels and speakers' voices
> >> would influence listeners' choice of [s] or [sh] as the category to
> >> which a particular step in the series belongs, and those predictions
> >> are clearly confirmed by the results.
> >>
> >> In addition to these stimulus characteristics and their predicted
> >> effects, I also have "listener" characteristics, namely, their gender
> >> -- there are 23 women and 24 men -- and their scores on Simon
> >> Baron-Cohen's Autism Spectrum Questionnaire -- it provides a total AQ
> >> score and scores on five subsets of questions for each listener.
> >>
> >> The interest in this study is how a listener's gender and their total
> >> AQ or subset scores influence the effects of stimulus characteristics
> >> on their categorization performance. In a mixed effects logistic
> >> regression model like this, I would ordinarily represent differences
> >> between listeners as a random effect, but here I also want to include
> >> the gender and total and subset AQ scores as fixed effects, in
> >> interactions with the fixed effects that represent the stimulus
> >> characteristics.
> >>
> >> I should add that I have an analogous data set collected from a
> >> different group of listeners, evenly divided between women and men,
> >> in which the stimulus characteristics are the same but the other
> >> listener characteristics are scores from the five subsets of the Big
> >> Five personality questionnaire. As the subsets measure what are
> >> supposed to be independent personality traits, there is no total Big Five
> score.
> >> So my question is more general; namely, how do I model the effects of
> >> these measures of listeners' personality traits while also capturing
> >> uncontrolled differences between listeners?
> >>
> >
> >
> >
> >
> > ------------------------------
> >
> > Subject: Digest Footer
> >
> > _______________________________________________
> > R-sig-mixed-models mailing list
> > R-sig-mixed-models using r-project.org
> > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> >
> >
> > ------------------------------
> >
> > End of R-sig-mixed-models Digest, Vol 155, Issue 8
> > **************************************************
> 
> --
> John Kingston
> Professor
> Linguistics Department
> University of Massachusetts
> Integrative Learning Center N434
> 650 N. Pleasant St.
> Amherst, MA 01003
> 1-413-545-6833, fax -2792
> jkingston using linguist.umass.edu
> http://blogs.umass.edu/jkingstn/
> 
> _______________________________________________
> R-sig-mixed-models using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models