Kevin Trenberth's latest paper, which appears in Wiley Interdisciplinary Reviews: Climate Change, is uncharacteristically and refreshingly blunt right from the first few words of the introduction:
Humans are changing our climate. There is no doubt whatsoever. There are arguments about how much and how important these effects are and will be in future, but many studies (e.g., see the summary by Stott et al.1) have demonstrated that effects are not trivial and have emerged from the noise of natural variability, even if they are small by some measures. So why does the science community continue to do attribution studies and assume that humans have no influence as a null hypothesis?
That question, the nature of the null hyopthesis in climate research, is what Trenberth wants to change. So provocative is his argument, that the journal commissioned two responses to appear along with his paper and put all three online outside a firewall. Neither of the responses is supportive, but neither do they manage to definitively kill Trenberth's proposal, which is that climatologists should turn their research design on its head. Instead of trying to disprove the idea that humans aren't be blame for a given change in climate, he says they should be trying to disprove that we are.
Too someone not trained in the scientific process, it all seems a bit backward. But the debate goes to the heart of the science at issue, and for those interested in the nuts and bolts of climate research, it's worth exploring. So I'll give it a try.
Before we get into Trenberth's paper, "Attribution of climate variations and trends to human influences and natural variability," it will be necessary to talk about this "null hypothesis" thing. Anyone with a science degree should be able to skip the next two paragraphs.
To start, it's important to understand that science isn't about proving things to be true, but proving some other idea to be false. This is because we can never know for sure if the statement "all swans are white" is true; even if many, many observations fail to turn up non-white swans, there still might be a black swan out there we haven't seen. But it is possible to know for absolute certainty that the statement is false; all we need to do is find one black swan. This "falsification" process is one of the ways science is different from other ways of knowing.
Most science that involves the standard scientific process of testing an idea requires that the researcher establish a null and an alternative hypothesis. The null hypothesis is what the research tries to falsify. The alternative is what's left if the null hypothesis is proven wrong.
So far in climate science, the null hypothesis is that humans are not to blame for climate change. Climate research is designed to test that idea. If it finds evidence that contradicts the hypothesis that humans aren't to blame, then it's falsified, leaving the alternative that humans are to blame. This sort of conclusion is accompanied by a degree of confidence. In most cases, the hull hypothesis is rejected if the results shown there is less than a 5% chance that it's correct.
But, says Trenberth, science has been pretty clear for a while now that humans are to blame, in general terms, for the global average temperature rise since the advent of the industrial revolution.
Given that global warming is 'unequivocal', and is 'very likely' due to human activities, to quote the 2007 IPCC report, should not the null hypothesis now be reversed? Should not the burden of proof be on showing that there is no human influence?
To Trenberth, a climatologist at the National Center for Atmospheric Research in Boulder, Colorado, this is not an abstract argument. There are some very real negative consequences of continuing to do things the way they've been done. Basically he argues that the existing null hypothesis and the nature of the changes the planet is experiences make it too easy to falsely conclude that humans aren't to blame (what's called a Type 2 error), and so in places where climate change is already having a serious effect (mostly the developing world), scientists aren't able to tell politicians what they need to be hearing: that things are bad and we need to do something about it.
Georgia Tech's Judith Curry, the first rebutter, is having none of it. The null hypothesis that we aren't changing the global climate is "trivially false," she writes, and so impossible to disprove. In fact, the whole idea of a null hypothesis is pointless when it comes to attributing causation for any specific change in climate, so she suggests just ditching the idea together in favor of a scientific exploration of
The key scientific question [which] is the importance of human influences relative to natural modes of climate variability.
Curry goes further, suggesting that Trenberth has strayed from science into the real of politics:
... the statement appears less about scientific analysis than about policy and winning a battle against the 'deniers' and reluctant politicians. In this sense, his essay comes across as a polemic. Trenberth is using the idea of reversing the null hypothesis as a metaphor for changing the political balance in the climate change debate.
... such strategies are likely to exacerbate skepticism and inflame the political debate, which can be counterproductive.
Finally, we have Myles Allen of the University of Oxford, who, while devoting a fair bit of his take on the subject to a sympathetic airing of where Trenberth is coming from, eventually decides that it's not yet to time to reverse course.
Only when the signal of anthropogenic influence on extreme weather becomes overwhelming (which looks to be a long way off at present) will it make sense to assume human influence has increased the odds of any weather event that occurs.
Allen agrees with Curry that what we're talking about is essentially a political issue, one revolving around the Precautionary Principle. But Curry's solution, argues Allen, is worse that Trenberth's.
