Nonlinear response of tropical lower stratospheric temperature and water vapor to ENSO

water vapor to ENSO Chaim I Garfinkel1, Amit Gordon1, Luke D Oman2, Feng Li3, Sean Davis4, and Steven Pawson2 1The Fredy and Nadine Herrmann Institute of Earth Sciences, Hebrew University of Jerusalem, Jerusalem, Israel. 2 NASA Goddard Space Flight Center, Greenbelt, MD, USA. 3 Universities Space Research Association, Columbia, MD, USA. 4 NOAA Earth System Research Laboratory, Boulder, CO, USA. Correspondence to: Chaim I. Garfinkel (chaim.garfinkel@mail.huji.ac.il)

use multiple linear regression (e.g. Crooks and Gray, 2005;Marsh and Garcia, 2007;Mitchell et al., 2015). An assumption underlying this method is that the response to these forcings is linear, i.e. that the response to a given magnitude El Niño is equal and opposite to that of a La Niña event of equal magnitude. Is this assumption really true? Second, Garfinkel et al. (2013a) found that EN events whose sea surface temperature anomalies peak in the Central Pacific (i.e. CP events) lead to dehydration regardless of season while events peaking in the Eastern Pacific (i.e. EP events) lead to boreal spring moistening. However, 5 EP events tend to be stronger than CP events, and it is not clear to what extent the difference found by Garfinkel et al. (2013a) reflects the intensity of the EN event or the flavor of the event. Third, to what extent is the tropical stratospheric response to ENSO governed by SST anomalies in the Indian Ocean sector that typically follow (though with diversity in their amplitude) ENSO? Finally, it has been suggested that SST variability in the Pacific Ocean contributed to the drop in water vapor in the 10 early 2000s (Rosenlof and Reid, 2008;Garfinkel et al., 2013b) possibly via ENSO (Brinkop et al., 2016), but this contribution has not yet been quantified except by one very-recent study (Ding and Fu, 2017). This paper will demonstrate that there are nonlinearities in the lower stratospheric temperature and water vapor response to ENSO. While typical EN events lead to tropical lower stratospheric cooling and dehydration in boreal winter, the spring response is nonlinear: strong EN events and LN lead to moistening while weak/moderate EN events lead to dehydration. We 15 clarify that discriminating between CP and EP events may not be crucial, and rather one should discriminate between very strong EN events and moderate EN events. As CP events tend to be weaker than East Pacific events (Johnson, 2013), it is easy to confuse a composite of CP EN events with a composite of moderate EN regardless of type. This nonlinearity apparently originates in the Indo-West Pacific response to EN, as warming in this region leads to moistening of the stratosphere in spring.
Finally, by comparing changes in water vapor concentrations between the early 2000s and late 1990s in a large ensemble of 20 model simulations forced with observed sea surface temperatures, we suggest that the deterministic component of the water vapor drop in the early 2000s was 0.14ppmv, approximately one-quarter of the observed drop.
A complete explanation of inter-annual variability in stratospheric water vapor, particularly that associated with El Niño (Bonazzola and Haynes, 2004;Hasebe and Noguchi, 2016;Konopka et al., 2016), requires consideration of changes in both temperature and air-parcel trajectories near the tropopause, and might also be influenced by changes in cloud ice (Avery et al., 2017). We cannot distinguish among these various effects in our simulations, since the model output necessary to run a La-25 grangian trajectory model was not archived. Nonetheless all of these effects operate in the simulations, and the simulated interannual variability in water vapor will arise from some combination of these effects.
