on the field; expected sign: + ) are chosen as such explanatory variables. Being aware that
these four variables only partly explain the differences in yields across farms we also add the
ten dummies of program participation in 1997 to the regression. The purpose to include these
dummies is to pick up differences in yield that eventual program participants exhibited even
before the introduction of the these programs. Hence,
10
(4)
Vi,1994 = Yθ + Yl AREA1994 + '2 RATIOl994 + Уз UV1994 + У AUH1994 + ∑ ( § kDk ,1997 ) + S i
k=1
Subsequently, we take the parameters estimated in of regression (4) and the values of
AREAl997, RATIOl997, UVl997, and AUHl997 to calculate the hypothetical yields we would have
observed in 1997 with no agri-environmental program in place:
10
(5) hi,l997 = c0 + cl AREAl997 + c2 RATIOl997 + c3 UVl997 + c4 AUHl99c + ∑( dkDk ,1997 )
k =1
Taking into account the stochastic nature of the coefficients estimated in regression
(4), we perform a Monte-Carlo simulation. In particular, we utilize the covariance matrix of
regression (4) to draw a sample of 2000 coefficient vectors1. These 2000 coefficient vectors
are used to calculate 2000 hi,1997 and to estimate 2000 times regression (3). The results of
these procedures are presented below.
1 the coefficients of the step1 regression are multivariate normally distributed as (b, COV), with dimension 15. In
order to draw the required sample from this multivariate normal distribution, we need the Cholesky-factorization
A of the (15x15)-covariance matrix COV; with this matrix, we can transform a vector y of 15 independent
realization of a standard normal distribution so as to conform to the distribution of the covariance matrix COV:
x = b + Ay with A = cholesky(COV) and y~N(0,1)