Do Bears Snack in the Woods?

Kerry Emanuel is the designated rebutter @ 538, and in an incredibly nice way strips Roger Pielke down to his skivvies before taking a bite

There is an even more significant problem with Pielke’s analysis. In a nutshell, he addresses trend detection when what we need is event risk assessment. The two would be equivalent if the actuarial data was the only data available pertaining to event risk. But that is far from the case; we often have much more information about risk.

unclothing Roger's nous as a political scientist.  Then enter the hungry ursine

Let me illustrate this with a simple example. Suppose observations showed conclusively that the bear population in a particular forest had recently doubled. What would we think of someone who, knowing this, would nevertheless take no extra precautions in walking in the woods unless and until he saw a significant upward trend in the rate at which his neighbors were being mauled by bears?

The point here is that the number of bears in the woods is presumably much greater than the incidence of their contact with humans, so the overall bear statistics should be much more robust than any mauling statistics. The actuarial information here is the rate of mauling, while the doubling of the bear population represents a priori information. Were it possible to buy insurance against mauling, no reasonable firm supplying such insurance would ignore a doubling of the bear population, lack of any significant mauling trend notwithstanding. And even our friendly sylvan pedestrian, sticking to mauling statistics, would never wait for 95 percent confidence before adjusting his bear risk assessment. Being conservative in signal detection (insisting on high confidence that the null hypothesis is void) is the opposite of being conservative in risk assessment.

When it comes to certain types of natural hazards, there are more bears in the woods. For example, there is a clear upward trend in overall North Atlantic hurricane activity by virtually all metrics, over the past 30 years or so, though the cause of this is still uncertain. But given that only about a third of Atlantic hurricanes strike the U.S.; hurricanes do damage during a very small fraction of their typical lifetimes; and only intense hurricanes (a small fraction of the total) do significant damage, the amount of hurricane data pertinent to U.S. damage is a tiny fraction of the entire database of North Atlantic hurricanes. Thus it is hardly surprising that the upward trend in U.S. hurricane damage is of only marginal statistical significance, and Pielke’s own analysis shows that it takes several decades for such a trend to emerge.

As Eli has been saying, on a proposition, 20:1 is really good odds, "statistical significance" doesn't mean that bet is ironclad or that anything less is not worth taking action on, and when you have the odds in your favor and the physics in your favor, double down.