More generally, the advantage in studying historical changes in water vapor and temperature in free running climate simulations is not to form a best estimate of the actual interannual variability; for that purpose, nudged experiments and/or Lagrangian 30 trajectory modeling are far better. Rather, the motivation is three-fold: one, assuming the model is capable of capturing interannual variability, the causes of trends or discontinuities (such as the drop in the early 2000s) can be better understood in a framework in which there is no possibility that changes in the observing or modeling system could have led to these trends or discontinuities; two, large ensembles of a free running model can be produced in order to better isolate the forced response from a single EN event from unrelated internal atmospheric variability not forced by anomalies at the ocean surface; three, 35 and relatedly, the observational record is not long enough in order to confidently conclude whether the response to ENSO is 5 nonlinear or to confidently separate the impacts of Indian Ocean SSTs from Pacific SSTs due to their strong covariability, and thus only by considering large model ensembles can these effects be confidently identified.
The data and methods are introduced in Sections 2 and 3. Section 4 demonstrates the nonlinearity of ENSO's effect on tropical lower stratospheric temperature and water vapor. In order to better understand the nonlinearities evident in Section 4, Section 5 considers more closely the strongest EN event covered by our model experiments -the event in 1997/1998 -and highlights the importance of the Indian Ocean. Section 6 considers implications of the interannual variability for the drop in the early 2000s and for the EN event in 2015/2016. The supplemental material discusses the linearity of the influence of ENSO on the BDC.

Data
We analyze the MERRA (Modern-era retrospective analysis for research and applications; Rienecker et al., 2011) reanalysis, the merged water vapor product from SWOOSH v2.5 (Davis et al., 2016), and output from atmospheric chemistry-climate general circulation models (GCMs) and coupled ocean-atmosphere GCMs on various time scales. The Goddard Earth Observing System Chemistry-Climate Model, Version 2 (GEOSCCM, Rienecker et al, 2008) couples the GEOS-5 (Rienecker et al, 2008;Molod et al., 2012) atmospheric general circulation model to the comprehensive stratospheric chemistry module StratChem Oman and Douglass, 2014). The model has 72 vertical layers, with a model top at 0.01 hPa, and all 15 simulations discussed here were performed at 2 • latitude x 2.5 • longitude horizontal resolution. The model spontaneously generates a QBO (Molod et al., 2012). The model vertical levels between 140hPa and 50hPa are located at 139.1hPa,118.3hPa,100.5hPa,85.4hPa,72.6hPa,61.5hPa,and 52.0hPa; output is plotted at standard pressure levels.
The convection scheme used in GEOSCCM is based on Relaxed Arakawa-Schubert (Moorthi and Suarez, 1992;Rienecker et al, 2008), and the cloud ice parameterization is described in Molod et al. (2012). Note that there is cloud ice in the version of the 20 model under consideration here up to 85hPa (as is shown below). To the extent that entry water vapor is controlled by large scale temperature patterns and the relatively crude ice parameterization in the current generation of the model, we expect that our model captures the response of water vapor to ENSO. That being said, more advanced treatments of ice clouds are currently under development, and hence similar studies must be performed as models improve.
A series of integrations were performed with the GEOSCCM, and they are listed in Table 1 and described below. They 25 fall into two classes: coupled ocean-atmosphere simulations, and historical-SSTs simulations with an atmospheric chemistryclimate general circulation model (AGCM). Both modeling frameworks have their advantages: coupled ocean-atmosphere simulations allow the model to self-consistently develop SST anomalies and teleconnections without violating energetic constraints, and also allow us to examine the stratospheric response to a wider range of ENSO events than have occurred in the historical record. On the other hand, simulations forced with observed SSTs can be more easily compared to the observed 30 response to ENSO.
