This study applies the technique of event attribution to the East African rainy seasons preceding the drought of 2011. Using observed sea surface temperatures (SSTs), and sea ice conditions with a state-of-the-art atmosphere model, the precipitation totals during late 2010 (the “short rains”) and early 2011 (the “long rains”) were simulated hundreds of times to produce possible distributions of precipitation. Alternative distributions of precipitation were produced, consistent with a world with neither anthropogenic forcings nor human influence on SSTs and sea ice. Comparing these modeled distributions to the observed rainfall, no evidence was found for human influence on the 2010 short rains, with their failure being affected by La Niña. However, human influence was found to increase the probability of long rains as dry as, or drier than, 2011. The magnitude of increase in probability depends on the estimated pattern by which human influence changed observed SSTs.
 In early 2011, the Greater Horn of Africa was impacted by a particularly severe drought. It consisted of the failure of two successive rainy seasons, known in Kenya as the “short rains” (typically October–December) and the “long rains” (March–June) [Hastenrath et al., 2011]. With inadequate rainfall in both the 2010 short rains and the 2011 long rains, this caused crop failures which led, along with the political situation, to famine in Somalia [Funk, 2011].
 With such an event, it is common for the question to be asked as to whether it could have happened in the absence of anthropogenic climate change. Though it is very rarely possible to declare an event is entirely caused by anthropogenic climate change because it would have been impossible without it, the growing research area of event attribution [Allen, 2003; Stott et al., 2013] aims to quantify how the probability of an event has changed. This involves defining a threshold, for example, a precipitation rate, which marks the point at which an extreme threshold has been crossed and for which exceeding it defines the occurrence of the event. This then allows the calculation of the fraction of attributable risk due to human influence on climate FAR=1−PNAT/PALL, where PALL is the probability of an event beyond the threshold in the actual world (with all known climate forcings, including those which are anthropogenic). PNAT is the probability in the “world that might have been,” where there were only natural forcings on the climate. It is also possible to calculate fraction attributable risk due to other factors, such as a particular mode of natural internal climate variability or other external forcings of the climate system, e.g., due to well-mixed greenhouse gases only.
 This study follows the technique first applied by Pall et al.  by using ensembles of climate models with prescribed patterns of sea surface temperature and sea ice and uses an expanded version of the experimental framework built for the study of Christidis et al.  based on the state-of-the-art HadGEM3-A climate model [Hewitt et al., 2011]. It then applies this to the regions of Kenya and Somalia to try to determine if anthropogenic climate changes affected the probability of the low rainfall in 2010–2011.
2 The ACE 2010 Method
Christidis et al.  developed the Met Office Hadley Centre ACE (Attribution of Climate-related Extremes) system for the study of extremes around the globe. It is built on the HadGEM3-A atmospheric model and is used to generate ensembles representing the climate with and without the effect of human influences. Simulations of the actual climate include both natural (solar and volcanic) and anthropogenic forcings (well-mixed greenhouse gases, aerosols, ozone, and land use changes) and use prescribed sea surface temperatures (SSTs) and sea ice data, e.g., from the HadISST observational data set [Rayner et al., 2003]. Simulations of the climate without human influences include only natural forcings and an estimate of the change in the SST and the sea ice due to anthropogenic forcings subtracted from the observations. The first experiment using this system was ACE 2010, which examined the extremes of the September 2009 to December 2010 period. Each ensemble of ACE 2010 consists of 100 members, simulating 100 possible outcomes for the atmospheric conditions, consistent with the prescribed SSTs, sea ice, and atmospheric concentrations.
 Event attribution then requires a further ensemble of runs to represent what could have happened in the “world that might have been,” here the “natural” world without anthropogenic influence on climate. The atmospheric changes in the model are implemented by removing the above anthropogenic forcings. Estimates of the SST changes (ΔSSTs) due to human influence on climate are obtained by employing estimates of the change in the SST driven by anthropogenic forcings from experiments with coupled general circulation models. In order to gauge uncertainty in the simulated patterns of the SST change, three candidate models were selected which had been already used to conduct attribution experiments and, thus, already included ensembles of transient simulations from 1860 with a variety of forcings, plus long control simulations without external forcings. These models were HadCM3 [Johns et al., 2002], HadGEM1 [Stott et al., 2006], and HadGEM2-ES [Jones et al., 2011]. For HadCM3 and HadGEM1, the 2000–2009 monthly mean SSTs from simulations with anthropogenic forcings only were taken, then the means of several hundred years for the same month from the control simulation were subtracted. The resulting ∆SSTs were then subtracted from the HadISST observed SSTs to provide three alternative realizations of the SSTs for the “world that might have been” without human influence on climate. The same process was also performed for HadGEM2-ES, though in this case, with no available anthropogenic-only simulations, this signal was reconstructed by subtracting the SSTs of the natural runs from those of runs including both natural and anthropogenic forcings. The control signal was then subtracted, and the rest of the above method followed. An example of the SST patterns obtained for the long rains season are shown in the Auxiliary Material.
