Rice, Carbon Dioxide, Climate Change, And Feeding The Future

This study investigates the relationship between rice yields, climate change, and carbon dioxide 6 (CO 2 ). We integrate gridded climate data in the growing seasons and Asian rice yield data reported 7 by the Food and Agriculture Organization with free air carbon dioxide enrichment (FACE) 8 experimental data. Using those data, we estimate prediction models of rice yields that evolve over 9 time and decompose effects of climate, CO 2 , and technological progress. The results show that 10 atmospheric CO 2 has significantly increased rice yields, with the contribution accounting for 29% 11 to 33% of the observed yield growth. The results also reveal that increases in temperature decrease 12 rice yields in parts of Asia, implying that both CO 2 mitigation and climate change are yield growth 13 depressing factors. The finding suggests a potential need for more agricultural research and 14 development investment to offset CO 2 mitigation and climate change effects. are the third stage of our estimation procedure. The significant Breusch-Pagan test at least at 1% significance level for the first stage results suggests the existence of heteroscedasticity, providing a justification of using the 3-step FGLS approach.


Introduction 19
Globally rice is the most widely consumed food crop (Food andAgriculture Organization 20 2014, 2016), and population growth is increasing rice demand. Without substantial rice 21 productivity increases, billions may struggle from malnutrition and food insecurity. Declines in 22 rice productivity also can cause conflict and societal instability (Brück and d'Errico 2019). 23 However, rice productivity is facing significant challenges. Agricultural research and development 24 (AGRD) funding is diminishing and limits the potential growth of rice yields (Alston and Pardey 25 2006;Andersen et al. 2018;Pardey et al. 2016). Some studies show that future rice productivity is 26 threatened by climate change (Ortiz-Bobea et al. 2021;Sinnarong et al. 2019), especially in Asia. 27 A large portion of Asian rice production is in mega deltas, which are vulnerable to sea-level rise 28 and storm surge (Wassmann et al. 2009). The above facts and findings highlight the importance of 29 investigating climate change impacts on rice yields and how rice production might adapt to better 30 feed the future. 31 There are several factors influencing rice and more generally crop yields. Climate 32 conditions have been found to be key determinants of average crop yields and their variability 33 (IPCC 2014;Jithitikulchai, McCarl, and Wu 2018;Pathak et al. 2018;Raza et al. 2019;Wang et 34 al. 2018). For rice, studies in Thailand and Bangladesh have found that temperature increases 35 reduce rice yields and increase yield variability (Sinnarong et al. 2019, Sarker, Alam, andGow 36 2014). Increasing temperature has been found as a cause of rice yield losses in Laos and India 37 (Ishimaru et al. 2016). Zhang, Tao, and Zhang (2016) show that rice yields are generally more 38 sensitive to high temperatures than cold temperatures in China. Collectively these findings suggest 39 the importance of developing integrated adaptation strategies to alleviate potential climate change 40 impacts. Simultaneously, an increase in atmospheric CO2 has been found to enhance rice yields 41 (Hasegawa et al. 2013;Kirschbaum 2011). Technical advances as stimulated by research 42 investments are another crucial productivity determinant and can offset possible negative impacts 43 of climate change (Chen, McCarl, and Chang 2012;Villavicencio et al. 2013). 44 Nevertheless, due to the high correlation between time and CO2 atmospheric concentration, 45 it is statistically challenging to separately identify CO2 and technological impacts on rice yields. 46 The conventional way of modeling yield growth due to technological advances is to use time as a 47 technical progress proxy (Attavanich and McCarl 2014;McCarl, Villavicencio, and Wu 2008). 48 However, given the steady advance in CO2, using this approach captures not only technological 49 advance but also the influence of highly correlated increasing CO2 (Attavanich and McCarl 2014). 50 As such, one can easily overestimate technological effects on yield growth due to correlated CO2-51 induced yield enhancement. Thus, it is important to partition out the CO2 effect. But to do this, one 52 must overcome a collinearity issue between time and atmospheric CO2 concentration increases 53 since the correlation of CO2 concentration with time is above 95% (see Section 3.1). Attavanich 54 and McCarl (2014) addressed this identification challenge for U.S. corn, wheat, sorghum, and 55 cotton by combining the USDA-reported data on yields with free air carbon dioxide enrichment 56 (FACE) experimental data. They found that estimating time as a proxy for yield growth rate 57 without considering CO2 effects upwardly biased their proxy technological progress rate estimates 58 by as much as 40%. 59 Herein we address the effects of climate, technological advances, and CO2 fertilization on 60 Asian rice yield growth. In addition, we forecast the impacts of projected climate change and CO2 61 concentration on yield growth under different climate change scenarios. Relevant implications of 62 our findings are discussed at the end of the paper. 63

