The impact of the El Niño-Southern Oscillation (ENSO) on temperature extremes is examined in both observations and coupled climate model simulations. HadEX2, a newly developed observed gridded dataset of climate extremes indices shows marked contrasts in seasonal composites of the monthly maximum value of daily maximum temperature during the cold and warm phases of ENSO. Extreme maximum temperatures are significantly cooler over Australia, southern Asia, Canada and South Africa during strong La Niña events compared to El Niño events and significantly warmer over the contiguous United States and southern South America. Two climate models are contrasted for their ability to capture these relationships given their very different simulations of ENSO. While both models capture some aspects of the observed patterns, the fidelity of the ENSO simulation appears to be crucial for simulating the magnitude and sign of the extreme maximum temperature relationships. The impact of future climate change on these patterns is also investigated.
 Sea surface temperature (SST) patterns are known to impact mean seasonal climate around the globe [Ropelewski and Halpert, 1987] with significant regional relationships [Alexander et al., 2002] in e.g. Australia and North America that have been exploited for predictability. Less is known about the impact of SST patterns on climate extremes and temperature extremes in particular. While many extreme temperature events are caused by extremes in atmospheric circulation e.g. blocking, the predisposition of these circulation patterns can be influenced by the underlying SST patterns. Some regional studies have been undertaken [Gershunov, 1998; Gershunov and Barnett, 1998; Nicholls et al., 2005; Renom et al., 2011] but until recently, the limited observed coverage of daily data made more global studies difficult in this regard.
Kenyon and Hegerl examined station-based indices and found that large scale climate variability and particularly different phases of the El Niño–Southern Oscillation (ENSO) appeared to influence temperature extremes globally, often affecting cold and warm extremes differently. The majority of seasonal SST variability on the global scale is modulated by ENSO with El Niño and La Niña phases having statistically significantly different distributions of temperature extremes in many regions but particularly marked around the Pacific Rim [Alexander et al., 2009].
 Only a few studies however have investigated how well climate models are able to capture these strong responses that are present in observations of extremes [Meehl et al., 2007a; Gershunov and Barnett, 1998]. The goal of this study is therefore to examine the ability of state-of-the-art coupled climate models to reproduce the relationship between extreme maximum temperatures and ENSO and to contrast climate models with very different simulations of ENSO in their ability to capture these relationships.
2. Method, Models and Data
 We use HadEX2 (M. G. Donat et al., Updated analyses of temperature and precipitation extreme indices since the beginning of the twentieth century: The HadEX2 dataset, submitted to Journal of Geophysical Research, 2012) and HadISST [Rayner et al., 2003] observational products to diagnose the relationship between extreme temperatures and ENSO over the recent past. HadEX2 updates the earlier HadEX [Alexander et al., 2006] and represents the most comprehensive global gridded land-based dataset to date of temperature and precipitation extremes indices based on daily in situ observations and is available on a 3.75° × 2.5° longitude-latitude grid. HadISST1 is a 1° × 1° longitude-latitude grid of globally complete monthly SST and sea-ice concentration derived from in situ measurements, digitized sea ice charts and passive microwave retrievals. The use of these gridded products allows for an easier comparison with model output.
 Two versions of the National Center for Atmospheric Research (NCAR) global coupled model are used to investigate whether state of the art models can reproduce some of the significant responses that are seen in the observational analysis. The Community Climate System Model version 3 (CCSM3) [Collins et al., 2006] is a circa-2005 model that contributed to Coupled Model Intercomparison Project (CMIP) Phase 3 [Meehl et al., 2007b]. Components include the CAM3 atmosphere at T85 (∼1.4°) horizontal resolution with 26 vertical levels and the POP ocean at a nominal resolution of 1 degree. The transient climate response (TCR) of 1.5°C was at lower end of CMIP3 climate sensitivities. The Community Climate System Model version 4 (CCSM4) is a direct descendent of the CCSM3, contributing to CMIP5, and includes improved component models [Gent et al., 2011], with a finite volume atmospheric grid of ∼1° horizontal resolution with 26 vertical levels and the POP ocean with a similar horizontal resolution to CCSM3 but an increase in the vertical resolution from 40 to 60 levels. It is a slightly warmer model with a TCR of 1.73°C. No flux adjustments are used in either model.
