Dear all,
I have a model that I am currently working on that is going through
covariate selection. There's one particular parameter, whose ETA-shrinkage
before adding any covariate was 25%. However, after adding a covariate
successfully, the ETA-shrinkage was increased to almost 80%. While the
model parameter estimation in this final model is reasonable, and there's
no warnings or errors in the output file, the increase in the ETA-shrinkage
basically means that the individual estimation falls back to population
estimation. Under such a circumstance, how should I decide if it's
reasonable to keep this parameter? Thank you.
Best Regards
--
Xinting
Covariate lead to increase in Eta shrinkage
it sounds like that covariate provides information for the parameter influenced
by eta.
you have taken something that was unexplained population variation and
explained part of it with the covariate.
this is usually a good thing.
if the covariate helps so much to predict the parameter that the eta becomes
very small and as a result gets a high shrinkage then this can happen when
there isnt much information left.
can you just do the regular hypothesis testing?
try removing the eta i.e. (0 FIXED) , reestimate and compare AIC? and whatever
other metrics you might use.
warm regards
Douglas
________________________________
Quoted reply history
Van: [email protected] [mailto:[email protected]] Namens
Xinting Wang
Verzonden: January 22, 2015 8:27 AM
Aan: [email protected]
Onderwerp: [NMusers] Covariate lead to increase in Eta shrinkage
Dear all,
I have a model that I am currently working on that is going through covariate
selection. There's one particular parameter, whose ETA-shrinkage before adding
any covariate was 25%. However, after adding a covariate successfully, the
ETA-shrinkage was increased to almost 80%. While the model parameter estimation
in this final model is reasonable, and there's no warnings or errors in the
output file, the increase in the ETA-shrinkage basically means that the
individual estimation falls back to population estimation. Under such a
circumstance, how should I decide if it's reasonable to keep this parameter?
Thank you.
Best Regards
--
Xinting
________________________________
Dear all,
After checking again the data, and ran the model with the specific ETA
fixed at 0, the OFV did not change. It seems that in the final model, the
variability associated is explained to the vast majority part by this
covariate.
My question here is, since I am building the final model using a step-wise
method, and the variability described by this ETA affects the model, is it
OK to ignore this ETA in the final model?
Quoted reply history
On 22 January 2015 at 19:55, Eleveld, DJ <[email protected]> wrote:
> it sounds like that covariate provides information for the parameter
> influenced by eta.
> you have taken something that was unexplained population variation and
> explained part of it with the covariate.
> this is usually a good thing.
> if the covariate helps so much to predict the parameter that the eta
> becomes very small and as a result gets a high shrinkage then this can
> happen when there isnt much information left.
> can you just do the regular hypothesis testing?
> try removing the eta i.e. (0 FIXED) , reestimate and compare AIC? and
> whatever other metrics you might use.
> warm regards
> Douglas
>
> ------------------------------
> *Van:* [email protected] [mailto:[email protected]]
> *Namens
> *Xinting Wang
> *Verzonden:* January 22, 2015 8:27 AM
> *Aan:* [email protected]
> *Onderwerp:* [NMusers] Covariate lead to increase in Eta shrinkage
>
> Dear all,
>
> I have a model that I am currently working on that is going through
> covariate selection. There's one particular parameter, whose ETA-shrinkage
> before adding any covariate was 25%. However, after adding a covariate
> successfully, the ETA-shrinkage was increased to almost 80%. While the
> model parameter estimation in this final model is reasonable, and there's
> no warnings or errors in the output file, the increase in the ETA-shrinkage
> basically means that the individual estimation falls back to population
> estimation. Under such a circumstance, how should I decide if it's
> reasonable to keep this parameter? Thank you.
>
> Best Regards
>
> --
> Xinting
> ------------------------------
>