In figure 5, the poverty incidences as the dependent variables are plotted against agricultural productivity (an explanatory variable). Agricultural value-added per worker (VA_agr) is initiated as Thai agricultural productivity in every graph. For graph A, the poverty rate goes in the opposite direction with agricultural value-added per worker. Employment in every sector in graph B is also negative non-linear relationship with the productivity. Graph C obviously shows that the linear estimation is improper since it will generate the higher error from mean than non-linear from. In graph D, purchasing power and agricultural value-added per worker are in the increasing trend. In every graph, those correlations are non-linear form. Log-linear model is likely the better model because it easier reflects non-linear functions i.e. quadratic and cubic functions. Furthermore, MWD test5 confirms that log-linear function is more appropriate than another since R-square of adjusted linear model is less than one of log-linear model (Gujarati and Porter, 2009). Therefore, all estimations are done in log-linear functional form.