Warming hiatus

Why did a decade of flat temperatures not mean the planet had stopped warming?

For a stretch of years in the 2000s, the global surface temperature record looked almost level. The IPCC's Fifth Assessment reported the 1998 to 2012 trend at about 0.05 degrees per decade, against roughly 0.12 for the longer 1951 to 2012 period. That is a real number in a real dataset, and it was widely read as the plainest possible evidence: whatever had been happening had stopped.

The interesting question is not whether that reading was right — it was not — but why it was so persuasive. What would have to be true about a measurement for a flat decade in it to not mean what it appears to mean?

Three things, and each is worth deriving rather than being told.

Start with the choice of starting point. 1998 was the strongest El Nino of the twentieth century, a year when the Pacific dumped an unusual amount of stored heat into the atmosphere. If you begin a trend line at an exceptional high, the following years must fall short of it before they can exceed it. Try the same calculation from 1996 or 1999 and the flatness largely dissolves. A trend that changes sign depending on which year you start it in is not describing the system; it is describing the noise.

Second, ask what the surface record is a record of. It measures air a couple of metres above land and the top layer of the sea. But the vast majority of the planet's accumulated heat — around ninety per cent — is in the ocean. So the surface temperature is not the planet's thermometer; it is one thin layer of a very deep container, and that layer exchanges heat with what lies beneath it. Ocean heat content measurements, importantly, showed no pause at all through the period. The energy was still arriving. The question was only where it was going.

And here the answer is genuinely physical. During periods with more La Nina-like conditions in the Pacific, strengthened trade winds push warm surface water westward and deepen the mixing that carries heat below the surface layer. During El Nino-like conditions the winds slacken and that stored heat surfaces. The 2000s were dominated by the first mode. So the surface paused while the ocean interior kept accumulating — and the honest caveat is that the relative contributions of the Pacific, the Atlantic and the Indian Ocean to that redistribution are still argued over.

Third, the record itself had measurable problems. Sea surface temperatures come partly from ships and partly from drifting buoys, and buoys read slightly cooler than ship intakes. As buoys proliferated through the 2000s, the mixture changed and imparted a spurious cooling to the blend — a bias Karl and colleagues corrected in 2015. Separately, the datasets simply had no coverage over much of the rapidly warming Arctic, and averaging methods effectively filled those gaps with the global mean. Cowtan and Way showed in 2014 that interpolating them properly raised the trend appreciably. Both are ordinary measurement problems, not adjustments to taste.

By 2015 the argument had been settled by events rather than by statistics: 2014, 2015 and 2016 each set records in turn, and no version of the record now shows the period as flat.

Three refinements.

First, the pause was never a pause in the energy budget. Nobody found a period during which the Earth stopped accumulating heat. The apparent flatness was in one output variable, and only after selecting a favourable start year.

Second, this is not a story about scientists explaining away inconvenient data. The hiatus generated a large and productive literature — on ocean heat uptake, on volcanic aerosols and a weak solar cycle in that decade, on the mismatch between model ensemble means and a single realisation of internal variability. The right lesson is that a short-term discrepancy is a research prompt, not a verdict.

Third, and most usefully portable: the mistake was applying a signal-detection question to a window too short to answer it. Roughly speaking, at the current warming rate, the year-to-year variability in the global mean is large enough that fifteen years is near the edge of what can distinguish a trend from a run of luck. The perception failed not because the data lied but because the window was chosen after seeing the data.

• How long a record must be before a trend in it is statistically distinguishable from variability.
• Why model ensemble means should not be expected to match a single realisation of the real world.
• What the ship-to-buoy correction actually did, and why its critics' predictions did not hold.

Every term the collection defines is gathered in the glossary.