Nothing grows in Texas
The Sacred Cowboys sang nothing grows in Texas. Last summer it came true. The Texan drought of 2011 was hotter and drier than any of the great droughts of the 1930s. Is it possible to diagnose what role climate change may have played? State Climatologist for Texas John Nielsen-Gammon asked this question last year in a preliminary post on the drought.
He concluded there was a strong anthropogenic evidence for warming but little evidence for its influence on the extremely low rainfall. The following analysis was developed without seeing his summary and happily, they have much in common. The big difference is the explicit treatment of non-linear warming.
I applied the methods for analysing step changes and attributing regional warming developed for SE Australia to the Texan climate. Was it possible to estimate the relative contributions of global warming for a single season? The results were more successful than I had anticipated. Tmax was 86% more likely to be due to anthropogenic warming and Tmin 95%. The method is unorthodox, so may be controversial. Details over the fold.
There are 49 stations in the USHCN climate data set for Texas with observations from 1895–2011 (October). These needed a bit of quality control. Each was checked for inhomogeneities for Tmax, Tmin and P. Some changes were clustered together and looked systematic – around 1921, 1933, 1958, 1990 and 1998. These coincide with similar changes in GISS temperatures for the latitudinal bands 24-44°N and 44-64°N. Other changes not associated with these dates or that were anomalously large were treated with suspicion. Where there was a large inhomogeneity before 1960, the series was truncated up to the inhomogeneity. If there was one after, the series was discarded. Nine stations were discarded, 21 others used unchanged and 19 shortened. No adjustments were made. Individual station anomalies were calculated for 1961–90 and averaged with no area weighting.
The GISS temperatures for 24-44°N and 44-64°N illustrating step and trend are shown below. All the shifts are statistically significant (to 1%) but don’t on themselves suggest climate variability or change. What they potentially show is regime shift. Also shown are counts of individual station shifts (out of a total of 49) along with the timing and magnitude of shifts in zonal mean average temperatures from GISS. Shifts interpreted as artefacts were those clustered around World War II, some of the early shifts, and most of those between 1960–1980. The cleaned up data sets were used to analyse shifts in both annual data (1895–2010) and summer data (1895–2011).
Over continental areas, Tmax and P are correlated, whereas Tmax and Tmin are more widely correlated in mid latitude regions. John N-G shows the relationship between Tmax and P for the Texas summer. Step change analysis showed that for annual temperatures, Tmin changed relative to Tmax by 0.5°C in 1990 and Tmax changed relative to P by 0.8°C in 1998. The period 1895–1990 was therefore selected as being stationary. The correlation between Tmax and P was -0.68 and between Tmax and Tmin was 0.58 for that period. Relationships between Tmax and P and Tmin and Tmax were then estimated using linear regression. Changes in Tmax and Tmin during non-stationary periods were then estimated from P and Tmax respectively and the residuals were assumed to be regional anthropogenic warming. The anthropogenic component of both Tmax and Tmin underwent a step change in 1990 of 0.6°C and 0.8°C respectively. The step and trend charts for Tmax, TmaxARW, Tmin and TminARW are shown below.
Interestingly Tmax shows a shift in 1998, but TmaxARW shows a shift in 1990 at the same time as Tmin and TminARW. This suggests that the increase in rainfall observed in recent decades has been suppressing an underlying warming pattern between 1990 and 1996. The earlier warm period of the 1930s and the cooler period from the late 1950s are interpreted as climate variability. According to this model, there is no trend in warming but there is an abrupt shift in about 1990, which is reflected in the zonal average temperature for 24-44°N (1987) and 44-64°N (1988). The timing isn’t exact but may be due to the effects of local variability.
So the model worked as it did for SE Australia, but I’m less sure of interpreting the results not knowing the regional climate that well. The northern hemisphere is more complex than the southern hemisphere in its interactions between oceans and land. Some of the dates lined up with known shifts in the North Atlantic Oscillation.
The next step was to try the model for summer. The correlations between paired variables were much higher, -0.82 between Tmax and P and 0.79 for Tmin and TMax. It was also harder to establish the stationary period: Tmin departed from Tmax in 1993 and Tmax from P in 1995. TmaxARW and TminARW both shifted significantly from 1996 (1%). A smaller shift in TminARW in 1977 points to a small amount of earlier warming but was only significant to the 5% level. I suspect a Pacific influence is operating there. 1996 was chosen as the date for the shift non-stationarity. Even though there is some uncertainty about when the non-stationary period commenced, the model was insensitive to that uncertainty. The results were checked using a slider that showed them graphically. The final results were also checked using the dates 1977, 1990 and 1996 for the shift date – they were within ±1% probability (they are discussed below).
The drought of 2011 produced a total P of 122 mm for Jun–August compared to a long-term average of 197 mm, Tmax of 37.5°C compared to 34.3°C and Tmin of 23.4°C compared to 21.0°C. Anomalies from the stationary period were -2.5 for P, 3.2°C for Tmax and 2.4°C for Tmin. Return periods from the long-term record were 370 years for P with skewness removed (cube root) and 1,500 without; about 440 for Tmax and 2,700 for Tmin. These numbers don’t mean a lot in themselves; they just indicate that within a stationary time series these events are very rare. Rainfall from January–August 2011 at 289 mm compared to a long-term average of 473 mm was roughly the same probability. The two wettest summers in the timeseries were 2004 and 2007, but this wasn’t the case for annual rainfall.
Illustrated below are the confidence levels for the trend and individual years for Tmax dependent on P and Tmin dependent on Tmax. The blue dots represent data during the stationary period and the red crosses Tmax and Tmin from 1996 onwards. They clearly show the warming signal, in Tmin especially. The step change in TmaxARW in 1996 is 0.7 °C and for TminARW is 1.0°C.
Two tests were run to compare the hypotheses of these anomalies being due to climate change as opposed to climate variability based on the codependence between Tmax and P and Tmin and Tmax. One was a Bayesian test that compared the likelihood of the anomalies being due to climate change versus climate variability. The other is a ratio test based on the tracking knowledge tests outlined by Roush (2005). Both compare pretty much the same thing – the likelihood of the anomalies for Tmax and Tmin according to the two competing theories.
The climate variability test was set up using the whole time series to determine the regression relationships between the variables, then testing the likelihood of the 2011 anomaly within that relationship. The climate change test used the stationary period to determine the regression relationship, then the step change values of 0.7°C and 1.0°C were removed from Tmax and Tmin respectively from 1996.
The Tmax anomaly as a function of P was 28% likely under climate change and 4.6% likely under climate variability. The Tmin anomaly as a function of Tmax was 83% likely under climate change and 4.0% likely under climate variability. The Bayesian relationship for Tmax being due to climate change compared to climate variability was 86% likely, a 6:1 ratio. For Tmin, the results were 95%, a 21:1 ratio.
Therefore, analysis of the 2011 Texas drought showed that the warming component was much more likely to be due to climate change rather than climate variability. For rainfall, there was no straightforward way to assess an anthropogenic component. If at some stage it was possible to identify an anthropogenic component in the reduction, then these odds would strengthen.
These results are being written up for journal submission. John Neilsen-Gammon has more recent updates on the progression of the drought and its impacts.