The sample selection model is based upon a bivariate or a multivariate structure, and distributional assumptions are in this context more severe than in univariate settings, due to the limited availability of tractable multivariate distributions. While the standard FIML estimation of the selectivity model assumes normality of the joint distribution, alternative approaches require less stringent distributional hypotheses. As shown by Smith (2003), copulas allow great flexibility also in FIML models. The copula model is very useful in situations where the applied researcher has a prior on the distributional form of the margins, since it allows separating their modelling from that of the dependence structure. In the present paper the copula approach to sample selection is first compared to the semiparametric approach and to the standard FIML, bivariate normal model, in an illustrative application on female work data. Then its performance is analysed more thoroughly in an application to Contingent Valuation data on recreational values of forests.
27 August 2012 in Journal of Agricultural Economics
Selectivity bias caused by protest responses in Contingent Valuation studies can be detected and corrected by means of sample selection models. This paper compares two methods: the Heckman 2-steps method and the full ML, applied to data on forest recreation - where WTP is elicited as a continuous variable. Either method has its own drawback: computational complexity for the ML method, susceptibility to collinearity problems for the 2-steps method. The latter problem is observed in our best fitting specification, with the ML estimator outperforming the 2-steps. In this application, overlooking the effect of protest responses would cause an upwards bias of the final estimates of WTP.
Distributional assumptions are crucial in the estimation of the value of public projects assessed by means of contingent valuation analyses, and it would seem obvious that tests for model specification should play an important part in the statistical analysis. It can be observed, though, that when the competing hypotheses are non nested, the choice of the model is often based on heuristic grounds, or, at most, on deterministic selection model criteria such as Akaike's (1973). In this paper we study two alternative, probabilistic, approaches to checking model specification, that, like Akaike's, are based on the Kullback-Leibler Information Criterion (KLIC): the model selection testing proposed by Vuong (1989) and the non nested model test proposed by Cox, in the simulated approach of Pesaran and Pesaran (1993). The three approaches are confronted by comparing their performance in selecting among different models applied to simulated contingent valuation data. Our results seem to warrant the use of the Cox test for medium-large size samples, while for small size samples its performance is less satisfactory. When the data set is small, use of a model selection method may be preferred to model testing. In this case, the Vuong model selection testing is recommended as an alternative to the deterministic approach of the Akaike's criterion.