Systematic non-response in discrete choice experiments: implications for the valuation of climate risk reductions

ABSTRACT

Discrete choice experiments (DCEs) addressing adaptation to climate-related risks may be subject to response biases associated with variations in risk exposure across sampled populations. Systematic adjustments for such response patterns are hindered by the absence of rigorous, standardised selection-correction models for multinomial DCEs, together with a lack of information on non-respondents. This paper illustrates an empirical approach to accommodate risk-related response patterns in DCEs, where variations in risk exposure may be linked to observable landscape characteristics. The approach adapts reduced form response-propensity models to correct for survey non-response, capitalising on the fact that indicators of risk exposure may be linked to the geocoded locations of respondents and non-respondents. An application to coastal flood adaptation in Connecticut, USA illustrates implications for welfare estimation. Results demonstrate systematic effects of risk-related response patterns on estimated willingness to pay.

KEYWORDS:

• Non-response bias
• response propensity
• coastal adaptation
• flood
• sea-level rise
• stated preference

Acknowledgments

This research is supported by the Northeast Sea Grant Consortium, via prime award NA10AOR4170086 to MIT Sea Grant (sub-award 5710003190). Opinions do not imply endorsement of the funding agency.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. For example, see Edwards and Anderson (1987), Loomis (1987), Bockstael et al. (1990), Whitehead (1991), Whitehead et al. (1993), Cameron et al. (1996a, 1996b), Messonnier et al. (2000) and Cameron and DeShazo (2005).

2. DCE questionnaires often ask each respondent to answer multiple choice questions, leading to panel structure in the data.

3. A general discussion of this issue is provided by Greene (2003, 183–189).

4. For example, unlike properly specified and applied Heckman-selection models, there is no formal econometric proof that guarantees that these approaches eliminate non-response bias.

5. One may also use the (inverse of) estimated response propensity as an observation weight prior to estimation of the second stage model (Groves 2006).

6. These variations were used to test hypotheses unrelated to non-response bias.

7. Surveys were only mailed to physical addresses, not including post office boxes.

8. Slight variations in attributes between survey versions are not expected to influence response propensity patterns.

9. Fitted response propensities are scaled up by 100 before incorporation into the second stage model.

10. This interaction, when included, leads to an insignificant coefficient estimate with a large standard error in both the conditional and mixed logit models. This large standard error on the interaction causes a non-trivial portion of the estimated marginal utility of income distribution () to overlap zero. As discussed by Daly et al. (2012) and Hole (2007), the result is an undefined mean welfare estimate for the sample, even within the conditional logit model. To avoid this problem, we do not include the interaction between the response propensity and program cost in the random utility model. The exclusion of this interaction has no significant effect on other aspects of the model.

11. Parallel WTP–space models would not converge with the response propensity interactions, hence all models are estimated in preference space. Models were tested with alternative numbers of Halton draws to evaluate stability.

12. The model correctly predicts response behavior for 72% of the observations.

13. Because the underlying implicit price for Hard and Wetlands is negative, a positive sign for the implicit price difference (associated with increased response propensity) implies a smaller negative WTP.

Additional information

Funding

This work was supported by MIT Sea Grant, Massachusetts Institute of Technology [grant number 5710003190].