[R-sig-ME] double repeated measures

John Maindonald john.maindonald at anu.edu.au
Thu Aug 10 23:28:56 CEST 2017

All that the simulations in the O’Hara et al paper show is the E[log(y)] does not equal log(E[y]), where E[] is expectation.
The O’Hara et al paper’s claim that "the transformations performed poorly” is just wrong, as it relates to what
their simulations might demonstrate.

See https://stats.stackexchange.com/questions/114848/negative-binomial-glm-vs-log-transforming-for-count-data-increased-type-i-erro/215080#215080
and the following paper that discusses the same sort of issue for RNA-Seq gene expression counts:
Law, CW, Chen, Y, Shi, W, Smyth, GK (2014). Voom: precision weights unlock linear model analysis tools for RNA-seq read counts. Genome Biology 15, R29. http://genomebiology.com/2014/15/2/R29

I have an immediate interest in the equivalent issue for glmer models, used
for insect dose-mortality data where the error is a version of over-dispersed
binomial, with the amount of over-dispersion greatest around 50% mortality
and reducing at high mortalities.  Working with transformed mortality and
an lmer() model does a much better job of modeling the within replicate
variation than anything that one can readily do with a glmer() model that
is set up to (strictly) handle only binomial error.  One possibility for adapting
glmer() may be to apply weights that are designed to “fix up” the within
replicate variance structure — my impression is, however, that this adjusts
both levels of the variance structure.  The attempt to incorporate observation
level random effects led (at least when I tried to fit a model that had random
slopes and intercepts) to a message that the model was over-parameterized.

The vignette cfAnalyses.html that can be found at
looks at this issue, plus outliers issues!  Practical data analysis can get
very messy!

John Maindonald             email: john.maindonald at anu.edu.au<mailto:john.maindonald at anu.edu.au>

On 11/08/2017, at 01:06, Dexter Locke <dexter.locke at gmail.com<mailto:dexter.locke at gmail.com>> wrote:


While this doesn't address the question about repeated measures, consider
looking at

O’Hara, R. B., Kotze, D. J., O ’Hara, R. B., & Kotze, D. J. (2010). Do not
log-transform count data. Methods in Ecology and Evolution, 1(2), 118–122.

for the left-hand side of the model.

- Dexter

On Wed, Aug 9, 2017 at 10:52 PM, Julie McIntyre <jpmcintyre at alaska.edu>


I am trying to fit a model to data that are recorded in a doubly-repeated
measures type design, and I'm having trouble with the syntax for lmer.  The
response is a count representing the number of animals harvested at each of
six locations during a hunting season.  Counts are recorded daily at each
location for a fixed period (about 45 days).  In addition, the daily counts
at each location have been repeated themselves over several years (about
30).  The dates of the counts are the same every year.  Additional
covariates are measured on a daily basis.

Graphically, for all locations there is a clear trend in count by day, with
some year-to-year variation.  There are also clear but weaker trends in
counts by year (for fixed day), with variation among locations.  The
general shape of the trend changes quite a bit depending on the day (e.g.,
early vs. late in the season).  That said, the main interest is in
understanding the influence of the covariates on harvest.

I believe the following code fits a random intercept and slope model to
daily counts within years, separately for each location.  This model fits
well, and allows testing of the covariate effects (X1 and X2).  However it
ignores the second layer of repetition and the trend in count by year,
within locations.

M1=lmer(log(Count+1)~X1+X2+Location+Day+(1+Day|Year), data=Harvest)

I would like to know the correct syntax to also include terms for the
repeated measurement by year, within locations.  This model might be close,
but I'm just not certain:


I'd appreciate any suggestions or advice.  Thank you!

Julie McIntyre

       [[alternative HTML version deleted]]

R-sig-mixed-models at r-project.org mailing list

[[alternative HTML version deleted]]

R-sig-mixed-models at r-project.org mailing list

	[[alternative HTML version deleted]]

More information about the R-sig-mixed-models mailing list