I like this figure a lot. I think it is clear and crisp; the colours have appropriate cultural meanings; the rise in global temperatures is captured well; the uncertainty in the decadal averages is nicely highlighted. Temperature series plotted with annual data can be very busy, especially if you are only really interested in long term trends.
The labelling of the decades is a nice touch: no more ink is used, but there is no need to run the eye back and forth between the axis and the data.
If I had any complaints, it would be that a) there is a little too much ink in the axes, but this is a personal preference, b) that the long “T’s” on the uncertainty bars might lead to an overestimate of uncertainty, and c) that those uncertainty bars need a separate explanation (what measure of uncertainty do they represent?)
Regarding that first point: Edward Tufte recommends using a high data:ink ratio. As a general principle, as much ink should go into showing the data as possible, compared with annotating or decorating the plot. There is a nice example of increasing the data:ink ratio here.
Comments and suggestions very welcome: you can see the comments policy here.
After the excellent feedback below, John has created a new version of the figure. It addresses many of the points that were made, but suggestions are always welcome. Bear in mind that it is impossible to please everyone, mind.
This post was originally posted at the previous incarnation of this blog at Posterous. There were lots of useful comments (thanks all), and I’ve attempted to transfer them below
clim8resistance (Twitter) responded:
I am not sure that ‘the colours have appropriate cultural meanings’. Red has a ‘cultural meaning’ of ‘danger’, rather than simply ‘warmth’. Yet the era around the 1910s is perhaps as dynamic as 1980/90/00, and is only blue by virtue of the somewhat arbitrary baseline period. The colouring cannot help but lead the viewer to a conclusion about the significance of post 1980 warming. It would be better to have all columns in the same colour.
Doug McNeall responded:
On reflection, I’m tempted to agree with you on the ‘appropriate cultural meanings’. This might be me, as the colours here are remarkably close to those on the hot and cold taps in my bathroom as a kid :) I reckon this could be helped by making the ‘warm’ bars closer to a warm orange than red.
In my own work, I’ve tried to avoid (e.g.) the red/green pairing, both for cultural *and* for access-for-the-colour-blind reasons. I think it’s important to pick neutral colours, for future greenhouse gas emission scenarios, for example
mewo2 (Twitter) responded:
I think in order to give meaningful feedback, it would be useful to know the intended audience and “delivery mechanism” for the figure. Is it intended for a scientific audience, in a paper or a presentation? Is it intended for the general public, either on a website or blog post (where the accompanying text is controlled) or as part of a press packet (where there are no guarantees about the context)? These considerations matter a lot when planning a figure.
I think the figure as given draws a little too much attention to an arbitrary baseline. In terms of data:ink ratio, the 1910s get four or five times as much ink as the 1970s, purely due to the choice of baseline. I would suggest replacing the current bars and uncertainty ticks with bars which cover just the uncertainty range. This would solve the issue of the eye being drawn to the baseline, as well as placing more ink in the centre of the uncertainty range, rather than at the ends as currently. There might be an issue with more ink being given to the decades with high uncertainty, but I think it would still be an improvement.
I’m not sure that the colours are really a problem. I think that the association of red with danger is much less problematic as part of the red/blue = warm/cold pair than it would be if paired with, say, green. However, the abrupt change in colour from blue to red at the (again) arbitrary baseline is a little problematic.
I agree that the axes are probably a little ink-heavy; there’s no real need for a top and bottom to the frame. I’d also move the caption to the bottom, rather than the side (possibly leaving the units on the side). Vertical text is never pleasant to read.
I’m not sure that you won’t still lead the viewer with colours. If colours were useful, it would not be necessary to demonstrate colour with height.
If you must use colours to differentiate, I would suggest filling the columns with gradients, starting with identical colours at the origin. That way you can’t be accused of implying ‘all cold good, all warm bad’.
