Those skeptical about the science of global warming have frequently alleged that the surface temperature record has "paused" since about 1998 and showed no significant warming trend (see this recent wattsupwiththat post as a representative example).

The mainstream climate science community has responded to this in several ways (see my italics in the quotes below highlighting the different claims).

Some deny that the pause exists or argue that recent extensions to the record show it to be a data error, see this news story from the Independent claiming:

A new study has found that global temperatures have not flat-lined over the past 15 years, as weather station records have been suggesting, but have in fact continued to rise as fast as previous decades, during which we have seen an unprecedented acceleration in global warming.

But other mainstream scientists accept the pause exists and seek explanations. A recent review in Nature starts with this admission:

For several years, scientists wrote off the stall as noise in the climate system: the natural variations in the atmosphere, oceans and biosphere that drive warm or cool spells around the globe. But the pause has persisted, sparking a minor crisis of confidence in the field. Although there have been jumps and dips, average atmospheric temperatures have risen little since 1998, in seeming defiance of projections of climate models and the ever-increasing emissions of greenhouse gases.

but continues

...Now, as the global-warming hiatus enters its sixteenth year, scientists are at last making headway in the case of the missing heat.

So some people don't believe there has been a pause and others are trying to explain the pause. My question is given the uncertainties have global surface temperatures shown no significant growth since about 1998?

Note for clarity I realise that most heat is not absorbed by the atmosphere and that climate change may well be continuing. That is relevant context but not the question. The question is about surface temperatures. So answer that first and provide context later.

  • 1
    @Tor-EinarJarnbjo The met office question is far too specific and didn't lead to good answers on the general question. This question is related but should allow scope for proper analysis of all the data and all the claims about it.
    – matt_black
    Jan 26, 2014 at 18:29
  • 1
    @GlenTheUdderboat I can't access your pdf. But I think answers will need to provide some good discussion on statistics, smoothing and noisy data and justify "significance" if they choose to use the term.
    – matt_black
    Jan 26, 2014 at 18:55
  • 2
    It's possible for both papers to be correct; They don't necessarily contradict each other. If both are correct, the atmosphere is warmer than we think it is and the deep oceans have absorbed more heat than we expected them to and therefore we are underestimating the rate at which global warming is taking place (at least over the last 16 years). Because the papers can both be true their accuracy should probably be queried in two separate questions. Since both papers were published last year, it's unlikely there have been any independent studies covering the same ground yet.
    – Ladadadada
    Jan 27, 2014 at 0:05
  • 2
    @Ladadadada I'm not particularly interested in the two papers: i simply used some representative cases to illustrate the point that people don't seem to agree on the data. What I'm interested in is what the data says and that has been discussed in many papers. What's the consensus and does it agree with the statistics?
    – matt_black
    Jan 27, 2014 at 0:22

2 Answers 2


The first thing to point out is that "no statistically significant warming" does not mean that there has been no warming, essentially it just means that there hasn't been enough warming to rule out the possibility that there has been no warming. If that sounds counter-intuitive, it is because it is, but that is the way frequentist statistical hypothesis testing works.

The way frequentist hypothesis tests work is broadly as follows: Say you have a hypothesis (H1) that you wish to support using a set of observations (X). Next you define a "null hypothesis" that is basically what you need to show to be false in order for your H1 to be true. For example, if you hypothesise that there has been some warming, then the obvious choice for H0 is that there has been no warming at all, i.e. the rate of warming is zero. You then calculate the p-value, which is the probability of observing a trend at least as large as that observed IF H0 is true. If the p-value is sufficiently small, say p < 0.05, this is taken as sufficient evidence that H0 is false so we say that "we reject the null hypothesis" or equivalently "the rate of warming is statistically significant" and otherwise "we fail to reject the null hypothesis" or "the rate of warming is not statistically significant".

