This paper investigates the effect of land fragmentation on farm efficiency in Kosovo utilising agricultural household survey data. To recognise heterogeneity among agricultural production systems in Kosovo, we estimate the technology separately for different groups or “classes” of farms, identified using latent class modelling. This approach separates the data into multiple technological “classes” according to estimated probabilities of class membership based on multiple specified characteristics, relating in this case to land fragmentation and market integration. The latent class frontier method is linked to the estimation of a multi-output multi-input production function, namely a directional output distance function, and to the estimation of Morishima elasticities of substitution, based on shadow price changes indicating allocative efficiency changes. The analytical results confirm that the usual approach of using one homogenous function to estimate fragmentation effects is misleading and can lead to inappropriate policy recommendations. Three distinct classes of farm households are identified, which show different levels of efficiency and the proxies for land fragmentation and market integration show different signs over these classes.