l LIBRARY
.4) LU 37 Michigan §tate
University

This is to certify that the
thesis entitled

Attribute Based Modeling of Recycling Preferences at
Michigan State University

presented by

David J Gebben

has been accepted towards fulfillment
of the requirements for the

MS. degree in Agricultural, Food, and
Resource Economics

 

 

Major Professor’s Sign’a’ture
37/2/05)

Date

 

MSU is an affinnative-action, equal-opportunity employer

 

A _‘-.-u-rdcnun.-.-.---.—.----u-.-----—-a-—a---.—-.-n-u--.--.--.-.-.—.---.-..---.-.---

-.-.-.-.-.-._._

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE
JUN i 7 201;

-An ‘3
“31" '

 

7"

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5’08 K.IProj/Acc&Pres/ClRC/Dateoue.indd

ATTRIBUTE BASED MODELING OF RECYCLING PREFERENCES AT
MICHIGAN STATE UNIVERSITY

By

David J Gebben

A THESIS

Submitted to
Michigan State University
In partial fulfillment of the requirements
For the degree of

MASTER OF SCIENCE

Agricultural Food and Resource Economics

2008

ABSTRACT
ATTRUBUTE BASED MODELING OF RECYCYCLING PREFERENCES AT
MICHIGAN STATE UNIVERSITY
By
David J Gebben

Environmental, economic, and political pressures increasingly make recycling
programs important parts of both public and private solid waste management strategies.
Recycling programs differ across communities, businesses, and public institutions yet
share common attributes such as the materials to be recycled, and the material separation
required of recyclers. Effective recycling programs should benefit from an
understanding of the recycling behaviors, attitudes, and preferences of the people the
program intends to serve.

This thesis examines recycling preferences though a stated preference choice
experiment performed at Michigan State University. We perform a trade-off analysis to
provide policy makers information about the trade-offs campus members are willing to
make between program attributes and costs. The first essay examines the preferences and
trade-offs for the campus populations: students, faculty, and staff. The second essay
compares two econometric methods for identifying segments of the population that have
distinct preferences. All else equal, we find that campus members want a recycling
program in which they can recycle more materials, in some sort of container, in locations
close to their dorm rooms or offices, with lower costs. In terms of the trade-offs between
program costs and the program attributes, all groups were willing to incur additional costs

to have recycling locations that are near their dorm rooms or offices.

ACKNOWLEDGEMENTS

First and foremost I would like to thank my advisor Dr. Frank Lupi for giving me
the opportunity to work on this data set and guiding me through the process of what it
takes to write a professional document. Without Dr. Lupi being picky and a stickler for
“Yes, but get it right,” this research would not be near the quality that it is. Any mistakes
or weaknesses are my own. I would also like to thank Drs. Michael Kaplowitz and Laurie
Thorp who agreed to take me on my final term here with funding to get this thesis
completed. Along with that I must thank the Office of the Vice President of Operations
and Finance for the funding it provided for my research assistantship. And Dr. Scott
Loveridge who helped guide me through my academic time here at MSU, and until that
last possible moment when on the day of my defense we finally got around to changing

the paperwork so that Dr. Lupi was on paper my major professor.

iii

Table of Contents

Chapter 1 - Introduction .................................................................................................. 1
Chapter 2 - Essay 1 ........................................................................................................... 3
Chapter 3 - Essay 2 ......................................................................................................... 32

iv

List of Tables

 

 

 

 

 

 

 

 

 

 

Table 2.1: Recycling Program Attributes and Attribute Levels 14
Table 2.2: Web-Survey Response Rates - - - 18
Table 2.4 Per-Person MRS for Each Campus Group and Aggregate MRS for
Campus 25
Table 2.4 Per-Person MRS for Each Campus Group and Aggregate MRS for
Campus 25
Table 3.1 Recycling Program Attributes 45
3.4 Interaction Effects Model - 48
Table 3.2 Tests on the Interaction Effects Models 49
Table 3.3 Parameter Estimates from the Interaction Effects Models 51
Table 3.5 AIC and BIC Results for Each of the LCM Models that Converged ........ 55
Table 3.6. MRS for Two-Segment Latent Class Models for the “Faculty and Staff
Dummies” Version and the “No Variables” Version 56
Table 3.8. Marginal Rates of Substitution Averaged Across Segments 58
Table 3.9 Latent Class Models 63

 

Table 3.10 Coefficients on Probability of Membership in Segment One from the
Two Segment LCM Staff and Faculty Dummy Variables 64

 

Table of Figures
Fig. 2.1. Example page of choice experiment a respondent might have seen. ........... 31

Fig 3.1 Example of a choice experiment question a respondent may have seen ....... 47

vi

Chapter 1 - Introduction

Environmental, economic, and political pressures increasingly make recycling
programs important parts of both public and private solid waste management strategies.
Recycling programs differ across communities, businesses, and public institutions yet
most recycling programs often share common attributes such as the materials to be
recycled, the location of collection sites, and the material separation required of recyclers.
Each of these attributes can affect program costs and revenues as well as the participation
and satisfaction of the recyclers to be served by the program. Designing effective
recycling programs should benefit fiom an understanding of the recycling behaviors,
attitudes, and preferences of the people the program intends to serve.