 Three natural ensembles, each of 100 members of the HadGEM3-A model with this pattern of SSTs and sea ice and including present-day natural forcings only, were then produced. In the atmosphere, anthropogenic forcings (from concentrations of well-mixed greenhouse gases, sulfate, black carbon and biomass aerosols and ozone, as well as from land use) were set to preindustrial (control) conditions. This enables a comparison to be made between the atmospheric conditions in this set of “natural” worlds and the set of “actual” worlds produced by the 100-member ensemble of HadGEM3-A with observed SSTs and sea ice and present-day anthropogenic and natural forcings.
3 Application to East Africa 2010–2011
 With the model data accumulated for ACE 2010, it was possible to then perform event attribution on events that occurred during the period covered by the experiment, i.e., September 2009 to December 2010. This meant that an analysis was possible on the short rains of 2009 and 2010 and on the long rains of 2010, but crucially not 2011. Consequently, a further experiment was set up providing ensembles of simulations with and without human influences of the climate from January to August 2011, to complement ACE 2010 and enable the assessment of the causes of the long rains failure in 2011.
 Within this technique, observations are required to correct any possible biases in the models and for calculation of FAR. Due to the scarcity of African in situ measurements, the latest version of the TAMSAT data set [Milford and Dugdale, 1990; Grimes et al., 1999], known as TARCAT v2.0 [Maidment et al., in preparation; Tarnavsky et al., in preparation], was selected to provide observed rainfall. This is based on geostationary satellite measurements beginning in 1983, provides high spatial resolution, and is temporally resolved down to 10 day periods (aka dekads). Model data were averaged over these dekads to enable direct comparison.
 The average precipitation over the area of interest was then calculated using the region defined in the work of Rowell , which covers Kenya and Somalia (this is a subarea of the East Africa region used by Christidis et al. ). As part of that work, Rowell  verified that the teleconnections during the short rains period were well represented in the majority of models examined. To verify the representation of precipitation in these experiments by the HadGEM3-A model, the variability of long and short rain precipitation in models and observations was examined, along with the ability of the model to reproduce dry events. It was found that while the short rains were relatively predictable, the range of model climatology is smaller than that found in the observations. In contrast, the dry events in the modeled long rains are not reliably predictable but exhibit the correct magnitude of variability of precipitation. Plots illustrating this contrast can be found in the Auxiliary Material.
 To enable a better comparison between observed and simulated rainfall, the bias between them was determined using the average precipitation in the Kenya-Somalia region over the 1983–2010 period. This was calculated for the TAMSAT observations and for five “validation” HadGEM3-A runs over this period, where the model has been run with observed SSTs from 1960 to 2010. The difference between the observed precipitation and the mean simulated precipitation was then subtracted from all other simulated precipitation figures on a dekad-by-dekad basis to remove the bias.
 Once the spatial averages were calculated and any biases corrected, the total precipitation was then calculated for the whole season in that region. The short and long rains are considered separately. With this quantity calculated for each ensemble member and for the observations, the precipitation values are binned to give distributions for the actual world (with all forcings) and the three realizations of the world that might have been (with natural forcings only and the three estimates of the ΔSSTs).
4 The 2010 Short Rains Results
 Figure 1a shows the probability distributions for the season in question (broadened to September–January to allow for shifts in the season onset), while Figure 1b compares them in a probability-probability (P-P) plot, where the number of natural-forcings ensemble members in each precipitation histogram bin is plotted against the number of all-forcings members.
 As can be seen, the three natural worlds show little shift in their distributions. A two-sided Kolmogorov-Smirnov (KS) test also shows no significant difference in the curves at the 10% level (P values for HadCM3 ΔSSTs: 0.4; HadGEM1: 0.7; HadGEM2: 0.4). As such, this shows no evidence for any effects of anthropogenic climate change on the short rains in 2010.
 It is clear from Figure 1a that the observed rainfall was well within the probability distribution for the 2010 short rains. It is widely reported [Ogallo, 1988; Nicholson and Selato, 2000] that this season is particularly affected by the El Niño Southern Oscillation (ENSO), as indicated by the Australian Bureau of Meteorology's Southern Oscillation Index (SOI). ENSO was strongly in the La Niña phase (SOI between 18 and 27, which is expected to give a drier short rains season) during the 2010 short rains, in contrast to 2009, where the phase was in El Niño (SOI between −15 and −7, giving a wetter season). This can be emphasized by plotting the actual-world model distributions for the short rains of 2009 and 2010 together in Figure 2, where the curves are significantly different from each other (P<0.01). Consequently, the failure of the 2010 short rains is ascribed to the effects of La Niña, rather than anthropogenic climate change.
