On 24/07/2020 10:39, Norman Lynagh wrote:
JGD wrote:

On 24/07/2020 08:31, N_Cook wrote:

But obviously not with the likes of Aviso implying , by continuing
to "fit" straight lines, that everything is hunky-dory.

No-one is remotely suggesting that as far as I'm aware (though
linearity is probably the least-worst generic option unless you have
a better model (not arbitrary function) that the data can be fitted
to).

But compounding one piece of arguably bad science (the linear model)
with another piece of bad or worse science (wild extrapolation of a
model with no justifiable connection to the data) is not good, to put
it mildly and lays the results wide open to exactly the criticism I'm
making.

It's the huge extrapolation which is the especially bad part of this.
Different data fits can be tried if you're interpolating values
within the approximate range of the dataset but that's clearly
irrelevant here if the aim is to estimate sea level in eg 2100.

What I'm slightly puzzled about is that there clearly must be
professional estimates of future sea level based on a range of
carefully researched models and which are presumably updated at
intervals. Why not devote your energies to publicising and explaining
these as new updates become available - that would be really
interesting?

Here is one example of the sort of problems that I encounter:

If you have 20 years of hourly wind data from what is normally a
relatively benign location, but the data includes 6 hours of high winds
resulting from the passage of a hurricane, how do you extrapolate the
distribution to predict the once in 50-year wind event? The short
answer is that it is impossible.

Bayesian analysis will get you the best answer provided that you are
able to specify *exactly* what your question is, what prior knowledge
and how much data you have.

The problem is that the winds produced by the passage of the hurricane
are not part of the same population as all of the other (benign) winds
and therefore the total wind environment cannot be described by any
single statistical function. In this case, what is necessary is to
determine the long-term hurricane climate of the location and work
backwards from that to produce an estimate of the once in 50-year wind
event.

Same sort of problem applies to the in service failure of things subject
to "preventative" maintenance but sometimes also expire on replacement
due to infant mortality. The decision of when to replace them to obtain
maximum efficiency is a distinctly non-trivial problem.

Filament light bulbs in awkward locations is the canonical example.

In predicting future sea-level rise what is first of all needed is to
determine the 'climate' of the events responsible for sea level rise.
These include, but are not limited to, simple melting of land-based
ice, expansion of warming sea water, ice-sheet/glacier collapse. In
order to quantify the effects of each of these it is necessary to
predict their frequency and/or rate of occurrence and the range of
their effects. This is not a trivial task! Taking measurements of the
past combined effects of these causes and trying to fit them to some
statistical function and then extrapolating that far into the future is
not sound science. By trial and error it is possible to find a function
that appears to fit the data well and which give the prediction that
you would like to see!

There is unlikely to be enough data to go anything beyond a quadratic
fit and there will be a huge uncertainty in the second order term.

--
Regards,
Martin Brown