Although several studies on the behaviors and preferences of recyclers have been
conducted for households and cities, there are few studies focused on large institutions
such as college campuses. This thesis addresses this gap by conducting a trade-off
analysis of attributes of a potential university recycling program. The data are from a
stated preference survey of university students, faculty and staff. The two main goals of
this thesis are:

1. To estimate campus members’ preferences for recycling program attributes and
their marginal rates of substitution between program attributes and program costs;

2. To Compare and contrast the performance of two econometric tools, interaction
effects and latent class models, for explaining consumer heterogeneity and

segmentation regarding recycling preferences.

To accomplish these objectives this thesis presents two essays analyzing
stakeholder recycling preferences. Essay 1 addresses the first objective and presents
estimated recycling program preferences for three campus groups: students, faculty, and
staff. The recycling program attributes are materials to be recycled, the location of
recycling receptacles, the degree of source separation required by users, the type of
recycling containers provided to users, and the cost of the program to the institution. The
data are from a web-based stated preference survey conducted from November to
December 2007. Random effects probit models are used to estimate the preferences, and
marginal rates of substitution (MRS) are derived for the individual user groups and for
the campus as a whole. The MRS are useful for comparing preferences across user
groups and represent the trade-offs stakeholders are willing to make between additional
program costs and the different possible attributes of the recycling program.

Essay 2 compares and contrasts two approaches for examining consumer
heterogeneity. The first approach relies on interaction terms to distinguish unique
segments of users with distinct underlying preference structures. With this approach the
researcher’s choice of the interaction terms, perhaps based on prior understanding of the
population, determines the possible market segments. The second approach uses latent
class models (LCM) to let the data reveal different underlying market segments.
Numerous variants of the LCM models are reported. The preferences and marginal rates
of substitution that are revealed by the LCM and interaction effects models are compared
to those of the interaction effects models, and the relative merits of the approaches are

discussed.

Chapter 2 — Essay 1

1. Introduction

Environmental, economic, and political pressures increasingly make recycling programs
important parts of both public and private solid waste management strategies. While
recycling programs differ across communities, businesses, and public institutions, most
recycling programs share some common attributes. These attributes include such things
as: the materials to be recycled, the location of collection sites, the design and type of
recycling receptacles, as well as the procedures and processes for material separation.
Designing effective recycling programs should benefit from an understanding of the
recycling behaviors, attitudes, and preferences of the people the program intends to serve.
Although several studies of these features focus on households and cities, there are not
many studies focused on large institutions. Solid waste management decisions facing
large institutions like universities, businesses and industries in many ways parallel those
of communities and cities. The design of such programs impacts the size and distribution
of such factors as collection costs, separation costs, operational costs, as well as potential
revenue streams from recycled materials. This paper focuses on developing an
understanding of the recycling behaviors, attitudes, and preferences of the various
constituencies of a new recycling program at a large, institution-- Michigan State
University (MSU), a Big Ten university.

MSU occupies a 5,200-acre campus with 2,100 acres in existing or planned
development. It is comprised of approximately 579 buildings, 46,000 students, 4800

faculty, and 6100 staff. The university generates more than 12,000 tons of solid waste per

year (Selke et al. 2006). As is typical for a major public university, MSU has several
characteristics that complicate the recycling planner’s tasks--buildings spread out over a
large area; building ages that may be more than 100 years old and others that may be
new; and a wide range of types of wastes, from radioactive materials to food stuffs to
office paper. Furthermore, planning for a new recycling facility and program requires
decisions about what materials to recycle; whether and where materials are to be
separated; how to collect such materials, as well as what facilities to build and staff. The
reported research is part of an effort to plan, develop and implement an integrated,
campus-wide recycling program that is cost effective and responsive to the desires and
preferences of the in situ stakeholders of a large institution—4m students, faculty and
staff of a tier-one, research-intensive university.

Public participation in the design and implementation of recycling programs has
been associated with improved recycling program performance (e.g., F012 1991, F 012 and
Hazlett 1991, Nyamwage 1996). Our study was designed to elicit university community
members’ preferences for recycling program attributes as part of the university’s efforts
to build new recycling facilities and develop a new recycling program. To do so, we
implemented a university-wide preference survey (also called a trade-off analysis or
choice experiment) that asked respondents to evaluate alternative recycling program
configurations. The Attribute Based Method (ABM) was used to identify the relative
importance that different program attributes have on people’s preferences for alternative
recycling programs. Using ABM allowed us to estimate the preferences people have for

alternative programs (combinations of program attributes) that the university could

implement. We included a cost constraint to facilitate examining individuals’ willingness

to make trade-offs between various program attributes and costs.