Estimation equation 64
To estimate the effect of climate, atmospheric CO2 concentration, and other factors on rice 65 yield, we conducted an econometric estimation using merged observed and FACE experiment 66 yield data, climate conditions, atmospheric CO2 concentration, input use data for fertilization and 67 irrigation amounts, a time trend proxying technological advance, a country fixed effect 68 individually, and its interaction term with climate variables. To do this, a rice yield production 69 function is assumed following Just andPope (1979) andChen, McCarl, andSchimmelpfennig 70 (2004). In particular, the rice yield functional form is expressed as: 71 We also introduced interaction terms between the country dummies and the climate variables, so 100 we could examine differential climate effects across countries. Finally, note that since our data 101 incorporates observed and FACE data plus some country data are missing, it does not fully have a 102 balanced panel data structure. 103

Estimation method 104
In the estimation, we used a three-step feasible generalized least squares (FGLS) method (as 105 discussed in Just and Pope 1979;and Attavanich and McCarl 2014). This approach is applied to 106 correct for heteroscedasticity and estimate the effects of climate and other explanatory variables 107 on rice yield variance. The empirical estimation procedure is as follows: 108 • First, we estimated ( ) = ( , ) + by using pooled ordinary least squares 109 (OLS) and obtained the residuals (̂). 110 • Second, we took the first step residuals and then estimated an equation where their square was 111 a function of . Namely, we estimated (̂2 ) = ℎ( , ) + and obtained fitted 112 values (̂2 ) and calculated √ ( (̂2 ) ). 113 • Third, we estimated ( ) = ( , ) + using weighted least squares (WLS) with 114 √ ( (̂2 ) ) as weights. 115 In turn, the values of coefficient vector ( ) from the third stage are those we report below. 116

Data description 117
A list of data sources that were used in this study is presented in Table S1. Additional notes on 118 the data are: 119 • The observed rice yield data and irrigation rate data were assembled for 31 Asian countries for 120 the time period from 1961 to 2015 and were drawn from the Food and Agriculture Organization 121 FAOSTAT database (2020). However, the observations are not complete for all countries in 122 all years. For instance, Central Asian countries are missing observations before 1992, and Syria 123 does not have observations after 1996. Overview of rice yields in each country over time is 124 presented in Figure S1. 125 • Growing season gridded climate data and fertilizer application data for each country were from 126 Jägermeyr et al. (2021). The gridded data were further aggregated to the country level by 127 calculating the weighted average based on each country's rice land acreage. 128 • The FACE experimental rice yield data arose from studies in Japan and China (Hasegawa et 129 al. 2013;Jing et al. 2016;Sun et al. 2014) and are presented in Table S2. 130 • The global CO2 concentration data came from the NOAA Global Monitoring Laboratory 131 (2020). Drought-based data were incorporated using calculation procedures for the SPEI 132 Global Drought measure from Vicente-Serrano et al. (2010). 133 • The climate projections data came from the KNMI Climate Change Atlas (2020), as we discuss 134 later. 135 Summary statistics on the above data are provided in Table S3. 136 5 Results 137

