[R] Questions about corARMA
spencer.graves at pdf.com
Tue Jul 8 09:51:17 CEST 2003
a. Have you studied "Mixed Effects Models in S and S-Plus" by Jose
C. Pinheiro, Douglas M. Bates (2000; Springer)? That book contains
examples that might answer your question.
b. Since you wrote "y(t)" and you say you are new to this list, I
feel compelled to confirm that you know that parentheses "(..)" signal a
function call in S-Plus syntax. I would write your expression "y[t] =
0.03*x1[t]+1.5*x2[t]", etc. (And I would avoid "t", because that is is
the R function for matrix transpose.)
c. I wouldn't try "stepAIC" with "lme" until I was reasonably
confident of the "lme" noise model.
d. From a pragmatic perspective, I would not play with the
correlation structure until I felt I had a reasonable model for the
fixed effects. The reason is simple: Lack of fit can mascarade as a
virtually nonstationary process, even if it is just something
deterministic plus a small amouth of independent noise. For example, if
x is slowly moving like, "x <- 1:99" or "x <- sin((1:99)/(20*pi))", then
"y <- x + 0.01*rnorm(99)" fit without "x" will appear as nonstationary.
e. To understand autocorrelation structure (though not the "lme"
function), I have gotten a lot from the book by Box, Jenkins and Reinsel
(1994) Time Series Analysis: Forecasting & Control, 3rd Edition
hope this helps. spencer graves
> I'm a new member here in the list. I am a graduate from
University of Georgia. Recently in doing analysis using lme
on a dataset, I found several questions:
> 1. How to express the equation when the correlation
structure is very complicated. For exmaple, if the fixed
is y(t)=0.03x1(t)+1.5x2(t)(I omitted "hat" and others). And
the model with corARMA(p=2,q=3) is proper. What will be the
> 2. Is is that any regression error will be stationary?
(Forgive me for my poor math background. This may be a simple
question to most people.) Since corARIMA is not available.
> 3. Why not make a function to automatically select the best
corARMA structure (setting max p and q and the computer takes
care of the rest)?
> 4. When the initial model (without considering correlation
structure) has many variables and some have no significance,
should I use stepAIC first to eliminate some variables? Or,
try corARMA with different combination of p and q, which may
make more variables significant without having to reduce
variable. I prefer the latter.
> 5. If the best corARMA model out of a model still contain
some insignificant variables, I would use drop1 (instead of
stepAIC) and then try all the possible corARMA structures
again. So my steps would be drop1, corARMA, drop1,corARMA,
til I get a model with all variables significant. If the
initial model has many variables, it would be a time-consuming
process. Is it proper to do so? If proper, it'll be wonderful
if a new function is developed to automatically do so.
> Thanks, Hanhan
> Xianglu Han
> 206 Environmental Health Science
> University of Georgia 30602
> Phone: 706 255 2308
> [[alternative HTML version deleted]]
> R-help at stat.math.ethz.ch mailing list
More information about the R-help