Now the first point to notice here is that H0 should be the hypothesis you are arguing against. So for mainstream science, which suggests there will be warming due to the greenhouse effect, the natural H0 is that there is no warming. The "Skeptics" on the other hand hypothesise there is no warming, yet they are using that as their null hypothesis as well. This is a grave statistical error as it means that hypothesis testing no longer functions as a sanity check, as the skeptics are assuming that they are right and requiring evidence to prove them wrong. Mainstream science on the other hand are assuming that they are wrong (H0 is true) and asking if the observations refute H0 (implying, but not proving that H1 is true).

Now for the second point. If the trend is not statistically significant, there are at least two reasons: Firstly H0 actually is true, and secondly H0 is false, but there is insufficient data to demonstrate that it is wrong. Consider flipping a two-headed coin four times. The traditional test for the bias of a coin will fail to reject the null hypothesis as even getting four flips in a row will happen by chance with a fair coin more that 5% of the time. This is because the power of the test (the probability of rejecting the null hypothesis when it is actually false) is not very high.

This is the case for the "not statistically significant" observed trend we have now, given the expected size of the anthropogenic trend and the noise in the data (weather), the power of the test is so low that it is not at all surprising that the result is not statistically significant. Easterling and Wehner have demonstrated that the climate will occasionally show decadal (or more) periods with little or no trend, and that this is also found in model simulations.

To add to this, the hypothesis test assumes that you are looking at an n-year period chosen at random. If you cherry pick the start and end dates, the power is even lower, unless you compensate for the implicit multiple hypothesis testing.

The quote from the Independent does not show that it is a "data error"

A new study has found that global temperatures have not flat-lined over the past 15 years, as weather station records have been suggesting, but have in fact continued to rise as fast as previous decades, during which we have seen an unprecedented acceleration in global warming.

Saying that temperatures have not "flat-lined" is not incompatible with the rate of warming not being statistically significant, because the latter just means we cannot rule out the possibility that the underlying rate of warming is zero. The problem is that most journalists, and an even larger proportion of climate skeptic bloggers don't really understand hypothesis testing.

Saying that the rate of warming is the same as that before is not incompatible with the rate of warming being not statistically significant either for much the same reason.

The comment about acceleration needs a bit more evidence though.

The pause in warming is interesting, it is well explained as a result of the effects of ENSO (see the paper by Foster and Rahmsdorf), and it is providing an interesting area for research in climate variability. This does not however mean that the underlying rate of warming has changed, or that carbon dioxide is not a greenhouse gas etc. So the two views are not actually contradictory.

To give a direct answer to the question, whether the warming is significant or not depends on the dataset you look at, how you choose the period in question, your statistical assumptions (e.g. taking into account autocorrelation and multiple hypothesis testing due to choosing the period after looking at the data etc.). Even then, it doesn't necessarily mean much unless you also look at the statistical power of the test.

  • 1
    I like the point about cherry picking the starting point but it cuts both ways since claims of a strong trend in previous periods are also subject to the criticism. Is there a clear statistical way to adjust for this in general?
    – matt_black
    Jan 28, 2014 at 0:13
  • 3
    @matt_black regarding the starting point for a period of a strong trend, I presume you mean the 1970s? This is not necessarily cherry picked as (i) statistical procedures will suggest a change point at that time (e.g. tamino.wordpress.com/2010/08/13/changes) and (ii) it was predicted before it happened (the initial concern about future climate change originated at time when it wasn't really warming in the late 60s and early 70s) - you can't cherry pick the future! However, if you pick a period long enough (climatologists typically use 30 years) the test has substantial power and ...
    – user18604
    Jan 28, 2014 at 8:12
  • 2
    cherry picking begins to have little effect on the outcome. Climatologists use such long periods as it means the effects of internal variability (e.g. ENSO) average out and have a (more nearly) negligible effect of the estimate of the trend. The simplest way of compensating for the cherry picking is to simply use a long enough period that it doesn't matter. The other approach is to make an adjustment for multiple hypothesis testing, but that is difficult. I would probably use simulations on AR data with similar properties to estimate the effect, I'm not sure there is an analytic solution.
    – user18604
    Jan 28, 2014 at 8:15
  • 2
    "can produce some ideas from trusted statisticians" Firstly tamino is a statistician, being one myself I can recognise he is also a rather good one. Secondly I find discussing science with those that use ad-hominems in place of rational argument (saying the stats are "bogus" is not rational argument) to be a waste of time. If you want to explain the specific errors in Taminos post, then do so.
    – user18604
    Feb 3, 2014 at 9:02
  • 2
    By the way the 1979 isn't a cherry pick, it is the date that satelite observations became available, so if you want to run the same analysis on multiple datasets (so as to be seen not to cherry pick the datset) you have to use a period that actually exists in all of the datasets. Of course if you had actually read the post, you would know that "Let’s use temperature data starting in 1979 (so we can include satellite data for the lower troposphere)".
    – user18604
    Feb 3, 2014 at 9:28

