The El Niño impact on maize yields is amplified in lower income teleconnected countries

We use a multiple-regime panel smooth transition regression to examine the economic and climatic sources of the nonuniform relationship between El Niño Southern Oscillation (ENSO) and maize yields around the globe. While the yield effect is predominantly observed in strongly teleconnected countries, it is amplified in lower income countries, which we attribute to possible lack of resilience to ENSO—induced weather shocks. Both El Niño-like and La Niña-like conditions result in maize yield reduction, but it is during El Niño events when maize yields drop by up to 20% in most affected countries. Because in many of these countries maize is an important agricultural crop, the presented results are of interest to researchers and policy makers in the areas of world nutrition and international aid. Moreover, because larger share of maize is produced by high income weakly teleconnected countries, the observed geographic heterogeneity of the El Niño impact offers possible benefits from global risk sharing. These findings also offer insights to climate change economics, as possible increased frequency of the ENSO cycle may negatively impact maize production in strongly teleconnected low income countries.


Introduction
The impact of the ElNiño Southern Oscillation (ENSO) events on global weather and crop yields can be nonuniform (e.g. Iizumi et al 2014, Hsiang andMeng 2015, Anderson et al 2017). The reasons for this are heterogeneity in the ENSO-weather linkages, as well as heterogeneity in countries' responses to weather shocks (e.g. Cashin et al 2017). Weather anomalies-e.g. excessive heat, droughts, floods-are more frequent in the tropics, where the majority of low income countries are located (Masters andMcMillan 2001, Hsiang 2010). Moreover, the agricultural sector, which is inherently linked with weather, constitutes a larger share of these economies (Schlenker and Lobell 2010). Finally, the climate resilience of these countries is not at par with that of high income countries (Dell et al 2012, Hertel andLobell 2014). This study aims to disentangle and examine climatic and economic sources of the aforementioned heterogeneity of the ENSO impact on maize yields.
A typical ENSO event forms during the boreal summer, and reaches its peak at the end of that year. It usually vanishes during the subsequent spring, although an ENSO event may span a two-year period as well. Immediate impacts of ENSO are felt in countries adjacent to the Pacific Ocean, where this event occurs. For example, its warm phase-ElNiño 3results in increased rainfall across the western tier of the American continent and droughts in the Asia-Pacific region. Its cool phase-LaNiña-manifests in weather conditions opposite those described above. ENSO events also also affect other regions-including the US (e.g. Barlow et al 2001) and Europe (e.g. Pozo-Vázquez et al 2005, Bulić andKucharski 2012, Ionita et al 2012)-via the teleconnections, which link the realized ENSO conditions in the equatorial Pacific to temporal changes in weather conditions around the globe (Ropelewski and Halpert 1987, Rasmusson 1991, Stone et al 1996, Richter et al 2015.
We focus on maize because of the direct effect of extreme weather on field crops, as well as the importance of maize as food and feed, and more recently as a source of renewable energy. Our interest in the ENSO phenomenon is twofold. First, we present the year-toyear variability of maize yields due to the ENSO cycle within the current climate. This allows us to envisage transitory effects of the ENSO events. Second, as extreme ENSO events (i.e. El Niño-like and La Niñalike conditions) may become more frequent (Cai et al 2014(Cai et al , 2015, we offer a snap-shot of the possible yield impact of climate change.
