estimated abundance error

[esgarbanzo]

Hello,
I'm using NPMR to assess spawning-habitat relationships for salmon. I
am getting good results using redd count data as my response data, but
I would like to be able to show predicted versus observed redd counts.
So, I'm using the "...matrix of estimated abundance.." function under
" evaluate selected model and for all but 3 of my sites it gives
reasonable numbers, but in 2-3 sites it gives -99.99999, which seems
to represent an error of some sort.

Does anybody know what this is caused by? I can't find anything in the
data matrices that looks erroneous.

Thanks!
Jeremy

[esgarbanzo]

OK, I guess I partially answered my own question. It is a missing
value indicator, but I am working with a data set that does not have
missing values. Any idea what's going on?

Thanks again!

[Bruce McCune]

Jeremy, you'll get missing values for estimates when the neighborhood
size for the estimate is smaller than the minimum neighborhood size
for an estimate. Assuming you have selected a particular model, and
you have the same training data set as was open for the model fitting
phase, try this: Run Evaluate Selected Model | Output options tab |
Minimum neighborhood size for estimate | change the method to manual,
then try a smaller value. As this gets smaller and smaller, more of
your missing values will turn into estimates, because you are
demanding less data for a given estimate.

OK, so that should give you the idea of how that control works, but
it is not necessarily the best solution.

Going back to the model fitting phase, one of the controls on the
Model Options tab is Maximum Allowable Missing Estimates (%). Unless
you have changed this, it will be set at the default value. If you
really must have an estimate for each data point, then set that to
zero. Or you can set it to whatever you like. A given model won't
necessarily produce that proportion of missing estimates, but a model
is acceptable if it produces less than that proportion.

Some small fraction of missing values is not necessarily bad. For
example, if you have one point all by itself in a distant corner of
the predictor space, you might get a better model for the remaining
points by not forcing an estimate on that point. It's worth looking
at the data points that turn up with no estimate, to try to figure
out why they seem to have so few similar data points.

Good luck!
Bruce