[R] [R-sig-ME] prediction intervals from a mixed-effects models?
spencer.graves at pdf.com
Wed Apr 16 05:03:47 CEST 2008
Thanks for the suggestion. The index of Gelman and Hill for
"prediction" for "multilevel model" mentions pp. 272-275 and 361-363.
P. 362 starts with "predicting a new unit in an existing group", which
sounds like what I want. Now all I need to do is study that book enough
to be able to do what it says.
Reinhold Kliegl wrote:
> I think the Gelman & Hill (2007) book has examples that look less
> complicated to me in comparison to what you describe (i.e. simply
> sample from the estimated distributions). I have some code for
> computation of power, also following examples in this book. Perhaps, I
> am overlooking something.
> On Sun, Apr 13, 2008 at 7:10 PM, Spencer Graves <spencer.graves at pdf.com> wrote:
>> How can I get prediction intervals from a mixed-effects model?
>> Consider the following example:
>> fm3 <- lme(distance ~ age*Sex, data = Orthodont, random = ~ 1)
>> df3.1 <- with(Orthodont, data.frame(age=seq(5, 20, 5),
>> Subject=rep(Subject, 4),
>> Sex=rep(Sex, 4)))
>> predict(fm3, df3.1, interval='prediction')
>> # M01 M01 M01 M01
>> # 22.69012 26.61199 30.53387 34.45574
>> # NOTE: The 'interval' argument to the 'predict' function was ignored.
>> # It works works for an 'lm' object, but not an 'lme' object.
>> One way to do this might be via mcmcsamp of the corresponding
>> 'lmer' model:
>> samp3r <- mcmcsamp(fm3r, n=10000)
>> Then use library(coda) to check convergence and write a function
>> to simulate a single observation from each set of simulated parameters
>> and compute quantile(..., c(.025, .975)) for each prediction level
>> However, before I coded that, I thought I would ask if some easier
>> method might be available.
>> p.s. RSiteSearch("lme prediction intervals") produced 3 hits including
>> 2 from James A Rogers over 3 years ago. In one, he said, "I am not
>> aware of any published R function that gives you prediction intervals or
>> tolerance intervals for lme models."
>> (http://finzi.psych.upenn.edu/R/Rhelp02a/archive/42781.html) In the
>> other, he provided sample code for prediction or tolerance intervals of
>> lme variance components.
>> R-sig-mixed-models at r-project.org mailing list
More information about the R-help