You are on page 1of 22

ACIAR project on Agricultural Policy Research

to Support Natural Resource Management in


Indonesia’s Upland Landscapes (INDOGREEN)

Capacity Building Workshop

Em Prof Jeff Bennett

Australian National University


Choice Modelling Method
Concept

▪ Stated preference technique useful when non-use


values require estimation OR where markets do not
provide sufficiently rich data to allow estimation
▪ Same conceptual base as dichotomous choice CV
▪ Arose because:
• CV was proving difficult to implement because of the
controversies
• CM provided some specific advantages over CV
▪ Based on Lancaster’s theory of demand:
• Goods and services are composed of numerous
characteristics or attributes that exist at varying levels in
each version eg. Car has attributes of size, power
economy, comfort etc. Different model cars have different
levels of each attribute
▪ Developed in commercial marketing literature in the
context of exploring the market potential for new
products
▪ Used extensively in the transport economics
literature
▪ Respondents asked to choose between alternative
environmental management options
▪ The options are described by a sequence of
attributes … relate to the non-market characteristics
of the good including use and non-use
▪ One option is monetary … the payment vehicle
▪ The options are differentiated by the attributes
taking different levels
▪ A choice may be between two and many options
…known as a ‘choice set’
▪ One option is the ‘do nothing new’ or ‘status quo’
▪ A CM questionnaire involves multiple choices
sometimes up to 20 for market applications but
usually 6-8 for unfamiliar environmental
applications
▪ Usually around 5 attributes
▪ Usually 3 -6 levels
▪ In making their choices, respondents provide
information on the trade-offs they are willing to
make between the attributes
▪ With one attribute being money, WTP for more of a
non-marketed attribute can be estimated … called
the implicit price
▪ To estimate the trade-offs, respondents need to face
choices that give exposure to the full range of
possible combinations of attribute levels
▪ With so many attributes taking on so many levels,
and being combined into choice sets, the number of
possible combinations is huge …the ‘full factorial’
▪ Need to sample across the full factorial for a
manageable number of combinations … a fraction
of the full factorial
▪ This is done using an ‘experimental design’
▪ Representative … orthogonal design, efficient
design
▪ First order effects explain 80+% of variation
▪ Interactions require more complex experimental
design … select attributes to ensure independence
▪ Beware of ‘causally prior’ attributes
▪ Often the number of choice sets from the
experimental design > cognitive capacity …
establish ‘blocks’ of choice sets to be used across
sub-samples
Steps

▪ Follows similar lines to CV


▪ Questionnaire involves:
• Outline of the issue
• Solution to the problem
• Mechanism for the implementation of the solution
• Choice sets
• Follow up to detect protest
• Socio-economic background
▪ Focus groups attribute choice plus communications
▪ Expert involvement to assist in developing the
attributes
▪ Research design involves a number of versions of
the questionnaire being prepared relating to the
‘choice set ‘blocks’ … need to be randomly
distributed across the sample
Analysis
Choice models

▪ Respondents when making choices provide a


sequence of yes no responses to the options put
forward
▪ A three option choice set yields three rows of data:
• Attribute levels, socio-economic characteristics of the
respondent, chosen (Y/N) for EACH option
▪ Model analyses the probability of yes being
selected as a function of the levels of the attributes
and the socio-economic characteristics
▪ Multi-nomial logit is the standard form used in the
modelling … uses maximum likelihood estimation
procedure
▪ Assumes that independent alternatives are
irrelevant to the choice made (IIA assumption … a
Gumbel distribution for the error term)
▪ Coefficients of the attribute variables are by
themselves meaningless as they are confounded by
the ‘scale parameter’
▪ U (A) = ASC + β1 A1 + β2 A2 + … + βa Sa + βb Sb …
U (B) = β1 A1 + β2 A2 + … + βa Sa + βb Sb …
U (C) = β1 A1 + β2 A2 + … + βa Sa + βb Sb …
▪ ASC : Alternative specific constant … accounts for
otherwise unobserved variation
▪ Sa and other socio economic parameters are interacted with
the ASC to prevent singularities
▪ Dividing coefficients ensure the cancellation of the scale
parameter and delivers the willingness to make the trade off
▪ Generic models: same coefficients across all options
▪ Alternative specific models: different coefficients
▪ Where the denominator is the parameter of the
money attribute, the implicit price is estimated
▪ If a linear model is estimated, implies constant
marginal utility over the range of levels the attribute
takes in the questionnaire … can’t extrapolate
outside that range
▪ To estimate compensating surplus:
• Substitute levels for the projected outcomes of the
proposed solution into the estimated equation
• Substitute levels for the status quo into the equation
• Subtract new from old and divide by money attribute
coefficient
▪ Test for IIA violations by dropping one option from
each choice set and re-estimate the choice model
▪ If models are significantly different, need to
consider an alternative form
▪ Sometimes respondent heterogeneity will explain
IIA breaches … overcome by ensuring the socio-
economic variables are included
Random parameter logit

▪ The parameter estimated for each attribute is a


distribution that is determined by respondent
characteristics.
▪ The distribution around the mean parameter
estimate is a function of respondents’ ages,
incomes etc
▪ Enables identification of differential preferences for
the same attribute across the sample
Issues
▪ Framing – provide the range of substitutes as
alternatives in the choice sets
▪ Trade-off is extended to involve other attributes as
well as money (eg employment) … impacts on
strategic bias
▪ Social aspects as well as environmental attributes
can be incorporated and values estimated
▪ Analysis is more complex than CV
▪ Concerns regarding the inclusion of a provision rule
to ensure incentive compatibility
▪ The cognitive capacity needed of respondents is
high. Concerns regarding response rates and self-
selection of respondents
▪ Use of visual aids to encourage comprehension
▪ Prospect of using ‘workshops’ rather than full
surveys
▪ Scale and scope issues
▪ Uncertainty … in information and preferences
▪ Distance effects
▪ Dimensions of the choice task … number of
alternatives, attributes, levels, etc etc

You might also like