1.1 Previous Literature

There is an extensive literature on municipal and regional recycling programs,
both on participation rates and the economics of recycling. Previous research suggests
that recycling rates of university campuses and cities may be increased by increasing
community recycling options, raising community members’ knowledge of recycling, and
by improving the convenience of recycling (F 012 1991). It has been observed that “cities
with higher rates of [recycling] participation and waste stream diversion place more
importance on citizen involvement in the policy initiation and program design decisions”
(F 012 and Hazlett 1991). Others have pointed out that increased participation in
community recycling seems to have resulted fiom community input regarding changes to
the recycling containers and location of collection points (Nyamwage 1996). One such
study found that “democratizing” the recycling program was associated with an increase
in the recycling program participation rates (F012 1991). Others have examined the
economics of household recycling resulting in the development of a model where
information about recycling policies can be examined (Choe and Fraser 1999).

Narrowing the focus to valuation studies of recycling programs, previous research
has examined drop-off recycling programs (Tiller, Jakus, and Park 1997) as well as
curbside recycling programs (Caplan, Grijalva, Jakus 2002; Aadland, Capland 2003).
This previous valuation literature focuses mainly on the waste stream that is created by

households with survey research examining household or municipal managers’ responses.

While household waste streams are significant, large institutions and companies also have
significant waste stream management concerns. This paper begins to address and assess
recycling program preferences of community members (i.e., stakeholders) of an
institution that shares characteristics of a small community, a corporation, and a public
entity.

Examining recycling program preferences at an institutional level has advantages
over looking at such data from the household level. Large institutions by their very nature
produce more waste than a single household, yet they are, in some important ways, a
single decision-making entity. In the context of institutional decision making, planners
may be able to identify and organize groupings of individuals, waste-stream producers, or
reduction opportunities to a greater degree than a municipal entity might be able to
manage. Furthermore, large institutions such as universities and large businesses often
manage their own waste streams, have centralized planning, and may be able to fashion
their own systems with greater ability to effect change than a typical city with numerous
autonomous households. As has been shown, institutions can implement resource
recovery and waste recycling with relative ease resulting in reduced pressure on solid
waste disposal sites (Mbuligwe 2002).

Universities are unique institutions for undertaking waste reduction and recycling
program efforts. University constituencies are inherently more educated than the general
population, and therefore may be better able to understand complex and integrative
problems as well as large-scale problems such as environmental pollution and climate
change (Hines 1987). In addition, the university setting provides a social and

psychological location for the development of environmental values and concerns in

community members (Ewert and Baker 2001). Studies of university students’ recycling
choices and motivations have found that university students’ pro-environmental attitudes
predicted their pro-environment behaviors and that increased environmental knowledge
added to the pro-environmental impact of attitudes on behaviors (Mienhold and Malkus,
2005). Jennings (2004) found the most important determinants of recycling behavior in a
university setting to be both attitudinal (i.e., requires greater education to create change)
and logistical (e.g., can be altered, a new recycling center can be constructed, etc.). Our
study will address preferences over some of these logistical factors associated with a
campus recycling program.

To date, the numbers of campus recycling studies are few relative to studies of
recycling of the population at large or of residential populations. Almost all previous
college campus recycling studies have looked at respondents’ recycling attitudes,
knowledge, and behaviors. This paper differs fi'om those studies in that we examine
specific recycling program attributes and analyze the trade-offs that respondents make
between recycling program characteristics when they choose among possible recycling
programs for the university. In the next section we discuss the methods for quantifying
attribute preferences and trade-offs.

2. Methods
2.1 Theory and background

In a choice experiment, the researcher asks each respondent to make a choice
between two bundles of attributes with an associated price. The choice problem mimics
actual market behavior, where consumers choose between different types of cars or shoes

with different attributes and prices (Holmes & Adamowicz 2003). The estimates and

preferences are then derived from the respondents’ choices based on the underlying
utility that the different bundles provide the individuals.

Attribute based methods of analysis seek to understand consumer choices in terms
of the attributes (or characteristics) of the alternatives that are chosen by individuals.
That is, goods or programs are seen as consisting of bundles of attributes that combine to
produce consumers’ utility (Louviere et al. 2000). The Attribute Based Method estimates
economic values for a technically divisible set of attributes, and the inclusion of price as
an attribute in the choice sets provides a constraint to the choice. Eliciting and
quantifying these trade-offs is one way our study differs from what other campus
recycling literature offers where the examinations tend to address behavioral and
attitudinal research.

Building on the attribute based model, this study uses Random Utility Theory to
understand and evaluate individuals’ choices among alternative recycling programs
(Adomowicz and Boxall 2001). Under the assumptions of Random Utility Theory (RUT),
the general model of choice starts with the indirect utility, U, a consumer is able to obtain
which is a function of income I, a vector of attribute levels, x, and individual preferences,
e.