The warm/cold thing from your experience as a kid is interesting. Warm and cold have universal meaning in such a case because people are roughly the same temperature, and so cold and warm correspond to difference above and beyond that point. But in the febrile atmosphere of the climate debate, no measure of ‘warm’ or ‘cold’ pleases all of the people all of the time. Perhaps colour is simply ‘too much information’.
etzpcm (Twitter) responded:
This is a great one to start with. It’s going to be a bit tricky to stay within the blog policies! I’m not sure whether you genuinely think this is a good figure – I suspect you are trying a wind-up to generate some interest and controversy – fair enough.
It was discussed back In January at Bishop Hill. The context was an article in the Daily Mail, and the Met Office response to that article.
Bishop Hill wrote “The Met Office have now responded, with a blog post that has a whiff of Bob Ward about it: “includes numerous errors in the reporting of published peer reviewed science”. The argument seems to be that if you take decadal averages it is still possible to obscure the plateau in the temperatures.
Or words to that effect.”
I see that I made a comment On the Met Office blog, that I am not allowed to repeat here. There are several other interesting comments on the Met Office blog and on the Bishop Hill thread.
On twitter, Jonathan Jones wrote:
Jonathan Jones @nmrqip 30 Jan
@richardabetts The apparent desperation to avoid admitting that the temperature rise has paused just makes @metoffice look shifty
And that figure was the subject of my first ever tweet.
liz_stephens (Twitter) responded:
This is mainly nitpicking, as I do think that it is clear and crisp, but I agree that the different colours are largely pointless, it’s pretty obvious what the trend is without including such evocative colours.
I’d also like to see what caption would be used with the plot, it is hinted that it might be something like “the evolution of global average temperatures, since the 1850s”, but as mewo2 says, it is giving attention to an arbitrary baseline which perhaps needs to be emphasised more.
The sceptic in me wonders (as a non-climate scientist as well) why 1961-90 is used as the baseline anyway, is this an IPCC defined period? Surely if you’re comparing a decade with a 30 year average then you would expect deviation from that anyway? So what counts as a significant deviation? Should there be error bars around the baseline? Better explanation of the baseline and the error bars might solve this of course.
I definitely agree with mewo2’s points about the data:ink ratio, although I don’t know that their suggestion would necessarily be an improvement for all users of the information as it might not be as good at drawing the eye / look too ‘graphy’ for it to be taken up and used by the media.
micefearboggis (Twitter) responded:
The original context for the diagram was this:
It was an attempt to get away from the usual end of year ritual in which each centre that produces a global average temperature announces that that year was the nth warmest on record, where n varies from press release to press release. The uncertainty on annual averages are such that individual years can’t usually be assigned a definitive rank. Non-overlapping 10-year averages can be unambiguously ranked and the ranking is the same for all the data sets we looked at.
1961-1990 isn’t special climatologically. It’s used as the baseline because that’s the WMO standard climate normal. It’s both arbitrary and unarbitrary.
pble24 (Twitter) responded:
Thanks for starting this, great topic, and good to see some Tufte references!
On the use of ink in the chart, I would offer the following:
1. I would remove the horizontal major grid lines in grey. With the yticks mirrored one doesn’t need to read right across the graphic so distance isn’t an issue. Also the ability to determine the absolute value of each decade is less important than the overall pattern of decadal variability, so removing major grid lines won’t obscure meaning, I would argue. Getting rid of major grid lines would thus provide more white space, making the graphic easier to read.
2. The major horizontal grid line at zero – I would remove this and let the bars ‘hang’, it doesn’t aid interpretation so is a waste of ink. At the very least I would suggest reducing the weight to barely visible – a hair line.
On the use of colour:
1. Yes, colour is very powerful, but in the context of global warming where it is interesting to see which decades are warmer than the baseline, using colour to distinguish between warmer than average and cooler than average decades is of value. The fact that people may see it as being ‘political’ is inevitable, but you can’t argue with the numbers/pattern, which is what matters. So, while not absolutely essential to use colour, I would stick with it.
xenongeosci (Twitter) responded:
I agree that the ink used to indicate a “cool” vs “hot” deviation from the baseline does not covey more information than does the baseline itself. Using blue vs red to show changes in trend would be more informative, such as blue for a cooler measurement than the decade before, etc. Do not know how noisy the graph may become…
Annotating error bars with decade notation is confusing at best. Label decades on the axis. If your data has anything interesting to show in terms of variation in error by decade, then label the error bars with error values. If there is not anything particularly interesting in the variation of error, then do not label it at all.