5 The 2011 Long Rains Results
 Applying the same technique to the 2011 long rains (again broadened to February–July to allow for shifts in season onset) results in the distribution in Figure 3a. Here it is immediately clear by eye that at least one of the natural models (that with ΔSSTs from HadGEM1) is substantially different from the all-forcings ensemble. With the aid of the P-P plot in Figure 3b, it is clear that the changes from the three natural worlds show a universal drying signal, although the magnitude varies considerably between the three different realizations. Using a one-sided KS test to determine the significance of this signal, two of the three shifts are significant at the 5% level, with the third (HadCM3 ΔSSTs) significant at the 10% level (HadCM3: 0.06; HadGEM1: 0.00; HadGEM2-ES: 0.04). As such, while this result is not categorical, there is some evidence that the failure of the 2011 long rains was more probable following anthropogenic climate change.
 Turning attention to the fraction of attributable risk, fitting a gamma distribution to each histogram gives a better estimate of probabilities beyond the dry threshold than merely counting events in bins. Using these to calculate FAR, the HadGEM2-ES ΔSSTs give a value of 0.46, attributing a doubling of the probability of possible droughts as bad as, or worse than, that which happened in 2011 to anthropogenic forcing. With HadGEM1 ΔSSTs, FAR is calculated as 0.99, suggesting the risk of such an event could be attributed almost entirely to anthropogenic factors. HadCM3 ΔSSTs give an FAR value of 0.24, indicating anthropogenic forcings caused only a small increase in the risk of such a drought. Thus, while there is evidence for an anthropogenic influence on the risk of the failure of the long rains, the large range in estimates of FAR indicate that large uncertainty remains. This could be better quantified in the future from a larger selection of ΔSSTs obtained from CMIP5 or a similar data base of climate simulations.
 This study has investigated the 2011 East African drought, which was brought about by failures of both the preceding rainy seasons. It was found that there was no significant change in precipitation during the 2010 short rains from what would be expected in a world without anthropogenic climate change and that the dry conditions in that season were much more strongly affected by the prevailing La Niña conditions. In contrast, evidence was found for an enhanced risk of failure of the long rains in 2011 due to human-induced climate change. An increased probability of low rains was found irrespective of the change in sea surface temperature pattern used for simulating the natural world. However, the statistical significance and the magnitude of the effect vary considerably.
 The fact that a drying signal is found for the long rains season in 2011 is a particularly interesting result when viewed in the context of other studies. McSweeney and Jones  showed that over the CMIP5 ensemble of simulations, there is a general wetting predicted in East Africa due to climate change. This was also true of the CMIP3 ensemble [Vecchi and Soden, 2007]. In contrast, Funk and Verdin  and Lyon and DeWitt  noted that recent observations show a drying trend in this region, a conclusion supported by Williams and Funk , who suggested that increased Indian Ocean SSTs would be expected to continue to reduce precipitation in East Africa. However, whereas Williams and Funk suggested that East African drying is linked to Indian Ocean SSTs warming faster than Pacific SSTs, the effect in this paper appears with more uniform warming of tropical Indian Ocean and Pacific Ocean SSTs, albeit with less warming in the Arabian Sea in two out of three of the coupled models used in this analysis. This can be seen in the ΔSST patterns in the Auxiliary Material, along with the mean precipitation changes brought about by those ΔSSTs. Taking the results for the short and long rains together, these results provide some additional evidence of the role of SST changes, both anthropogenic and natural in origin, in driving regional precipitation trends [Hoerling et al., 2006; Tierney et al., 2013].
 The model reliability results produced in the Auxiliary Material of this paper emphasize interesting differences between the short and long rains. The short rains appear to be relatively reproducible from year to year, given knowledge of the variability in SST patterns over that time. This is demonstrated in the greater probability of failure of the short season rains of 2010 due to La Niña. For the long rains, year-to-year variability in SST patterns appears to provide little predictability. However, the model used here represents the variability of precipitation in the region well and shows systematic shifts to distributions with lower precipitation with SSTs that have generally warmed due to human influence on climate. Thus, there is a long-term climate signal on the long season rains, albeit with a large uncertainty in calculated FAR.
 While these results support a clear conclusion that La Niña had a substantial influence on the 2010 short rains and provide intriguing evidence for a climate effect on the risk of the failure of the long rains in 2011, further work is needed to understand the mechanisms involved and to address the major uncertainties of this study. These are likely to derive from the difference in ΔSSTs between models, and in particular, it would be interesting to understand better the relative role of greenhouse gas forcing and other anthropogenic forcings in driving the precipitation responses seen in this study. A system with many more realizations of the natural world would enable the calculation of a mean value of FAR with a standard error, giving robustness to the result. The CMIP5 ensemble of simulations gives a good starting point for this. Further improvements are also expected to be made by increasing the spatial resolution of the simulations. It is anticipated that this will improve their response to the SSTs. Until then, this analysis provides evidence that a drought in East Africa such as seen in 2011 has become more probable as a result of anthropogenic climate change.
 The authors would like thank David P. Rowell and Michael Vellinga for their advice on techniques for studying African meteorology, as well as Daniel M. Mitchell for suggestions on representing the precipitation statistics. This document is an output from a project funded by the UK Department for International Development (DFID) for the benefit of developing countries. The views expressed are not necessarily those of DFID.