The model configuration for the coupled ocean-atmosphere simulation is described in Li et al. (2016). The ocean model is the Modular Ocean Model version 5 (Griffies et al., 2015) with 50 vertical layers, and the ocean horizontal resolution is about 1 • latitude by 1 • longitude. We consider the last 240 years of 340 year-long simulation in which greenhouse gas (GHG) and ozone depleting substance (ODS) forcings are fixed at 1950 levels. Figure 1 compares the 2 meter temperatures over the Niño3.4 region to those over the (top) Indian Ocean and (bottom) Indo-Pacific warm pool region in the coupled model and in MERRA reanalysis data. The model simulates stronger ENSO events than have occurred, similar to the bias in a previous version of this ocean model (Dunne et al., 2012;Capotondi et al., 2015). Biases in climatological zonal wind stress and SSTs 5 in the Pacific are shown in Figures 3 and 4 of Li et al. (2016); briefly, SSTs in the tropical West and Central are too warm, consistent with zonal wind stresses that are not sufficiently easterly. Regardless of these biases, the tendency of EN events to lead to a warmer Indian Ocean is well captured by the model (Figure 1ab). The connection between ENSO and the the Indo-Pacific warm pool region is similar in both the ERSSTv5 dataset (Huang et al., 2017) that is used in Figure 1 and for the Met Office Hadley center observational database (Rayner et al., 2006) (not shown).

10
The foundation of the AGCM ensemble are the simulations discussed by Garfinkel et al. (2015) and Aquila et al. (2016), though several recent integrations have been added as summarized in Table 1 (Rayner et al., 2006) and from the National Climatic Data Center (Reynolds et al., 2002) (Aquila et al., 2016); for these seven integrations we discard the seasons 1991/1992 and 1992/1993 and the years 1991, 1992, and 1993 from consideration, as the eruption of Mt. Pinatubo had a large impact on the BDC and tropical temperatures in our simulations (Aquila et al., 2016;Garfinkel et al., 2017), and appears to have led to moistening in observational data as well (Fueglistaler, 2012;Dessler et al., 2014). In 1994 the difference in entry water vapor between these seven integrations and the other integrations is less than 0.05ppmv (not shown). Four of these seven integrations 25 also include time varying solar forcing. All simulations considered are summarized in Table 1. These simulations have been performed for various purposes and differ in the forcings included and in the physical parameterizations, but they all include changing SSTs and sea-ice.
GEOSCCM model output is compared to temperatures from MERRA and water vapor from SWOOSH v2.5. Temperatures from MERRA are interpolated to the same 2 • latitude x 2.5 • longitude degree grid used for the GEOSCCM simulations. In 30 order to isolate the interannual variability, we detrend timeseries for the AGCM simulations and for reanalysis/observations. Anomalies are computed as follows. A monthly climatology over the full duration of each model experiment, reanalysis product, and observational dataset is computed, and is then subtracted from the raw fields to generate monthly anomalies.
The model climatology is computed separately for each model simulation due to differences in the forcing agents and model components used.

35
ENSO events are identified based on November through February seasonal mean SST anomalies in the ERSSTv5 dataset (Huang et al., 2017) with a 1981-2010 base period. LN events are identified when SST anomalies in the Niño3.4 region (5S-5N, 170W-120W) are more negative than -0.5K, while EN events are identified when SST anomalies in this region are larger than 0.5K. LN and EN events are further categorized into four groups similar to Hurwitz et al. (2014): Central Pacific (CP) EN, 5 characterized by positive SST anomalies in the Niño-4 region (5S-5N, 160E-210E), and Eastern Pacific (EP) EN, characterized by positive SST anomalies in the Niño-3 region (5S -5N, 210E-270E), as well as CP and EP LN events, characterized by negative SST anomalies in the same two regions. EP LN events are identified when the Niño3 anomaly is 0.1 K less than the Niño-4 anomaly. Similarly, EP EN events are identified when the Niño-3 anomaly is 0.1 K larger than the corresponding Niño-4 anomaly. CP EN and CP LN events are identified analogously. All remaining years, either because they are neutral 10 ENSO or because the Niño-3 and Niño-4 anomalies are within 0.1K, are categorized as "other events". The years included in each composite are listed in Table 2.
Most ENSO events peak around December and decay through the following spring. Hence, we focus on the response of the lower stratosphere during the period from November through June.