U=u(I,x, e) (1)

In our choice experiment, respondents were asked to choose between two bundles

of attributes, each with a price. Therefore, the indirect utility for an amount p paid for an

attribute would be:

U1=u(I-p1,x1,el)- (2)

where, utility U 1 is a function of xl a vector of attributes provided to the consumer at
prices p,. In our choice experiments, the participants were presented with two recycling

program choices, and were asked to choose the program that they preferred. Therefore,

the indirect utility associated with the second recycling option can be expressed as:

U2 =u(1-P2,x2,82)- (3)
Following RUT, an individual would select (be willing to pay pl ) for the bundle

of goods xl if the utility is greater than the utility from x2 at p2 (i.e., Ul >U2 ). So the

probability that a respondent would say Yes when asked if x1 is preferred to x2 , may be

expressed as:

Pr(yes) = Mm (161,1 - 121,621) 2 u2062.1 - 172.62 )]

P (4)
r( yes) 2 Pr[Au > 0].

Using the basic approach embodied in (4), we can parameterize the utility functions to
quantify preferences. Let utility be given by the sum of observable and unobservable

components:
“ij = vij + eij . (5)
where Vij represents the observable portion of utility; ea. represents the unobservable

portion of utility; j represents alternatives, and i represents people. With that, we can then

modify (4) so that:
Pr(yes) = Pr[vi]’ + ey' > Vik + eikI (6)
The observable portion of utility is made up of the following characteristics:

Vij=axj+flUi_Pj)a (7)

where x j is a vector attributes associated with alternative j, a is a vector of estimable
parameters, ,6 is an estimable parameter, I ,- is income of respondent i, and p j is the

price of alternative j.
Substituting (7) into (6) we can write the probability of a yes answer to a
preference question comparing j versus k (where a respondent prefers program j to

program k) as follows:

Pr(yes) = Prlarxj +fl(1i-pj)+eij > ark +/3(Ii -pk)+eik]
Pr(yes) = Pr[(ax j wan-(13p; -flpk) > eik -eij] (8)
Pr(yes) = Pr[a(Ax )—fl(Ap )> eik -eij]-

It is typically assumed that the marginal utility of income does not change from one
choice to another. Since the utility of income does not change from one state to another,

we can then drop the AI,- tenns as in the second line of (8). This is because the income

(and other individual-specific variables) does not vary for the individual as he or she
makes choices. We can next make the assumption that the difference in the error terms is
normally distributed allowing for a model that gives the probability of an individual
choosing alternative j over alternative k :

Pr(yeS) = P1'l01(Ax)- [3(AP) > e] = (PMAx) - MAM], (9)
where (I) is the cumulative distribution function (cdf) of the standard normal distribution,
assuming that a = 1. Equation (9) can be estimated using the maximum likelihood
estimation procedure for a probit model. Due to the discrete nature of the choices, 0
cannot be identified in a probit, so one can take it as if 0' = l or simply recognize that all

parameters are identified up to the unknown variance term. As such, model comparisons

lO

that involve parameter ratios are fully identified since the unknown variance term will
cancel. We report such calculations below.

The Marginal Rate of Substitution (MRS) allows the comparison of magnitudes
for the preference of one attribute over another. Of particular relevance is the MRS

between an attribute and money. The calculation of the MRS was a / ,6 , where a is the
coefficient on the attribute variable and ,8 is the negative of the cost coefficient fi'om the

regression output. This MRS measurement provides a ratio that expresses in common
units people’s willingness to make monetary trade-offs for changes in an attribute. The
MRS expressions can be compared across the populations and from one attribute to

another.

11

2.2 Choice Experiment Survey Design

As mentioned, a choice experiment survey was conducted that elicited
preferences over possible recycling programs. The survey contained several sections
including background information and questions about the university’s plans to develop a
new recycling program. In the survey, each respondent was presented with three sets of
two possible recycling programs and respondents were asked to select their preferred
program among each pair. The recycling programs were described by several attributes
and each program had a cost associated with it. Adequately describing the program and
the attributes was crucial to ensure that when each respondent made their choices, each
respondent has in mind the same good. It is well-known that stated choice surveys should
be as realistic as possible so that the respondents are able to provide accurate, realistic
answers (Carson et al. 2001). Therefore, we now review the efforts that went into the
attribute selection and the way that attributes were conveyed to respondents.

The survey instrument was developed in three phases that centered on focus group
studies, survey design, and pretesting/revision of the survey (Kaplowitz et al. 2004). A
series of key informant interviews were conducted with university administrators,
consulting engineers, faculty active in university environmental affairs, as well as student
activists. These were conducted so that the survey targeted appropriate recycling program
attributes and levels as well as to design an instrument that adequately informed
respondents about the various recycling program attributes and the possible levels for
each attribute. The focus groups were qualitative studies conducted to gauge the

recycling knowledge, behaviors, and attitudes of MSU. The focus group results then

12

guided subsequent interviews and the design and implementation of the survey itself
(Hansen et al. 2007).

The result of our focus groups and interview efforts with university
administrators, consulting engineers, and university community members was the
identification of five program attributes with various possible levels for each attribute.

The program attributes consisted of the materials to be recycled (paper, cardboard, non-

deposrt containers, and deposrt containers ), the type of recycling container indivrduals
would be provided, the location where recyclables would be taken, the amount of waste
stream separation required by individuals, and the amount the program would cost the

institution. (See Table 2.1 and discussion below).

