Unrealistic spatial smoothing of richness measurements?

mpinsky commented 9 years ago

Something I've been pondering:

Are the hierarchical models estimating an unrealistically smooth richness surface across each survey? They seem to be predicting nearly constant richness within a survey, and only large differences across surveys. This seems to happen because we have no co-variates included, and yet I am hesitant to base our estimations on covariates that may or may not be correct (and in particular, that may or may not capture the true heterogeneity in the system). I would be especially hesitant to include temperature as a covariate, because that is what we want to test at a community-level, not at an individual-species level.

Some things to think about or check:

Map each region using it's own color bar (probably in separate graphs). It may be that each region looks smoother than it really is, just because the colorbar is stretched so much.
Include latitude as a covariate?
Divide each survey into smaller sub-regions (e.g., include a region co-variate)?
Use a Chao estimator for richness per 1 degree grid cell (e.g., the methods in EstimateS). While a different approach, this is a much faster calculation than the hierarchical model.

rBatt commented 9 years ago

I agree that they're smoothing over space. That's something that needs to be worked on. I'm not inclined to try other estimators. I've been through several, and it didn't seem that the others were useful for our problem. At this time I don't think it's appropriate to keep trying other approaches, this takes a lot of time and I don't take it lightly. I'm going to stick with the MSOM for now.

Your first 3 suggestions in the bullet points are really good, though. Thanks! I'll definitely refer to them when updating the MSOM to include all years. Thanks @mpinsky

rBatt commented 8 years ago

I think we've resolved the underlying cause of this issue. It's just an effect of the MSOM hierarchy. There's not something wrong with the analysis, but it is important to interpret the analysis carefully.

rBatt commented 8 years ago

Reopening because it'll be useful to see if this is resolved when we combine yeras

rBatt commented 8 years ago

I think we're going to have this problem in time and space now that we're combining years.

Binary data are just so hard to deal with. Even in state space framework, you still can't tie the estimates back to the true state nearly as effectively it seems.

For now, I think the thing I'll try is to just include more covariates.

@mpinsky You mentioned substrate type? I wonder if that could find its way into trawlData ... and I wonder what format the data are in. Hoping for a gridded raster ....

mpinsky commented 8 years ago

Yes, a gridded raster exists. Lauren Rogers at NOAA (larogers@stanford.edu) knows which one we used.

On Thu, Dec 10, 2015 at 10:27 PM, Ryan Batt notifications@github.com wrote:

I think we're going to have this problem in time and space now that we're combining years.

Binary data are just so hard to deal with. Even in state space framework, you still can't tie the estimates back to the true state nearly as effectively it seems.

For now, I think the thing I'll try is to just include more covariates.

@mpinsky https://github.com/mpinsky You mentioned substrate type? I wonder if that could find its way into trawlData ... and I wonder what format the data are in. Hoping for a gridded raster ....

— Reply to this email directly or view it on GitHub https://github.com/rBatt/trawl/issues/27#issuecomment-163827224.

rBatt commented 8 years ago

@mpinsky sent the rugosity data. It's been added to the trawlData package.

I also noticed today that surface temperature isn't really very well correlated with bottom temperature. That could be a potential indicator of environmental heterogeneity as well, and it varies over space and time.

Lat and lon could both be used to create spatial differences. And depth.

rBatt / trawl

Unrealistic spatial smoothing of richness measurements? #27