Skip to content

lmer/glmer standard error interpretation and visualization

3 messages · Colin Wahl, Ben Bolker

#
Hello,
I am in the process of finalizing figures for my thesis on stream
invertebrate distributions among watershed and riparian types. See
below for additional information on the design. I'm having difficulty
including standard errors from the lmer modeling as error bars in the
figures. Here is the table I've created from the lmer output:
estimates of %EPT and St Error are back transformed from logits and
converted from fractions to percents. Estimates are also absolute (not
relative to the intercept).

 Watershed ? ? Effect ? ? ? ? ? ? ? ? ?Estimate ?St. Error? z score ?p value

 Forested  ? ? Intercept: F vs. 0 ? ?   28.23 ?    59.6 ? ? -2.346 ? ? 0.019*
 ? ? ? ? ? ? ?? ? ? Riparian: ?F vs. NF ? 16.017 ?   62.3 ? ? -1.436 ?   0.151
 Cultivated ? ?Watershed: C vs. F ? ?1.351 ?   65.3 ? ? -5.297 ? ?<0.000*
 ? ? ? ? ? ?  ? ? ? Riparian: ?F vs. NF ? ?1.555 ?    69.2 ? ? ?1.071 ? ?0.284
 Developed ? Watershed: D vs. F ? ?0.175 ?    66.8 ? ? -7.714 ?  <0.000*
 ? ? ? ? ? ? ?? ? ? Riparian: ?F vs. NF ? ?0.292 ?    70.9 ? ? ?1.391 ?   0.164
 Grassland ? ?Watershed: G vs. F ?  28.94 ?   66.6 ?     0.05 ?   ?0.960
 ? ? ? ? ? ? ?  ? ? Riparian: ?F vs. NF ? ? 1.967 ? ? 70.7 ? ? -2.595  ?? 0.009*

The st. errors are huge. I initially used standard error calculations
in excel for error bars (stdev(x)/sqrt(n(x))), which look very
reasonable, and are reflective of significant differences.

Does anyone have any advice to offer for visualizing these glmer
results? Should I use the huge model St. Errors? My inclination is
yes, because they are used to calculated significant differences, but
28 + or - 59.6 with a significant p value seems ridiculous.

Thank you,
Colin Wahl
M.S. Candidate
Dept. of Biology
Western Washington University
#
Colin Wahl <biowahl at ...> writes:
[snip]
How did you back-calculate the standard errors?  It simply doesn't
make sense to compute plogis([standard error]) to get the standard
error on the response scale; you can either use the delta method as
one of the variants of predict.glm() does [i.e. multiply the standard
error by the *derivative* of the link function], or calculate the
confidence intervals on the link scale (i.e. estimate plus/minus CI)
and back-transform them (they will not in general be symmetric).

  This is not an lmer issue, this is a general issue with generalized
linear models, or any other model that works on a transformed scale
and for which one wants to backtransform the parameters.
#
Ben Bolker <bbolker at ...> writes:
PS  this is why epidemiologists spend so much time learning about
odds ratios and log-odds -- you can back-transform from logit effects
to odds-ratio effects, but once you get there there's just not any
perfect way to transform back to a probability scale in a way that
is completely general ... http://lesswrong.com/lw/8lr/logodds_or_logits/