 

1 Michigan has a deposit on certain types of beverage containers.

l3

Table 2.1: Recycling Program Attributes and Attribute Levels

 

Attribute

Attribute Level

 

Materials

Location

Containers

Separation

Cost

Paper, Cardboard, Non-deposit containers, Deposit containers
Paper, Non-deposit containers, Deposit containers

Paper, Deposit containers, Cardboard

Paper, Non-deposit containers

Paper, Non-deposit containers, Cardboard

Paper, Deposit containers

Paper, Cardboard

Floor/Classroom
Building
Drop-Off
Central

No bin

Multiple small bins
One large bin

A single hanging bag

One stream
Multiple streams
Complete separation

$50

$150
$300
$500

 

In the survey, respondents were told that MSU was designing a recycling program

and that their feedback was needed on the type of program to be implemented. To

introduce survey respondents to the attributes included in the recycling program,

respondents were first presented with a description and sometimes a photograph of the

each program attribute and the possible attribute levels. We then provided a Likert scale

question about the attribute for the respondent to answer. Asking a question about the

attribute immediately after providing information about the attribute and its levels was

done to get respondents to read the information rather than just skip over the text. After

14

informing the respondents about all of the possible program attributes and their levels, we
then presented the three pair-wise program choice experiment questions (see figure 1 for
an example of one of the choice pairs).

Note that there are many possible combinations of all the attributes, yet each
respondent only saw three of the many possible pairs. The attributes pairs were then
varied across respondents following an experimental design. The experimental design
consisted of a mix of pairs from a main-effects design for these attribute levels and
random pairing of the attribute levels. By varying the attribute pairs across respondents,
the statistical techniques outlined above can be used to identify the effect that each
attribute level has on the probability that an alternative is preferred.

The survey allowed for the collection of data in a simulated market setting by
asking respondents to indicate choices from realistic sets of alternatives that included
costs (McFadden 1986). Since the respondents were asked to make choices under a
constraint, cost, it was crucial that we explain how program costs would be paid. The
program costs were costs to the institution, and the construction costs would be paid for
through a capital bond mechanism, as was explained to the respondents. Further,
respondents were told that when the costs of the program were incurred, it would leave
less money for the institution to use in other areas. In the choice questions, the cost to the
institution was expressed in per-person terms. Hence, the cost associated with the
program is not a direct cost to the respondent, rather it is an opportunity cost to the
campus - as more money is spent on the recycling program, less money is available to

spend on other campus items. That is, implementation of a more expensive project would

15

translate into a higher per person expenditure (cost) and less money would be available to
spend on campus elsewhere.
2.3 Survey Administration

Most stated preference surveys have been done in person, or through a mail
survey. Both methods have their advantages and disadvantages. Both methods are
relatively expensive to implement. In person surveys offer greater control of information
because the survey interviewer may provide immediate feedback, while mail surveys
have the advantage of being able to reach a much larger number of potential respondents
than a team of interviewers would be able to contact. As in other fields of survey
research, web surveys are increasingly being used for stated preference studies including
choice experiments and other ABMs. Given that the study population (university
students, faculty, and staff) all had access to and use e-mail and the web for everyday
university business, the survey was designed for, and implemented as a web based
instrument. Doing so allowed for a low cost and a highly sophisticated study design. The
method also allowed each respondent, depending on their status as student, staff, or
faculty, to receive only those questions relevant to them and it allowed the researchers to
monitor and control the information viewed. The web-based method also allows the
researcher to randomize the pairs for the respondent. In this way, the web-based survey
design utilized some of the pros of both in person surveys and mail surveys. (A copy of
the web survey is available at http://recycle-
survey.msu.edu/Default.aspx?uid=96 1 085352397 .)

The sample list for the study was drawn from the university’s official lists of

faculty, staff and student. The registrar provided a random list of the email and mailing

l6

addresses of about 30% of each group — students, faculty, and staff. Those on the list
received either a postcard or an email inviting them to participate in the survey. We were
then able to track the participation rates and responses of each group. This degree of
control and information is unique relative to other web based surveys since we could
track who responded and we knew characteristics of the entire population. An initial
invitation was sent to all members of the sample in November 2007 informing them of
the study and providing them with a link to the survey. Up to two additional invitations to
participate were sent to potential respondents during November and December 2007.
That is, those members of our sample that did not participate in the survey after the first
invitation were asked or reminded to participate a second time, and finally,
nonrespondents were contacted a third time and reminded/asked to participate. The
overall response rate for this study was 25% with students responding at a lower rate than
faculty or staff. Table 2.2 presents the differences across the three strata of university

community members.

17

Table 2.2: Web-Survey Response Rates

 

 

All campus Faculty Staff Students
Percent
. 25% 38% 42% 20%
responding
Number. 3896 563 883 2450
responding

 

In our choice experiment, each individual is asked to respond to more than one
stated preference question making it likely that there are unobserved characteristics
specific to individuals that induce correlation across individuals’ responses. Therefore,
we estimate the model using a random effects probit model (Wooldridge 2002). In
random effects models, the error term is treated as two separate components. One
component is the unobservable portion that is unique to the individual respondent. The
second component is the random component across all individuals and their responses

(Boxall et al. 2003). This utility model can be written as:
Auy=a(AXj)-fl(Apj)+/Ii+eg', (10)
where ,u, is the unique individual error term, and eij is the random error term across the

whole population sample.

