Hello,
Regarding the 0.12degC number. In our table 5, we presented results from matches between individual ship and buoy observations. Globally, there were 21,870 matches with a mean SST difference of of 0.12degC and a standard deviation of 0.85degC. This gives a standard error of 0.01degC to two decimal places.
The question is, what is the appropriate uncertainty – standard deviation or standard error – to assign to that mean difference when adjusting the ship data as a whole? (the “as a whole” bit is key). I think that the appropriate uncertainty is the standard error of 0.01degC rather than 0.85degC. Here’s why…
We know that individual ships will be biased differently and that every measurement from any one of those ships will have additional random errors that change from one measurement to the next. We can estimate the uncertainties associated with these two factors and this has been done several times. A large component of the standard deviation of the difference between the SSTs measured by ships and SSTs measured by buoys will come from those two factors. If the population of ships *as a whole* was unbiased (i.e. the individual ship biases had a mean of zero) then averaging all the measurements together from all the ships would have a mean close to zero. You would expect it, in fact, to be within a few standard errors of zero.
However, the mean of all the observations is 0.12C, which suggests there is some common bias across the fleet of ships. The uncertainty of the mean of that distribution is the standard error. It’s this common bias and its uncertainty we are interested in removing before combining ship and buoy data. The uncertainties associated with the other types of error that affect single observations are also factored in through the uncertainty estimates (in the case of HadSST3) or via the relative weights given to ship and buoy observations (in ERSST).
To make it clearer we can ask what the same numbers tell us about the bias in a single solitary ship observation. In that case, our best estimate of the bias in that single solitary ship observations is still 0.12degC, but in this case the uncertainty would be the standard deviation of 0.85degC which is close to the uncertainty associated with errors of a single ship observation estimated in other ways. As I mentioned before, that uncertainty is already included.
The question has also been raised as to whether it is better to adjust the ship data using the 0.12degC or to adjust the buoy data. There are arguments both ways. Adjusting the ship data brings the biased ships into line with the unbiased buoys. On the other hand adjusting the buoy data brings the buoys into line with the ship data, which constitute the majority of the historical record. Either way, when these are presented as anomalies relative to the 1961-1990 base period, you have to take into account the fact that data from the climatology period was mostly ship data. Subtracting 0.12C from the ship data would cool the climatology by 0.12C (more or less) which would mean that the anomaly for a drifting buoy observation relative to that climatology would increase by that amount. If you are looking at anomalies, the net effect of adjusting the ship data is the same as the net effect of adjusting the buoy data.
When we made HadSST3, one of the tests we did was try both. There was, as expected, little difference between the two choices. See part 2 of the HadSST3 paper, section 4.4 “Exploring the sensitivity of bias adjustments”
http://www.metoffice.gov.uk/hadobs/hadsst3/
Cheers,
John