 We compare these two models as they have markedly different ENSO simulations and as such provide a useful case study for the importance of the simulation of ENSO on extreme temperatures. The ENSO period in CCSM3 is overly biennial compared to the observations, with an SST pattern in the tropical Pacific that is too meridionally confined along the equator [Deser et al., 2006]. A number of improvements to the convective parameterizations in development of the CCSM4 [Neale et al., 2008] have led to a much more realistic ENSO period of 3–6 years and improvements in seasonality and pattern [Deser et al., 2012]. We use one ensemble member of the 20th Century experiment [Meehl et al., 2006, 2012] for each model as daily maximum temperatures are available for only a single member of the CCSM3. The improved fidelity of the CCSM4's ENSO pattern in the tropical Pacific can be seen in Figure 1 in which the sea surface temperatures are composited for cold minus warm events for the DJF season.
 Identical techniques are used to analyse the SST and daily temperatures from the models and observations. We use TXx, the maximum value of daily maximum temperature in each month, to represent extreme temperatures [Zhang et al., 2011]. As an extreme measure, this therefore represents a challenge for each climate model to reproduce. Strong ENSO events are diagnosed when monthly Niño3.4 (5°S–5°N, 170–120°W) SSTs are greater or less than 1 standard deviation from the long-term average. For each season we composite TXx for ENSO events from 1950–1999. This period is chosen due to the joint availability of data and model experiments during these years. In the observations, the method results in a total number of 16 cold and 16 warm ENSO months analysed for the DJF season. The number of ENSO events in the models is similar to that observed and we focus on the DJF and SON seasons as these are periods when the observed relationships are strongest. We first test the ability of the models to reproduce the TXx extreme index, as it is crucial that the model can simulate extreme events in general before any further understanding can be gained. Figure S1 of theauxiliary materialshows the pattern of TXx averaged over each season for the entire 1950–1999 period for both observations and models. The largest values of observed TXx are found over subtropical land areas and the tropical oceans and the models are able to reproduce the observed climatological features both in magnitude and pattern. We test for statistical significance using both a modified t-test (with reduced degrees of freedom to account for autocorrelation in the timeseries) and a Kolmogorov-Smirnov test. All results are given at the 5% level.
 Results from the observations (Figure 2a) show very strong statistically significant opposite responses from La Niña to El Niño events. Extreme maximum temperatures are statistically significantly cooler over Australia, western Canada, south Asia, South Africa and parts of South America during strong La Niña events than strong El Niño events. Opposite signed relationships are found over the USA and Siberia. These patterns are qualitatively similar to the results of Kenyon and Hegerl  based on slightly different observations, indices and seasons.
 The CCSM3 composites (Figure 2b) show only weak changes in maximum temperature extremes from La Niña to El Niño events. This is likely related to the poor simulation of the ENSO pattern, seasonality and frequency in CCSM3 described above, biases which are likely to affect the ENSO teleconnections to the extratropics for extreme events. In some regions, for example the contiguous United States (USA), the CCSM3 composite is opposite in sign to that observed. In comparison, the CCSM4 composites (Figure 2c) are much improved over the CCSM3, with more realistic amplitudes and patterns and statistical significance. The CCSM4 captures the correct sign of ENSO teleconnections over the USA and Australia, which are explored further below, however its simulation over Greenland and North Eurasia is deficient in comparison to CCSM3. Deser et al.  find similar biases in the northern high latitude CCSM4 ENSO teleconnections using seasonal mean surface temperatures, related to errors in the sea level pressure patterns.
 To highlight the contrast between La Niña and El Niño events further, box plots of TXx over Australian and USA land points are created for all monthly events in the DJF and SON seasons (Figure 3). The Australian and USA areas are chosen as they are regions well known for ENSO teleconnections on seasonal timescales and they also have good coverage in the HadEX2 dataset. In both seasons, the observed distribution of TXx during La Niña events is significantly different to the distribution during El Niño events over Australia. The mean value of TXx is 1.36°C colder during La Niña with the 5th percentile of TXx close to 2°C colder. The 95th percentile is 1.17°C warmer during an El Niño compared to a La Niña. Over the USA the differences are less clear although in both DJF and SON, distributions of TXx are statistically significantly different during an El Niño compared to a La Niña. The mean value of TXx is 0.91°C warmer during La Niña in DJF but it is the upper extremes of the distribution that show the most difference i.e. at its 95th percentile TXx is 1.78°C warmer during a La Niña compared to an El Niño. The mean, 5th and 95th percentiles of TXx are all warmer in SON during a La Niña compared to an El Niño in the USA (1.76°C, 1.61°C and 0.99°C warmer respectively).
