Impact of precipitation and increasing temperatures on drought trends in eastern Africa

In eastern Africa droughts can cause crop failure and lead to food insecurity. With increasing temperatures, there is an a priori assumption that droughts are becoming more severe, however, the link between droughts and climate change is not sufficiently understood. In the current study we investigate trends in long-term agricultural drought and the influence of increasing temperatures and precipitation deficits. Using a combination of models and observational datasets, we studied trends, spanning the period from 1900 (to represent the 5 pre-industrial era) to 2018, for six regions in eastern Africa in four drought-related annually averaged variables — soil moisture, precipitation, temperature and, as a measure of evaporative demand, potential evapotranspiration (PET). In standardized soil moisture data, we found no discernible trends. Precipitation was found to have a stronger influence on soil moisture variability than temperature or PET, especially in the drier, or water-limited, study regions. The error margins on precipitation-trend estimates are however large and no clear trend is evident. We find significant positive trends in local temperatures. However, 10 the influence of these on soil moisture annual trends appears limited. The trends in PET are predominantly positive, but we do not find strong relations between PET and soil moisture trends. Nevertheless, the PET-trend results can still be of interest for irrigation purposes because it is PET that determines the maximum evaporation rate. We conclude that, until now, the impact of increasing local temperatures on agricultural drought in eastern Africa is limited and we recommend that any soil moisture analysis be supplemented by an analysis of precipitation deficit. 15 Copyright statement. TEXT

1 mational value is limited. We therefore analyse evaporative demand, which is also referred to as "potential evapotranspiration" PET. PET is the amount of evaporation that would occur if an unlimited supply of water were available, which is calculable or available for both observations and model simulations and is a function of temperature, humidity, solar radiation and wind speed.
We investigate evaporative demand as a means to study the influence of temperature on soil moisture, however, for regions 5 that are irrigated or where irrigation is being considered, evaporative demand itself can be regarded as more relevant than soil moisture as a measure of drought tendency.
Whilst attribution studies specifically for the east African region have not previously used soil moisture or PET to explore drought, PET has been used in various attribution or trend studies outside our region of study, to explore for example, the influence of climate change on the hydrological cycle in China (e.g. Yin et al., 2010;Li et al., 2014;Fan and Thomas, 2018), 10 trends and variability at sites in West Africa Obada et al. (2017) and compound events of low precipitation and high PET in Europe Manning et al. (2018).
Summarizing, the objectives of this study are to (i) consider the attribution question "do increasing global temperatures contribute to drier soils and thus exacerbate the risk of agricultural drought (low soil moisture) in eastern Africa?" and (ii) to investigate if global-warming driven trends in precipitation or local temperature via PET explain any emerging trend in 15 agricultural drought. Our approach to attribution comprises the following steps: (1) Definition of the study variables and explanation of the study regions, (2) Description of observational data and detection of trends in observations (3) Model evaluation including description of the models, (4) Attribution of trends in models, (5) Synthesis of the results. Assessments will be based on both observations and climate and hydrological model output on the annual time scale. We will illustrate the method using examples of recent droughts in eastern Africa. 20 The outline of the remainder of the paper is as follows: In Section 2 the chosen study regions are presented followed by a description of the datasets used in the study. In Section 3 we describe the stepwise approach to attribution used in this paper, including assumptions and decisions made and illustrative examples. In Section 4, the results are synthesized per region.
Finally, the discussion and conclusions are presented in Sections 5 and 6.
2 Study variables, region and datasets 25 In this section, we present the chosen study variables and study regions in eastern Africa and the datasets used to provide the variables to be analysed. Brief descriptions of the projects from which the datasets originate are provided in the supplement.

Study variables and region
We analyse four different variables: soil moisture, precipitation, temperature, and PET. We average these variables over six regions, as trend analyses of time series of regionally averaged quantities are more robust than the same analyses for point 30 locations. This is especially true for precipitation, which shows small-scale spatial variability if the time period is not long  after Pricope et al. (2013), which were also used to define the study regions.
enough to sufficiently sample the distribution from multiple precipitation events. It is however necessary to select homogeneous zones, so that the signals present are not averaged out.
The focus of the study is on eastern Africa -Ethiopia, Kenya and Somalia (including the Somaliland region). We selected six regions based on precipitation zones, in which the annual mean precipitation and seasonal cycle are homogeneous (Fig. 1a Table 1. Data is annually and spatially averaged over the study regions.