Relationship between rice yields, global CO2 concentration, and climate 138
The relationships between rice yields and the key variables are presented in Figure 1. 139 Specifically, Figure 1a shows that global CO2 concentration is positively correlated with regional 140 Asian rice yields. This also indicates identifying CO2 fertilization effects independently of time 141

Estimation of independent variable impacts on rice yields 159
To avoid inappropriate estimates caused by non-stationary data, we conducted pre-modeling 160 unit root tests to examine data stationarity. The test method followed Levin-Lin-Chu ( Levin et al. 161 2002) and Im-Pesaran-Shin unit root tests (Im et al. 2003). The test specifications used the 162 inclusion of individual intercepts and the inclusion of both individual intercepts and trends. Table  163 S4 presents the test statistics and indicates that the null hypothesis of non-stationarity under 164 different test specifications and methods is rejected and significant at all the usual testing levels, 165 suggesting that the data doesn't need to be differenced in order to make it stationary. 166 The estimation results with different model specifications and the final model are presented 167 in Tables S5 and S6, respectively. Overall, the results show that rice yields are significantly 168 increasing with time, reflecting the effect of technological advances. Also, note that excluding the 169 CO2 atmospheric concentration variable causes the linear time coefficients to be much larger, 170 showing that omitting the CO2 fertilization effect upwardly biases the estimates for time as a proxy 171 for technology effects by 56% at the mean. Similarly, the linear time coefficients are larger when 172 excluding fertilizer application amounts and irrigation land coverage variables. In addition, we 173 found significant and positive CO2 concentration effects, indicating a rice yield response to a CO2 174 enriched atmosphere which is the well-known C3 CO2 crop growth stimulation effect (see the 175 review in Korres et al., 2016). Overall, fertilizer application, irrigation rate, CO2 concentration, 176 and the dummy for FACE observations all significantly increase rice yields. Due to the 177 heterogeneous climate effect across different Asian regions that were present in Figure 1, the 178 overall coefficients on precipitation, average maximum temperature, and SPEI are not significant, 179 but a number of their interaction effects with the country-dummies are significant. Detailed 180 estimation results appear in Table S6. 181

Projected effects of climate change 182
To better illustrate the impact of CO2 concentration and temperature on rice yields, we used 183 Model 2 of Table S6 to

The CO2 fertilization effect on the historical rice yield growth 198
To further investigate the CO2 fertilization effect, we used Model 2 in Table S6 to conduct in-199 sample rice yield growth projections. The estimation, including and excluding the CO2 fertilization 200 effect, reveals the role CO2 concentration plays in the rice yield projection model. Specifically, we 201 projected yields using our best-fitted equation under two cases. The first case holds CO2 202 concentration constant at the 1990-level, and the other case uses the evolving observed 203 atmospheric CO2 concentration levels. Figure 3 presents the projected rice yields under these 204 cases. The result shows that regional 1961 to 2015 rice yield growth varies from 47% to 81% when 205 CO2 concentration is held at 1990-levels; in contrast, that increases to 79% to 123% under the 206 evolving observed CO2 concentrations. This finding implies that the CO2 fertilization effect has 207 been responsible for 32% to 42% of the regional rice yield growth over that period. 208  Table S6 and show that rice yield 213 growth is much smaller without the CO2 influence. In particular, depending on region, yield growth ranges 214 from 47% to 81% when CO2 is held constant and is much higher with actual CO2 (79% to 123%, i.e., a