Empirical investigation of the relationship between ENSO and agricultural production can be traced back to the mid-1980s (e.g. Handler 1984, Nicholls 1985, but the turning point in the economic assessment of this climate phenomenon was the 1997/ 1998 ElNiño event, which caused several billion US dollars worth of damage in the US alone (Adams et al 1999). Since the late 1990s, a large body of literature has addressed the ENSO impact on cereal production in different regions of the world, including the Americas, Australasia, and Africa (Phillips et al 1998, Naylor et al 2001, Amissah-Arthur et al 2002, Adams et al 2003, Stige et al 2006, Roberts et al 2009, Anderson et al 2017. A considerable portion of the existing studies rely on simulated, rather than observed, yield data (e.g. Phillips et al 1998, Adams et al 2003. While simulated observations offer some benefits, such as availability over a longer time span and for smaller geographic areas, their main drawback is that such data largely ignore any adaptive actions that crop producers may have taken in response or in anticipation of climatic shocks (Miao et al 2016). A relatively smaller group of studies, which rely on observational data, focus on a subset of countries, often just a single country (e.g. Cane et al 1994, Podestá et al 1999, Naylor et al 2001, Selvaraju 2003, Anderson et al 2017. Moreover, most previous studies consider ENSO as a categorical variable and use basic statistical methods to compare yield distributions during the event years with those during neutral years (e.g. Legler et al 1999, Podestá et al 1999, Selvaraju 2003, Iizumi et al 2014. As in the foregoing instance, studies that rely on a regressionbased framework, or a continuous proxy variable for ENSO, are scarce and often focus on small group of countries or regions within a country (e.g. Tack and Ubilava 2013, Anderson et al 2017). Finally, while recent research has addressed the climatic heterogeneity of the ENSO effect (Hsiang and Meng 2015), none of the existing studies assess the heterogeneous impacts of ENSO conditional on the economic development of the affected countries, which is the focus of this research.
We contribute to the literature in several directions. First, we present a global regression-based analysis using observational data on sea surface temperature anomalies in the Niño3.4 region (SST), a continuous series that proxies ENSO, and countrylevel maize yields. Second, we allow for asymmetries in yield responses to SST, and geographical heterogeneity of these effects. Third, in examining the heterogeneity, we allow a continuum of the effects ranging across distinct regimes, based on climatic and, importantly, economic characteristics of countries in consideration. We do so by adopting a panel smooth transition regression (PSTR) framework of González et al (2017) -a modeling approach that nests commonly applied linear and discrete specifications as its special cases and offers a parsimonious framework for examining potentially complex relationships between the variables of interest.
We find that SST anomalies, indeed, considerably affect maize yields in the lower income, strongly teleconnected countries. This finding is consistent with Hsiang and Meng (2015), who examined cereal grain yield responses to SST anomalies. In addition, we also find that the effect is particularly strong, both statistically and economically, during positive SST deviations, i.e. in ElNiño-like conditions, and is negligible during negative SST deviations, i.e. in LaNiña-like conditions. This finding is consistent with Cashin et al (2017), who employ a dynamic multi-country framework to find the asymmetric international macroeconomic transmission of ElNiño and LaNiña related weather shocks. Thus, not only the ENSO cycle has geographically heterogeneous effects on maize yields, but these effects are also asymmetric.
Several implications follow from these findings. First, because in many affected low-income countries maize is one of the more important crops, the local adverse effects of ENSO shocks, particularly as they relate to household incomes, can be substantial. Second, because a larger share of maize is produced in the high income countries that are hardly impacted by ENSO shocks, their global effect on the maize yield can be seen as economically negligible, which implies little pressure on the world prices of this commodity. This is consistent with Ubilava (2018), who examined commodity price responses to ENSO shocks. Third, the asymmetry of maize yield responses to positive and negative sea surface temperature deviations-a finding that echoes existing studies (e.g. Iizumi et al 2014, Cashin et al 2017)-indicates that, for example, a negative impact of an ElNiño event in the teleconnected countries is not subsequently balanced out by a positive impact of a LaNiña event. It then follows that in the short term, proactive global risk sharinge.g. international aid to low income countries-is particularly crucial during ElNiño years. In the long term, in the absence of adaptation, if ENSO cycles were to become more frequent, we can expect an increase in the incidence of bad crops and reduction of the unconditional mean of maize yields, particularly in the lower income countries that are directly affected by ENSO events.