Datasets
For the four study variables, we use as many datasets as readily available over the study area, provided that (i) the data are sufficiently complete over a time period long enough to be used for trend calculations and (ii) the model data pass the validation tests (see Sect. 3). For this purpose, we decided to use time series of 35 years and longer. As the focus of this paper is on annual time scales, using monthly data is sufficient. The observational and model datasets used in this study are shown in Fig. 2 and 5 listed in tables 2 and 3 below. For brief descriptions of the projects from which these data originate, please see the Supplement.
Note that we use the data as it is available without applying any additional bias correction. Some of the data has undergone bias correction within project of origin, as described in the Supplement.
For observations of precipitation and daily mean near-surface temperature, we use gridded observational data sets and reanalyses.

10
For soil moisture and PET, no direct observations meeting the above criteria exist. Instead, we use observational estimates of soil moisture and PET resulting from various combinations of observational forcing data and models (see Fig. 2a).
Concerning soil moisture, observational series are few and generally too short to use for trend analysis and they do not correlate well with reanalysis or model data over eastern Africa . It is therefore important to use multiple observationally forced model estimates to span the large uncertainties from inter-dataset differences. There being no a priori 15 reason to favour one soil moisture dataset over another, we treat all resulting soil moisture datasets equally. For all soil moisture data sets, observed and modelled, we use the topmost layer (see Fig. 2 for the depth of the topmost layer) provided by each dataset and scale each time series to have a standard deviation of 1 in order to make comparisons in trends possible. An exception to this is weather@home where the available soil moisture variable is an integrated measure of all four layers of soil moisture in the model, including the deep soil.

Methods
In this section we describe first the method we use for detection and attribution of trends in the four variables, including model validation and the synthesis of observational and model results. Furthermore, in subsection 3.2 we describe the assumptions    and decisions that are made concerning the data/model setup and in subsection 3.3 we provide an example of how the method is applied to real data.

Detection and attribution of trends
We use a multi-method, multi-model approach to address attribution. We use global mean surface temperature (GMST) as a measure for anthropogenic climate change for calculating trends. We calculate trends for all variables, regions and datasets 5 and synthesize results into one overarching attribution statement for each of the four variables in each of the six regions. We use this method, following the approach applied in earlier studies on drought in eastern Africa (e.g., Philip et al., 2018a;Uhe et al., 2018) and other drought-and heat-attribution studies (e.g., Philip et al., 2018b;van Oldenborgh et al., 2018;Kew et al., 2019;Sippel et al., 2016), which represents the current state of the art in extreme event attribution. The method is extensively In this study, for transient model runs and observational time series, we statistically model (i.e., fit) the dependency of annual means of the different variables on GMST, (the model GMST for models, and GISTEMP surface temperature GMST (Hansen et al., 2010) for observations and reanalyses) as follows:

5
After inspection of whether a Gaussian or a General Pareto Distribution fits the observational or reanalysis data best, we use the following distributions: for soil moisture: a Gaussian distribution that scales with GMST, focussing on low values, for precipitation: a General Pareto Distribution (GPD) that scales to GMST, analyzing low extremes for temperature: a Gaussian distribution that shifts with GMST, focussing on high values, and 10 for PET: a Gaussian distribution that scales with GMST, focussing on high values.
When the distribution is shifted, a linear trend ↵ is fitted by making the location parameter µ dependent on GMST as with ↵ in [units of the study variable]/K. When the distribution is scaled, which keeps the ratio of the location and scale parameter /µ invariant. In each case, the fitted distribution is evaluated twice: it is not necessary to use a fitting routine as the large amount of data permits a direct estimation of the trend. This also provides an opportunity to check the assumptions made in the fitting, notably that the values follow an extreme-value distribution and 25 that the distribution shifts or scales with the smoothed GMST. We calculate trends for the time series of spatially and annually averaged data of all four variables and all six regions for all datasets by dividing the difference in the variable between the two ensembles by the difference in GMST. We only use results from model runs if they pass two different validation tests -a qualitative test on the seasonal cycle and a stronger test on variability. For soil moisture, due to the difficulties in obtaining reliable soil moisture measurements (e.g., Liu and Mishra, 2017) and the differences between the observational (reanalysis) datasets, we cannot assume that observational or reanalysis data are more accurate than model data. Therefore we simply use the soil moisture model data if the model input - PET and precipitation -passed the validation tests.

