A note about model selection and tests for non-nested contingent valuation models
Introduction
Non-market goods can be evaluated by means of contingent valuation models, where the parameters of the underlying discrete choice model are estimated by maximum likelihood, and the value of the mean, or median, willingness to pay (WTP) is calculated as a function of the estimated parameters. It is well known that the parameter estimates, and hence the estimates of mean and median WTP, depend crucially on distributional assumptions, and therefore model specification should play an important role before policy implications are derived from any estimated model.
When comparing two different non-nested models for the willingness to pay, an applied researcher might follow one of two distinct paths: apply model selection criteria to come up with a ‘better’ model, in the sense that it fits the data better, or apply a non-nested test to see whether, given an alternative model and the data, there is evidence against the model of the null hypothesis (Gourieroux and Monfort, 1995). Although the purpose of carrying out a non-nested hypothesis test is not to come up with a ‘better’ model, the applied researcher might well decide to use either one or none of the models for further inference based on the results of the test. In contingent valuation analyses, non-nested competing models are generally assessed by means of selection criteria. For example, McFadden (1994) used the highest likelihood and the Akaike information criteria to choose among models. While some doubts as to the effectiveness of such methods have been raised (Reiser and Shechter, 1999), we are not aware of any alternative approach in this field; and, it might be added, that testing procedures for non-nested models are few in discrete data modeling in general. As reported by Pesaran and Weeks (2000), the extra computational difficulties that implementing the Cox test entail may explain why this path has been so neglected. However, the probabilistic framework suggested by Vuong (1989) for model selection has also not been much used, notwithstanding its computational simplicity.
The purpose of the present paper is to provide some guidance to the applied researcher as to what might be the outcome of the two different paths —Vuong’s testing approach to model selection and the Cox test for non-nested hypothesis — as well as of the widely used Akaike’s criterion. The experiment will involve the application of standard models for contingent valuation (and for discrete models in general) to simulated data.
The structure of the paper is as follows. Section 2 gives a brief background about the different procedures. Section 3 describes the experimental setting. The results of the simulation exercise are presented in Section 4. Section 5 contains our conclusions.
Section snippets
Methods
We are interested in comparing pairs of competing parametric families of conditional densities of Yi given Xi, given byA wide variety of criteria are available in the model selection literature to select the model that best fits the data. For instance, based on the Kullback–Leibler Information Criterion (KLIC), Akaike (1973) suggested comparing the values of the log-likelihood penalized by the number of parameters and selecting the model that yields the
Experimental design
We adopt the censored regression model proposed by Cameron and James (1987) and Cameron (1988) where the latent variable (WTP) is a known function of the regressors and an additive error term and what we observe is a YES–NO answer to a hypothetical payment question. For our experiment, we considered different DGPs for the WTP, obtained from a linear functional form for the deterministic part of the model, which is common to all experiments, and an additive error term that is varied across
Results
Table 1 reports the results of the pairwise model comparisons for each of the three DGPs (Normal, Extreme Value, translated Lognormal) and the three candidate models: probit, logit and weibit. The upper section of the table reports the percentage of times that a given conclusion is reached for each one of the three approaches, while the lower section reports which model yields the lowest MSE for the estimated mean and median WTP. The purpose of this study is to assess the three approaches in
Conclusions
When the selection of a model is based on model selection criteria or Vuong’s test, one should bear in mind that the ‘best fit’ model could still give a very poor fit if both models are heavily misspecified. Of course, it could be possible that, from some prior knowledge, the researcher is able to narrow down the pool of candidate models to those that are somehow ‘close’ to the truth: for an approach in this sense, see Carson and Jeon (2000).
From an operative point of view, in many
Acknowledgements
We would like to thank Karim Abadir for helpful comments on a previous draft of this paper.
References (14)
A new paradigm for valuing non-market goods using referendum data: maximum likelihood estimation by censored logistic regression
Journal of Environmental Economics and Management
(1988)- et al.
Information criteria for selecting possibly misspecified parametric models
Journal of Econometrics
(1996) Information theory and an extension of the maximum likelihood principle
Econometric Model Selection: A New Approach
(1989)- et al.
Efficient estimation methods for closed-ended contingent valuation surveys
The Review of Economics and Statistics
(1987) - et al.
On overcoming informational deficiencies in estimating willingness to pay distributions
- et al.
Statistics and Econometric Models
(1995)
Cited by (18)
Learning from inferred foregone payoffs
2015, Journal of Economic Dynamics and ControlA drift correction optimization technique for the reduction of the inter-measurement dispersion of isotope ratios measured using a multi-collector plasma mass spectrometer
2014, Spectrochimica Acta - Part B Atomic SpectroscopyCitation Excerpt :The reason for this difference among the Pb ratios is not known. In principle it would be possible to determine the probability density function of S [19–22,25] and then re-derive the necessary MLE equations for a range dependent merit function. The low number of data points (a minimum of N ≥ 7, spread out, as uniformly as possible, over the run time) makes it very likely that a statistical hypothesis test of an = 0 will prove inconclusive, with the recommendation that more data points be acquired [25].
A path-size weibit stochastic user equilibrium model
2013, Transportation Research Part B: MethodologicalValue relevance of comprehensive income and its components in front of the net income
2011, Revista de Contabilidad-Spanish Accounting ReviewCan unbiased be tighter? Assessment of methods to reduce the bias-variance trade-off in WTP estimation
2011, Resource and Energy EconomicsCitation Excerpt :Clarke (2007) demonstrated that the distribution-free test is more efficient than the Vuong test if the empirical distribution of the individual log-likelihood ratios is leptokurtic. The low power of the Vuong test was shown in an experimental setting by Genius and Strazzera (2002); therefore we applied both tests to select between the Gaussian copula with extreme value marginals and the Bivariate Normal model. The empirical distribution of the likelihood ratios is leptokurtic, and, in line with Clarke's findings, the Vuong test does not lead to a strong rejection of the null hypothesis that the two models are equivalent (chi-square = 1.62, p-value = 0.105); while the Clarke test indicates the Gaussian with extreme value marginals (B6 = 233, n = 376, p-value = 0.000) as the better model between the two.