 The boxplots from the NCAR models overlap those of the observed, thus producing credible magnitudes in extreme maximum temperatures over Australia and the USA during ENSO events. Similarly to Figure 2, the CCSM4 composites are more consistent with the observed distributions compared to CCSM3, with the CCSM3 showing little contrast in TXx between La Niña and El Niño events in either season or region. The CCSM4 captures the relative cooling over Australia during La Niña in both the mean quantities and the tails, although it overestimates the contrast slightly. The CCSM4 also captures the observed cooling in the mean of the distribution over the USA during El Niño events.
 Given the CCSM4's demonstrated ability to reproduce many of the observed features, the relationship of TXx with ENSO under future climate change is examined by analysing the last thirty years of the 21st Century in the Representative Concentration Pathway (RCP) 8.5 [Moss et al., 2010] experiment. ENSO events are determined in a similar way to the historical period, with El Niño and La Niña events both spread equally across the period. Figure 2d demonstrates that the main features highlighted in the historical relationships are also found under future warming. An interesting amplification of the historical difference between La Niña and El Niño events occurs in extreme maximum temperatures over Australia (Figure 3) in the CCSM4 RCP8.5 simulation. This suggests that for a similar magnitude of El Niño event, the extreme maximum temperatures over Australia will be warmer in the future relative to present day and larger than that expected from mean warming alone, with the caveat that this result is from a single model.
4. Discussion and Conclusions
 The hottest days in seasonal composites of extreme maximum temperatures are found to be significantly influenced by the phase of the El Niño-Southern Oscillation in both observations and models. During the peak of La Niña events, regions such as Australia, Canada, southern Asia and South Africa experience cooler values of TXx, the maximum value of daily maximum temperatures in each month, than during El Niño events. In other regions, such as Siberia and the contiguous United States, warmer extreme maximums occur. These general features are similar to ENSO teleconnection patterns found for seasonal mean temperatures [e.g.,Deser et al., 2012], although some regional differences may occur for some extremes [Gershunov and Barnett, 1998]. We examined two global coupled models for their ability to reproduce these relationships by compositing TXx for ENSO events over 1950–1999. The models produce a similar number of ENSO events to the observations over this period and create climatological extreme maximum temperature patterns of a similar magnitude and spatial structure. However the fidelity of the ENSO simulation appears to be crucial for simulating the spatial patterns and variations in magnitude of extreme maximum temperatures during ENSO events. It is clear that the improved physical parameterizations in the CCSM4 contributes to more realistic ENSO characteristics and results in a much improved simulation of variability in maximum seasonal daily maximum temperatures over that of the previous CCSM3 model version. Similar relationships hold in a simulation of future climate change, but with a suggestion of much warmer extremes for El Niño events over Australia. Future work will extend the analysis here to the CMIP5 multimodel dataset and investigate potential mechanisms leading to projected changes in these relationships, such as changes in base state or circulation [e.g., Meehl and Arblaster, 2003]. Further understanding of the influences of SST patterns on climate extremes will aid in reducing uncertainty in prediction and projections of extreme events.
 J.A. is supported by the Office of Science, Biological and Environmental Research, U.S. Department of Energy, cooperative agreement DE-FC02-97ER62402, and the National Science Foundation. The National Center for Atmospheric Research (NCAR) is operated by the University Corporation for Atmospheric Research under sponsorship of the National Science Foundation. L.V.A. is supported by Australian Research Council grant CE110001028. This research used computing resources of the Climate Simulation Laboratory at the National Center for Atmospheric Research (NCAR), which is sponsored by the National Science Foundation; the Oak Ridge Leadership Computing Facility, which is supported by the Office of Science of the U.S. Department of Energy under contract DE-AC05-00OR22725; and the National Energy Research Scientific Computing Center, which is supported by the Office of Science of the U.S. Department of Energy under contract DE-AC02-05CH11231. We thank Jerry Meehl, Wasyl Drosdowsky, Eunpa Lim and two anonymous reviewers for their constructive comments.
 The Editor thanks the two anonymous reviewers.