R help - Jan 2004 - spectrum

Dear R users

I have two questions about estimating the spectral power of a
time series:

1) I came across a funny thing with the following code:

data(co2)
par(mfrow=c(2,1))
co2.sp1<-spectrum(co2,detrend=T,demean=T,span=3)
co2.sp2<-spectrum(co2[1:468],detrend=T,demean=T,span=3)

The first plot displays the frequencies ranging from 0 to 6
whearas the second plot displays the same curve but with
frequencies between 0 and .5 although it is based on the same
data (length(co2)=468). Why does the selection (co2[1:468])
determine the vector of frequencies and plot it obviously
incorrectly?

Generally, is it possible to choose the vector of frequencies at
which the spectral density is estimated and plotted?

2) How can the significance of the peaks be tested?


Thanks for any help!!


Andreas


------------------------------------------------------
This mail was sent through IMP at http://mail.unibe.ch

On Thu, 22 Jan 2004, Andreas Pauling wrote:
> I have two questions about estimating the spectral power of a
> time series:
> 
> 1) I came across a funny thing with the following code:
> 
> data(co2)
> par(mfrow=c(2,1))
> co2.sp1<-spectrum(co2,detrend=T,demean=T,span=3)
> co2.sp2<-spectrum(co2[1:468],detrend=T,demean=T,span=3)
> 
> The first plot displays the frequencies ranging from 0 to 6
> whearas the second plot displays the same curve but with
> frequencies between 0 and .5 although it is based on the same
> data (length(co2)=468). Why does the selection (co2[1:468])
> determine the vector of frequencies and plot it obviously
> incorrectly?
It is correct!  co2[1:468] is not a time series: you should be using 
window() to subset time series.  So you supplied a vector which does not 
have a frequency (or a start or an end) and default values are used.

Did you consider actually looking at co2[1:468]?

> Generally, is it possible to choose the vector of frequencies at
> which the spectral density is estimated and plotted?
Wait a minute: spectrum() does not estimate a spectral density but calls 
helper functions to do so.  Please consult their help pages.  The answer 
is a qualified `yes'.
> 2) How can the significance of the peaks be tested?
Not an R question, and not well enough defined for a short answer.  See 
any good book on time series.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595