Curry argues that because framing a scientific question in terms of hypothesis tests makes it very important where the burden of proof lies and deciding that is not a purely scientific question, the solution is to abandon hypothesis tests. This seems to be throwing the baby out with the bathwater. She questions whether the IPCC's statement 'most of the observed increase in global average temperatures since the mid-20th century is very likely due to the observed increase in anthropogenic greenhouse gas concentrations' is 'useful' because it assigns a probability ('very likely', meaning less than 10% chance that the statement is false) to an imprecise statement, 'most of the observed warming'.
There is nothing imprecise about 'most': it means more than half. As it happens, this wording was introduced to replace the (vaguer but more evocative) phrase 'contributed substantially' in a nice example of the IPCC review process making its conclusions both more specific and less emotive. As Curry observes, an infinite number of statements could have been made, ranging from 'it is extremely likely that the anthropogenic increase in greenhouse gases has caused some warming' (not very informative, since an infinitesimally small warming is of no policy relevance) to 'it is about as likely as not that greenhouse-gas-induced warming exceeds the total observed warming' (which indicates the size of the greenhouse signal, but understates our confidence in attribution). Far from being a 'poor choice', in Curry's words, 'most' was chosen for precisely the reasons she advocates: large enough to be policy relevant, while small enough for the null hypothesis 'not most' to be rejected at an informative confidence level.
Allen is probably correct when he writes that Trenberth's proposal is unlikely to find a much support among his peers. By raising it, though, Trenberth may help climatologists around the world, including those charged with overseeing the next assessment from the IPCC, explore more thoughtfully the connection between how science is framed and the political context in which it will be presented. And I suspect that's probably what Trenberth was really hoping for.
Trenberth, K. (2011). Attribution of climate variations and trends to human influences and natural variability Wiley Interdisciplinary Reviews: Climate Change, 2 (6), 925-930 DOI: 10.1002/wcc.142
How about, "all of the visible swans in my kitchen are white". Could you prove that?
Bill, you would never get funding for that project!
I disagree that science is all about proving other ideas to be false. The reason is twofold: one is that there are an infinite number of hypotheses that "compete" with any given hypothesis, and the likelihood that the one you are showing up as competing is the right competitor almost zero. So disproving one or the other means quite literally nothing unless you have good reasons for positively preferring competing hypotheses. Hence you have to show that they are preferable choices, and so science is at least about demonstrating some hypotheses.
Second, science has always had as a major task showing that hypotheses are correct, since day one. We may have logical concerns about validation of hypotheses, but we have always done it and despite Popperian issues, always will. In like fashion, we generate our generalisations through induction. If science were all about falsification, it would grind to a halt in minutes.
Judith Curry is exactly right in this case (though her mis-analysis of the last ten years of the BEST data is frighteningly bad).
The null hypothesis that humans have no effect is trivially false. We are surely influencing climate, if only in the tenth decimal point of mean temp. The important scientific question is not answered by presenting data that rejects such an always-false null hypothesis. We can reject such a null hypothesis at whatever p-value we choose, just by making our data set large enough.
The real question is "How big is the effect?" not "Is there an effect?" This is a parameter estimation problem (with confidence intervals expressing the statistical uncertainty of the results), not a null-hypothesis-testing problem.
I feel like any "official" declaration of AGW as the null hypothesist would give denialists fodder â not that that should be a strong factor in decision-making, of course.
Coming from Australia where swans are black... well, we do like to have our choices served up as black or white. I don't know where the tipping point is that takes us from having to prove AGW is real to that being the null hypothesis lies but going by the cream of our scientific institutions it does seem to have already been passed. I certainly believe that energy and emissions policies should not be based on the assumption that AGW is an unproven hypothesis.
It is not a matter of hypothesis-testing!!!! It is a matter of estimating the magnitude of anthropogenic effects.
"The real question is "How big is the effect?" not "Is there an effect?" This is a parameter estimation problem (with confidence intervals expressing the statistical uncertainty of the results), not a null-hypothesis-testing problem."
And the answer is: Between 2 and 4.5 C per doubling of CO2, with a most likely answer around 3C per doubling.
Now, given that this has been answered, what's next?
Which way the null hypothesis goes matters because we construct statistical tests differently depending on this sort of thing. But, there are a few things that I'm not comfortable with.
First, there is John's point. Science doesn't really do the scientific method. More to the point, when we do, it is for specific purposes. A woodsman uses a chainsaw to cut down the forest, but if all you had was a chainsaw you wouldn't be able to cut down very many forests. The hypothesis testing part of science is a bit mounted in a drill as part of a toolkit.