Rich_Jay (Twitter) responded:
In general, if I were producing a plot like this I wouldn’t include the coloured boxes and just plot the uncertainty bars. However, I can see the advantage of the boxes if there were many more time points and you were looking at longer term trends. I would also but the decades on the x-axis because at the moment I don’t think it is easy on the eye (although this is obviously a personal preference).
rjhd2 (Twitter) responded:
As @mewo2 suggested, I think replacing the bars and uncertainty ticks with bars which cover only the uncertainty range is the way forward. They would remove the sudden change in colour at the reference period, and more clearly show the change over the entire time range. As these are anomalies, the period over which they are calculated is sort of irrelevant, so highlighting it both by the zero-line and the bar colour causes focus on that, whereas the focus is probably best placed to be elsewhere.
To address the problem of the decades with high uncertainty given more ink if only showing the bars over the uncertainty range, could some level of shading be added. Under the assumption that the uncertainties are Gaussian, perhaps a darker shade at the most likely values, and fading further from these going out. The central values would then be highlighted clearly along with their spread. If the shade at the peak could be linked to the size of the uncertainty this would ensure eyes are not overly drawn to the least certain data. A single colour (even greyscale) could then be used to avoid any “cultural associations”. This visual-weighting of displays has been discussed on http://andrewgelman.com/2012/08/graphs-showing-uncertainty-using-lighter-inte… To keep the scientific certainty of what the errorbars mean, perhaps thicker lines could be drawn on the faded shading to show the location of the standard uncertainties.
The only complication with all this is that it may require a lot more explanation…
micefearboggis (Twitter) responded:
Thanks for the feedback so far. If you fancy replotting the data, the values used were (format, year1, year2, median, 97.5% uncertainty limit, 2.5% uncertainty limit):
1850,1859, -0.390, -0.291, -0.487
1860,1869, -0.364, -0.269, -0.451
1870,1879, -0.226, -0.140, -0.306
1880,1889, -0.255, -0.162, -0.342
1890,1899, -0.322, -0.218, -0.416
1900,1909, -0.391, -0.288, -0.484
1910,1919, -0.436, -0.328, -0.534
1920,1929, -0.302, -0.193, -0.402
1930,1939, -0.147, -0.043, -0.244
1940,1949, -0.086, -0.029, -0.139
1950,1959, -0.173, -0.129, -0.218
1960,1969, -0.124, -0.079, -0.169
1970,1979, -0.077, -0.033, -0.121
1980,1989, 0.084, 0.131, 0.038
1990,1999, 0.231, 0.283, 0.178
2000,2009, 0.404, 0.460, 0.348
These are anomalies relative to the 1961-1990 average. The global average temperature for 1961-1990 is often quoted as 14.0C. The uncertainty on that figure is around 0.5C, much larger than the uncertainties on annual anomalies.
aparnellwork (Twitter) responded:
Hi Doug. Interesting plots. Just to add to the other points above. I would also take issue with the lack of labelling of the error bars. I come across so many scientists who simply ask ‘what’s the error?’ They have no understanding that you should care that the error bars might be 1 standard deviation, a 90% confidence interval, or a 95% confidence interval. It’s clear from the comment above that it’s more like the latter, but that isn’t at all clear from the graph as it stands. If you double those error bars (i.e. if you plotted 1sd) most of the signal probably disappears
As a final point, why are we using a bar graph to plot time? Surely a line plot is more appropriate? This example is nice: http://andrewgelman.com/2012/08/graphs-showing-uncertainty-using-lighter-inte…. I see that someone else has suggested it too!
etzpcm (Twitter) responded:
Colours are useful if they provide information. But here they don’t – the colours are entirely unnecessary, as the amount of information presented in the figure is so small.
The only thing the colours add is spin, creating the misleading impression that until 1970 the earth was ‘cold’, and since 1980 it has suddenly become ‘hot’.
