CAMP RESOURCE XVII JUNE 24-25, 2010 Empirical steps towards a research design in multi- attribute non-market valuation
Dec 18, 2015
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