We decompose yield difference relative to a reference level into components attributable to (1) efficiency difference, and movements along the frontier due to (2) land quality, to (3) land size, and to (4) other inputs. The production frontier is built using nonparametric methods requiring no specification of the functional form of the technology. We analyze the contributions to yield relative to a reference unit in terms of the quadripartite decomposition finding that results depend on the choice of the unit of reference. If the reference unit is chosen to be the mean, land size contributions are found to be negatively correlated to yield with usual finite moments regression methods. Also nonparamteric correlation confirms the negative sign of the relationship. If the reference unit is chosen to be the median instead, land size contributions are found to be negatively correlated to yield with usual finite moments regression methods. But nonparametric correlation is not statistically significant because many farmers have no contribution to production difference from their different land sizes. Integrated squared density difference tests show in both cases efficiency has a major role in shaping the distribution.