5
We perform only a qualitative validation of the seasonal cycle. For each region, each variable, and each model we check that the seasonal cycle resembles that of at least one of the observational datasets, in both the number and the timing of peaks. If the seasonal cycle is very different, we do not use the time series for that specific combination. This is the case for the original GCM precipitation in region NK for weather@home and in regions NK and CK for MIROC (the seasonal cycle is improved in the adjusted dataset, so we still use the time series in soil moisture) and for temperature in region SS for EC-Earth (we do 10 not have adjusted data to check so we do not use this model-region combination for soil moisture or PET).
The second validation test is on the model variability in precipitation and PET (variability relative to the mean for variables that scale with GMST). If the model variability of a specific variable in a specific region is outside the range of variability calculated from observations or reanalyses, we do not use that specific dataset for that specific region and variable. For temperature, we relax the validation criteria on variability as it became clear during the analysis that the trend in soil moisture 15 does not depend strongly on temperature and the trend in temperature agrees between models and observations. In two of the  models is evident (we see a small increase in variability with time for Berkeley, a small decrease for CRU and no consistency between the models). This does not affect the trend much, which is evaluated for the centre of the distribution. Trends are presented as change in a variable per degree of GMST warming. We show trends rather than probability ratios, 5 which conveniently results in finite ranges in confidence intervals for all variables. This is not the case for the probability ratio, where, for example, strong trends in temperature imply that mild extremes of the 2018 climate (e.g., a 1-in-20-year event) would have had a chance of almost zero around 1900, resulting in very large probability ratios and extensive extrapolation of the fit beyond the length of the dataset.
We synthesize the trends of all data that passes the validation tests in the following manner, see also Fig. 5. The observational 10 (reanalysis) estimates are based on the same natural variability: the historical weather. They also cover similar time periods.
The uncertainties due to natural variability (denoted as solid blue in the synthesis figures) are therefore highly correlated.
We approximate these correlations by assuming the natural variability to be completely correlated, and compute the mean and uncertainties as the average of the different observational estimates. The spread of the estimates is a measure of the representation uncertainty in the observational estimate and is added as an independent uncertainty to the natural variability 15 (black outline boxes). This results in a consolidated value for the observations (reanalyses) drawn in dark blue. In contrast, model estimates have more uncorrelated natural variability: totally uncorrelated for coupled models; and largely uncorrelated for SST-forced models (the predictability of annual mean precipitation given perfect SSTs is low in eastern Africa). We approximate these correlations by taking the natural variability to be uncorrelated. The spread of model results can be compared by the spread expected by the natural variability by computing the 2 /dof statistic. If this is greater than one, 20 there is a noticeable model spread, which is added in quadrature to the natural variability. This is denoted by the white boxes in Fig. 5. The bright red bar indicates the total uncertainty of the models, consisting of a weighted mean using the (uncorrelated) uncertainties due to natural variability plus an independent common model spread added to the uncertainty in the weighted mean.
Finally, observations and models are combined into a single result in two ways. Firstly, we neglect model uncertainties 25 beyond the model spread and compute the weighted average of models and observations: this is indicated by the magenta bar.
However, we know that models in general struggle to represent the climate of eastern Africa, so the model uncertainty is larger than the model spread. Therefore, secondly, we use the more conservative estimate of an unweighted average of observations and models, indicated by the white box around the magenta bar. This gives more weight to the observations. 30 We made the following assumptions and decisions about the data and model set-up in addition to completing the model evaluation to attain data of sufficient quality.