As discussed in the introduction it is well known that EN forces an intensified BDC, and associated with an accelerated BDC 15 are colder tropical lower stratospheric temperatures and less water vapor. Here we consider the response to ENSO without regressing out the influence of the BDC on water vapor except where indicated, as regressing out the BDC misrepresents the net impact of ENSO on the lower stratosphere. We consider two alternate diagnostics of the BDC: the tropical diabatic heating rate and the mean age; the main text shows results for tropical diabatic heating rate, and the supplemental material shows mean age. Details of the mean age calculation can be found in Garfinkel et al. (2017). 20 A QBO is spontaneously generated in all simulations considered here. The QBO phase is not coherent among these experiments (i.e. the phase does not match observations), and hence the impact of the QBO on e.g. tracer distribution (e.g. Liang et al., 2011) is averaged out when considering the ensemble mean. As the QBO does impact tracer distribution in observations, however, we linearly regress out variability associated with the zonal wind at 50hPa two months prior before considering the response to ENSO.

25
Due to the very slow vertical motions in tropical tropopause layer and relatively faster horizontal motions, entry water vapor is sensitive to the coldest regions in the tropics and not just zonal mean temperatures (i.e. the cold point, Mote et al., 1996;Hatsushika and Yamazaki, 2003;Fueglistaler et al., 2004;Fueglistaler and Haynes, 2005;Oman et al., 2008). We therefore include isotherms corresponding to the coldest region in the tropics on Figures 6, 7, and 8. The climatological cold point is enclosed with a green contour, and the corresponding contour during EN is enclosed in magenta. Temperature anomalies at 30 85hPa resemble quantitatively those at 100hPa, and we therefore show 100hPa anomalies only for brevity.
The adjusted R 2 (eq 3.30 of Chatterjee and Hadi, 2012) is used to quantify the added value in using a polynomial best fit (e.g. instead of a linear best-fit (e.g. H 2 0 ∼ c * EN ) . The adjusted R 2 takes into account the likelihood that a polynomial predictor will reduce the residuals by unphysically over-fitting the data. While in principle the polynomial fit could be preferred if the adjusted-R 2 for the polynomial fit is larger by any amount as compared to the linear R 2 , we elect to be conservative and demand that the adjusted R-squared for a polynomial fit exceed the R 2 for a linear fit by 33%. Note that the 33% criteria is subjectively chosen, though results are similar for a slightly modified criteria.

Linearity of the ENSO effect in the tropical lower stratosphere
We now consider the seasonality and linearity of the ENSO effect in the tropical lower stratosphere. Figure 2 shows the 5 response of temperature, water vapor, and the BDC to ENSO in the coupled ocean-atmosphere run, from November through June. Figure 3 is comparable but for the AGCM integrations, and Figure 4 is comparable but for MERRA and SWOOSH data.
The slope and uncertainty of the linear least-squares best fit is indicated on each panel for integrations where a linear best-fit is deemed satisfactory (see the methods section), while the adjusted R 2 is indicated when a parabolic fit is preferred. Different colors are used to distinguish CP from EP events. 10 We begin with temperature changes in boreal winter. EN leads to strong cooling of the tropical lower stratosphere ( Figure   2ad, 3ad), while LN leads to warming relative to the climatology. This temperature response is consistent, to zeroeth order, with the changes in the BDC associated with ENSO: EN leads to an accelerated BDC while LN leads to a decelerated BDC (Figure 2cf and 3cf; see also the supplemental material). In November through February, the relationship between ENSO and lower stratospheric conditions are linear; that is, the impact of EN and LN events of similar strength is equal and opposite. The 15 magnitude of these effects, as quantified by the best-fit line, appears to be slightly weaker in the AGCM ensemble as compared to the coupled ocean-atmosphere runs, and this could be because of the difference in the nature of ENSO events or decadal variability. The large spread in values for a given event in Figure  While the relationship between ENSO and lower stratospheric conditions is linear in boreal winter, it is nonlinear for both water vapor and temperature in boreal spring (bottom two rows of Figure 2 and 3). Namely, a parabolic (e.g. H 2 0 ∼ a * EN 2 ) 25 fit better describes the relationship between ENSO and water vapor and between ENSO and lower stratospheric temperature than a linear fit (Figure 2ghjk and 3ghjk). Hence, strong EN events lead to less cooling than what might have been expected given a linear best-fit, and consistent with this, the strongest EN events lead to more moistening than might have been expected based on a linear best-fit line. This is especially evident in figure 2hk, where the strongest EN events lead to spring moistening.