Forecasting climate change impact on the future rice yield growth 217
We also constructed a forecast of rice yields for the years 2016 to 2050 under selected climate 218 change scenarios again using Model 2. We did this for 25 countries, dropping those with small 219 rice production or limited observations (refer to Figure S1). The required covariate values were 220 retrieved from several sources. We used CMIP5  The projected Asian rice yields under different climate scenarios are shown in Figure 4. 232 The projection indicates that rice yields are expected to grow slowly, even fall in the last two 233 decades of this century under the RCP 4.5 scenario as mitigation proceeds. In contrast, the rice 234 yields are expected to have a significant increase under the RCP 8.5 scenario due to a higher 235 atmospheric CO2 concentration. The projected Asian rice yields under the RCP 6 scenario are 236 between the RCP 4.5 and RCP 8.5 scenarios. 237 The result in Figure 4 can be further broken down into the regional level (Table 1) and 246 country-level (Table S6) from 2015 to 2050 and 2100 by RCP scenarios. When we formed 247 estimates by Asian subregion, we aggregated the results using weighted averages based on the 248 ratio of country rice production level to total regional rice production in 2015. Our modeling result shows that 32% to 42% of that historical yield increase arose due to 264 increased atmospheric CO2 concentrations. However, the yield growth was depressed by climate 265 conditions in the form of high maximum temperatures in most Asian countries. Thus, technological 266 developments are not doing as well in increasing rice yields as might otherwise appear, and success 267 with climate-based mitigation policy directed toward atmospheric CO2 concentration reduction 268 will lower yield levels. This finding implies that future rice yields are subject to the opposing 269 forces of climate change and CO2 concentration mitigation. Namely, if CO2 -driven climate change 270 proceeds, yields will grow due to increasing atmospheric, yields will grow due to increasing 271 atmospheric CO2, but the changing climate will reduce that yield growth rate. On the other hand, 272 if substantial CO2 mitigation occurs and atmospheric concentrations stabilize or fall, yield growth 273 will slow and possibly decrease. But this will be offset by lesser degrees of climate change and 274 their yield implications. 275 Ideally, we would have liked to include agricultural research and development (AGRD) 276 expenditures in this analysis, but we could not do so due to the time length of our data set. In 277 particular, the effects of AGRD investment on rice yields are not instantaneous and, in fact, involve 278 a long lag. Specifically, Alston and Pardey (2006)  directions. First, while we used a fixed-effect model to consider systematic country effects, we did 288 not incorporate data that could reflect both country-specific differences and commonalities, such 289 as soil quality, degradation, etc. Adding these as independent variables would help further identify 290 the role of CO2, climate, and other factors. Second, we would like to have a longer AGRD 291 expenditure series, so we could better study links between investments, climate, CO2, and yield 292 growth. Third, a further study could assess the effects of long-term climate events (e.g., El Niño, 293 decadal climate variability, etc.) on rice yields and develop corresponding climate change 294 adaptation.  (2020), Meinshausen et al. (2011) FACE yield and CO2 data 1998-2003, 2007-2008, 2010 Hasegawa  Note: The null hypothesis is that the data is non-stationarity. The test used the data excluding FACE observations to remain a panel data structure. The panel data unit root test was conducted in R using purtest() from package plm. Adjusted R 2 0.054 0.940 0.943 0.945 0.944 0.951 Note: significance levels are marked as follows: *p<0.1, **p<0.05, ***p<0.01. The above results are the third stage of our estimation procedure. The significant Breusch-Pagan test at least at 1% significance level for the first stage results suggests the existence of heteroscedasticity, providing a justification of using the 3-step FGLS approach.  (1) and (2) in this table are the same as Models (5) and (6) in Table S5 but with detailed interaction term coefficients. The above results are the third stage of our estimation procedure. The significant Breusch-Pagan test at least at 1% significance level for the first stage results suggests the existence of heteroscedasticity, providing a justification of using the 3-step FGLS approach.  Table S6 when evaluated with the CMIP5 projection data. The percentage is the total yield growth since 2015. Darker and lighter green indicate larger and smaller growth rates, respectively. The regional weights, which are based on the ratio of country rice production level to total regional rice production in 2015, are used for calculating the regional weighted average projection.