Assumptions and decisions
1. As was shown by Funk et al. (2015), the CenTrends precipitation dataset includes many different sources of precipitation data and more stations than most other datasets. We therefore assume for precipitation that the CenTrends dataset is superior to other datasets over our region of study. We therefore only use the CenTrends dataset for observations of precipitation.
2. In general we use the longest time series of data available. We make exceptions in the starting year if necessary, based 5 on visual inspection of abrupt changes due to data limitations toward the beginning of the time series. 3. We do not know of a realistic soil moisture dataset that covers a long-enough time period to calculate trends. Therefore we do not select simulations based on evaluation criteria other than selecting runs based on PET and precipitation evaluation 10 in the input variables.
4. As models do not share a consistent set of soil moisture levels, we take the top level of each model, assuming that this is the most comparable level across models. We checked for LPJmL -the only selected ISIMIP hydrological model that has more than one level available for soil moisture -that the variability does not change by much when integrating over multiple levels instead of using level 1. 15 5. Within the ISIMIP project, variables required by the hydrological models, including temperature and precipitation, were bias corrected and the adjusted data was used to calculate PET and to drive the hydrological models to output soil moisture. In the synthesis, however, we present results for temperature and precipitation based on the unadjusted data, on principle that this better spans the range of model uncertainty in temperature and precipitation. The bias correction applied in ISIMIP aims to conserve the original trend. In accordance, we find little change in trend for most time series, 20 see Section 5.
6. Instead of PET, RefET(reference evapotranspiration) was available for the MERRA dataset. RefET can be converted to potential evapotranspiration by multiplying its value to a reference crop coefficient. We assume using RefET instead of PET does not influence the overall conclusions and we do not convert RefET into PET. relatively similar to each other. The forcing data is the same for the years 1860-2005 and only differs for the most recent years, from 2006 onwards. In general, however, using different scenarios can be seen as an advantage, as a greater range of scenario uncertainty will be spanned.
(a) We use RCP6.0 in ISIMIP as this choice resulted in the largest number of simulations and RCP8.5 in EC-Earth as this was the only scenario available. 8. Trends are calculated or extrapolated using all data up to 2018 and between the pre-industrial era (1900) and the present 5 (2018). Weather@home is an exception where trends are calculated between two stationary climates of the present and the pre-industrial era. Differences in trends can arise due to different time periods and lengths of datasets, which are generally shorter for observations and reanalyses than for model simulations. However, we consider the use of all available observational and reanalysis data and different model framings to lead to a more complete and robust attribution statement. 10 9. We analyse Jan-Dec annual means. Based on the seasonal cycles of precipitation and temperature, for all regions except for region WE (which has a single rainy season) we could also have chosen to analyse Jul-Jun annual means instead.
The influence of this choice on the trends is low (see also Sect. 3.3). 10. For consistency in the method, we fit the variability as a constant over time for all data. In both observed time series and simulations we see very little or no trend in variability up to 2018. 15 11. If for observational data a Gaussian fit is the best fit, we also fit model data to a Gaussian, even if a GPD is a better fit for that data. In doing this we avoid erroneous comparisons between the variable mean and variable extreme. We checked for model runs in which this disparity occurs, but found that in most cases the trend calculated from fitting model data to a GPD was not very different from the trend calculated from fitting model data to a Gaussian.

Illustrative examples 20
In this section we show an example to illustrate the method of detection of trends in precipitation data, as people often initially experience droughts as reduced or failed rainy seasons. For this purpose, we calculated return periods and risk ratios of recent droughts defined as low-precipitation events on the annual time scale, see Table 4. Note that the risk ratios are calculated from The results show, for instance for region WE, that in CenTrends the trend in precipitation between 1900 and 2018 is -0.09 mm/day/K (95% confidence interval (CI) -0.51 to 0.14 mm/day/K). With a change in GMST of 1.07 K and a mean precipitation in 1900 of 3.2 mm/day this is similar to a change of 3%. This means that if an event with the same precipitation amount as in the year 2009 had happened again in 2018 it would have been a one in 30 (95% CI 2 to 400) year event in 2018, whereas in 1900 it would have been a one in 80 (95% CI 30 to 1400) year event, corresponding to a probability ratio of 2.5 (95% CI 5 0.2 to 380). A return period that decreases in time indicates that such extreme droughts are becoming slightly more common, however, in this example we see large uncertainties consistent with no change. Note that the trend and probability ratio are not significantly different from zero at p < 0.05. The results for all regions are summarized in Table 4. We note that the trends calculated for the Jan-Dec events and for the Jul-Jun events in regions NK and SS respectively are not significantly different.
This supports the decision to analyze Jan-Dec annual extremes only. there is a tendency that the positive trends in PET are also stronger and vice versa; (ii) the uncertainty in precipitation trends is high compared to the trend magnitudes. This is one of the reasons why a clear relation with soil moisture trends is not evident; 10 (iii) no clear relation between local temperature trends and soil moisture trends is evident.
Looking at the different hydrological models we conclude that the trend in PCR-GLOBWB PET, which uses the Hamon PET scheme that depends only on temperature, is generally higher than the trend in in EC-Earth PET, which uses the more-complex Penman-Monteith PET scheme that additionally depends on humidity, wind and radiation. Using this more complex scheme can influence the trend in soil moisture, especially in wetter regions.
15 Table 4. Trends, return periods and probability ratios of equivalent events in the year 2018 and 1900 for three recent drought events registered in the EM-dat database (2005, 2009 and 2010/2011), based on annual average precipitation (mm/day) from the CenTrends dataset. 95 % confidence intervals are given between brackets. For each study region impacted by the events, the annual precipitation for the event year (P rcp, used to define the event magnitude) and the 1900-2014 climatological precipitation average (ClimP rcp) is given. The asterisk (*) denotes that Jul-Jun is taken instead of Jan-Dec to define a year.   The analyses of the individual model runs, stratifying by GCM or hydrological model, do not lead to a clear conclusion on the relation between the trends in precipitation, temperature, PET and soil moisture. We therefore turn to the analysis of the synthesized values, see Table 5 and Fig. 6 for a summary of the outcome and Fig. 5 and Figs. S1 to S6 in the Supplementary Information for synthesis diagrams. The table gives a concluding interpretation of the synthesized results shown in Fig. 6. For soil moisture we find no significant synthesized trends: there is practically no change in region EE and no trend to a 20 small positive non-significant trend in regions WE, NS, NK, CK and SS.
For precipitation, regions WE and NK show a positive but non-significant trend, in region NS there is a small positive trend, regions EE and CK show no trend and region SS a negative non-significant trend.
As expected from global climate change, the local annually averaged temperatures all have a significant positive trend, with best estimates between 1.0 and 1.3 per degree of GMST increase. Related to this, trends in PET are also positive in four of the six regions but lower than for temperature and generally with larger confidence intervals. The regions NK and CK are the exceptions. Although weighted averages show positive trends, models show tendencies opposite to observations. This incompatibility renders the results uncertain.