The AGCM runs capture this effect as well, as the 97/98 EN also leads to moistening (the most extreme EN event in figure It does not matter whether the ENSO event is categorized as a CP or EP event, as the red, black, and blue dots all indicate the same relationship between ENSO and water vapor. However, the strongest EN events tend to be EP in both nature and in the coupled ocean integration, and hence the nonlinearity is less detectable for CP events. This difference in strength also explains why the compositing approach of Garfinkel et al. (2013a) to characterizing the impact of EP events and CP events can mislead: the atmospheric response to a composite of EP events may differ from the response to a composite of CP events because the 5 events included in the EP composite are stronger, not because of the specific pattern of the SST anomalies.
The response to ENSO in GEOSCCM can be used to inform the interpretation of the observed response to ENSO (Figure 4).
EN leads to an accelerated BDC and a colder lower stratosphere in reanalysis data in January and February, and these changes are statistically indistinguishable from the response in GEOSCCM. More importantly, the qualitatively different behavior for the 1997/1998 event as compared to moderate EN events in the model experiments is also evident in observations in March 10 through June, and hence we recommend caution in generalizing about the tropical lower stratospheric temperature and water vapor response to EN events from the observed anomalies in 1997/1998. However, the relatively short data record limits the confidence with which we can identify nonlinearities in observational/reanalysis data, and none of the linear best-fit slope estimates for SWOOSH water vapor are statistically significant in either winter or spring. net effect of this warming of the climatological cold point region is that the cold point shifts to the east while warming during 97/98 (the magenta isotherm is 0.7K warmer than the green contour in Figure 6). In contrast, during other EP EN events, 30 roughly half of the climatological cold point region warms while the other half cools, and the net effect is that the coldest region shifts east but does not warm or cool overall for typical EP EN events (the green and magenta isotherms in Figure 7 correspond to the same temperature). The eastward shift in Figure 6b and 7ab is consistent with the shift in the Lagrangian cold point evident in Figures 8 and 9 of Bonazzola and Haynes (2004) and figure 8 of Hasebe and Noguchi (2016). In boreal spring, there is broad-scale warming over most of the equatorial band for the 97/98 event (Figure 6cd), while the temperature anomalies are similar to those in winter for moderate EN events (Figure 7cd). A similar effect is seen in the MERRA reanalysis (not shown). The net effect is that in boreal winter and especially spring, the 97/98 event led to warming of the cold point and moistening of the stratosphere relative to other EP EN events.

5
The changes in tropical temperature in GEOSCCM for the 97/98 event and for other events are summarized in Figure 8, which shows the temperature averaged from 5S to 5N from 300hpa to 50hPa. The overall quadrupole structure is similar to that in Liang et al. (2011) and Garfinkel et al. (2013b), and there is an eastward shift of the cold point region. The model captures the warming pattern in reanalysis (compare Figure 8 and S1). Most pertinently, there are clear differences between the changes in 97/98 and those in other EN years: the tropospheric warming is more pronounced and widespread in 97/98 from March 10 through June. The net effect is that the cold point region warms in 97/98 but not in the other EN years.
It is important to emphasize that this nonlinearity in the temperature and water vapor response does not involve stratospheric dynamics. The changes in the BDC appear to be mostly linear in Figures 2, 3  This suggests that the Central and East Pacific responses cannot explain the difference in stratospheric response. In contrast, these two events differed quite dramatically in the Indian Ocean (and more generally in zonally averaged tropical temperature).