In this answer, I focus on the Cowtan and Way paper, which seems to be causing some of the dynamics of this debate (e.g., the The Independent article mentioned by you).

I am assuming that when you speak of 'significant', you mean 'statistically significant'. There is another meaning to the word, and if you meant that, then this answer isn't very useful. I'm also assuming that you would like to take "no upward trend" as your null-hypothesis.

Given the amount of extrapolation done by Cowtan and Way, I suggest that their analysis is not suitable to answer your question in the negative [this is awkwardly phrased because you already have a 'not' in your question], but rather more aimed at providing a trend (i.e. point) estimate only. (Indeed there doesn't appear to be a claim about significance in either the abstract or the conclusion of the paper.)

However, in their paper (p. 11) they do provide us with:

Dataset          Trend +/- sigma
Hybrid s = 1.0   0.119 +/- 0.076

which might be used to answer your question in the positive (at least if this dataset/period is the only data admitted).

With some further assumptions, that translates to a p-value of about 6%. Given all the extrapolation going on, I'd suggest that objectively sigma should have been estimated higher, and therefore I'd suggest that the p-value is also higher. I don't know what significance level (to which the p-value should be compared) is conventional or justified in this area, but I wouldn't be surprised if it were 5% or less.

Summary: The Cowtan and Way paper isn't, and doesn't provide reason for us to be, conclusive either way with regards to your question. (That is: It can't give "significant" and it can't give "not significant".) If their data/period would be the only available, then their analysis would suggest: no significant upward trend.

I have noted that such things are hotly debated. Perhaps it is good to state my position. I don't care. (And I don't follow this debate.)

  • 2
    I am unsure of this: if one experiment is inconclusive, then we should look at other papers for answers -unless this is the only relevant paper, in which case your answer should note that.
    – Sklivvz
    Jan 26, 2014 at 21:56
  • There isn't any data mining involved in C&W (as discussed in chat), but +1 otherwise.
    – user18604
    Jan 27, 2014 at 20:21
  • The C&W paper is only one of many. It would be good to put it in the context of others especially since it implies a certain unreliability in the surface record that might also apply to times when the record clearly showed warming. The biggest problem with the paper might just be the degree it was overreported by journalists eager to rebut climate skeptics.
    – matt_black
    Jan 28, 2014 at 0:20
  • 2
    C&W doesn't suggest that the surface record is unreliable, the problem is that there aren't any permanent weather stations in the Arctic (for obvious reasons), but you need to compensate for this somehow as otherwise the dataset is systematically biased low. C&W wasn't over-reported, it genuinely is that good a paper. The real problem is that the significance of the pause in warming has been way-over-reported, particularly by the skeptic media as suggesting that there is a problem with mainstream climatology, which simply isn't true for the reasons given in my answer.
    – user18604
    Jan 28, 2014 at 11:44
  • 2
    Of course C&W and GISTEMP suggests more warming than the other datsets, but that is because they are the only ones that do something to compensate for the known bias caused by ignoring the existence of the Arctic (which is what the other datasets effectively do). However, looking at all of the datasets and understanding the reasons why they don't all say exactly the same thing is something I would wholeheartedly recommend. If someone only shows one dataset it is a perfectly reasonable thing to do to ask yourself why.
    – user18604
    Jan 28, 2014 at 11:47

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .