estimates to compute predicted values for the entire sample. Therefore, definition of farming skills is
modified to mean the contribution of farmer attributes to technical efficiency if the farmer were a self-
cultivator.
We include district-level dummy variables to account for broad regional agro-climatic
characteristics. Since only irrigation variables are available as farm level measures of land quality, farm-
level variation in land quality that is not correlated with irrigation may be left out. It can be argued that
some of this residual land quality reflects high skill levels of the farmer, such as the skillful management of
soil quality, and legitimately belong in the predicted skill index. If the remaining unobserved land quality is
correlated with observed farmer characteristics, the predicted skill estimates are biased. The implications
of this problem are discussed in detail in under specification problems.
The last stage in constructing the skill index requires the estimation of equation [48] for the self-
cultivators. Prediction efficiency equations are commonly estimated by regressing the technical efficiency
estimates obtained from the stochastic frontier estimation on a set of explanatory variables [Pitt and Lee
1981, Kalirajan 1981, Kirkley et al 1998]. Several recent studies [Kumbhakar et. al 1991, Reifschneider
and Stevenson 1991, Huang and Liu 1994, Battese and Coelli 1995] have pointed out that such a two-step
method is fundamentally incorrect because the dependent variable in the second step is assumed one-
sided, non-positive and identically distributed in the first step. A more appropriate method is to estimate
the parameters of equation [48] jointly with those of the stochastic frontier using maximum likelihood
methods using consistent distributional assumptions on the error structure. We use the version of this
method proposed in Battese and Coelli [1995].
The frontier-based skill index has its share of problems that include 1) strong distributional
assumptions on the error structure, 2) inconsistency of the technical efficiency estimator, and 3)
endogeneity problems. The distributional assumptions are common to most maximum likelihood methods.
similar boats and share the same waters. Although there is some merit to such an inference in a controlled setting
28