Data description
Data on country-specific maize yields, area harvested, production, and area allocated to agricultural land, were sourced from the online platform of FAOSTAT. The series span the 1961-2017 period. After discarding problematic countries (i.e. those with missing observations or those displaying constant yields for more than three consecutive years), we retained 67 countries representing 91% (93%) of global maize production (area harvested). Figure 1 illustrates the geographical distribution of the expected yields 4 in the sample of countries.
We obtained data on country-specific growing seasons from Villoria and Delgado (2017), who use the global information on the pixel-level planting and harvesting dates of Sacks et al (2010) to construct monthly growing season weights for each country. Thus, a weight of zero, in a given month, implies that the month is not part of the growing season; a positive weight denotes the share of pixels in a country for which the month is part of the growing season. We define the growing season as an uninterrupted sequence of months for which the weights exceed 0.5.
We sourced the sea surface temperature (SST) series from the online portal of National Oceanic and Atmospheric Administration (NOAA). The SST anomalies are measured as deviations ( • C) from the 1980-2010 base period in the Niño3.4 region, a rectangular area bounded by 120 • W-170 • W and 5 • S-5 • N, and are reported at monthly frequency. To obtain a measure of the relevant SST anomalies for country i in period t, we averaged these observations over the maize growing season of that country.
Following Hsiang et al (2011), we use ENSO teleconnections as a potential source of geographical heterogeneity in the effect of the SST anomalies on maize yields. However, the approach that we use here is slightly different from that of the aforementioned study. For each country i, first we calculate correlations between the growing season average temperature and growing season average SST anomaly, denoted by Another potential source of geographical heterogeneity in the ENSO-yield relationship is the level of economic development of a country. The reasoning here is similar to that put forward by Dell et al (2012), albeit in the context of ENSO-related year-to-year weather variation. Thus, for each country, we first calculate the average GDP per capita (in terms of 2010 US dollars) over the 2001-2015 period. We then obtain the measure of economic development, denoted by λ i , as the natural logarithm of this average. That is, G ln  Figure 2 offers insights on the associations among key variable of interest: high maize yields in higher income temperate countries that tend to be weakly teleconnected with ENSO; low maize yields in lower income tropical countries, which may be strongly or weakly teleconnected with ENSO. Put differently, while there is a strong positive correlation between country income levels and their latitudinal distance from the equator, there appears to be negative, albeit weak, relationship between the measures of economic development and teleconnection.

The econometric model
We estimate the relationship between the annualized measure of the SST anomaly, s it , and the natural log of a crop yield in country i at time t y , it , in a fixed effects setting. A basic representation of this model is given by: where ε it is an error term with zero mean and constant variance; β represents a semi-elasticity measure of yield with respect to a 1 • C change in s it ; μ i is a fixed effect that controls for country-specific time invariant factors; δ i (t) denotes country-specific trend, which here is modeled as a quadratic trend.
