Climate shocks, food price stability and international trade: evidence from 76 maize markets in 27 net-importing countries

The data used for estimation, described below, have a panel structure with 76 sub-national markets (denoted by k) in 27 countries (denoted by i) observed during the 2000–2015 marketing years, denoted by t. With these data, we estimate the following regression:

$$\begin{eqnarray}\begin{array}{rcl}{{CV}}_{{ik},t} & = & {\alpha }_{1}{I}_{i,t}+{\alpha }_{2}{Y}_{i,t}+{\alpha }_{3}{M}_{i,t}\\ & & +{\boldsymbol{\beta }}{{\boldsymbol{Z}}}_{i,t}+{\mu }_{i}+{\delta }_{k}+{\phi }_{t}+{\epsilon }_{{ik},t},\end{array}\end{eqnarray} \tag{ 1 }$$

where CV_ik,t is intra-annual CV of real monthly prices of maize in kth market of country i at year t.

Our main interests are the parameter estimates of α₁, α₂, and α₃. The data used to estimate these parameters are: the national annual ratios of net imports (i.e. imports minus exports) to domestic consumption, I_i,t; the annual absolute yield deviations from historical trends¹ in each country i, Y_i,t, which captures the effects of both positive and negative domestic supply shocks on domestic maize price instability; and the country-level stock-to-use ratio at the beginning of the marketing year t, M_i,t. The ${{\boldsymbol{Z}}}_{i,t}$ and ${\boldsymbol{\beta }}$ are vectors of control variables and their corresponding parameter estimates that have direct effects on domestic price stability, namely, exchange rate variability (Cho et al 2002), food aid (Barrett et al 1999, Lentz et al 2005) and social conflict (Bellemare 2015). In addition, as discussed below, the robustness of our results is explored by controlling for other variables that could be correlated with both price variability and import ratios, including physical trading distance, degree of import diversification, and per capita income (Luan et al 2013, Jeon and Ahn 2017).

We also exploit the panel nature of our data to control for unobservables. Specifically, the terms μ_i and δ_k denote country and market fixed effects, which control for time-invariant, unobserved country and market heterogeneity. The term ϕ_t is a year fixed effect that controls for the unobserved shocks affecting all the markets within a given marketing year. These account, for instance, for the excessive volatility that characterized the period 2005–2012 (Abbott 2012). The term _ik,t denotes residuals, assumed to be normally and independently distributed, as well as uncorrelated with the independent variables. Equation (1) is estimated using a fixed effects, panel data regression estimator—also known as Least Squares Dummy Variable—as described in Greene (2011, p.287).

Table 1 describes data sources and definitions of the regression variables in equation (1). Table 2 reports their descriptive statistics. The dependent variable is annual CV of real monthly prices of maize (in 1982–1984 US dollars) obtained from the FAO GIEWS database (FAO 2017). We focus on 76 retail or wholesale markets in 27 countries in the FAO GIEWS database that are net importers of maize. These countries are in Asia, Africa, and Latin America. The data are available from 2000–2015 (see appendix tables S1–S2, figure S2). Cumulative maize imports during 2011–2015 of these countries accounted for nearly one third of maize global imports. Figure 1 shows the 2000–2015 averages of intra-annual CV of real monthly maize prices, net import ratios, and stock-to-use ratios at the beginning of the marketing year for all the focus countries. At a glance, this map suggests that, generally speaking, countries with relatively low imports (e.g. those in Africa and Asia) tend to display more variable prices.

Zoom In Zoom Out Reset image size

A different perspective on these data is in figure 2, which plots the 2000–2015 average intra-annual CV of real monthly maize prices against import ratios in all the focus countries² . This figure shows that African and Asian countries have low import ratios (less than 0.1) and highly variable maize prices (Morocco is an exception). China, on the other hand, has a low import ratio and also the lowest levels of price variability among the focus countries, probably reflecting the effects of price stabilization policies (Pieters and Swinnen 2016, Chen et al 2018). Compared with the countries in Asia and Africa, the countries in Americas have higher import ratios and lower intra-annual CVs of their maize real prices.

Zoom In Zoom Out Reset image size

To predict the variability of future maize prices in each market, we proceed as follows. First, we obtain time series of maize yields (in tonnes per hectare) for the period 2006–2050 from two alternative sources: the yield-climate response function estimated by Moore et al (2017) and two of the global crop models from the GGCMI-AgMIP archive (Elliott et al 2014, Rosenzweig et al 2014)³ .

