search

PalEON-Project / Shiny

Code for R Shiny App which will allow us to share the stats composition maps.
0 stars 1 forks source link

Making 'absence' clearer #10

Open SimonGoring opened 9 years ago

SimonGoring commented 9 years ago

With the plotting a lot of it shows constant presence. Is it worth rebuilding the binomial model used elsewhere to add a predicted absence layer to the data, where we just clip using <- NA for cells with predicted absence?

Can we set an arbitrary threshold for values? Say < 0.05? This is an issue I've sort of run into elsewhere with this dataset and clearing it up would be helpful.

@paciorek

paciorek commented 9 years ago

I don't understand - the stat model makes predictions and those are never 0 but can get very close to zero. It's hard to fit a stat model and predict exact zeros.

What binomial model are you referring to?

We can of course clip out cells that are water or outside our core states.

On Mon, Jul 27, 2015 at 2:39 PM, Simon notifications@github.com wrote:

With the plotting a lot of it shows constant presence. Is it worth rebuilding the binomial model used elsewhere to add a predicted absence layer to the data, where we just clip using <- NA for cells with predicted absence?

Can we set an arbitrary threshold for values? Say < 0.05? This is an issue I've sort of run into elsewhere with this dataset and clearing it up would be helpful.

@paciorek https://github.com/paciorek

— Reply to this email directly or view it on GitHub https://github.com/PalEON-Project/Shiny/issues/10.

SimonGoring commented 9 years ago

I was referring to the two-step model that Xiao had built. I though there was a binomial GAM as the first step that predicted whether a tree was there or not, and then a second GAM over top of it that predicted continuous biomass.

I don't mind clipping out water, it's just that when you look at the map for Beech (for example) we know that there's really no beech in most of Minnesota, it would be nice if we could make those cells gray, and not part of the continuous color scale for composition. Presumably those values are very close to zero, but not actually zero.

paciorek commented 9 years ago

the composition model is very different in structure from biomass model, so there's nothing like that that we can make use of in plotting composition estimates.

The question is how to determine where we 'know' things are zero to override what the model estimates based on the data. My inclination is just to report what the model estimates, but I may have my statistician blinders on.

On Fri, Jul 31, 2015 at 10:01 AM, Simon notifications@github.com wrote:

I was referring to the two-step model that Xiao had built. I though there was a binomial GAM as the first step that predicted whether a tree was there or not, and then a second GAM over top of it that predicted continuous biomass.

I don't mind clipping out water, it's just that when you look at the map for Beech (for example) we know that there's really no beech in most of Minnesota, it would be nice if we could make those cells gray, and not part of the continuous color scale for composition. Presumably those values are very close to zero, but not actually zero.

— Reply to this email directly or view it on GitHub https://github.com/PalEON-Project/Shiny/issues/10#issuecomment-126752499 .

SimonGoring commented 9 years ago

I agree with this in principle.

There are several reasons we might want the Shiny app to show 'zero' values, and where a person might want a 'zero' cutoff. For one, from the standpoint of a forest ecologist it is potentially confusing to see near-zero values of beech in western Minnesota. Does that mean there is some, or is it just the model?

The other situation is where we're trying to decide range margins from the posterior data. Where is the margin when it just peters out to a background value of 0.01?

On Fri, Jul 31, 2015 at 1:50 PM, Christopher Paciorek < notifications@github.com> wrote:

the composition model is very different in structure from biomass model, so there's nothing like that that we can make use of in plotting composition estimates.

The question is how to determine where we 'know' things are zero to override what the model estimates based on the data. My inclination is just to report what the model estimates, but I may have my statistician blinders on.

On Fri, Jul 31, 2015 at 10:01 AM, Simon notifications@github.com wrote:

I was referring to the two-step model that Xiao had built. I though there was a binomial GAM as the first step that predicted whether a tree was there or not, and then a second GAM over top of it that predicted continuous biomass.

I don't mind clipping out water, it's just that when you look at the map for Beech (for example) we know that there's really no beech in most of Minnesota, it would be nice if we could make those cells gray, and not part of the continuous color scale for composition. Presumably those values are very close to zero, but not actually zero.

— Reply to this email directly or view it on GitHub < https://github.com/PalEON-Project/Shiny/issues/10#issuecomment-126752499>

.

— Reply to this email directly or view it on GitHub https://github.com/PalEON-Project/Shiny/issues/10#issuecomment-126782668 .