"Gus Gassmann" <[EMAIL PROTECTED]> wrote in message
[EMAIL PROTECTED]">news:[EMAIL PROTECTED]...
> Mountain Bikn' Guy wrote:
>
> > >
> > > 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.
> >
> > My working assumption is that the explanatory effect of any (originally)
> > previous observation is irrelevant. The model does not need to take into
> > account the value of previous observations. Given this situation, I
would
> > think that using time series techniques and methods would be
inappropriate
> > (even though I do collect the data sequentially over time).
>
> Not inappropriate. Perhaps "overkill", perhaps not statistically
significant,
> but certainly appropriate.

Understood. Dave at Autobox made the point that time series analysis can be
considered a superset of standard multiple regression methods.

>
> > > Note that your future observations still arrive /in order/, not
> > > shuffled. It doesn't help you to pretend otherwise.
> >
> > The fact that future values may arrive in order is, again, irrelevant. I
> > will use the model to predict a future value out of order (ie, in the
same
> > way oen would use any linear regression equation) so again, I do not
think I
> > should/could use time series methods.
>
> The reason for using time series models is simple: To make autocorrelation
> work _for_ you, by improving the forecasts. Just ask yourself this
question:
> In predicting the next value, will you be able to make a better forecast
> (with a tighter forecast interval) knowing the current value? If yes, time
> series
> analysis should be used. If not, or if you don't care about forecast
intervals,
> use any method you like.

This makes a lot of sense to me. I do not think we can/will get a better
forecast using other values of Y. Therefore, I am not going to use Y in the
regression equation as I would in autoregression.
>
> > The dilemma is that time series methods seem inappropriate, but there
are
> > also violations of most of the assumptions for using std regression
methods
> > as well unless I take actions to offset these (for example, resorting
the
> > data). Maybe I need to go completely to non parametric methods... ?
>
> Why do I get the impression that you are grasping at straws?
> What assumptions are you talking about, how are they violated,
> and why does re-sorting the data fix whatever problems you got?

Resorting doesn't change anything -- this was just an effort to explain my
situation. What is happening is that I am/was reading too many conflicting
things and therefore, trying to achieve a situation where I would not
violate *any* assumptions of any of my statistical techniques -- an
impossibility, to be sure.But this discussion has cleared some things up and
I now feel comfortable that I do not need to use autoregression techniques
even though I have an autocorrelated time series.

Maybe Dave at Autobox will post some of his thoughts because they were
helpful to me. (Thanks Dave!)

>
>
>


.
.
=================================================================
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