In summary, apart from the (a) the binning to hide the decline, (b) the deceptive use of colour, (c) the arbitrary meaningless baseline, (d) the unexplained and misleading ‘error bars’, this is a great figure!
Doug McNeall responded:
Thanks for an almost perfect demonstration of a breach of the comments policy. Phrases like ‘hide the decline’, and words like ‘deceptive’ imply an intent to decieve, which is not helpful. If you think something is hidden, point to it in another graphic, or make your own and I’ll link to it.
We’ll be looking at graphics from the ‘sceptical’ end of the climate spectrum soon, and I won’t tolerate this kind of language when we talk about them either. I’m interested in constructive criticism, not a slanging match about motives.
I’ll leave this comment here as an example, but the next one gets squished. The first part is useful, thanks! The second part, not so much.
Doug McNeall responded:
Well, I come across a bit finger-waggy in that previous comment, don’t I.
Thanks all for the comments so far, some excellent points. I’ll update this post with some alternative graphs, and hopefully some code.
Martin, I take your point about including context for the graph; I’ll remember to be clearer in the future.
Ben I think your point about the reference standard temperature of humans is very well made! There are interesting discussions to be had about standard reference periods for climate here. It turns out that there are lots of adults who weren’t born during the 1961-1990 period. I find this terrifying.
Andrew, the wider point about the representation of uncertainty certainly chimes with me.
John, thanks for being thick skinned enough to offer this graph up in the first place.
oceanCO2 (Twitter) responded:
I’d prefer clearer grid lines on the improved image, and more specific date labels (at a glance it’s not clear which set of bars correspond to each 50-year label). Also, having gaps between the bars seems a little odd for this continuous data.
The removal of the red/blue colours and reduced emphasis on the y-axis ‘0’ are definite improvements, as others have mentioned!
mewo2 (Twitter) responded:
I have to say, I much prefer the new version. The removal of the arbitrary baseline has made things a lot more readable. However, I think the original time axis, with per-decade labels near the data, was a lot more readable. Perhaps the decade labels could be placed in the center of the data bars, in place of the fairly redundant lines marking the medians. It also seems quite sad to lose colour completely, even though I agree that the red/blue division on the original was not ideal. Even if the bars were to be monochromatically shaded, it might add some life to the figure, and help draw the eye to the data rather than the axes, etc.
Also, and not intending to start a Bayesian bunfight, the phrase “Expect decadal average to fall within shaded area 95% of the time” strikes me as profoundly odd. How many copies of the 1920s do we expect to experience? I think I’d prefer something like “Decadal averages fall within shaded areas with 95% certainty”, but I have to admit ignorance as to the precise statistical meaning of the HadCRUT uncertainty estimates.
pble24 (Twitter) responded:
I think the figure now lacks a little structure, through having removed too much ink. I would prefer at least y axis and ticks to be reinstated, and possibly the complete frame, but just not too heavy. This may help in interpreting the decade labels too, which people have commented on. Underscore the point made wrt the title ‘Expect…’
Delete this comment
8 days agojdannan responded:
I think the graph is improved but the subtitle is horrid (though to be fair there probably isn’t a simple way of referring to confidence intervals which will not mislead at least 95% of readers). Also, given that the end points of the bars are explicitly referenced, it seems odd to make them fade out.
mewo2 (Twitter) responded:
Expect reader to misunderstand uncertainty estimates 95% of the time.
Mark Brandon responded:
Just out of interest, Leo Hickman has just shown that this very figure as original was used to brief for an episode of Question time
The cover note on the figure was:
Over short intervals of a few years, natural climate variations can temporarily mask long term warming. 1998 is the joint-warmest year on record but globally, the
decade 2000 – 2009 was significantly warmer than any previous decade in the
record, going back to 1850. Individual years (except 2005 and 2010) have not been
quite as warm as 1998 and this is quite normal because the human-forced warming trend is taking place on top of natural variations. We will always see such short-term fluctuations but, to confirm long term changes of climate, we need to look at records several decades in length.