The foregoing specification assumes a linear, i.e. symmetric, relationship between the SST anomaly and maize yields. But empirical evidence indicates that this relationship is likely asymmetric (e.g. Iizumi et al 2014, Anderson et al 2017). To allow for asymmetries, we interact s it with a Heaviside indicator function, ), that takes the value one when s it 0, and zero otherwise. In addition to asymmetries, it is plausible that yield responses to ENSO shocks vary across countries. We consider two sources of such geographical heterogeneity. One, motivated by Hsiang and Meng (2015), is based on the association of a given countryʼs weather with the SST anomalies in the Niño3.4 region. That is, ENSO shocks are likely to have greater impact on strongly teleconnected rather than weakly teleconnected countries. The other, motivated by Dell et al (2012), is based on the economic development of a country. That is, not only high income countries are less prone to ENSO-related weather shocks, but they are also more resilient to (i.e. better equipped to deal with) these shocks, compared to lower income countries. Unlike the aforementioned studies, we do not assign countries to the different groups, however. Instead, we introduce nonlinear (smooth transition) functions of the economic development and the teleconnection measures-i.e. λ i and τ i -given by g c ; , ), respectively, and as such, allow for a continuum of effects across different possible regimes. Finally, to account for possible lag dependence in yields, we add lagged dependent variable to the regression. This results in the following heterogeneous panel specification: , and j , 0, 1, 2 j b = , is the associated vector of the parameters, and where the heterogeneity stems from the interaction of location-specific transition functions with the regimespecific effects. These transition functions are designed to be bounded by zero and one, wherein γ and c are smoothness and centrality parameters that govern the shape of these functions 5 . A notable distinction of the present study from the previous studies is that the threshold (i.e. the inflection point) that divides countries into poor versus rich countries (as in Dell et al 2012), or teleconnected versus weakly affected countries (as in Hsiang and Meng 2015), is not predetermined, but rather it is estimated in the nonlinear regression setting. As a final note, SST anomalies are mediated to the crop yield variability primarily via the weather variables (i.e. temperature and precipitation), but possibly also via other channels (e.g. extreme climatic events not captured by the conventional weather variables, expected prices that affect crop producers' decision making, etc). Therefore, the estimated parameters here capture all channels through which SST anomalies manifest in crop yield variability (refer to Technical appendix for further discussion on this matter). Table 1 presents the estimated effects of SST anomalies associated with each of the four regimes, while the estimated transition functions are illustrated in figure 3.

Results and discussion
Several key findings emerge here. First, the relationship between SST anomalies and maize yields is most evident in the strongly teleconnected countries, but is largely absent in the weakly teleconnected countries. A positive 1°C SST deviation may result in 11%-21% yield reduction in the strongly teleconnected countries. This finding is expected and, in effect, it validates the weather link between SST anomalies and maize yields.
Second, the relationship between SST anomalies and maize yields is amplified in lower income countries. The difference in yield reduction, due to a positive 1°C SST deviation, is 10% between the lowest income and highest income countries in the sample. Because we account for the teleconnection in the regression setting, this finding suggests that there may be additional ENSO-related factors (e.g. extreme weather events not measured by temperature and precipitation that we applied to obtain the teleconnection measure) that result in yield reduction in lower income countries, and that these countries are less resilient (or unable to effectively adapt in the short term) to ElNiño-like conditions, compared to high income countries.
Third, while it is a positive deviation in SST (i.e. the El Niño-like conditions), rather than a negative deviation in SST (i.e. the La Niña-like conditions), that strongly impacts yields, both these events are potentially damaging in the teleconnected countries. Notably, the effect of a negative SST deviation is qualitatively and quantitatively similar in countries from both income groups. In the teleconnected countries, a negative 1°C SST deviation results in 5% (statistically insignificant) yield reduction. That yield reduction during ElNiño conditions is not balanced out by yield increase during LaNiña conditions is an important finding, as it implies that on average ENSO cycles are yield-reducing in the affected countries.
Fourth, the transition between the low income and high income regimes is smooth. That is, majority of the countries in consideration fall between the estimated extreme regimes. On the other hand, a switch between weakly teleconnected and strongly teleconnected countries happens almost instantaneously, at approximately 0.4. The effect of these results are illustrated in figures 4and 5. A positive 1°C deviation in SST results in 10%-15% yield reduction in the affected higher income countries, and as much as 20% yield reduction in the affected low income countries that are predominantly located in southeastern tier of Sub-Saharan Africa, as well as Central America, South Asia, and Australia. A complete list of countries is presented in the appendix table D1.
Several implications follow from these results. First, on a local scale, we find clear indication that ElNiño-like conditions are most detrimental to maize yields in low income countries. In these countries maize is an important element of the agricultural sector (as measured by its share of area harvested relative to the total agricultural land). Thus, adverse effects of ElNiño shocks on household incomes are likely to be substantial. Indeed, majority of recent famines in Sub Saharan Africa have been in some way associated with ENSO events.