5
We can identify the following relationships between different variables: (i) Precipitation trends have a (small) influence on soil moisture trends in regions WE, NS and NK; (ii) in regions WE, EE, NS, NK and CK, temperature and PET have no discernible influence on soil moisture trends; (iii) in region SS, the non-significant negative trend in precipitation does not lead to lower soil moisture and neither do the trends in temperature or PET.

Discussion
In this section, we discuss the interpretation of our results in the light of how choices and assumptions made may have influenced the outcome and we compare previous studies on similar topics.
We study drought trends on annual as opposed to sub-annual time scales, as long-term drought presents a greater risk for food security. We define the annual period to be from January to December. This definition is a natural choice for each of our On the annual time scale, we do not see strong explanatory relationships between the trends in the four studied variables.
To gain insight in the relationships between the variables, we additionally looked at correlations on a sub-annual time scale.
Simple correlations between monthly precipitation, temperature, PET and soil moisture (not shown) support the conclusions of Manning et al. (2018) on the influence of precipitation and PET on soil moisture at dry sites in Europe. They found that at 15 water-limited sites the influence of precipitation on soil moisture is much larger than the influence of temperature, via PET, on soil moisture. In our study, we find the same for the driest regions and the driest months in the wetter regions, and for the more temperature-based PET schemes.
Looking at seasonal cycles -monthly means averaged over recent decades -a comparison between seasonal cycles of the different variables shows that the seasonal cycle of soil moisture is similar to that of precipitation in all six study regions. 20 In contrast, the inverse seasonal cycle of temperature is not similar to that of soil moisture. Whether the PET seasonal cycle reflects elements of the soil moisture cycle or not depends on the PET scheme used: temperature-or radiation-based schemes show a seasonal cycle that is similar to that of temperature, whereas more advanced schemes reflect a mixture between the seasonal cycles of precipitation and temperature.
We thus conclude that the influence of precipitation on soil moisture is higher than that of temperature or PET. This is 25 supported by the synthesized results that show negligible or no trends in soil moisture and precipitation whereas the trends in temperature and PET are strongly positive.
If temperature has, via PET, an influence on trends in soil moisture, we expect to see that the positive trend in temperature is coupled to a drying trend in soil moisture. As we average over the annual scale, we may miss parts of the season when this effect is strongest. Therefore we selected a region and period outside the rainy season, in which the seasonal peak in 30 temperature corresponds to a dip in soil moisture (region CK, months Feb-Mar), to inspect sub-annual trends (not shown).
Even then, we find that there is no negative trend in soil moisture accompanying the positive temperature trends.
While improving the data with respect to some characteristics, an additional uncertainty arises from the bias correction of the GCM data prior to use in the hydrological model. The bias correction in ISIMIP was set up to preserve the long-term trend, but it also decreases the daily variability by truncating extreme high values (e.g., in precipitation) (Hempel et al., 2013). The most important element for our analysis is that it also increases the daily variability by removing excessive drizzle, which is often present in GCM precipitation data. Prudhomme et al. (2014) noted that such a statistical bias correction can influence the signal of runoff changes but that the effect generally remains smaller than the uncertainty from GCMs and global impact models. By far the largest difference we found in our analysis between trends in original and bias-corrected data was for temperature for 5 IPSL in region NK: we found 1.9 K/K (95% CI 1.8 to 2.1 K/K) for the original trend and 1.4 K/K (95% CI 1.3 to 1.5 K/K) for the trend in bias-corrected data. All other differences were smaller and non-significant.
A study by Rowell et al. (2015) discussed the possibility that climate model precipitation trends in East Africa are influenced by inability of the models to represent key physical processes reliably. In attribution studies on drought, especially for this region, it is therefore high priority to extend model evaluation techniques to assess models' representation of key physical 10 processes. The approach taken in this paper has been to apply simple evaluation techniques on the seasonal cycle and frequency distributions of readily available data and that results from models passing validation tests represent the status of our current knowledge. Rainy seasons in this region are governed by large-scale processes, such as the shifting of the ITCZ and ENSO dynamics. The ability of a model to capture the seasonal cycle in precipitation and temperature thus provides some assurance that large-scale physical processes are reasonably well described by the model. We see the tests we perform as a minimum re-15 quirement for model validation. However, to improve the performance of models and to understand the discrepancies between models and observations, a much more thorough investigation into the models' representation of physical processes and feedbacks is required, such as demonstrated by James et al. (2018) and encouraged by the IMPALA (Improving Model Processes for African Climate) project (https://futureclimateafrica.org/project/impala/).
