RAE of WTP for Area Flooding

RAE of WTP for Area Flooding

Figure 13.1. Plot of kernel-smoothed distribution of the Relative Absolute Error of WTP estimates for Area Flooding. True data generating process (Continuous line NL): MXL-e dashed-dotted line; Dashed line MNL; Dotted line MNL-Asc.

SQ or the alternatives different from the SQ beyond what can be explained on the basis of the attributes values alone. We found that in our samples the conditional logit model, that ignores any source of SQ effect, produces the lowest estimates of benefits from provision of externalities. While from the societal viewpoint such a conservative estimate would guide investments in a cautious way, it would still represent a sub-optimal resource allocation, as many potentially beneficial proposals would fail the Pareto efficiency test by providing too low a benefit estimate.

Following other authors (Haaijer, 1999; Kontoleon and Yabe 2003), we have argued that there are very good reasons for investigating the existence of SQ effects in the application of choice-experiments, and that these reasons might be particularly compelling in non-market valuation of environmental goods.

RAE of MRS for Area Flooding

RAE of MRS for Area Flooding

Figure 13.2. Plot of kernel-smoothed distribution of the Relative Absolute Error of WTP estimates for Area Flooding. True data generating process: MXL-e (dashed-dotted line). Continuous line NL; Dashed line MNL; Dotted line MNL-Asc.

We examined three specifications that can be used to account for these effects: the conventional logit model with alternative-specific constant, the nested logit model and and the less conventional mixed logit with error components and alternative specific constant.

Secondly, we reported how we observed different forms of statistical evidence of SQ effects in two separate studies on preferences for water management attributes, which include important public goods, such as number of areas protected by flooding and number of households protected from odour and flies. While in a study we observe that all three specifications accounting for SQ afford similar statistical performance and WTP estimates, in the other application we observe that the mixed logit with error component and alternative-specific constant statistically dominates the nested logit and MNL-Asc, but this dominance does not implies statistically different estimates.

Finally, we investigated the effects of mis-specification using in turn the three SQ data generating processes by means of Monte Carlo experiments over a plausible range of sample sizes. The results of the experiments suggest a number of points.

First, when SQ effects are a concern, the use of simple conditional logit specifications may produce strongly biased estimates for the taste parameters. These will also produce biased welfare measures.

Secondly, when the true DGP is mis-specified, the MXL-e specification generally provides a good performance in our Monte Carlo experiments. Such performance is not matched neither by the NL model nor by the MNL-Asc model when the true DGP is MXL-e.

In conclusion, our empirical results confirm the existence of a systematic effect of the status-quo alternative on choice selection. This was previously discussed and evidenced in general terms by Samuelson and Zeckhauser (1988) and Hartman et al. (1991). Such effect was examined more specifically in the context of choice-experiment in market research by Haaijer (1999) and Haaijer et al. (2001) and addressed in environmental economics by Hanley and Wright (2003), and Li et al. (2004) by means of nested logit models.

We find that a less usual specification, namely the MXL-e consistently achieves better results than MNL with an alternative-specific constant for the SQ and NL specifications. The MXL-e model is parsimonious, yet, it captures SQ effects in both the systematic component of preference via alternative-specific constant, and the unobserved heterogeneity associated with hypothetical changes described by unfamiliar attribute levels. It also breaks away from the restrictive independence of irrelevant alternatives.

Of course the usual caveats pertaining to Monte Carlo results apply here. Namely, these results might be not very general and perhaps they are due to the particular data employed in this study. Nevertheless we find quite plausible that a specification that accommodates status-quo effects simultaneously in both the stochastic and deterministic component of utility outperforms specifications that only address one at the time.

Further research should investigate how general these preliminary results are, and how status-quo effects can be related to the various features of the experimental design, investigating — for example — the relationship between choice-complexity and degree of familiarity with attributes levels defining the status-quo vis-a-vis the proposed changes.

Was this article helpful?

0 0

Post a comment