"Mountain Bikn' Guy" <[EMAIL PROTECTED]> wrote in message 
news:<VmjZ9.63329$Ve4.6714@sccrnsc03>...
> My dependent variable fits at least one definition of a time series: "If you
> take a sequence of equally spaced readings, this is called a time series."
> Furthermore, there is very strong autocorrelation (near 1) in the dependent
> variable -- when tested in the order the data is collected. However, I can
> randomly resort all the data (dependent plus independent variables) so that
> there is no longer any autocorrelation and this does not affect the
> predictive ability of the independent variables. So I'm thinking that I am
> not dealing with a time series. Any thoughts?

Time series really just means collected over time. What matters is 
the extent to which the time order results in a correlation structure
in the data.

If you randomly shuffle any time series, *of course* you lose 
the correlation structure between adjacent or near-adjacent elements, 
because the correlated elements are no longer adjacent. Why would you
think that would somewhow mean the original data was uncorrelated?

> Any arguments in favor of using time series analyses?

Because it can take into account the effect of serial correlation?

Note that the "predictive ability" *is* affected in the sense that
a prediction interval for the next observation will not have the
correct coverage probability, because you have ignored the explanatory
effect of the immediately previous observations. 

There is a kind of bias in the following sense: look at your fitted 
values for the observations that follow "high" previous values, and
look at your fitted values for the observations that follow "low"
previous values. You should see that the fitted values in the first 
group tend to be a bit low, and the fitted values in the second
group tend to be a bit high.

> Knowing that I _can_ remove the autocorrelation, 

You haven't removed the relationship to the observations that 
*were* immediately preceeding, you've just made them all different 
distances away now, so that if you use a measure only designed to 
pick up the structured correlation you used to have, it naturally
can't see the correlations that are still there, because you 
changed the structure. You haven't removed the structure, you've
simply hidden it.

Note that your future observations still arrive /in order/, not
shuffled. It doesn't help you to pretend otherwise.

> ... can I proceed to perform
> parametric regression analysis without actually randomly sorting the data
> and treat this as a non-time-series analysis?

You can do what you like. 

Whether it is a sensible thing to do is another matter.

Glen
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to