QUESTION:
If anybody has a handy
way of dealing with this problem please let me know (with references please).
I thought this would be straightforward, but
everybody I talk to gives me a different answer. I have two
short time series (of say 15 points) x and y
and I want to test the significance of a regression of y on x.
I need to correct for
the inflated degrees of freedom caused by the non-independence of the data.
The series are too short
for ARIMA modeling, and I don't want to use first differences or detrended
series
because that will eliminate most of the
signal I am trying to capture. Any hints how to do this ? Also, what
can I do if one or both series has missing
data, or if one of the series consists of only a fraction of the
points contained in the other.
ANSWER:
AFS has spent
considerable time and interest in the problem you refer to. Our time series
work uses the
tools of Transfer Functions to identify the
form of the relationship. We extend the Transfer Function to
detect pulses, level shifts, seasonal pulses
and time trends. Furthermore assumptions regarding
constancy of variance of the errors is tested
and the analysis modified to reflect the results of the
hypothesis testing. Finally, gulp !, we speak
to the issue of constancy of model/parameters leading to final
models. I suggest that you download the data
for the CASE STUDY and attempt to use whatever
spreadsheet tool or limited time series tool
like SAS or RATS or SHAZAM that you have access to.
These products have serious limitations as
they relate or don't relate to your problem. AUTOBOX is
geared and focused directly on the problem
you refer to. My opinions are somewhat biased but reflect 30
years of work in time series analysis and
some 21 years of developing expert systems to aid users like
yourself. With respect to the paucity of data
(15 values), I would like to say that you might do well by
setting up a starting model and allowing
AUTOBOX to evolve from there by its rigorous tests of
necessity and sufficiency. When the number of
values are small it sometimes is impossible to be purely
empirical and things can go better with a
nudge in the right direction.
P.S. If you have missing
data this can be treated via INTERVENTION DETECTION or OUTLIER ANALYSIS.
P.S.S The need for
differencing and or time trends can be identified as part
P.S.S.S. There is no
need to adjust for inflated degrees of freedom as the model ultimately
incorporates a
gaussian process thus all Central Chi-Square
tests follow.
CLICK HERE:Home Page For AUTOBOX