Coding the survey responses followed typical conventions. We estimated a
random effects probit regression using the standard statistical software package STATA
®. The probability of an individual choosing the first bundle of attributes (Program A) as
opposed to the second bundle of attributes (Program B) was treated as a “yes” and coded
as a “l .” Because the recycling program attribute levels, with the exception of costs, are

all dummy variables, not all levels can be identified. For each attribute, one level is set as

18

the baseline and is omitted from the estimation. The coefficients on the remaining
attribute levels then require an interpretation in relation to the base level. Thus, a
negative coefficient implies that that the attribute level is less preferred than the baseline
level for that attribute, while a positive value implies that attribute level is more preferred

than the baseline.

19

3. RESULTS

3.1 Preferences

The respondents were informed that university recycling programs could focus on the
following materials: paper, cardboard, deposit containers, and non-deposit containers.
The results, presented in Table 2.3, not surprisingly, indicated that people prefer
recycling programs with more materials for recycling rather than recycling programs with
fewer materials. Also not surprisingly, deposit containers are a less preferred program
recycling material for respondents than the other recycling program materials. We
speculate that this is due to the residual value that a deposit bottle or can retains in
Michigan because of the state bottle bill requiring refundable deposits on carbonated

beverage bottles and cans.

20

Table 2.3 Estimated Coefficients from the Random Effect Probit Model for
Recycling Program Attribute Preferences

 

 

Students" Faculty“ Staff***
Coef. p-value Coef. p-value Coef. P>z

Materials

Paper, NDC, DC, CB (baseline)

Paper, NDC, DC 033 0.000 -0.40 0.000 -0.33 0.002

Paper, NDC, CB 016 0.003 -0.22 0.014 -0.31 0.006

Paper, NDC -0.45 0.000 -0.45 0.000 -0.56 0.000

Paper, DC, CB -0.36 0.000 -0.46 0.000 -0.37 0.001

Paper, DC 057 0.000 -0.66 0.000 -0.69 0.000

Paper, CB -0.72 0.000 -0.63 0.000 -0.67 0.000
Location

Floor/Classroom (baseline)

Building -0.12 0.009 -0.14 0.062 0.05 0.595

Drop-off -0.68 0.000 -0.91 0.000 -0.95 0.000

Central -0.80 0.000 -0.86 0.000 -0.94 0.000
Containers

No Bin (baseline)

One 0.50 0.000 0.52 0.000 0.41 0.000

Bag 0.43 0.000 0.32 0.000 0.26 0.002

Multiple 0.49 0.000 0.49 0.000 0.38 0.000
Separation

Separate all (baseline)

Multiple streams 0.06 0.121 -0.01 0.858 -0.07 0.388

No separation -0.07 0.113 -0.17 0.013 -0.25 0.007
Cost (by 10005) -2.20 0.000 -l.80 0.000 -l.70 0.000
Sigma U 0.20 0.000 0.22 0.000 0.23 0.000
Rho 0.03 0.000 0.44 0.000 0.05 0.023
Log Likelihood -2746.66 -1088.11 -622.75

 

DC = Deposit Containers NDC= Non-Deposit Containers CB = Cardboard
* Number of Students: 2450 "Number of Faculty : 883 *** Number of Staff : 563

21

Respondents also learned that the university recycling program was considering
alternative material collection locations. The possible collection stations included floor or
classroom, building, drop-off points on campus, and a central location. The results for
respondents’ location preference indicate that respondents would like the more
decentralized options (on floor or classroom; in building). Students seemed to prefer
being able to recycle on their dorm building floor while faculty preferred a floor or office
level collection of materials. Interestingly, university staff seemed to prefer a single drop
point per building for recycling materials. Never the less, there was a clear preference in
all groups for recycling options that were closer to offices or dorms when compared to
the less convenient options of drop-off locations or a central location.

In recycling programs, recyclable materials may be source separated into multiple
streams, one comingled stream, or could be completely separated. Students did not show
significant preferences one way or another regarding separation of material. Faculty and
staff however did show a significant preference towards complete separation of
recyclables at the source compared to no separation. This result was not expected as it
was thought that complete separation at the source is less convenient for recyclers than
one comingled stream.

Survey respondents were also informed that the university recycling program had
an array of options for either providing or not providing individual collection bins to
campus community members. These options included a set of small bins, one large bin, a
bag that hangs on a door handle, or no university-provided recycling container. As the

results in Table 2.3 show, there was a strong preference by all groups for some sort of

22

university provided collection bin over the baseline comparison of no bin. The results
also show that for all groups the bag was the least preferred collection bin.