Moore et al (2017)'s function is a meta-analysis regression of a large database on crop yield changes compiled by Challinor et al (2014). This function relates changes in temperature, precipitation, CO₂ concentration, and on-farm management to changes in crop yields of several crops. The parameter estimates of this response function are global (see appendix section S1); therefore, variability in yield responses at any resolution other than global (e.g. gridcell or country) comes solely from regional variability in climate and the other explanatory variables. Moore et al (2017) used this function, jointly with data of future climate, to estimate high-resolution yields for the entire world. In an analogous manner, we use the their parameter estimates to predict country-level maize yields for each year in 2006–2050.

For this, we use data on temperature and precipitation from the ensemble of five global circulation models included in the GGCMI-AgMIP archive (Hempel et al 2013, Warszawski et al 2014), aggregated to the growing season of each focus country by Villoria et al (2018). Moreover, we obtain predictions with and without the effects of CO₂ fertilization, using the data on temperature to infer CO₂ concentrations as indicated in Moore et al (2017) (see appendix section S1 for details)⁴ . We also obtain projected yields with and without CO₂ from the GGCMI-AgMIP archive (Elliott et al 2014, Rosenzweig et al 2014), aggregated both to the country-level and to the growing season of each country using the GGCMI-AgMIP interface by Villoria et al (2016). To be consistent with the treatment of the maize yield shocks used to estimate equation (1), we calculate absolute projected maize yield deviations from their trends (see footnote 1). We denote these absolute deviations as ${Y}_{i,t}^{* }$, for t in 2006–2050.

Second, we combine these future yield deviations with the parameter estimates of equation (1); this procedure yields the corresponding predicted intra-annual CV of real monthly prices of maize in each focus market for each year in 2006–2050. In the analysis below, we will contrast our predicted scenarios of price variability against historical price variability. The most obvious measure of historical price variability in our study are the observed intra-annual CV of real monthly maize prices used to estimate equation (1). However, historical maize yield variability explains only some of the observed variance of past real maize prices while our predictions are solely driven by yield fluctuations. In order to circumvent this difficulty, we use absolute deviations from trend (see footnote 1) of observed maize yields during 1961–2014, denoted by ${Y}_{i,t}^{{\rm{FAO}}}$, to predict the historical intra-annual CV of real monthly prices of maize in each focus market. These are directly comparable with the future predictions.

Formally, we get the historical (${\widehat{{CV}}}_{{ik},t}^{{FAO}}$) and future (${\widehat{{CV}}}_{{ik},t}^{* }$) predictions using:

$$\begin{eqnarray}\begin{array}{rcl}{\widehat{{CV}}}_{{ik},t}^{{\rm{FAO}}} & = & {\widehat{\alpha }}_{1}{\bar{I}}_{i}+{\widehat{\alpha }}_{2}{Y}_{i,t}^{{\rm{FAO}}}+{\widehat{\alpha }}_{3}{\bar{M}}_{i}+\widehat{{\boldsymbol{\beta }}}{\bar{{\boldsymbol{Z}}}}_{i}\\ & & +{\widehat{\mu }}_{i}+{\widehat{\delta }}_{k}+{\widehat{\phi }}_{t=2015},\,t\in \{1961,...,2014\},\end{array}\end{eqnarray} \tag{ 2 }$$

and

$$\begin{eqnarray}\begin{array}{rcl}{\widehat{{CV}}}_{{ik},t}^{* } & = & {\widehat{\alpha }}_{1}{\bar{I}}_{i}+{\widehat{\alpha }}_{2}{Y}_{i,t}^{* }+{\widehat{\alpha }}_{3}{\bar{M}}_{i}+\widehat{{\boldsymbol{\beta }}}{\bar{{\boldsymbol{Z}}}}_{i}\\ & & +{\widehat{\mu }}_{i}+{\widehat{\delta }}_{k}+{\widehat{\phi }}_{t=2015},\,t\in \{2006,...,2050\},\end{array}\end{eqnarray} \tag{ 3 }$$

where hat indicates that these are the parameters estimated in equation (1), or their predicted values. The rest of the variables are set to their means over 2000–2015. For example, ${\bar{I}}_{i}$ is the average import ratio of country i during 2000–2015. In addition, ${\widehat{\mu }}_{i}$ and ${\widehat{\delta }}_{k}$ are intercept estimates specific to each country i and market k, and ${\widehat{\phi }}_{2015}$ is the estimated value of the fixed effect for year 2015. In other words, our predictions of both historical and projected CV of maize prices assume that the rest of the variables remain constant, and therefore, they exclusively capture the effect of historical and projected yield variability on price variability.

