Re: point me in the right direction for time series analysis?

On 24 Nov 2005 09:47:01 -0800, "amorphia" <spam.ontoast@xxxxxxxxx>
wrote:

> Hi Art,
>
> Sorry, it's so easy to think you've been clear when you have overlooked
> the most basic thing!
>
> I want to analyse each infant separately. So forget for a moment that
> there will be more than one infant and just think about the analysis
> for one infant. The infant plays the game. The target goes round and
> round for several minutes, moving in and out of the capture zone.
> During that time, the infant can be squeezing the sensor, at various
> strengths or not at all. For each baby two time series are generated:
> squeezing, and distance from catching zone.
>
> Given that, I think this paragraph from the original post should now
> make more sense:
>
> I want to demonstrate that the babies squeeze more often the closer the
> object is to the the catching area. In other words, I want to test for
> a correlation between two time series variables which are both strongly
> autocorrelated: distance of target from catching zone, and strength of
> squeeze.
>
> >From what I have managed to glean from the web, although I haven't yet
> managed to get hold of a decent book on time series analysis, cross
> correlation is what I need. I think I had better get a decent book on
> time series analysis!

I have never dealt with data entirely like these, but here
is what comes to mind for me. I know I have seen analyses
that were done this way --

Cross-correlation is a start, and lagged cross-correlations
might be enough to let you illustrate your point. "Testing"
is always a difficulty for time series, so start first with the
illustration.

- You have a series that runs (say) 1-1000 for each variable.
Correlate those, A1-A1000 with B1-B1000.

Now, look at a whole bunch of *lagged* correlations, and
see whether the original correlation is higher than any
of the others, and lies at a peak of correlations.

To use the full N, you can offset-and-wrap: For series B,
renumber 1-990 to 11-1000, and re-use 991-1000 as 1-10.
Correlate the revised set of pairs.

You get 100 correlations if you shift B (or A) by 10 at a time
for a series of 1000. - You could do more, but you probably
don't need more than that.

The tricky part might be: How to set up the file.
Does SPSS give autocorrelations with very large lags?
Does it allow wrapping to complete the variables?
I expect it might do the first, but not the second. So --

I think I would create a file with variables called (say)
A1, B001, B011, B021, ..., B981, B991.

Here, the Bnnn indicates which 'nnn' the lagged series
starts at; each line has one value of A and 100 values of B
with the different lags. The file I want to analyze will
have 1000 such lines (one for each value of A).

I know that in Fortran, I could read in both whole series;
I would duplicate the values of B, so that I had a longer
vector, B(1) to B(2000), to make the wrap-around easy to do;
The first records should have
A(1), B(1), B(11), B(21), ... B(991)
A(2), B(2), B(12), B(22), ...

I would then write out, in a loop for k=1,1000
m_end= k+990
A(k), (B(m), m=k, m_end,10)
- The final "10" tells Fortran to use every 10th value
when looping from k to m_end, which is gives the values
in my example.

I don't know if SPSS allows that syntax on XSAVE
within a loop, but I would try that first. Otherwise,
that vector could be created (new variable) by an
explicit loop before the XSAVE.

Hope this helps.

--
Rich Ulrich, wpilib@xxxxxxxx
http://www.pitt.edu/~wpilib/index.html
.

Relevant Pages

  • ANN: Time Series Analysis Software Tools
    ... STSA, The Statistical Time Series Analysis Toolbox for O-Matrix. ...
    (sci.math.num-analysis)
  • Re: Time series
    ... I might have to take up some projects based on time series analysis. ... could you all suggest me some books which talks more about the ... The good news is it's free ... ...
    (sci.stat.consult)
  • Re: some basic questions on time series analysis
    ... May be you'll can find some answers about time series analysis at ... some persons feel that there exists an attractor for the ... >> following points on the graph of Correlation Dimension vs. Embedded ...
    (sci.nonlinear)
  • Re: Hypothesis testing, two different rates
    ... The correlation between to time series can be computed. ... significance of a correlation coefficient using "standard procedures" ... In specific ARREST RATE is significantly related to OD rate in the ...
    (sci.stat.edu)
  • Re: Statistics of cross-correlation
    ... That won't happen for two "random time series" if it ... you will get some positive correlation, ... Do you think that this would be a valid test? ...
    (sci.stat.consult)
Loading