where X is a vector of time varying observable independent variables, α is an unobserved individual specific effect of the ith farmer that cannot be explained by Xit, and uit is the classical error term. When αi is not correlated with explanatory variables (random effect) or with initial mental health status, Eq. (1) could be estimated by generalised least squares (GLS) approach (Roy and Schurer, 2013). However, it is more likely that αi is correlated with independent variables of the model, which will result in biased coefficient estimates (Wooldridge, 2005). By applying the CRE model,