It is still unknown how vegetation will respond to substantial increases of CO 2 concentration. Two counteracting effects 20 -physiological (restriction of stomatal openings leading to decreased evapotranspiration) and structural (increased leaf area leading to more stomata and increased evapotranspiration) responses -are expected, but their net effect is unknown (e.g. Wada et al., 2013). So-called 'dynamic vegetation models' include these CO 2 effects and there are indications that these models show a weaker response of drought to climate change (Wada et al., 2013;Prudhomme et al., 2014). In this study our selection of hydrological models is restricted by the variables we require, however, out of the four ISIMIP hydrological models 25 that match our criteria, one (LPJmL) uses dynamic vegetation modeling. The soil moisture response to increasing GMST in LPJmL simulations is mid-range amongst the ISIMIP results. The PET response for LPJmL simulations is, however, somewhat on the low side of the ISIMIP results. It has not been verified if this behaviour is linked to dynamic vegetation modelling, but with confidence intervals generally overlapping with the synthesized model outcome, there is no exceptional difference.
The approach taken in this paper towards uncertainty has been to 30 -Perform a multi-model and multi-observation analysis that summarises what we know at the present moment, using readily available data and methods.
-Apply simple evaluation techniques to readily available data, treating datasets that satisfy evaluation criteria equally and rejecting the others.
-Communicate uncertainties from synthesis. A simple 'yes' or 'no' is not appropriate in this analysis where there is no clear significant positive or negative trend. Rather, the uncertainties (confidence intervals) and their origin (e.g. natural variability or model spread) are given.
In the long term, a trend in PET only has meaning for crop growth if there is water available for evaporation. Much of eastern Africa is in a water-limited evaporation regime. In the case that irrigation would be locally applied, more water would become 5 available for evaporation, shifting the situation away from a water-limited regime and towards an energy limited regime. A trend in PET seen in our analyses (especially if the analysis using different schemes produces a robust PET trend) could then signify a trend in real evaporation and would therefore be accompanied by an increase in irrigation water demand. Note that irrigation is not accounted for by the models or reanalysis datasets used here.
Previous studies have shown that both the PET scheme and the input data used for calculation of PET can have a large influ-10 ence on PET values (Trambauer et al., 2014;Wartenburger et al., 2018). We confirm this using the CLM-ERA-PT (Priestley-Taylor), CLM-WFDEI-PT and CLM-ERA-PM (Penman-Monteith) datasets (not shown). In our study regions, PET values are consistently higher when using PM then when using PT. The differences in trends in PET using ERA or WFDEI input or using PT or PM input are sometimes significant. However, comparing study regions, there is no consistency in the difference; in four out of the six regions the PM data shows a higher trend than the PT data and in four out of the six regions WFDEI data shows 15 a higher trend than the ERA data.
There is some evidence that warm spells are increasing in length, particularly in Ethiopia and northern Somalia/Somaliland region (Gebrechorkos et al., 2019), as is the number of consecutive dry days in some parts of eastern Africa, which may have an impact on drought length and increase the rapidity of onset and the intensity of drought (Trenberth et al., 2014). However, the overall impact on crops and food security during long-duration droughts on annual timescales is probably insensitive to 20 this.
It is possible that increasing temperatures have a negative impact on food security during droughts in ways that are beyond the scope of this study, e.g., decreased immunity of livestock, or increased water demand for cooling and water supply (Gebrechorkos et al., 2019, and references therein). In addition, in regions suffering from recent meteorological drought, non-meteorological factors such as increasing population and land-use changes also play a role in worsening the declining 25 vegetation conditions, even after precipitation returns to normal (Pricope et al., 2013).

