proxies. Pant [1983] includes the value of farm implements. However, these proxy variables may have
an independent effect on land tenancy that cannot be separated out by including them directly in an OLS
regression. For example, education may have a positive effect on leasing by improving access to market
information or a negative effect by improving non-farm outside options, two effects that are quite
independent of the positive skill effect. Therefore, there is no a priori basis to identify any of the farmer
characteristics as indicators of skill.
In this paper, we combine these two approaches by interpreting farming skill as the contribution of
a farmer’s observed characteristics to the farm’s productive efficiency. Farming skills are defined as a
weighted average of the tenant’s observed characteristics (such as age, education, household
demographics, types of non-farm activity and wealth). Since there is no a priori basis to assign values to
the weights, they are computed by estimating the predicted contribution of farmer characteristics to farm
efficiency. The simplest method is to include a vector of farmer characteristics directly in a production
function. We use a conceptually similar method, but give a more appropriate structure to the estimates by
using a stochastic production frontier method.
The stochastic production frontier, first proposed by Aigner et. al [1977] and Meeusen and van
der Broeck [1977] is a production function with two error terms, a random disturbance term, ε i, that is
independent and identically distributed, and a one-sided non-negative error term, ui, that is distributed
independently of εi. 20 The former captures both the effects of unobserved stochastic factors
(e.g.weather shocks) and errors due to mis-specification of the model, and the latter represents “technical
inefficiency” of the farmer or, more precisely, the ratio of the observed to maximum feasible output,
where maximum feasible output is determined by the stochastic production frontier [Lovell 1993]. This
structural model is more appropriate for our purposes because it explicitly recognizes the non-linearity of
20 See Greene [1993b] for an extensive survey of this literature
25