[R-sig-Geo] Location level analysis of spatio-temporal data

[chris english]

Sai,

There is something like take it from the end to the beginning and back, or
beginning to end and back to
the beginning and end and back, and ask yourself what does my data
recommend. The end is how I suppose you
generally anticipate analyzing/modeling the data, and the beginning being
your understanding of how the data
was collected and its vagaries.

If you sampled your data, say .60 in a train set and reduced each variable
to its quantile median
     quantile(your_data[[x]][, 'your_one_of_many_variables'], type =8)[[3]]
# without na.rm = TRUE anticipating running this through caret 'gbm'
and then examine influence of time as against space among your variables.

Space can be characterized in a lot of dimensions, and then you have time
which seems often to be like a sampling rate upon those
potentially many spatial characteristics, But give it a try anyway and find
if time isn't up there in position one or two in variable importance with
spatial variables after you run your data through gbm.

For further reading I would suggest the work of Emmanual Parzen and also
his work in collaboration with Subhadeep (Deep) Mukhopadhyay
on why quantile median might be your special friend.

As ever, I probably shouldn't comment as I know little and there are much
better informed scientists here. The interesting claim to be
wrestled from the Parzen/Mukhopadhyay material is that the data (that you
have) informs the sufficient statistics to be found and that specific
domain knowledge is not necessary to such. This, is the claim, is the power
of the quantile median analysis.

Does this relieve you of answering your question of whether to apply
spatio-temporal upon the whole set, or time series upon a sampling point;
hard to say but Parzen/Mukhopadhyay say, give me your data and I'll give
you your sufficient statistics. It will, I suspect, at the very least
confirm
or disprove the proposition that your data is spatio-temporal (since you
don't say what it is) as a received notion, which is a good starting point
in any case.

My thoughts,
Chris


On Sat, Apr 28, 2018 at 11:16 PM, Sai Kumar Popuri <saiku1 at umbc.edu> wrote:

> Hi,
>
> I am new to spatio-temporal analysis and trying to understand some basics.
> Suppose I have a large spatio-temporal data. I can either fit an advanced
> model with a spatio-temporal covariance structure or I could fit a time
> series model at each location separately. When are these two approaches
> similar? When is the second approach justified?
>
> Thank you,
> Sai
>
>         [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-Geo mailing list
> R-sig-Geo at r-project.org
> https://stat.ethz.ch/mailman/listinfo/r-sig-geo
>

	[[alternative HTML version deleted]]