Second, because maize is mostly supplied by the countries that are hardly affected by ENSO shocks (see Table 1. Estimated effects of the SST anomalies. 5 Equation (2) represents the modeling framework, put forward by González et al (2017), that can be seen as an extension of univariate smooth transition regressions in the time series setting Luukkonen et al (1988), Eitrheim and Terasvirta (1996) or as a generalization of the panel threshold model of Hansen (1999) with potentially gradual, rather than instantaneous, switches between two (or more) regimes. See Technical appendix for more details on the transition functions, as well as the model selection, estimation, and evaluation specificities.
appendix figure D1), the global effect on maize yields is small, suggesting little pressure on the prices. This finding is consistent with Ubilava (2018), who found little evidence for cereal grain price responses to ENSO shocks. Thus, because a considerable share of major maize exporters are largely unaffected by ENSO shocks, there is an opportunity for global risk sharing, be that via trade or international aid, from higher income/weakly teleconnected countries to lower income/strongly teleconnected countries.
, which is related to the degree of a countryʼs teleconnection with the SST anomalies in the Niño3.4 region (measured as the geometric average of the correlation between SST anomalies and temperature, and SST anomalies and precipitation over the 1961-2014 period). Third, the asymmetry of maize yield responses to positive and negative SST deviations suggests that, for example, a negative yield impact of an ElNiño event in the strongly teleconnected countries is not subsequently balanced out by the positive yield impact of a LaNiña event. To the extent that extreme ENSO events are anticipated to become more frequent due to climate change (Cai et al 2014(Cai et al , 2015, this finding suggests that, in the absence of adaptation, ENSO cycles can become more damaging for the strongly teleconnected countries in the future. To that end, the aforementioned international involvement in mitigating the negative impacts of ENSO shocks is to become more crucial with time.
This study is not without caveats and limitations. These are primarily related to the 'data issues.' In the presented analysis, we rely on the national yield data. While these data give us sufficient information to conduct an econometric analysis and answer the key question of interest-does economic development of a country play a role in the climate-production nexus? -more detailed yield data could possibly unveil additional intricacies of this relationship, which may be camouflaged in the national aggregates. We retain this as one of the interesting and important directions for future research.

Conclusion
We find that the effect of ENSO events, as measured by SST anomalies, is most pronounced in lower income strongly teleconnected countries. We also find that the ENSO effect is asymmetric: it is evident during ElNiño-like conditions, and virtually nonexistent during LaNiña-like conditions. These findings point to potential benefits of risk sharing-over time and across space-particularly as it relates to food programs directed to low income countries. In absence of adaptation, the issue is likely to become more crucial with climate change, due to a possibility of more frequent ENSO cycles.

Acknowledgments
We would like to thank Weston Anderson, Jesse Tack, and Nelson Villoria for their valuable suggestions on the earlier version of the manuscript. We are also grateful to the editor and the anonymous reviewers for their helpful comments on the manuscript.

Appendix A. Smooth transition functions and estimated effects
where i p is a transition variable and s p is its standard deviation, the role of which is to make the smoothness parameter, γ, unit-free. In practice, it is sufficient to consider the m=1 and m=2 cases of this transition function (González et al 2017). When m=1, we have a single centrality parameter, c, and for relatively low ) is a monotonically increasing function, which takes the value 0.5 when c i p = . It converges to a constant when 0 g  , and it becomes a Heaviside indicator function of c i pthat mimics a discrete jump across regimes as g c , ; , ). When m=2, we have two centrality parameters, c 1 and c 2 , and for relatively low values of c g , ; , i g p g ( )is an inverse bell-shaped function, which reaches the minimum at c c 2 1 2 + ( ) and attains its maximum at both the low and high values of s it . As in the previous instance, c g ; , i p g ( ) converges to a constant when 0 g  , while as g  ¥, it becomes a Heaviside indicator function with three regimes, of which the outer regimes are identical to each other and different from the 'inner' regime.
