CORVALLIS, Ore. – A type of statistical model developed just a few years ago, which so far has found use in such tasks as biomedical research or identifying credit card fraud, also appears to be extraordinarily accurate in predicting the biological impacts of climate changes, a new study suggests.
These “random forest” models – a data analysis technique that has nothing to do with forestry – greatly outperformed five other types of statistical or machine learning models that can evaluate how changes in climate may affect species distributions, Oregon State University researchers concluded.
The findings were just published in Global Change Biology, a professional journal, in a study that could help improve the accuracy of projections about the impact of climate change, and the way temperature or precipitation changes will force shifts in where species can live and survive.
The study was also earmarked as one of particular significance by the Faculty of 1,000, a group of leading scientists that highlights the most important recent papers in various fields of science.
“To anticipate the effects of climate change and identify conservation strategies that might mitigate its impacts, we have to identify the modeling approaches that work best,” said Joshua Lawler, a research associate in the OSU Department of Zoology. “There are dramatic differences in the projections of various models. Given a certain degree of climate change, some models might predict the range of a certain species would double while another model suggests it will be halved.”
“What we know now is that some of these approaches are clearly more accurate than others.”
By far the best approach is a “random forest” model, the study showed, which is actually an averaging approach based on the cumulative predictions of hundreds or thousands of subordinate models. Given current climate conditions and using a test sample of 100 different species of mammals, this type of model, on average, accurately predicted the presence of a species 86 percent of the time and the absence of a species 99 percent of the time.
Approaches shown to be less accurate included generalized linear models, classification tree models, artificial neural networks, and others.
“Many people have suggested that because of the uncertainties involved, these computer models have no value,” Lawler said.
“But in fact, they can be quite accurate, errors can be identified, and we’re continuously working to determine which approaches work best and provide the results that are most dependable,” he said.
There are already some uncertainties involved in the predictions about climate change, the OSU scientists said, and it’s important to not compound that uncertainty when considering biological impacts.
The use of random forest models is quite new, researchers say, only being developed five years ago. Existing random-forest applications have been in areas such data “mining,” prediction of health problems, or analysis of consumer buying patterns – credit card companies can use this approach to identify unusual purchases, contact their customers and ensure that cards are not being fraudulently used.
Until just recently, however, this type of model had never been used in ecology. When it was, it turned out to be extremely accurate – which is good, because many other modeling approaches had proved to be inconsistent and undependable.
“The issues involved are pretty important,” said Andrew Blaustein, a professor of zoology at OSU and co-author on the study. “Knowing how species will adapt to climate changes will be important for protecting them, avoiding extinctions and planning for the changes. Farmers will face an influx of new pest species. And the changing range of animal species could have implications for human health, such as the expanding range of mosquitoes that can transmit malaria or the rodents that carry hantavirus.”
Global temperatures are expected to rise 2.5 to 10 degrees in the next century, the OSU scientists said, and both plant and animal species are already moving. Some bird and insect species are expanding their range toward the polar regions and up mountain slopes, and in some cases abandoning parts of their historic habitat. And some species appear to be breeding earlier in the season.
The ability of models such as this to predict the future will never be perfect, Lawler said.
“When you are talking about biological impacts, there are other variables besides climate to consider,” he said. “At present, these models don’t account for the interactions between species, their ability to move, the timing of certain events, or their ability to adapt to new environmental conditions.”
All of those variables, he said, will have to be considered in addition to climate change alone, and may either mitigate or compound its effects.
“With improved models such as this, we still may not be able to predict exactly where a certain species will go to live, or whether it will survive,” Lawler said. “But we should be able to tell with some accuracy which areas of the world, in general, will be hardest hit in terms of species shifts and biological impacts. That’s pretty valuable information in itself.”
Collaborating on this research were scientists from the Forest Service and the EPA. Funding was provided by the Nature C