BBC: remove errors bars for better headline

2008-08-27

In this article from the BBC Richard Black claims “This year appears set to be the coolest globally this century”. There is no basis for this claim, and moreover the very notion of picking warmest and coolest years amounts to bickering about global warming.

Black appears to be making this claim on the basis of looking at column 2 of the HadCRUT data. Here’s a graph, freshly minted from the Google Chart API:

The data is taken from HadCRUT, here’s a relevant extract:

2000  0.238  0.249  0.227  0.333  0.144  0.238  0.233  0.334  0.143  0.334  0.143
2001  0.400  0.411  0.388  0.495  0.304  0.400  0.394  0.495  0.304  0.495  0.303
2002  0.455  0.466  0.445  0.553  0.358  0.455  0.450  0.553  0.358  0.553  0.357
2003  0.457  0.468  0.447  0.556  0.359  0.457  0.452  0.556  0.358  0.556  0.358
2004  0.432  0.444  0.421  0.530  0.335  0.432  0.426  0.530  0.334  0.530  0.334
2005  0.479  0.490  0.469  0.580  0.378  0.479  0.473  0.581  0.378  0.581  0.378
2006  0.422  0.432  0.412  0.517  0.327  0.422  0.416  0.518  0.326  0.518  0.326
2007  0.404  0.414  0.394  0.501  0.307  0.404  0.398  0.502  0.307  0.502  0.307
2008  0.281  0.292  0.270  0.428  0.134  0.281  0.275  0.429  0.134  0.429  0.134

The format of the data is described here, by Hadley.

In the graph the red line is the best estimate, the pink lines shows the combined 95% uncertainty from all sources. You can get more, or possibly just different, graphs from Hadley.

The first thing to notice is that Black’s claim is false if you include the year 2000. Okay so technically I know that “this century” starts in XX01 but I also know we all celebrated the beginning of the millennium in 2000 and we accepted then that although 2001 was technically the beginning of the millennium (and hence the century) it was much hipper to celebrate 2000. So that deserves a mention at least.

But really my gripe is about not observing the error bars. The uncertainty in the data is such that the error bars all overlap! The data does not support the claim that 2008 is warmer than 2005 for example; if we take as our null hypothesis that these two years are the same temperature then we cannot reject it with any confidence. The same is true about any other pair of years (except possibly for 2005 versus 2000; we might be able to claim that 2005 was warmer than 2000).

Neglecting 2000, as Black obviously does, the data are consistent with a constant anomaly of +0.4°C. That’s just an example, many other temperature series would be consistent with the data, including ones which make 2005 the coolest year.

And that’s the problem with trying to “rank” years. The uncertainties in the data are all so large compared to the yearly changes that it’s totally meaningless to talk about the warmest year or the coolest year. We just don’t know.

Of course, if Richard Black had thought about the uncertainties in the data then he would’ve had to say “latest HadCRUT data shows 2008 about as warm as any other year this century”, and that’s not a very controversial thing to say. All this dramatic concentration on the yearly, monthly, daily ups and downs of global temperatures, greenhouse gas levels, what-have-you is nonsense. It’s just talking about the weather while the planet burns.

About these ads

4 Responses to “BBC: remove errors bars for better headline”

  1. Francis Davey Says:

    I’m always skeptical about “error bars” because they are frequentist not Bayesian artefacts. Having said that, the point remains good, any reasonable Bayesian inference from this would not think it very likely that the temperature in 2008 globally was really lower than any other temperature this century.

    The idea of a “global” temperature is a little unsound anyway, since the world isn’t in some kind of thermal equilibrium. It is the kind of figure more useful for sound bites and much less useful for working out where we really are.

    As you say, its all nonsense.

  2. drj11 Says:

    How does a Bayesian express uncertainty about their computations?

  3. Francis Davey Says:

    The result of a Bayesian analysis will give you a probability that a value is x or within a range [x,y] (if indiscreet). So you could give a 95% probability error bar if you like (though that would imply a 1/20 chance you were wrong).


Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

Follow

Get every new post delivered to your Inbox.

%d bloggers like this: