[R-sig-Geo] Building a prediction raster when the statistical model was built from sampling units of different sizes

[chris english]

I have requested another environmental data scientist  to look at this
thread of conversation as he is better versed in your area of science and
my comments may have been ill-informed or confused /confusing or worse,
both.

But now I understand your research interest, likelihood of disease vector
instances (among striped skunks) given land cover. Is the disease rabies or
some other ailment of concern? I just ask because if rabies, there could be
a large, healthy population given proper resources to sustain them, without
necessarily indicating an areal propensity to rabies. Keeping with rabies
though, you might find you are exploring extremes (here meaning very, very
few are disease vectors) as against general additive models, hence lots of
skunks and relatively very few vectors. Combined in some way that I have
yet to think through, GAM and extreme might lead to the environment more
likely to produce vectors.

So, lets throw land cover under the bus for the moment as it is probably
immaterial, same for area, and ask what environmental conditions are more
likely to generate the vectors of interest, whatever the particular disease
condition is. This indeed would be an interesting study.

Chris
On Jun 9, 2016 18:38, "Nelly Reduan" <nell.redu at hotmail.fr> wrote:

> Hi Chris,
>
>
>
> Thank you very much for your answer. My objective is to build a predictive
> map of capture success of striped skunks at large scale in order to
> delineate areas of high abundance of striped skunks that are susceptible to
> promote disease transmission. My GAM was used to assess the relationship
> between proportions of land cover types and capture success within trapping
> sites. Then, I estimated a capture success value within homogeneous grid
> cells of 2km x 2km from my GAM that was built from capture success data
> associated with trapping sites of different sizes. I also think that the
> approach is quite problematic. So, I am trying to view whether or not there
> are solutions to minimize bias in the predictions.
>
>
>
> Thank you very much for your time.
>
> Have a nice day.
>
> Nell
>
> ------------------------------
> *De :* chris english <englishchristophera at gmail.com>
> *Envoyé :* mardi 7 juin 2016 23:41:09
> *À :* Nelly Reduan
> *Cc :* Help R-Sig_Geo
> *Objet :* Re: [R-sig-Geo] Building a prediction raster when the
> statistical model was built from sampling units of different sizes
>
> Nell,
>
> I'm still trying to understand what sounds to me like an embedded data
> reduction. Please understand, I don't trap skunks (striped or otherwise). I
> have, on many occasions, observed them in the wild but I am cautious not to
> make them scared due to the known, smelly results.
>
> Understand as well that I am not versed in capture success, per se, I just
> examine data and wonder if it contains generalizable or perhaps surprising
> properties.
>
> So if it were me, contrary to my nature and practice of observation,
> trapping striped skunks in a 24 km^2 study area of a given land use/land
> cover, and I trapped 15 skunks over a period of 2 months, I would deem that
> 15 observations for that study area/time period.If, in my 32 km^2 study
> area of slightly different land cover and different resource availability I
> got 70 in two months, I'd have 70 observations. This is how I would view
> it, and quite probably within the accepted science I would be going about
> it all wrong.
>
> As you present the matter, at least as I understand it, the number of my
> hypothetical captures always reduces to one dimension (the study area),
> irrespective of the above variance between study areas. This approach
> puzzles me as it seems that information about desirable resource
> distribution (from the point of view of the skunk) gets lost, and 'capture
> success' becomes murky, at least for me.
>
> As my wife always says, "It depends on what the research question is."
>
> What is the research question in this case?
>
> My apologies to you and capture science if I have completely misunderstood
> as I all too often do.
>
> Chris
>
> On Tue, Jun 7, 2016 at 5:53 PM, Nelly Reduan <nell.redu at hotmail.fr> wrote:
>
>> Hi Chris,
>>
>> Thank you very much for your answer.
>>
>>
>> They are striped skunks that have been captured. In my data, all striped
>> skunks that have been captured within a same trapping site have the same
>> capture success. Thus, each of 50 trapping sites was assigned to one
>> capture success. If, I group trapping sites together, I reduce the sampling
>> size. As the actuel sampling size (50 trapping sites) is rather small, can
>> this cause problem for predicted data estimates?
>>
>> Thanks very much for your time.
>>
>> Have a nice day.
>>
>> Nell
>>
>>
>> ------------------------------
>> *De :* chris english <englishchristophera at gmail.com>
>> *Envoyé :* jeudi 2 juin 2016 06:10:32
>> *À :* Nelly Reduan
>> *Cc :* Help R-Sig_Geo
>> *Objet :* Re: [R-sig-Geo] Building a prediction raster when the
>> statistical model was built from sampling units of different sizes
>>
>>
>> Hi Nell,
>>
>> Just a couple of questions. Trapping sites range from 24km^2 - 236km^2,
>> and there are 50 such sites, looking at the 50 sites might there be a way
>> to bin them reasonably into trap area groups?
>>
>> Your 50 observations suggests one thing was trapped and thereafter
>> trapping was discontinued. Is this correct?
>>
>> And just for general information, what was being trapped?
>>
>> Sticking closer to your data, you might consider GAM(ing) the bins and
>> summing the resultant GAMs. Need to think some more on the predictive
>> raster aspect. Sorry for an essentially inconclusive answer.
>>
>> Chris
>>
>
>

	[[alternative HTML version deleted]]