All groups showed a significant and negative sensitivity to the program costs,
That is, programs that cost the university more money were less likely to be preferred.
Because the discrete nature of the dependent variable in the probit estimation makes it
impossible to identify the underlying variance within each model, one cannot easily
compare the strength of the preferences across user groups. However, by taking
parameter ratios, as we do in the next section, one can make such comparisons. The
lower rows of table 2.3 provide the estimated panel error terms which indicate that in
each case the panel error structure was significant and has a statistically better fit than a

model that ignored the panel nature of the data.

3.2 Marginal Rate of Substitution

The Marginal Rate of Substitution (MRS) between an attribute and the per-person
program cost informs us of the rate that a person is willing to trade off that attribute for
additional program costs. The MRS are useful because unlike the individual parameter
estimates, the MRS can be used to compare preferences (willingness to make trade-offs)
across attributes and across user groups. Table 2.4 illustrates the per-person MRS for the
attributes for each of the three user groups. The final column also shows the population-
level MRS (labeled “All campus”). These latter values provide a measure of the rate at
which the total campus community would be willing to substitute program costs for
attribute levels. The costs associated with these bundles represent the opportunity costs

that the campus would be willing to incur to get a particular attribute level relative to the

23

baseline level for that attribute. These total campus values were computed by scaling up
from the per-person. Specifically, for each group the per-person MRS were multiplied by
population size and summing all these for students, staff, and faculty. In all cases, the
baseline MRS value is left blank as that is the level that the other values are either

moving away or towards.

24

Table 2.4 Per-Person MRS for Each Campus Group and Aggregate MRS for
Campus

 

Per-person MRS All Campus
Students Faculty Staff MRS

Materials

Paper, NDC DC, CB, (baseline)

Paper, NDC, DC -148 -215 -187 -9000

Paper, NDC, CB -71 -122 -180 -5000

Paper, NDC -201 -244 -322 -12000

Paper, DC, CB -161 -250 -213 -9900

Paper, DC -259 -361 -397 -16000

Paper, CB -326 -343 -383 -19000
Location

Floor/Classroom (baseline)

Building -52 -76 28 -2600

Drop-off -305 -498 -545 -20000

Central -361 -469 -542 -22000
Containers

No Bin (baseline)

One 224 280 232 13000

Bag 194 173 148 10600

Multiple 221 267 216 12700
Separation

Separate all (baseline)

Multiple streams 28 -7 -42 1000

No separation -30 -95 -142

-2600

 

DC = Deposit Containers, NDC = Non Deposit Containers, CB = Cardboard
Numbers for the all campus MRS are rounded and reported in 10005.

25

One of the most striking results from the MRS is that people are very willing to
incur additional program costs in order to have a recycling collection location closer to
them. The MRS indicates that location is the program attribute that when compared to
any of the other attributes, stakeholders are willing trade off the most for -- especially for
faculty or staff. The MRS values for waste separation also indicate that the preferences of

faculty and staff lean more towards source separation than single stream separation.

4. CONCLUSIONS

Across all three of the institution’s subpopulations, we found a strong preference
for recycling programs that include more recyclable materials as compared to those with
fewer materials. The results show that if a recycling program had to be limited to only
three of the four possible materials, the first material to be given up by respondents would
be deposit containers. This makes sense given the state’s deposit bill. If the campus
recycling program had to be limited to two materials for recycling, the MRS based on the
model results show that the stakeholders preferred choice for recycling materials would
be paper and non-deposit containers. Of course, if cost were not a factor, the overall
results show support for including as many recyclable materials in the program.

The study’s results support recycling program designs that reduce the
inconvenience of recycling and those that place recycling collection containers in close
proximity to respondents work/study/living areas. These findings are in line with
previous studies on recycling behavior (Kelly et al. 2006). These authors reported

significant difference in respondents’ occupation (staff or student) and their on-campus

26

recycling behavior which “appeared to be predominantly a contrast between
undergraduate students recycling sometimes on campus against general staff and
postgraduate students (and to a lesser extent, academic staff) recycling on campus
frequently” (Kelly et al. 2006). While our study found that students, faculty and staff
cannot be treated as the same groups since each population has preferences that are
significantly different statistically, we did find that the preference orderings were similar
across groups.

Program cost was highly significant and negative across all groups, as was
expected. Cost was explained as an opportunity cost, the coefficient on price and
significance therefore indicates the sensitivity to the opportunity cost of a new program
for the various programs. Our results indicate that students are more sensitive to the
opportunity cost of campus spending on recycling programs than either staff or faculty.
Put differently, the faculty and staff are more willing to spend MSU’s money on
recycling programs than students are.

A waste stream management plan that focuses only on source reduction ignoring
the subsequent user behaviors is not a comprehensive plan. A comprehensive plan needs
to consider full examination of the waste stream. Policies aimed at reducing the waste
generated can inadvertently redirect the consumer to purchasing alternative products that
do not meet as stringent requirements (Choe & Fraser 1999). While our study looked
only at the end user options portion of the stream, hopefully it will assist policy makers in
devising strategies that in the end will be more comprehensive.

Source separation can increase revenue, but collection can increase cost. We

found that people were willing do source separation, or showed no preference regarding

27

how source separation was done. However, we also found that people showed a strong
preference towards recycling closer to their dorms or offices. The design implication of
these findings suggests the collection of materials in multiple separate streams in multiple
locations within each campus building, rather than investing in specialized, central
separation facilities.

