For a fully parametric approach, you might want to use of zero-inflated
beta distribution (e.g., as available in gamlss package), which is designed
for zero-inflated proportions. Or for a semi-parametric approach, you
could estimated a sequence of quantile regression estimates (e.g., in
package
Joana and any others: You cannot obtain a valid or useful measure of
relative importance of predictor variables across multiple models by
applying relative AIC weights or using model averaged coefficients unless
all your models included a single predictor (which, of course, is not what
is usually
Hello All: This post was motivated by the earlier posts this week
regarding CCA/NMDS/RDA etc and dissimilarity measures. I have often
thought that the usual thinking on double zeros for species
abundance/composition comparisons across sites has confused several issues
and seems driven by an
Response back to test.
Brian
Brian S. Cade, PhD
U. S. Geological Survey
Fort Collins Science Center
2150 Centre Ave., Bldg. C
Fort Collins, CO 80526-8818
email: ca...@usgs.gov brian_c...@usgs.gov
tel: 970 226-9326
On Tue, Apr 23, 2013 at 9:04 AM, Ouren, Douglas our...@usgs.gov wrote:
Chris: It may make sense to compare relative importance of predictors
across regions. But it makes NO SENSE to base measures of relative
importance of predictors in your models on AIC weights. AIC weights apply
to an entire model (1 to many predictors) not to individual predictors
within a
Jonathan: I wonder if your solution might be as simple as using one of
your model forms but allowing different intercepts for different plots by
modeling them as fixed effects (using a categorical variable for plot).
This would allow your time series model (whatever specification you use)
to
Travis: I wonder if you can modify the example from predict.lm to do
something comparable (saw this posting recently) with mixed effects models
from glmer().
?predict.lm
Offers this example, which seems to meet the request
x - rnorm(15)
y - x + rnorm(15)
predict(lm(y ~ x))
new - data.frame(x =
You ought to be very careful about using residuals from one analysis as the
response variable in another analysis as the inferences about the second
analysis will almost certainly be flawed. Best to try and do this another
way if at all possible.
Brian
Brian S. Cade, PhD
U. S. Geological
You could try estimating the conditional cumulative distribution function
with quantile regression by estimating a large interval of quantiles (e.g.,
0.01 to 0.99 if your n is large enough). Quantile regression will readily
handle skewed and heterogeneous responses. Some finessing required to
Laura: I think you need to include an interaction of Year + newWater
(i.e., Year:newWater) if you want trends across Year to be able to differ
by newWater categories (the common regression modeling approach of allowing
separate slopes and intercepts among different categorical groups in a
common
Just to echo Bob O'Hara's comment and elaborate a bit more - Don't model
average the regression coefficients, especially if you are considering
models with and without interactions among the predictors. Follow the link
provided by Bob to Cade (2015. Model averaging and muddled multimodel
Teresa: There probably are no simple short cuts here - you need to
investigate the correlations structure for each of your possible
comparisons. You can use the variance inflation factor function vif() in
the car package for glms, which includes an extension for categorical
predictors. I
Peter: Your question is not quite clear to me. I thought at first you
might be talking about quantile regression but then you mentioned the 50%
quantile (which is not the mean) of the predictor and binning. So I'm not
sure exactly what you are after. But under the presumption that you might
You might want to reconsider whether it make any sense to model average the
individual regression coefficients. See Cade (2015. Model averaging and
muddled multimodel inferences. Ecology 96: 2370-2382).
Brian
Brian S. Cade, PhD
U. S. Geological Survey
Fort Collins Science Center
2150 Centre
Ellen: Not sure why these differences would occur but note that with two
groups of 3 observations each there are only 6!/(3!3!) = 20 possible
permutations. Not sure where your 720 came from. Also, I would not expect
a permutation test for homogeneity of dispersions to be very useful with
such
Ellen: If you are running permutation procedures with data that have very
small sample sizes in each group (your two groups of n = 3 each yields only
6!/(3!3!) = M = 20 permutations under Ho), then you just have to live with
the fact that the smallest possible P-value is 1/M (= 0.05 for your two
Beta regression can be used for modeling proportion (or percentage) cover
data, but there are some issues with using it if you have many values of
0.0 or 1.0. A much more flexible approach that I've used is to use
quantile regression with the proportion response (y) data logit
transformed. Much
While the previous responders have provided some useful advice, it was a
bit misleading. The linear model for continuous responses does not
automatically assume a normal distribution (of the errors, of which the
residuals are an estimate). A specific way of estimating the conditional
mean in the
You might want to consider modeling multiple quantiles in a quantile regression
(package quantreg) as a more general, highly flexible way of modeling variation
in a response (y) as some function of predictors (X) without having to specify
a distributional form for the conditional response.
Think I miss sent this just to Phillip Dixon so reposting.
Rich: Just to expand on Phillip Dixon's reply a bit. You can always estimate
the median in the log transformed scale, with for example quantile regression,
and then back-transform to the original concentration scale without bias or
20 matches
Mail list logo