Something I have been bothered about for some time now, is how we best discuss climate change in the context of extreme events. Given the devastation from Hurricanes Harvey and Irma, damaging floods in South Asia and Nigeria, and the potential of more damage from Hurricane Jose, you’d think there would be a clear way to discuss the relationship to climate change that was also consistent with the scientific evidence.
However, there is clear pressure – from some – to avoid discussing the link to climate change because that supposedly politicises catastrophic events (FWIW, I think this argument is itself – rather ironically – political). There are also those (who I probably don’t need to name) who continually point out that the IPCC is clear that we cannot yet detect an anthropogenic climate change signal in many of these events. The problem, though, is that we have some prior knowledge as to how we expect climate change to influence these events and waiting until it is obvious that it has done so, before we take it seriously, seems rather unsatisfactory.
This has been a somewhat lengthy introduction to what I wanted to mention, which is one of James Annan’s posts in which he discusses a new paper by Michael Mann, Elisabeth Lloyd, and Naomi Oreskes called Assessing climate change impacts on extreme weather events: the case for an alternative (Bayesian) approach. Their argument is that rather than using a frequentist approach when trying to assess climate change impacts on extreme weather events, a Bayesian approach should be used.
I’m not an expert at Bayesian analysis, so I’m not going to try and explain it. Instead, I’ll try to explain what I think is wrong with the frequentist approach. This is partly based on a couple of James’s other posts. Essentially, the standard frequentist approach is to assume some kind of null hypothesis and to only reject this if you detect a statistically significant signal in your data. In the climate change context, the null hypothesis would normally be that there has been no change in some type of event. If you do detect some change, then one can try to attribute that to anthropogenic influences (i.e., detection and attribution).
In some sense, it’s a two-step process; detect some signal and then perform some kind of attribution analysis. If you do not detect any signal, then you do not reject the null hypothesis that there has been no change, and the process stops before any attempt at attribution. The problem, though, is that we know that our climate has changed due to anthropogenic influences and we can be pretty confident that this will have influenced, and will continue to influence, weather events. Therefore, the frequentist null hypothesis of no change is immediately wrong and the frequentist test will regularly return a result that is essentially incorrect (i.e., we will conclude that there has been no change even if we are pretty confident that there must have been some kind of change).
As you may have already noticed, though, the obvious other problem is that being confident that there must be some kind of change does not – itself – tell us anything about how it probably did change. Events could get stronger, or weaker. Events could become more, or less, frequent. Maybe they’ll move and start to occur more in regions where they were once rare.
However, in many cases we do have prior knowledge/understanding of how these events will probably change under anthropogenically-driven warming. There will probably be an increase in the frequency and intensity of heatwaves and extreme precipitation events. We expect an increase in the intensity and frequency of the strongest tropical cyclones, even though we might expect a decrease in the overall number of tropical cyclones. It seems, therefore, that we should really be updating our prior knowledge/understanding, rather than simply assuming that we can’t say anything until a statistically significant signal emerges from the data.
So, given that we are confident that anthropogenic climate change is happening and that this will almost certainly influence extreme weather, using a technique that will return no change until we eventually have enough data to detect changes, seems very unsatisfactory. That said, I don’t have a good sense of how to effectively, and properly, introduce a more Bayesian approach to discussing these events. If anyone has any suggestions, I’d be happy to hear them. Similarly, if anyone thinks I’m wrong, or am confused, about this, I’m also happy to hear that. I will add that this post was partly motived by Michael Tobis’s post which takes, I think, a slightly different line.
Assessing climate change impacts on extreme weather events: the case for an alternative (Bayesian) approach, by Mann, Lloyd and Oreskes.
More on Bayesian approaches to detection and attribution, by James Annan.
The inevitable failure of attribution, by James Annan.
Detection, attribution and estimation, by James Annan.
Neptune’s revenge, by Michael Tobis.
I had intended to highight this Realclimate post by Rasmus Benestad, which makes a related argument. Climate change has to impact the probability density function of weather events and, therefore, must impact the probability and intensity of those that are regarded as extreme.