It’s time to provide the answer to my recent challenge. For those who don’t remember, the challenge was to look at three graphs and decide how many sudden, non-climatic shifts were in them. I’ll now provide the same image as before, but with the breakpoints indicated:
I can’t say I’d have ever guessed the locations of any of those breakpoints. The only apparent breakpoint my eyes found was in the third graph, around ~145. The shift in data there is far more apparent than any of those supposedly found in the other series.
But that doesn’t mean there’s a problem. Sometimes our eyes are bad at finding breaks in patterns. The best way to tell if there’s a problem is to check how serious the differences in each piece are. Fortunately, BEST provides us the information to do so. For each station, it provides raw data and an adjusted version, saying:
In many cases, raw temperature data contains a number of artifacts, caused by issues such as typographical errors, instrumentation changes, station moves, and urban or agricultural development near the station. The Berkeley Earth analysis process attempts to identify and estimate the impact of various kinds of data quality problems by comparing each time series to neighboring series. At the end of the analysis process, the “adjusted” data is created as an estimate of what the weather at this location might have looked like after removing apparent biases.
With that understanding, we can begin looking at the stations used in this challenge. The first station was ANNA-1E, a station in Illinois. This station is listed as having five empirical breakpoints (and one record gap), but only one in the period the challenge covered. This breakpoint was found at September, 1979. Correcting for it shifts the graph as such (black-raw, red-adjusted):
If you have trouble seeing the difference, don’t feel bad. The shift is only one tenth of a degree (.11). It’s hard to imagine what non-climatic influence would shift the data by such a small amount. It’s even more difficult to imagine how an automatic process could find one if it did.
Things are no better when we look at BROOKPORT DAM 52, the second station used in this challenge. This station is given as having five empirical breakpoints (and two station moves). This is the effect of the three in the period covered by this challenge:
The first breakpoint, at June, 1968, has a notable effect of .38 degrees. I’m not convinced a non-climatic bias of that size could be confidently identified, but it’s far more plausible than the .11 of our previous example.
Unfortunately, three years later in October, 1971, a second non-climatic bias was supposedly found. This bias was .27 degrees, in the opposite direction. That means aside from a three year period, the net effect of the breakpoints is once again .11 degrees. That lasts for a while.
Then, in November, 1980, another non-climatic bias was supposedly found. This bias was .13 degrees in the same direction as the second, further offsetting the first. When combined, the three biases are given as having a net effect of .02 degrees.
The third record chosen for this challenge was MAKURAZAKI, a random one I picked from Japan.
I’m skeptical any of these “empirical breakpoints” reflect anything at all, much less discernible, non-climatic biases. I struggle to see any reason why temperature records should be broken apart for effects considered to be that small. My impression is it’s just over-fitting and a waste of processing power.
I could be wrong, of course. Maybe there are legitimate reasons for thinking these breakpoints are necessary. However, I can’t help but notice the adjustments to the (full) first station cause its stated trend to increase from -.43°C / Century to .42°C / Century; the second from .96°C / Century to 2.27°C / Century.
Given a similar pattern of adjustments exists for Illinois as a whole, it’s difficult to see how assigning all these breakpoints could be the best idea.