Climatic Change

, Volume 112, Issue 2, pp 525–533

Projected temperature changes indicate significant increase in interannual variability of U.S. maize yields

A Letter
  • Daniel Urban
  • Michael J. Roberts
  • Wolfram Schlenker
  • David B. Lobell

DOI: 10.1007/s10584-012-0428-2

Cite this article as:
Urban, D., Roberts, M.J., Schlenker, W. et al. Climatic Change (2012) 112: 525. doi:10.1007/s10584-012-0428-2


Climate change has the potential to be a source of increased variability if crops are more frequently exposed to damaging weather conditions. Yield variability could respond to a shift in the frequency of extreme events to which crops are susceptible, or if weather becomes more variable. Here we focus on the United States, which produces about 40% of the world’s maize, much of it in areas that are expected to see increased interannual variability in temperature. We combine a statistical crop model based on historical climate and yield data for 1950–2005 with temperature and precipitation projections from 15 different global circulation models. Holding current growing area constant, aggregate yields are projected to decrease by an average of 18% by 2030–2050 relative to 1980–2000 while the coefficient of variation of yield increases by an average of 47%. Projections from 13 out of 15 climate models result in an aggregate increase in national yield coefficient of variation, indicating that maize yields are likely to become more volatile in this key growing region without effective adaptation responses. Rising CO2 could partially dampen this increase in variability through improved water use efficiency in dry years, but we expect any interactions between CO2 and temperature or precipitation to have little effect on mean yield changes.

Supplementary material

10584_2012_428_MOESM1_ESM.doc (26 kb)
Table S1Quantiles of temperature in the training data and future period as projected by 15 GCMs. (DOC 26 kb)
10584_2012_428_MOESM2_ESM.pdf (22 kb)
Figure S1Aggregate yield at each year in the 1950–2005 training period. Detrending both the observed and fitted aggregate yields by regressing on year + year2 gives an R2 between the aggregate yield trends of 0.39, meaning we are explaining roughly 40% of the variance in aggregate yield residuals with weather effects. The large drop in observed yields in 1993 that our model misses was a result of heavy flooding. (PDF 22 kb)
10584_2012_428_MOESM3_ESM.pdf (44 kb)
Figure S2(a) The percent changes in the coefficient of variation of predicted yields in 2030–2050 relative to 1980–2000 as a function of mean growing season temperature in the base period. Each point in the scatter corresponds to an individual county. Cooler counties exhibit bigger increases as they move away from the optimum growing season T. (b) The same quantity as (a) but for the total yield variance, which includes the variance of predicted yields as well as the variance of the model residuals, with the latter assumed to remain constant in future climate. (c) The fraction of observed variance explained by the model in each county, plotted against the mean T in that county. Because the model explains much less variance of the time-detrended yields in cooler than in warmer counties, the residual variance contributes a greater relative share to the total variance. Thus, although the variance of predicted yields goes up more in cooler counties, the total variance is projected to go up more in the warmer counties. (PDF 43 kb)
10584_2012_428_MOESM4_ESM.pdf (34 kb)
Figure S3Scatterplot of the square of residuals from the training data fit as a function of temperature, overlaid with a local linear nonparametric fit. The fit shows a slight increasing trend in residuals at higher temperatures, but the trend is weak, with a correlation between the fitted values and data points of 0.0036. (PDF 34 kb)
10584_2012_428_MOESM5_ESM.pdf (19 kb)
Figure S4A quantile regression of yield anomalies on the same set of predictors as the ordinary least squares (OLS) regression shows little distinction in the shape of the distributions of quantile yields as a function of temperature. The higher quantiles show slightly less dropoff at higher temperatures, but combined with the small amount of variance in the residuals explained by temperature, it seems the effect of trends in unexplained variance is small. (PDF 18 kb)
10584_2012_428_MOESM6_ESM.pdf (76 kb)
Figure S5The distribution of log-fitted yields of the panel data more nearly matches the distribution of actual panel yields than does the linear fit, adding support to our choice to regress log of yields rather than actual yields on the weather and county predictors. (PDF 75 kb)

Copyright information

© Springer Science+Business Media B.V. 2012

Authors and Affiliations

  • Daniel Urban
    • 1
  • Michael J. Roberts
    • 2
  • Wolfram Schlenker
    • 3
  • David B. Lobell
    • 1
  1. 1.Environmental Earth System Science, Center on Food Security and the EnvironmentStanford UniversityStanfordUSA
  2. 2.Department of Agricultural and Resource EconomicsNorth Carolina State UniversityRaleighUSA
  3. 3.Department of Economics and School of International and Public AffairsColumbia UniversityNew YorkUSA