The foregoing modeling setup offers a number of possible effects to be estimated. More specifically, at the extreme, four possible effects can be identified as illustrated below. In addition, a wide range of effects can also be available, for instance, when c g 0 ; , 1

Appendix B. Linearity (homogeneity) testing
Because PSTR is nonlinear in parameters, we cannot directly test restrictions in the model. This issue, better known as the Davies' problem (Davies 1977, 1987, is due to the unidentified nuisance parameters. Put simply, the PSTR is not identified if the true model is homogeneous (González et al 2017). As a workaround, we can test the homogeneity hypothesis in an auxiliary regression framework, which involves Taylor series expansion, initially put forward for nonlinear time series models (Luukkonen et al 1988, Terasvirta andAnderson 1992) and subsequently adopted for the PSTR (González et al 2017).
To obtain the auxiliary regression, we set up a linear fixed effects model, and interact the independent variables with polynomials of the candidate transition variables (i.e. i l and i t ). This results in a auxiliary regression of the following form: . If we fail to reject both of these hypothesis, the homogeneous model is then considered to be consistent with the underlying data. Otherwise, if the rejection of H 0 ¢ is the strongest, we set m=1, and if the rejection of H 0  is the strongest, we set m=2 in the transition function.
The results of the homogeneity test are presented in panel(a) of table B1. As shown, homogeneity is strongly rejected when either i l or i t are used as transition variables. In both instances, the heteroskedasticity and autocorrelation consistent variants of the test favor m=1 in the transition function. Therefore, we sequentially picked each of the transition variables and set m=1 to estimate the two-regime PSTR model. That is, we first estimated a PSTR using either i l or i t as a transition variable. We then tested the esti-  the specification given by equation (2) to generate the main results of this study.

Appendix C. Remarks on the use of mediators and controls
In the applied regression, the parameters are total effects of SST anomalies maize yield variability. The potential channels through which SST impacts yields include the key weather variables, i.e. temperature and precipitation, but may also include additional links through which ENSO can influence crop yield variability. These could be extreme climatic events not captured by the conventional weather variables, as well as non-weather variables, e.g. expected prices, that affect crop producers' decision making. To the extent that these variables are influenced by ENSO, they are the outcome of the natural experiment, and represent mediators in the ENSO-yield relationship. That is, they are post-treatment variables, whereby treatment is an ENSO event. By not including them in the regression setting, we are not inducing the omitted variable bias (Hsiang andBurke 2013, Smith andUbilava 2017). Quite the contrary, if we were to include these variables in the regression setting, they would serve as 'bad controls' and would introduce the so-called 'post-treatment bias' to the estimated coefficients (Angrist and Pischke 2008, Acharya et al 2016, Montgomery et al 2018).
Notably, in the regression setting, we could 'safely' use other controls, i.e. variables that are uncorrelated with SST anomalies but are known to affect yields. By doing so, however, we would merely improve the efficiency of the estimated coefficient. To that end, we can think of the currently reported standard errors as their upper bound. In summary, because the goal here is to estimate the total effect of SST anomalies on maize yields, there would be little benefit (i.e. efficiency gains) and possibly more harm (i.e. post-treatment bias) if we were to control for additional factors in the model. On the other hand, because there are no known phenomena or events that simultaneously cause ENSO cycle and crop yields (as well as factors affecting crop yields), the currently specified model does not suffer from the omitted variable bias.
Appendix D. Heterogeneity of El Niño impacts across countries  Note: The table entries are in percentage terms The columns headed by Export include the share of maize exports relative to the world exports (measured as the average of the shares over the 2001-2015 period); the columns headed by Area include the share of maize area harvested relative to the area of agricultural land in a given country (measured as the share of averages over the 2001-2015 period). The values underneath the Export and the Area headings indicate the sum and the average of the entries in these columns. The data on exports are obtained from USDA/FAS Production, Supply and Distribution online database. The other data sources are presented in the Data Description section.