[R-lang] Re: High collinearity in logit linear mixed effects modelling
T. Florian Jaeger
tiflo@csli.stanford.edu
Sat Jun 26 08:59:05 PDT 2010
Hi again,
you can but it may change the meaning of the individual contrasts. See
eg the recent discussion on the archives of this list or the tutorial
on coding by Maureen Gillespie on our HLP lab blog
hth
Florian
On Saturday, June 26, 2010, Zhenguang Cai <s0782345@sms.ed.ac.uk> wrote:
> Hi Florian,
>
> I have one more question. I am going to use centering for the predictors. I am wondering whether there is any restriction as to what type of predictors that centering apply to? For instance, can I apply centering to categorical predictors with more than 2 levels (e.g., 4 levels)?
>
>
> Thanks,
>
> Zhenguang
>
> T. Florian Jaeger wrote:
>
> it shouldn't really tell you anything new in addition to what you know already, but you can do if you also include variables that indicate the experiment.
>
> in any case, just to be clear, i would consider it problematic to only /report/ the results from the combined analysis if you know that one of them returns zero.
> Florian
>
> On Fri, Jun 25, 2010 at 5:42 PM, Zhenguang Cai <s0782345@sms.ed.ac.uk <mailto:s0782345@sms.ed.ac.uk>> wrote:
>
> Thanks for all the advice and comments. I think I can also follow
> John's advice by pooling the two experiments together (which I did)
> and see the factor I am interested in will improve the model fit
> when added second. Is that OK?
>
>
> Zhenguang
>
> T. Florian Jaeger wrote:
>
> well, than you have one null result and one result =). It's not
> the strongest case, but at least you can argue that one of your
> experiments seems to suggests that it is indeed one of the
> predictors (rather than the other) that's driving the effect.
>
> crucially, for the result to be what you probably are looking
> for, the factor you're interested in must be the one that if
> added /second /still improves the model significantly.
> Florian
>
> On Fri, Jun 25, 2010 at 5:19 PM, Zhenguang Cai
> <s0782345@sms.ed.ac.uk <mailto:s0782345@sms.ed.ac.uk>
> <mailto:s0782345@sms.ed.ac.uk <mailto:s0782345@sms.ed.ac.uk>>>
> wrote:
>
> yes, I did two cross-language/dialect structural priming studies,
> with the same design and the same materials. In one experiment,
> Mandarin was the response language in one experiment and
> Cantonese
> was the response language in the other(primes were either in
> Mandarin and Cantonese in both experiments).
>
>
> Zhenguang
>
> T. Florian Jaeger wrote:
>
> Dear Zhenguang,
>
> what do your mean by Experiment 1 and 2? You have two
> different
> data sets?
>
> Florian
>
> On Fri, Jun 25, 2010 at 9:12 AM, Zhenguang Cai
> <s0782345@sms.ed.ac.uk <mailto:s0782345@sms.ed.ac.uk>
> <mailto:s0782345@sms.ed.ac.uk <mailto:s0782345@sms.ed.ac.uk>>
> <mailto:s0782345@sms.ed.ac.uk
> <mailto:s0782345@sms.ed.ac.uk> <mailto:s0782345@sms.ed.ac.uk
> <mailto:s0782345@sms.ed.ac.uk>>>>
>
> wrote:
>
> Dear Professor Trueswell,
>
> Thanks for the advice. I did that and found that P2 can be
> subsumed
> by P1 but not the other way round. I think that means
> something.
>
> My further question is that we always at least have to
> determine 1)
> whether P2 can be subsumed by P1 (i.e., whether the
> addition
> of P2
> can significantly improve model fit) and 2) whether P2
> can be
> subsumed by P1 (i.e., whether the addition of P2 can
> significantly
> improve model fit). Is that correct?
>
> Zhenguang
>
> John Trueswell wrote:
>
> Zhenguang,
>
> If Experiment 1 and Experiment 2 are similar
> enough, you
> could
> combine
> the data from the two experiments and model the entire
> set (keeping
> Experiment as a predictor in the model, to see if that
> matters).
>
> John Trueswell
>
>
>
> On Thu, Jun 24, 2010 at 5:17 PM, Zhenguang Cai
> <s0782345@sms.ed.ac.uk
> <mailto:s0782345@sms.ed.ac.uk> <mailto:s0782345@sms.ed.ac.uk
> <mailto:s0782345@sms.ed.ac.uk>>
> <mailto:s0782345@sms.ed.ac.uk
> <mailto:s0782345@sms.ed.ac.uk> <mailto:s0782345@sms.ed.ac.uk
> <mailto:s0782345@sms.ed.ac.uk>>>>
>
> wrote:
>
> Dear R-language people,
>
> I realized that this is probably a question
> that has been
> frequently asked
> already, so sorry for spam to some people.
>
> I found high correlation between two
> predictors (P1
> and P2)
> (r = .8). So
> following Florian's advice, I did model
> comparisons
> to try
> to exclude one of
> the predictors. However, I am not sure whether
> I did
> things
> in the right
> way.
>
> Step 1 (to determine whether P2 can be
> subsumed by P1)
>
> M0<-
> lmer(Data~1+(1|Subject)+(1|Item),family='binomial')
> M1<-
> lmer(Data~P1+(1|Subject)+(1|Item),family='binomial')
> M2 <-
> lmer(Data~P1+P2+(1|Subject)+(1|Item),family='binomial')
>
> anova (M0, M1)
> anova (M1, M2)
>
>
> Step 1 (to determine whether P1 can be
> subsumed by P2)
>
> M0<-
> lmer(Data~1+(1|Subject)+(1|Item),family='binomial')
> M1<-
> lmer(Data~P2+(1|Subject)+(1|Item),family='binomial')
> M2 <-
> lmer(Data~P2+P1+(1|Subject)+(1|Item),family='binomial')
>
> anova (M0, M1)
> anova (M1, M2)
>
>
> In Experiment 1, I found P2 can be subsumed by
> P1 but not
> the other way
> round.
>
> However, in Experiment 2, I found P1 and P2 can be
> subsumed
> by each other.
> How to resolve this?
>
>
> Thanks,
>
> Zhenguang
>
> --
> The University of Edinburgh is a charitable body,
> registered in
> Scotland, with registration number SC005336.
>
>
>
>
> -- The University of Edinburgh is a charitable body,
> registered in
> Scotland, with registration number SC005336.
>
>
>
> -- The University of Edinburgh is a charitable body,
> registered in
> Scotland, with registration number SC005336.
>
>
>
> -- The University of Edinburgh is a charitable body, registered in
> Scotland, with registration number SC005336.
>
>
>
>
> --
> The University of Edinburgh is a charitable body, registered in
> Scotland, with registration number SC005336.
>
>
More information about the ling-r-lang-L
mailing list