The 1997/1998 event led to remarkable impacts in the Indian Ocean: warm anomalies exceeded 2C locally over the West Indian Ocean and enhanced convection over Africa was anomalously strong even for EN (Webster et al., 1999;Su et al., 2001). Sea surface temperatures north of the equator were anomalously warm throughout 1998 as well (Yu and Rienecker, 2000). The cold 25 point moves toward India over the course of boreal spring (e.g. Bonazzola and Haynes, 2004;Garfinkel et al., 2013a) and thus warming in this area can impact water vapor. This difference in near surface conditions in the Indo-Pacific and Niño3.4 region is quantified in Figure 1. El Niño events are followed by warming throughout the Indo-West Pacific (Figure 1ac). Conditions during the 1982/1983 event are shown with a red diamond, and during the 1997/1998 event with a large red x. Despite largely similar anomalies in the Niño3.4 region, the 1997/1998 event was characterized by remarkably warm anomalies in the  Pacific that lie in the tail of the warming generated spontaneously in the coupled ocean-atmosphere model.
The importance of Indian Ocean SSTs for entry water vapor is quantified in Figure 9, which shows the regression coefficient between 85hPa water vapor and 2meter temperatures from 5S to 5N at each longitude grid point. We show both the regression coefficient in the annual average with no lag between water vapor and surface temperature and in boreal spring with 2meter temperatures leading water vapor by two months (Garfinkel et al., 2013a). The black curve shows the regression after linearly 35 regressing out the BDC and the QBO from the water vapor, and the blue curve regression after linearly regressing out the QBO from the water vapor.
In the annual average, warmer near-surface temperatures over the Central and Eastern Pacific lead to dehydration of the stratosphere in all three data sources (black curves in Figure 9ace), though during boreal spring warming in the eastern Pacific leads to moistening of the stratosphere two months later. More importantly however, stratospheric water vapor is most sensitive 5 to variability in the Indian Ocean basins and the Warm Pool region, with warmer temperatures in this region leading to enhanced water vapor in all three data sources in boreal spring (and if the BDC influence on water vapor is regressed out, also in the annual average). While the importance of a large regression coefficient in a given region depends on the magnitude of near surface temperature variations in that region, results are similar if correlations are examined (not shown). In summary, an ENSO event that more efficiently warms the mid-troposphere (such as the 1997/1998 event) by modifying 20 SSTs in the Indian Ocean can more efficiently moisten the stratosphere. Strong EN events tend to have a stronger impact on the Indian Ocean than more moderate events (cf. Figure 1), and hence their impact on the tropical lower stratosphere in the boreal spring and early-summer is more pronounced, which ultimately leads to nonlinearity in the connection between EN and the tropical lower stratosphere.
6 Implications for the drop in the early 2000s and the 2015/2016 EN event 25 It has been suggested that SST changes in the Indo-Pacific contributed to some of the drop in water vapor after the year 2000 (Rosenlof and Reid, 2008;Garfinkel et al., 2013b) via ENSO (Brinkop et al., 2016), and here we consider whether the AGCM simulations simulate a drop. Before proceeding, it is important to mention that the 1997/1998 El Niño was followed by nearly three consecutive years of strong La Niña conditions -the Niño3.4 index in the ERSST5 dataset did not drop below -0.5K until March 2001 -which was then followed by weak El Niño conditions from 2002 through 2004. As discussed above, strong La 30 Niña events also lead to moistening of the stratosphere, while weak El Niño lead to dehydration. The net effect is