Geoscience Reference
In-Depth Information
agriculture in the United States. They report negative climate
impacts using a 'farm land' model, but a positive outcome using
a 'crop revenue' approach. Their findings highlight the impor-
tance of taking adaptation factors into account when evaluating
climate effects.
19.2
Description of the model
Statistical model
A statistical model relating yield per decare to meteorological
data is employed. The relationship between yield per decare,
Y, and temperature, T, precipitation, P, and a time trend, τ, is
assumed to be linear. Temperature is measured in GDD. The
equation is
Y
=+ +
αβ γ
T
P
+
θτω
+
ijt
ij
ij
ijt
ij
ijt
ijt
where i is the index for crop (potatoes, wheat, oats and barley),
j is the county index and t is the time index denoting annual
observations from 1958 until 2001. ω
ijt
is the error term.
*
GDD
is defined as the annual sum of degrees accumulated above 5°C
threshold. Through an ordinary least squares (OLS) regression
we seek to correlate variations from year to year, in yield per
decare, to the variability in GDD and precipitation. The esti-
mated parameters are
ˆ
,
ˆ
ˆ
ˆ
where the indices are
αβγ
,
nd
θ
τ
ij
ij
ij
left out for simplicity.
We were unable to take an explicit account of a number of
non-climate factors. However, a time trend variable which was
included in the regression runs to account for general long-
term time trends, which may have been influenced by a number
of other factors. Examples of such influences are technologi-
cal change and innovations (e.g. improvements in agricultural
inputs and/or practices, and/or changes in production patterns),
increased productivity due to other climate variables, and a
fertiliser effect from increased CO
2
concentration in the atmo-
sphere. As an alternative to the time trend we included CO
2
concentrations in some of the regressions (see Annex 19.3 for
a closer description of this model variant). Sunlight is another
important weather variable for crop yields since it provides
*
We assume that the error variances are constant and that the errors are not
autocorrelated. Given that these assumptions are fulfilled, the ordinary least
squares estimators are the best linear unbiased estimators. Checking the
Durbin Watson statistic for some country cases revealed no indications of
autocorrelation problems.
