CAMP RESOURCE XVII JUNE 24-25, 2010 Empirical steps towards a research design in multi-attribute non-market valuation.

CAMP RESOURCE XVI IJUNE 24-25 , 2010

Empirical steps towards a research design in multi-

attribute non-market valuation

Plan of the talk

Focus on SP data and survey developmentExperimental designs evolved from orthogonalityChoice of elicitation method (incentive compatibility)Gumbel Heteroskedasticity vs Utility

HeteroskedasticityGeneralised logitDecision heuristics as systematic components of

heterogeneity ATTRIBUTE ATTENDANCEMaybe also....Order effects in repeated choicesContext effectsSubjective scenario conjectures

Basic generic question

What should I think about when starting a new SP survey/study for multi-attribute non-market valuation?

What do recent research results suggest?1st issue is whether the new SP data are need to

enrich existing RP data, or they are to serve stand-alone

If data enrichment the typical concern is to supplement the existing RP data and break away from multicollinearity (need special experimental designs “pivoted” on existing data)

Assume the study is only SP: typical steps?

Define research questionDraft survey, decide on:

specify provision rules, policy deliverables, preference elicitation mode, incentive compatibility, administration mode (face2face, CAPI, web-based, phone

supported, paper and pencil, etc. ) info to deliver, feedback on respondent’s understanding of this info info on attribute processing, scenario conjectures etc.

Run focus groupsGet starting designs (Orthogonal on the Diff.)

Typical steps (Cont’ed)

Simulate data, estimate specification of interests and welfare estimates for scenarios of interest (here you find if the data you collect give you back the spec you need, e.g. Animal welfare study) example, plot

Run pilot(s)Amend draft survey Obtain priors to optimize sample size use

Prior on parameter estimates (beta hats) Prior on specific functional form (Choice probabil.)

ExpDes: One shot vs sequential

One shot Use priors and select design criterion (or

combination of design criteria and their respective weights) D-efficiency, S-efficiency, C-efficiency, Minimum

entropy, Minimum complexity, etc.

Sequential Determine size of sampling waves Decide Bayesian rule to adopt to embed sequential

learning Each previous phase “informs” design of all following

stages Same sample size can give 1/3 more accuracy

Elicitation Methods

Pair-wisePair-wise + status quoFull RankingRatingBest-worst

Inter agency

Joint decision-making and group interactions diadic (e.g. Couples, Beharry et al.) or triadic (couples + child, Marcucci et al.))

Consensus seeking with interaction (e.g. Connected business solutions, location decisions, etc.

Types of heteroskedastic effects

Gumbel error heteroskedasticity Common form sigma=exp(z’theta), so that >0 z= vector of choice-task related effects (e.g. measures

of choice complexity, Swait and Admowicz 2001, DeShazo and Fermo 2002)

Utility heteroskedasticity Var(U1,2) Var(Usq) (Scarpa et al. 2005, Hess and Rose

2009) Common form additional error component

Both forms are likely to co-exist, and SQ choice-task often induce the latter

Scale and utility effects in logit

njtnjtnnnnnjt xU ])1([

( ) 1,..., ; 1,..., ; 1,..., ,njt n njt njtU x n N j J t T

( ) 1,..., ; 1,..., ; 1,..., ,njt n njt njtU x n N j J t T Taste heterogeneity, MXL

Scale heterogeneity, S-MNL

Generalised Logit, G-MNL

G-MNL to WTP-space

njt n njt n njt njtU x p

* * /njt n njt njt njt nU x p “WTP Space”

“Utility Space”

Set gamma =0 and phi=1

From –beta_n/phi_n

From WTP_n

Attribute processing: non-attendance

Either Ask people which attributes they attended to Yes/no to attendance to each attribute Or Likert scale

Or infer it from observed sequence of choices Zero constrained latent classes Variable selection model (spike model in CV)

Recent evidence: attendance may not be the same across all the sequence of choices (choice-task non attendance)

From a WTP estimate of Euro 790/year down to Euro 20/year!!! For preservation of mountain land landscaped

Variable selection (spike model equivalent)

Conclusions

Multi attribute research design is becoming increasingly complex

Need to simultaneously address many issues before one can retrieve “unconfounded” utility structures

Respondent interaction and feedback are increasingly becoming as validating and informative

Order effects

WTP estimates depend on the order at which you estimate them in the sequence of choices

Learning effects?Strategic response effects?Heterogeneity?

Order effects in WTPs

0.00

1.00

2.00

3.00

4.00

5.00

6.00

7.00

8.00

9.00

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

Question Order

WT

P p

ou

nd

s

Marginal WTP (odor)

Marginal WTP (color)

Confidence intervals

Context effects in choice-tasks with 3 alternatives

From Rooderkerk, van Heerde and Bijmolt, 2009

Scenario adjustments

Proposed scenarios may be mis-construed or subjectively adjusted (e.g. Risk latency in micro-risk (Cameron and DeShazo))

Subjective perception of Status-quo attribute levels versus objectively measured ones (Marsh et al.)

Conclusions

Multi attribute research design is becoming increasingly complex

Need to simultaneously address many issues before one can retrieve “unconfounded” utility structures

Respondent interaction and feedback are increasingly becoming as validating and informative