[R-sig-ME] [ADMB Users] getting standardized coefficients in admb
Ellen Robertson
robertsonep at gmail.com
Tue Feb 9 17:52:34 CET 2016
Thanks, yes I'm not really sure what the best approach is. I'm leaning
towards scaling only my predictor variables for binomial/poisson glmer
models (and possibly scale by 1/sd of the response as you mentioned) . For
my lmer models, I could do the same or I could scale both predictor and
response variables. My lmer coefficients are extremely different depending
on which I do: if I scale the lmer response variable I get a beta=-0.25 and
if I don't scale the response I get beta=-12.54.
Best,
Ellen
On Tue, Feb 9, 2016 at 11:03 AM, Ben Bolker <bbolker at gmail.com> wrote:
>
> Actually, now that I look at the article more carefully it turns out
> that those paragraphs are mostly focused on *mixed* models, and don't say
> too much about how the argument generalizes (so to speak) to the
> generalized-linear (mixed) model case. Scaling the *estimated parameter*
> by 1/standard deviation of the response is not insane (you can't scale the
> response variable *before* you fit the model in a GLM, that doesn't make
> sense), but doesn't have the same nice interpretation as in a linear
> model. In general the link functions do put the parameters on a simple,
> dimensionless scale, but I'm not sure about a sensible, general way to
> compare among parameters of models fitted with *different* link functions.
>
> On 16-02-09 09:49 AM, Ellen Robertson wrote:
>
>> Thanks very much for your response and for pointing me in the direction
>> of that article. Cheers, Ellen
>>
>> On Sun, Feb 7, 2016 at 4:46 PM, Ben Bolker <bbolker at gmail.com
>> <mailto:bbolker at gmail.com>> wrote:
>>
>> (I think this is more appropriate for r-sig-mixed-models, but I'm
>> leaving ADMB users cc'd for this last response.)
>>
>> It's not obvious to me whether there's a simple analogue of
>> standardizing by response variance in the GLMM world. I suppose you
>> *could* still standardize by predictor variance, or you could decide
>> that the link functions (log for NB/Poisson, logit for binomial)
>> effectively standardize the prediction side of the model. It looks
>> like the last section of Schielzeth's 2010 MEE paper "Simple means
>> ...", "Extensions", discusses this issue, but I haven't read it
>> carefully/absorbed it/tried to implement that in a function.
>>
>> cheers
>> Ben
>>
>>
>> On Sun, Feb 7, 2016 at 2:47 PM, Ellen Robertson
>> <robertsonep at gmail.com <mailto:robertsonep at gmail.com>> wrote:
>> > Ben,
>> > Sorry for the delayed response. In my earlier email, I was
>> referring to
>> > your post on
>> >
>>
>> http://r-sig-mixed-models.r-project.narkive.com/1EtbqR8T/r-sig-me-standardized-coefficients-in-glmer-model
>> > where you talk about using a function similar to the 'lm.beta'
>> function for
>> > getting standardized coefficients from lmer models
>> ('lm.beta.lmer') .
>> > I'm trying to get standardized beta coefficients from
>> different types
>> > of glmer models (poisson, binomial, Gaussian) so that I can
>> compare the
>> > effect sizes from each of these (I'm using all three of these
>> different
>> > types of glmer models within a piecewise structural equation
>> model and want
>> > to be able to compare the strengths of different paths). I know
>> that with
>> > continuous response/predictor variables I can just scale
>> everything before
>> > running the model and that will output standardized beta
>> coefficients. But
>> > I am unsure of do this with non-continuous variables (such as a
>> binomial
>> > response variable)? You show (in the link above) how to scale
>> binomial
>> > predictor variables (change them to numeric, 0/1, rather than
>> > male/female..and then scale)...but how would you do this with a
>> binomial
>> > response variable which has to be 0/1? I tried your
>> "lm.beta.lmer" function
>> > and it worked when I had 2 predictors in my model but for some
>> reason it
>> > didn't work with only one predictor variable. I also wasn't sure
>> if it
>> > would work with poisson/binomial models or if it only worked with
>> lmer.
>> > Thanks for any help you can give. Cheers,
>> > Ellen
>> >
>> >
>> >
>> >
>> > On Wednesday, November 25, 2015 at 5:37:25 PM UTC-5, Ben Bolker
>> wrote:
>> >>
>> >> I meant to respond to this earlier (maybe I did, and maybe it
>> fell
>> >> through the cracks).
>> >>
>> >> Ellen, it's not clear whether you're asking about generic
>> ADMB models
>> >> or about glmmADMB models: if the latter, then
>> >> r-sig-mix... at r-project.org <mailto:r-sig-mix... at r-project.org>
>> is probably the more appropriate venue.
>> >> If the former, then I'm not even sure what you would mean by
>> >> "standardized coefficients", as it would probably depend on the
>> model.
>> >>
>> >> Can you give a link/reference for "Bolker's code for beta.lmer
>> for
>> >> glmer models"?
>> >>
>> >> The very generic answer to your question is that you can
>> either (1)
>> >> scale/center your continuous input variables *before* running
>> the model
>> >> or (2) adjust the coefficients afterward, based on the means and
>> >> standard deviations of the parameters. This
>> >>
>> >>
>> >>
>>
>> http://stackoverflow.com/questions/23642111/how-to-unscale-the-coefficients-from-an-lmer-model-fitted-with-a-scaled-respon/23643740#23643740
>> >>
>> >> gives a function that rescales parameters -- it should be
>> ecumenical
>> >> (i.e., apply to any set of coefficients from a linear or
>> generalized
>> >> linear model, no matter what software it was fitted with).
>> >>
>> >>
>> >> On 15-11-25 05:30 PM, Johnoel Ancheta wrote:
>> >> > Is this possible?
>> >> >
>> >> > On Mon, Nov 23, 2015 at 7:31 AM, Ellen Robertson
>> <rober... at gmail.com <mailto:rober... at gmail.com>>
>> >> > wrote:
>> >> >
>> >> >> Hi everyone,
>> >> >> Is it possible to get standardized coefficients from admb
>> models? I
>> >> >> know about lm.beta for linear models and saw Bolkner's code for
>> >> >> beta.lmer
>> >> >> for glmer models....but I have been unable to get standardized
>> >> >> coefficients
>> >> >> from my admb models. Thanks for your help,
>> >> >> Ellen
>> >> >>
>> >> >> --
>> >> >> You received this message because you are subscribed to the
>> Google
>> >> >> Groups
>> >> >> "ADMB Users" group.
>> >> >> To unsubscribe from this group and stop receiving emails from
>> it, send
>> >> >> an
>> >> >> email to users+un... at admb-project.org
>> <mailto:users%2Bun... at admb-project.org>.
>> >> >> For more options, visit
>> >> >> https://groups.google.com/a/admb-project.org/d/optout.
>> >> >>
>> >> >
>> >>
>> >
>>
>>
>>
[[alternative HTML version deleted]]
More information about the R-sig-mixed-models
mailing list