Re: Statisically Insignificant

Hi Lou. Hope all is well with you.

But then what is the risk or payoff for time series correlations?

That can be hard to quantify. What is your reward for publishing papers
which make true statements? It's OK not to quantify that; however,
as a consequence you don't have a basis (within decision theory)
for choosing one action over another.

If it's a medical decision, you might want 99.9999% assuredness.
If it's just to show there's some common signal 70% might do.

Well, I suspect that medical decisions are often made with much
less assurance, even in this day and age. Whether or not such
decisions are warranted depends on how bad are the consequences,
which changes by orders of magnitude depending on the problem;
fixed numbers like "95%" just don't enter the picture at all.

Bottom line: Do you know of any way to assess risk or payoff
abstractly. I realize this might be almost a ludricrous question, but
maybe there is some aspect of risk analysis that can be extended in some
meaningful way to such abstract, data-analysis situations.

How much does it cost you to be wrong (or right)?
If you don't have an answer for that, you don't have a basis for
choosing one action over another within the realm of decision theory.
You can, of course, either refrain from choosing (not always possible)
or fall back on some extra-theoretical consideration such as
"The boss said so".

In data analysis, "not choosing" could be something like
presenting the data and computed probabilities such that an
interested party (such as you at a later date) can impose their
own utility function on the problem and make their own decision.

Even without making a decision, it is often possible to make
statements like "Model A is better supported by the data & prior
information than model B". That's probably more interesting and
useful than "Model A is right" when both A and B are known
to be wrong a priori (common enough in data analysis).

For what it's worth,
Robert Dodier

.

Loading