To aid with interpretation, figure 3 reports our regression results as elasticities evaluated at their sample means, with 95% confidence intervals generated by the Delta method (appendix section S3). The underlying parameter estimates are reported under column 'Model 4' in appendix table S3. These elasticities measure percentage changes in the intra-annual CV of monthly real maize prices given one percentage change in a given independent variable.

Zoom In Zoom Out Reset image size

Starting with the effects of yield shocks, our estimates suggest that a 1% increase in the absolute deviation of annual maize yields from their historical trend, is associated with a statistically significant increase of 0.09% in the intra-annual CV of real monthly maize prices. Also, a 1% increase in maize stocks relative to maize consumption at the beginning of the marketing year, is associated with a reduction of the intra-annual CV of real monthly maize prices of 0.22%. Regarding imports, an increase of 1% in the ratio of maize imports to total consumption is associated with a decrease in the intra-annual CV of real monthly maize prices of 0.29%. In short, our findings are aligned with the expectation that domestic maize supply shocks are a source of maize price fluctuations. They also confirm that larger import ratios and stock inventories at the beginning of the year are associated with lower price instability in the focus countries.

In appendix table S4, we show that these findings are robust to estimations using different time periods, to the omission of years with either high or low international prices, to alternative calculations of domestic supply shocks and to additional controls such as physical import distances, import diversification and per capita income. In appendix table S5, we show that these results are robust to estimation focused on different geographic regions, to the inclusion of countries with high or low import ratios, and to the inclusion of countries with large or short distances to their trading partners.

Figure 4 displays, for each focus country⁵ , the density distributions of market-average intra-annual CV of real monthly prices of maize in the historical period (1961–2014), ${\widehat{{CV}}}_{{ik},t}^{{\rm{FAO}}}$ from equation (2), and in the projected period until mid-century (2006–2050), ${\widehat{{CV}}}_{{ik},t}^{* }$ from equation (3). The price projections in this figure were obtained using projected yields assuming no effects of increased CO₂ fertilization under RCP 2.6 and 8.5 ⁶ . The most consequential finding in this figure is that the two sources of projected maize yields produce conflicting views of future scenarios of price instability. In particular, the GGCMI-AgMIP crop models predict that, in 17 out of 24 focus countries, maize prices will become more variable over the coming decades, which is evidenced by the more right-skewed density curves over the projection period 2006–2050. In contrast, the yield response function from Moore et al (2017) predicts that maize prices will become less variable in almost every focus country, relative to the historical densities (appendix figure S8 shows a case for Kenya).

Zoom In Zoom Out Reset image size

The conflicting views on prices are rooted in very different projections on future yield variability from the two modeling frameworks. Specifically, visual inspection of figure 5 reveals that the GGCMI-AgMIP crop models predict increases in maize yield variability in most countries. In contrast, the meta-analysis yield response function from Moore et al (2017) predicts density functions that are less spread than the density of historical yields.

Zoom In Zoom Out Reset image size

We surmise that this reduction in maize yield variability owes to the fact that nearly 70% of the data points in the underlying database used by Moore et al (2017), are for the US and China (supplementary table 1 in Challinor et al 2014). Countries in tropical regions, which are more sensitive to warming (Porter et al 2014, Challinor et al 2014), are much less represented. As noted by Moore et al (2017, p.3) 'our meta-analysis results are more optimistic in tropical areas' than the GGCMI-AgMIP crop models—in fact, their projected yield losses are 5 to 50% lower than those projected by the GGCMI-AgMIP crop models within the tropical regions (see supplementary figure 6 in Moore et al (2017)). Such a reduction in yield variability is at odds with the expectation that future yields will become more variable because of more variable temperature and precipitation⁷  (Chen et al 2004, McCarl et al 2008, Porter et al 2014, Müller and Robertson 2014, Haile et al 2017). In light of this discussion, we will use the prices projected using GGCMI-AgMIP yields in the discussion that follows. We notice, however, that these data have some well known caveats. For instance, the GGCMI-AgMIP crop models use simplistic assumptions about the distribution of soils, sowing dates, varieties and fertilizers, which can affect the temporal dynamics in the yield projections (Müller et al 2017). Nor do these models capture weather-related pest or disease outbreaks that could constitute short-term yield shocks (Rosenzweig et al 2014, Müller et al 2017).