Geoscience Reference
In-Depth Information
energy for photosynthesis. However, as meteorological stations
were unable to provide relevant proxy data (i.e. cloud cover
observations) for the complete period of our study, we were not
able to include this variable in the analyses.
We carried out regressions at the national level by merging
county data into two different variants of the model. In the first
model variant, we allowed different constant terms for each
county, whereas we assumed that the marginal effect of changes
in weather data was the same for all counties. This model vari-
ant implies that there are differences in the yield level across
counties, but no differences in the marginal yield of changes
in the weather (i.e. GDD and precipitation). This is modelled
through an additive dummy variable for each county with the
exception of Akershus/Oslo, which is taken as the reference
county. In the second model variant, different constant terms
are retained, but in addition we allow for a shift in the marginal
effect (slope) of annual precipitation by adding a multiplicative
dummy variable to the precipitation variable for each county.
The latter model variant implies that there are systematic dif-
ferences between counties with respect to the level of yield per
decare for a crop, as well as with respect to the marginal effect
on yield of changes in precipitation, but with no differences in
the marginal effect of changes in GDD. The different treat-
ment of GDD and precipitation is based on regression results
at county level, which indicated that there is a larger variance
in the marginal effect of precipitation across counties than in
temperature (GDD).
Variants of the
model
The main model contains GDD, annual precipitation and a
time trend as independent variables, and was employed on each
crop at county level and at national level. However, a number of
model variants were tested on the crop yield and weather data
before ending up with this model. The chosen model produced
more significant coefficients and a better fit to the data than
the alternatives. The model variants included growing season
precipitation, carbon dioxide concentration (in different data
formats), frost events in the spring (in different data formats),
fertiliser use for the latter part of the estimation period, and
logarithmic or quadratic weather variables.
*
See Annex 19.3 for
a more detailed account of the model variants that were tested.
*
Thompson [22] advocates the use of quadratic terms for weather variables.
Parry and Carter [16] also find changes in climate to have non-linear effects.
