The Blackboard:

Roger Pielke Jr. and John Tierney have recently been comparing published predictions of AGW to empirical data. In response, Gavin Schmidt suggested that comparisons should include more than 8 years, and pointed out that computations for Hansen et al. 1988 scenarios A, B, C began in 1983, as Gavin said last May, “giving us just over two decades of comparison with the real world.”

Of course, Gavin is correct. We should look at longer time frames before making our mind up. So, let’s do that!

The graph below graph compares the Hansen et al. 1988 to measurements of the global mean surface temperature measured over land & ocean. It suggests that Hansen et al. 1988 scenarios over-predicted the temperature in the years after transient experiments were initiated.

In this figure:

• Data are GISS Land-Ocean and HadCRUT III data which include both land and ocean measurements.
• I adjusted the baseline year for the anomaly so the zero point matches the average for years 1974-1983 inclusive is zero. For purposes of plotting, year zero is the midpoint of year 1978; so 1984 corresponds to 5.5 on the chart.
• The comparison is made using data collected only after 1984.

We see:

• Scenarios A & B over-predict warming as measured over the oceans.
• Scenario C falls between the HadCRUT III and GISS Land Ocean data; matching the HadCRUT data almost perfectly.

Gavin Schmidt has suggested Scenario B represents the best match to forcing due to greenhouse gases (GHGs) that actually occurred during the period after 1984; the forcing in Scenario C was less than what occurred. Scenario A represents more forcing. So, in principle, scenario B should track the actual changes in temperature best.

Main Observation:
When comparing Scenario projections A, B & C, to land-ocean type data, it appear Hansen et al. 1988 over-predicted the real world warming that occurred in years following publication of his paper. This is the main observation based on this data. The remaining discussion are details for those who wish to know more.

(Note: I have done no statistical analysis of this; the observation is based on eyeballing the graph only. You can find other graphs to eyeball appear at Climate Audit and Real Climate. The main difference between my graph is that I focus on post 1984 data only.)

---

Details regarding graphs.

Why use Land-Ocean and HadCrut III data instead of Met station or Crut data?
I compared the computational results to Global Mean Surface Temperatures (GMST) that include the ocean because the averages from computations include GMST over both the land and ocean. Comparison to land only and disregarding the 70% of the earths surface covered by water strikes me as not quite an apples-to-oranges comparison, but it is at least oranges-to-lemons.

Why reset the baseline?
After selecting the data for comparison, I recognized that the Scenario predictions, the GISS data and the HadCRUT data all use different baselines to define their anomalies. The baseline for the Scenarios is “perpetual 1958″ computed by a simplified model with a shallow ocean; the fuller ocean appears to have been added for the transients– see pate 9360 or the Hansen paper. The GISS data zero point is years 1951-1980 inclusive. The HadCRUT baseline appears to be 1961-1990.

I wanted bring all series to similar baselines. I chose the 10 year period 1974-1983 inclusive as my baseline for the following reasons:

1. The period does not include any time after Hansen initiated computations for Scenarios A, B & C.
2. Choosing a baseline starting in 1974 ensures a decade and a half of computational years after the onset of computation (1959); Gavin Schmidt advised against using computations in the first decade or so after initiation of computations in this blog post.)
3. Prior to 1984, all scenarios shared similar forcings thought to approximate the real world forcing. So, assuming the models are accurate, one would expect the model projections and the empirical data to represent different ensembles of the same random process during this period.
4. Ten years is a round number. Though arbitrary, it so seems a fairer selection than, say 13 or 9.
5. A period of of 10 years seems minimally sufficient to reduce the variability due to weather when computing the average anomaly between 1974-1983. (I’d prefer a 20 year baseline, but it is not possible to select one while also excluding the decade or so of computations that are likely affected by the cold start issue, and also restricting the baseline to years before 1984.)

After selecting the time period for my baseline, I shifted the baseline for the anomalies as follows:

• For scenarios A, B & C, I calculated the average anomaly based on years 1974-1983, inclusive. That gave me an average anomaly of 0.130 based on 30 years. I subtracted this value from all anomaly values for later years, and plotted them.

  I used identical anomaly values for all scenarios because the almost same forcing was applied to each of the three scenarios up to 1984. Up to that year, they all represent different realizations of the same case. (Note: Forcing for the Scenarios do differ from each other slightly during this period. )

• For the two experimental data sets, I calculated the average anomaly based on years 1974-1983 individually. Then, I subtracted that value from each set of data.

This essentially forces the average temperature of computations and each experimental sets to match during the 1974-1983 period.

Why force the Linear regression through zero?
To an extent, this decision is based on the assumption that, after selecting similar baselines, GISS II computations on average, do, on average, predict temperature variations accurately during that period. Permitting an arbitrary intercept would permit “jumps” in data matching at some year after 1978, which would be reasonable if we believe the GISS II computations did not, on average, match the empirical data in 1974-1983.

Did Hansen’s projections really over-predict the warming?

It appears the GISS II model over-predicted warming. That said, warming was predicted and warming occur.red

It’s also not clear this visual comparison between means very much. The 24 year time period after 1983 is short given the amount of scatter in the data and the total increase in temperature. The two factors together mean results might differ if I selected 1988, the first year after the manuscript was submitted, as the initial date. (I based the choice of 1984 on the advise by Schmidt at Real Climate.)

Worse, because the time period is short, and there is a fair amount of variability, I can ginn up graphs that make Hansen et al.’s 1988 projections look both better or worse than these. For example, even starting with 1984, if I don’t shift the baselines to match on average between 1974-1983, the model projections to overpredict the temperature rise by a larger amount.

Both the unflattering cases and the flattering ones involve making judgments about displaying the data that seem plausibly fair to me.

As far as I can tell, even with 24 years of data, this exercise appears to be a cherry picker’s dream!

---

Data Sources:

1. The GISS Land Ocean Data are from http://data.giss.nasa.gov/gistemp/tabledata/GLB.Ts+dSST.txt.
2. The direct link to the HadCRUTIII data is Real Climate

Update: Note that Scenario ABC data at Real Climate were reported to be digitized when made available, see Gavin’s response to my request.

Previous Post:
« Temperature Variability: Does it Reduce the Greenhouse Increment of 33C?

Next Post:
How Gavin’s Weather vs Climate Graphs Compare to Hansen Predictions? »