[R-sig-ME] Method to transform fixed effect parameter estimates back on to the response scale [after glmer() with family=Gamma(link = "log")] ??

[Steve Denham]

Hi Dan,
While that would be nice, it just isn't the case that for a gamma distributed variable that there is an increase of some fixed amount for a given change in the predictor.  What actually happens is that for a given change in the predictor, there is a fixed MULTIPLICATIVE change.  Say that coefficient for the fixed effect is 2.  Then for a unit increase in the fixed effect, I believe the response would be exp(2*(unit change)) larger (It is early and caffeine hasn't kicked in yet, so that may be incorrect).  Unit changes at larger values will have a greater effect on the response than at smaller values. Steve Denham
Director, Biostatistics
MPI Research, Inc.
 
      From: Daniel Newman <dan.newman86 at gmail.com>
 To: r-sig-mixed-models at r-project.org 
 Sent: Thursday, July 16, 2015 1:41 AM
 Subject: [R-sig-ME] Method to transform fixed effect parameter estimates back on to the response scale [after glmer() with family=Gamma(link = "log")] ??
   
Subject: Method to transform fixed effect parameter estimates back on to
the response scale [after glmer() with family=Gamma(link = "log")] ??

Dear lme4 experts,

I am using lme4 to model human reaction-time (RT; in milliseconds)
responses. My model includes both nested and fully crossed random
intercepts, and fixed effect “predictor” factors.

lmer() seems to work quite well for this, and is nice since I can use the
fixed effect parameter estimates (beta and Std.Errors) for interpretation
to say that for every unit of the “predictor” that changes, we expect a
change of ~beta units in RT on the response scale (milliseconds). Pretty
happy with the results….BUT…

glmer() may be the better option since the response distribution has a
positive skew with no zero values (typical RT distribution), and glmer()
allows the same model specification as my lmer() model, except using
family=Gamma(link = "log") to account for the skewed response distribution.

This seems to work well and gives very similar results to the equivalent
lmer() model, but with somewhat improved residual plots, so I guess the
glmer() with family=Gamma(link = "log") is a more valid approach since it
explicitly accounts for the shape of response distribution.

PROBLEM: Is there a way to transform the fixed effect parameter estimates
and Std.Errors from the glmer() with family=Gamma(link = "log"), back to
the response scale (i.e. back to RT in milliseconds). It would be nice/aid
interpretation to say that for every unit of the “predictor” that changes,
we expect a change of ~beta units on the response scale.

Thank you so much for your time!!

Cheers
Dan


--

    [[alternative HTML version deleted]]

_______________________________________________
R-sig-mixed-models at r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models


   

	[[alternative HTML version deleted]]