that ENSO was in a phase that leads to enhanced water vapor during 1998, 1999, and 2000 and in a phase that leads to reduced water vapor from 2002 to 2004. It has already been documented that QBO and BDC variability are key ingredients for the observed drop (Randel et al., 2006;Fueglistaler, 2012;Fueglistaler et al., 2014;Dessler et al., 2014). Note that the QBO phase in these GEOSCCM experiments does not match that observed, and the specific wave events that drove the accelerated BDC in late 2000 are not nudged to occur in these free-running GEOSCCM simulations either. Hence we do not expect to be able to capture the full magnitude of the drop. However these experiments can be used to quantify the contribution of SSTs to the difference in water vapor between 2002 through 2004 and 1998 through 2000, and with these caveats duly noted we now proceed.  (2012) and Hasebe and Noguchi (2016). The mean value is approximately one-quarter of the total drop (which equals 0.62ppmv in the deep tropics if we apply the same definition to SWOOSH data, though as shown by Fueglistaler et al. (2013) the different satellite products that underly the SWOOSH data disagree as to the magnitude of the drop.) As discussed above, the rest of the drop is associated with BDC 15 and QBO variability which these GEOSCCM simulations are not expected to capture. Hence, our GEOSCCM simulations suggest that SST changes contributed to the drop (in agreement with Rosenlof and Reid, 2008), but were not the major forcing factor, consistent with Garfinkel et al. (2013b), Brinkop et al. (2016), and Ding and Fu (2017). Note that these integrations also simulate a drop after 2011 (Urban et al., 2014;Gilford et al., 2016), suggesting that part of this drop was forced by SSTs as well. 20 Finally, five of the integrations have been extended to the near present and hence include the 2015/2016 El Niño event. This event was comparable in strength in the Niño3.4 region to that in 1997/1998, and while it satisfies the criteria we adopt for an EP event, it was less strongly eastern Pacific-focused as compared to the 1997/1998 event. We now consider the evolution of water vapor in those integrations in Figure 11. Note that these simulations are forced with time varying SSTs and sea ice only.
The model simulates a 0.5ppmv increase in H2O in 2016 (annual average) as compared to 2015, approximately 70% of the 25 observed increase. Hence, the model is clearly capable of capturing the enhanced stratospheric water vapor following strong EN events. The seasonal evolution of the change is shown in Figure 5c, and the increase in water vapor occurs in the March after the EN event has already begun to decay. The moistening in 2016 is comparable to that in 1998 (cf. figure 5ac). Note that the QBO phase in GEOSCCM does not match that observed, and hence we are not surprised that model misses the observed pronounced drying that occurred in mid-2016 and late-2016 due to the QBO disruption (Tweedy et al., 2017). In summary, Tropical lower stratospheric temperature and water vapor changes have important implications for both stratospheric and tropospheric climate as well as stratospheric ozone chemistry (SPARC-CCMVal, 2010;World Meteorological Organization, 2011. Hence, it is crucial to understand interannual changes in this region in order to correctly interpret future changes.
Analysis of a series of chemistry-climate atmospheric model in two distinct configurations -coupled to an interactive ocean 5 model and forced by historical sea surface temperatures -yielded the following conclusions: 2. There is no appreciable difference in the tropical lower stratospheric response to Central Pacific versus Eastern Pacific El Niño events, if one controls for the amplitude of the El Niño event. As Eastern Pacific El Niño events tend to be stronger, however, the nonlinear effects discussed above are pronounced mainly for events of this type.
