Web landshape.org

First Time At Niche Modeling?

This is a blog on the power of numeracy. My first book — Niche Modeling — is now in print.

The first six chapters are tutorial topics in R programming and theoretical topics in niche modeling: functions, data, spatial, topology, environmental data collections, and examples. The last six chapters are about using niche modeling to detect errors: bias, autocorrelation, non-linearity, long term persistence, circularity and fraud - useful information for any biological modeler.

About Niche Modeling

Filed under: Uncategorized — admin @ 12:00 am

I have always been fascinated with prediction.

As an undergraduate I made stock predictors on the first PCs and lost money in 1987.

Studied maths, statistics and started a PhD in ecological prediction.

Developed betting systems and lost money.

Studied algorithms for predicting species distributions and developed GARP which other people used for cool things like finding new species of Gecko in Madagascar.

Developed automated trading systems for FOREX in 2002 and lost money.

So I know a few things about prediction, and more about how not to do prediction. In addition, in this blog I hope to pass on a few, and help people to predict better. Like predicting the risk to poultry from Bird Flu using GIS spatial analysis. Or monitoring the health of different types of hydrocoral polyps on reefs. The possibilities are endless.

There is a new thing in business called the Predictive Enterprise, based around not only system management, but proactive management, with decisions based on predicting potential demand. There are prediction markets, where groups trade on events such as when Microsoft will release Vista, or the number of hurricanes in a season.

There are grand prediction enterprises, such as predicting climate a hundred years into the future and thousands of years into the past. One of these, the reconstruction of temperatures for the last 2000 years, has just seen a major reversal, with the release of a National Academy of Sciences report on page 107 diplomatically stating:

“Some of these criticisms (by McIntyre and McKitrick) are more relevant than others, but taken together, they are an important aspect of a more general finding of this committee, which is that uncertainties of the published reconstructions have been underestimated.”

We have yet to see what the flow-on will be to that other grand enterprise, the prediction of future temperatures due to burning fossil fuels.

Overall, I would say people generally have exaggerated confidence in their predictions. I have been working on ways of validating models that include more of the uncertainties, such as Monte Carlo methods of normalizing for known biases. However, generally, overconfidence is the norm.

There is a blog.

There is prediction software.

And because prediction means being aware of the present, a controversial topics aggregator.

Enjoy.

2 Comments »

  1. Hi, I’m interested in doing some niche modelling for a species that I’m studying. I’ve casually followed the use of ecological niche modelling and GARP over the past few years and am now starting to read up about it more. The more I read, however, the more I realize there are so many other newer alternative predictive tools available, such as the new WhyWhere algorithm or some of these novel techniques evaluated in the Elith et al. 2006 paper (e.g., MAXENT, BRT, mars-comm, etc.). It’s difficult to know which one to use? How would you go about deciding which way to go? If I felt that they all were very useful, I might feel more comfortable just sticking with DesktopGARP, but that was one of the lowest performing methods according to the Elith et al. paper. Any advice? Thanks!

    Comment by Rob — July 14, 2006 @ 10:00 pm

  2. Good question Rob but the answer has a lot of facets that I will try to do justice to next week in a longer reply. Think about the assumption in your question “It’s difficult to know which ONE to use?”. There are a number of dimensions across the models you might want to explore - classical vs heuristics, new vs. well researched, plus they more rightly exist in a benefit/cost framework, i.e. an optimal method in your situation might be something really simple that is average.

    I think the OpenModeller group found a bug in the rule archiving part of DesktopGARP that increased the accuracy of the GARP algorithm to the medium group along with GLM and some others. You might want to look into that version but they would have the full story on that though. Then there is the fact that you get higher accuracy with a broader range of variables, particularly monthly temperature and rainfall instead of annual averages reported in S06. Time could be more efficiently spent getting predictor variables optimized.

    I also don’t think the validation process used in Elith et.al. tells the full story either. An in-range random sample from highly autocorrelated data does little more that test goodness of fit. The real test of a niche model is on out-of-range problems, such as detection of new species or invasive species problems, of which there are many documented successes for DesktopGARP. Then there are the types of model not examined, based around k-means, such as reviewed here. I am not saying it is a flawed study, but that it is not complete.

    I am the first to acknowledge that results can sometimes be dissapointing. Its not an exact science yet. You have to do a lot of testing.

    Comment by davids — July 14, 2006 @ 10:18 pm

RSS feed for comments on this post. TrackBack URI

Leave a comment