Conclusions
In this first multi-model, multi-method attribution study using several drought estimates in eastern Africa, we address the recurring question on whether increasing global temperatures exacerbate drought. Previous attribution studies for the eastern Africa region have examined drought from a meteorological perspective (precipitation deficit) and have found no clear trends above 30 the noise of natural variability. In this study, we examined trends in eastern African drought from an agricultural perspective (soil moisture) as well as the meteorological perspective (precipitation, temperature and PET) for six regions in eastern Africa.
We also investigate whether global-warming driven trends in these meteorological variables can be seen to contribute to trends towards drier soils. In this section, we draw conclusions for each variable in turn and make recommendations.
Out of the four studied variables, soil moisture is most closely related to food security via crop health. In standardized soil moisture data, we found no discernible trends. The uncertainties in trends from model runs were found to be large and there are no long observational runs available. This emphasizes that the use of an ensemble of models is imperative. Due to the large 5 uncertainties in both soil moisture observations and simulations, we conclude that soil moisture cannot be relied upon on its own as a drought indicator and it is therefore important to examine other drought indicators as well. Besides, soil moisture also has a physical lower limit: once the soil is dry it will remain dry. In water limited regions an analysis of precipitation is thus a helpful addition.
Precipitation was found to have a stronger influence than temperature or PET on soil moisture variability, especially in the 10 drier study regions (the significant positive trend in temperature is not reflected by a decrease in soil moisture). However, the confidence intervals on precipitation trend estimations are large and no clear trend is evident.
As expected from the increase in global temperatures, we find significant positive trends in local temperatures in all six regions. The synthesized trend is between 1.0 and 1.3 times the trend in GMST, which corresponds to a local temperature rise of 1.1 to 1.4 degrees from pre-industrial times to 2018. However, the influence of this on annual soil moisture trends appears therefore appropriate to use in attribution studies in eastern Africa, supplementing the analysis of soil moisture. The outcome of previous studies that have focussed on precipitation deficits only (e.g., Philip et al., 2018a;Uhe et al., 2018) are thus still relevant and compare well with our results, that no consistent significant trends on droughts are found.
Finally, communication of the uncertainties in the analyses of soil moisture, precipitation, temperature and PET (and any drought indicators) to policy makers, the media and other stakeholders is crucial. Without insight into the uncertainties in 30 synthesized trends in the different drought indicators, conclusions become meaningless and results can easily be misinterpreted. Competing interests. We declare that there are no competing interests.
Acknowledgements. We would like to acknowledge funding for this work from the Children's Investment Fund Foundation (CIFF) grant 10 #1805-02753. For their roles in producing, coordinating and making available the ISI-MIP model output, we acknowledge the modelling groups and the ISI-MIP coordination team. We would also like to thank the volunteers running the weather@home models as well as the technical team in OeRC for their support. MERRA-2 data were developed by the Global Modeling and Assimilation Office (GMAO) at NASA Goddard Space Flight Center under funding from the NASA Modeling, Analysis, and Prediction (MAP) program; the data are disseminated through the Goddard Earth Science Data and Information Services Center (GES DISC), preliminary data have been made available thanks