3. The very strong El Niño event in 1997/1998 followed by more than two consecutive years of La Niña led to enhanced lower stratospheric water vapor. As this period ended in early 2001, entry water vapor concentrations declined. We quantify this effect using a large ensemble of AGCM simulations with imposed SSTs, and find that the deterministic part of the water-vapor drop arising from these imposed SSTs is about one-quarter of that actually observed, in agreement with the recent estimate of Ding and Fu (2017)  This study leaves several unanswered questions. First, the ENSO amplitude in the ocean model used here for our coupled ocean-atmosphere simulations is too large (Capotondi et al., 2015), and mean state biases are also present (e.g. figures 3 and 4 of Li et al., 2016); the results from GEOSCCM presented here need to be confirmed with other models. Second, it is not clear mechanistically how upper tropospheric warming over the Indo-West Pacific leads to moistening of the stratosphere in boreal spring. However this effect appears to be consistent with recent suggestions that mid-tropospheric warming can directly lead to 5 a warmer cold point tropopause and wetter stratosphere (Dessler et al., 2013(Dessler et al., , 2014. Third, and relatedly, we cannot provide a complete explanation of how El Niño modulates stratospheric water vapor. Inter-annual variability in stratospheric water vapor, particularly that associated with El Niño, depends both on a 'sampling effect' (i.e. changes in the residence time in the coldest regions of the tropical tropopause layer) and a 'temperature effect' (Bonazzola and Haynes, 2004;Hasebe and Noguchi, 2016;Konopka et al., 2016). These two effects cannot be distinguished in our simulations, since the model output necessary to run 10 a Lagrangian trajectory model was not archived. Nonetheless neither is prevented from operating in the simulations and the simulated interannual variability in water vapor will arise from some combination of the two. In addition, direct injection of cloud ice may be important for stratospheric water vapor during El Niño: Avery et al. (2017) find enhanced cloud ice in Calipso data in December 2015 during the most recent strong El Niño event. We therefore briefly consider whether the model can capture this effect in Figure 11b, which shows tropical cloud ice between 5S-5N at 100hPa. Our GEOSCCM simulations 15 capture a jump in tropical cloud ice at 100hPa of around 0.5ppmv associated with this event, in general agreement with Calipso data (Avery et al., 2017), and even at 85hPa cloud ice increases by 0.05ppm in the zonal mean. The spatial distribution of the change in cloud ice at 85hPa in December 2015 is shown in Figure 11c; the pattern of anomalous ice matches that found in Calipso data (see Figure 1 of Avery et al., 2017). While the ice cloud parameterization in this version of GEOSCCM is crude, the qualitative agreement between Calipso and GEOSCCM suggest that direct injection of ice may not be an insignificant 20 pathway for stratospheric water vapor during strong El Niño events, and this effect should be explored as models improve.
More generally, entry water vapor may be influenced by physical processes that are missing or poorly-represented by the current generation of climate models, and hence all results shown here with regards to water vapor should be re-evaluated as models improve.
However, the nonlinearity of the lower stratospheric response in temperature and water vapor to El Niño is robust and appears 25 to depend on large scale circulation and temperature anomalies, which we expect our model to capture. Hence caution must be exercised when deciding on a methodology for analyzing the tropical stratospheric response to El Niño.  Li et al. (2016) historical SSTs SST+sea ice   1982/1983, 1986/1987, 1991/1992, 1997/1998CP El Niño 1994/1995EP La Niña 1984/1985, 1985/1986, 1995/1996/2008CP La Niña 1983/1984,1988/1989, 1998, 2008  (left) temperature at 85hPa, 5S-5N; (center) water vapor at 85hPa, 5S-5N; (right) diabatic heating rate at 70hPa, 5S-5N. For all quantities, the data has been detrended (see section 3) and the component of the variance linearly associated with the QBO at 50hPa two months prior has been regressed out before data is stratified by the Niño3.4 index (see section 3). Winters categorized as Central Pacific ENSO are in black, Eastern Pacific ENSO are in blue, and all other years in red. A linear least-squares best fit is shown in each panel, and the slope is indicated, when a linear fit is deemed satisfactory (see section 3). When a polynomial fit better describes the dependence on ENSO, we show the R 2 for a linear fit and adjusted R 2 for the polynomial fit (see section 3).      Figure 7. As in figure 6 but for all ENSO events except 1982ENSO events except /1983ENSO events except and 1997ENSO events except /1998   (e-f) for SWOOSH water varpor and MERRA 2 meter temperatures. The longitude bands corresponding to the Indian Ocean, Niño3, and Niño4 regions are in color. The left column is for annual averaged quantities and the right column is for March through June water vapor with T2m two months prior. We show results regression coefficients after first regressing out the effect of the QBO at 50hPa from the water vapor anomalies (black) and also after regressing out the effect of the QBO at 50hPa and the BDC from the water vapor anomalies (blue).