There can not be a null hypothesis that says "humans create climate change" or one that says "climate change is not human caused" because those two sentences are crappy hypotheses. A hypothesis has to be more refined, narrowly defined (usually) and it has to refer to specific things changing or being different under different conditions or treatments as measured a certain way.
Made up example:
Observation: CO2 sampled at 1000 feet over two years in northern Wisconsin is highly depleted in C-14. Null hypothesis: C-14 in the sampled gas is at X percentage in accordance with expectation of gas cycling through the upper atmosphere and the conversion of N in to C-14 by cosmic rays. The study shows; The data show less C-14 than expected. The assumption is that less C-14 would occur if Y percent of the C is from fossil sources. The null hypothesis, that C-14 would be at a certain level, is rejected. The implications of the test, not the test itself, lead us to conclude that it is most likely that C is being added to the atmosphere from an ancient (C-14 depleted) source.
At a higher level, a concept like "null hypothesis" has value, but it is now an expanded version of actual null hypotheses that one would actually form and test and put in a paper. It is more of an overarching or amalgamated idea or set of suppositions or assumptions that the preponderance of evidence leads us to be comfortable with.
The preponderance of evidence tell us that ancient carbon released into the atmosphere by humans has warmed the planet. Any research program investigating atmospheric gasses, temperatures, climate forcing for the recent past, present, and future needs to be designed with this assumption. Metaphorically, we have a "null hypothesis" that human generated greenhouse gasses are warming the atmosphere. That's reasonable and appropriate. But it is really more like a baseline assumption.
There are some types of studies (or really, communications about what is known) that would definitely benefit from what Trenberth is suggesting. For example, we have a very good measurement of the magnitude of global warming over the last 30-40 years. So whenever someone asks the (silly) question "yes, but what's the rate since 1998/2005/last year" it should only be answered in the form "Is it different from the longer known trend" - that is, the null hypothesis is what we already know, not some artificial lack of knowledge. This is a Beysian viewpoint.
This approach is not necessarily useful in all cases, but it seems remarkably applicable to some of the worst areas of denial (the temperature record, the sea level record, etc.)
Greg, it is not true that null hypotheses of this kind have value. Please look at my comment 4, or any advanced statistics text. Such a precise null hypothesis is guaranteed to be false if our data set is large enough, and rejecting it means absolutely nothing interesting.
"There can not be a null hypothesis that says "humans create climate change" or one that says "climate change is not human caused" because those two sentences are crappy hypotheses"
They're also not null, they are proscriptive, just in the negative sense. It's saying THERE IS SOMETHING doing it, "not-humans".
A null would be "there is no change in climate".
But then you have to make your hypothesis. E.g. "The sun did it". Estimate a mechanism, "GCRs being deflected by solar magentosphere", then work out the change that would occur from that and test that.
If you can't accept that method to 95% confidence (i.e. can't definitely reject your hypothesis), then your hypothesis is not proven. Null hypothesis there is all about the rejection of the hypothesis made.
But to test against the hypothesis "humans aren't doing it", you have to check that there's less than a 5% chance of humans are doing it and, if you can't reject "humans aren't doing it" to a 95% confidence limit (e.g. accepting a greater than 5% chance that humans ARE doing it), then you have not proven that hypothesis.
Oddly enough, that would make it impossible to reject humans doing it. It's certainly a much greater than 5% chance humans are causing climate change now.
Opposite to what deniers complaining about the null hypothesis want to happen.
Bill, one could prove that all the visible swans in the kitchen are white, but such "proof" has little scientific value or merit.
We use science to learn about the world around us, beyond what our eyes can see or our hands can grasp. The predictive value of our past observations is critical. Based on your proof, we would have no idea if the neighbor's kitchen had all white swans or even any swans at all. Knowing your kitchen is full of white swans has no value beyond a passing mention over a cup of coffee.
We need to learn what is likely and what aspects are related to each other. Think of going in to a doctor's office with a set of symptoms. The doctor needs to know what ailments are vastly more likely to produce those symptoms and treat those ailments first, rather than treat something that may affect 0.1% of the population.
If ACG is true, then the climate is changing due to humans. If ACG is not true, however, then the climate is still changing (since change over time seems to.be the nature of climate), albeit not due to humans. Speaking as a non-scientist, this raises questions. For example, if past climate change is not completely circular, how can we ever know that human activity is causing climate to deviate from baseline changes? If climate change is linear then from where to obtain a baseline to show deviation is occurring? It is an honest question, and I am not suggesting we cannot know; I just don't understand how.