[R-lang] Re: % of total variance accounted for by each predictor?

[Xiao He]

Hi Dr. Jaeger,

Thank you for the explanation! With regard to your question, I am indeed
running a *logit mixed model*, with binary outcomes (Corret / incorrect),
four predictors (iv1, iv2, iv3, iv4) and two random effects (subject and
item).  I don't know how I should put the various information into the
function you wrote. It'd be great if you could provide a brief explanation
to that.

On the meantime, I will read the article you suggested. Thanks in advance!

Best,
Xiao





On Tue, Aug 3, 2010 at 9:10 PM, T. Florian Jaeger
<tiflo@csli.stanford.edu>wrote:

> Hi Xiao,
>
> I am traveling but if you let me know what the problem is I will try to
> help. Since I can't tell from this email what type of model you're running,
> let me be clear about one thing: To calculate an R-square for *linear*mixed models, you don't need that function - it's for mixed logit models
> (and, as I think I not on that page, I am actually a it skeptical as to the
> usefulness of the Nagelkerke R2).
>
> To calculate R2s for linear mixed models you only need to correlate the
> predictions against the actual outcome and square the results (=cor(outcome,
> fitted(lmer(outcome ~ .....)))^2). But be careful in the interpretation of
> the numbers. Most likely most of the variance the model captures will be due
> to the random participant intercepts. It's not uncommon that for, e.g. an R2
> of 50%, 45%+ of that come from the random intercept, while the remaining
> percent are due to the combined effect of the fixed effect predictors.
>
> The same, of course, applies to the Nagelkerke R2. If you really need such
> a measure, you might want to have a look at the following paper for an
> example of how one might report the results (though I am curious about what
> other people think of this):
>
> *Jaeger, T.F.* 2010. Redundancy and reduction: Speakers manage syntactic
> information density. *Cognitive Psychology 61(1), 23-62. *
>
> There is a footnote on the Nagelkerke R2.
>
> Florian
>
> On Tue, Aug 3, 2010 at 6:38 PM, Alex Fine <afine@bcs.rochester.edu> wrote:
>
>> Hey Xiao,
>>
>> Basically, yes, I think that's what the function requires.  There could be
>> something non-obvious about how the formula has to be specified.   That's
>> all I can think of....Probably the best thing to do would be to post a reply
>> on the blog, let dear florian answer it when he gets a chance, and then
>> future users who run into similar problems will see the answer there.  You
>> could probably just post your last message, verbatim.
>>
>> alex
>>
>>
>> Xiao He wrote:
>>
>>> Hi Alex
>>>
>>> Sorry that I didn't include any information about my data or code that I
>>> used :)
>>>
>>> I did look at the codes on the page, but they are bit opaque to me.
>>> Mainly, I don't understand the syntax of the codes. For example, what
>>> exactly should the "f" look like? For example, I have a data sheet where the
>>> column "correct" corresponds to the binary responses obtained from
>>> participants, and "iv1", "iv2, "iv3", and "iv4" are four columns
>>> corresponding to the 4 predictors that I want to evaluate, and "subject" and
>>> "item" are the two columns for the random effects. The name of the data
>>> sheet is "data".
>>>
>>> Do I need to do the following? =>  my.lmer.nagelkerke
>>> (correct~iv1+iv2+iv3+iv4 + (1|subject) + (1|item), data)
>>>
>>> This is what I did, but it is wrong. I hope you can help me with this.
>>> Thank you :)
>>>
>>>
>>> Best,
>>> Xiao
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://mailman.ucsd.edu/mailman/private/ling-r-lang-l/attachments/20100804/fbfdccde/attachment-0001.html