Students showed no preference regarding the form of source separation. The
faculty and staff showed waste stream separation preference, and so along with that, the
logical choice of recycling containers would be multiple small bins. This would allow for
one bin for paper, one bin for cardboard, and one bin for plastic bottles and containers.
All of the groups preferred having some sort of bin to collect the materials to no bin.
Either a bag to or multiple small bins for collection would be greatly preferred to no bin
by a great deal.

Areas for fisture research include incorporating socioeconomic features into the
model or comparing the colleges to each other for preference differences. This would fall
in line with other campus recycling research and potentially help target information
campaigns on campus. Our survey also had a number of Likert scale questions in the data
set that were used to help explain the various attributes. For this paper we concentrated
only on the choice experiment questions, however, there is the potential to examine if
respondents who ranked items as “highly important” also show a strong correlation
towards recycling attribute choices in the choice experiment. Another potential area for
research would be to examine if department major has any significant effect on responses

to the choice experiment.

28

References

Aadland, Capland. 2003. Willingness to Pay for Curbside Recycling with Detection and
Mitigation of Hypothetical Bias. American Journal Agricultural Economics, Vol
85 No (2) pp. 492-502.

Adamowicz, V., and P. Boxall. 2001. Future directions of stated choice methods for
environment valuation. Paper read at Choice Experiments: A New Approach to
Environmental Valuation, April 10, 2001, at London, England.

Caplan R.J., Grijalva, T.C., Jakus, RM. 2002. Waste not or want not? A Contingent
ranking analysis of curbside waste disposal options. Ecological Economics.Vol
43, No. 2-3 pp. 185-197.

Carson, R.T., N.E. Flores and NF. Meade. 2001. Contingent valuation: Controversies
and Evidence. Environmental and Resource Economics. Vol 19. No 2. pp.l73-
210.

Choe, C. and Fraser, Iain. 1999. An Economic Analysis of Household Waste
Management. Journal of Environmental Economics and Management.Vol 38. No
2. pp 234-246.

Cropper, M. I., Deck, L., Kishor, N., and McConnel, K.E. 1993.Valuing Product
Attributes Using Single Market Data: A Comparison of Hedonic and Discrete
Choice Approaches. The Review of Economics and Statistics. Vol. 75, No 2. pp.
225-232.

Ewert, A. and D. Baker. 2001. Standing for Where You Sit: An Exploratory Analysis of
the Relationship between Academic Major and Environment Beliefs.
Environment and Behavior Vol. 33. No. 5. pp. 687-707.

F 012, DH. (1991). Recycling Program Design, Management, and Participation: A
National Survey of Municipal Experience. Public Administration Review. Vol 53.
No. 3. pp. 222-231.

Folz, D.H., Hazlett, J .M. 1991. Public Participation and Recycling Performance:
Explaining Program Success. Public Administration Review. Vol 51. No.6. pp
526-533.

Hansen, L.T. McMellen, C. Olson, L. 2007.Recycling Attitudes andBehaviors on a
College Campus: Use of Qualitative Methodology in a Mixed-Methods Study.
Working Paper. MSU.

Holmes, TR, and Adomowicz, W.L. 2003. Attribute-Based Methods. In: Champ, P.A.,

Boyle,K.J., Brown, T.C., (eds.) A Primer on Nonmgrket Valuation. Kluwer
Academic Publishers: Dordrect/Boston/London pp. 171-220.

29

Hines, J. M., Hungerford, H. R. and Tomera, A. N. 1986. Analysis and Synthesis of
Research on Responsible Environmental Behavior: A Meta-Analysis. The Journal
of Environmental Education. Vol.18 No. 2 pp.l-8.

Jennings, M. E. (2004). An Instrument to Measure the Recycling Attitudes and Beliefs of
Undergraduate Students at a Large Northeastern University. Online Submission,
Education Resources Information Center: 22.

Kaplowitz, M.D., Hadlock, T.D., Levine, R. 2004. A Comparison of Web and Mail
Survey Response Rates. Public Opinion Quarterly. Vol. 68 No.1 pp. 94-101.

Kelly, T. C., I. G. Mason, M. W. Leiss and S. Ganesh. 2006.University community
responses to on-campus resource recycling. Resources, Conservation and
Recycling. Vol. 47. No. 1. pp. 42-55.

Louviere, J..l., D.A. Hensher, and J .D. Swait. 2000. Stated Choice Methods: Analysis and
Application. Cambridge: Cambridge University Press..

McFadden, D. 1986. The Choice Theory Approach to Market Research. Marketing
Science. Vol. 5, No. 4 pp. 275-297.

Mbuligwe, S. E. 2002. Institutional solid waste management practices in developing
countries: a case study of three academic institutions in Tanzania. Resources,
Conservation and Recycling. Vol. 35. No. 3. pp. 131-146.

Nyamwage, M. 1996. Public Perception of Strategies for Increasing Participation in
Recycling Programs. The Journal of Environmental Education. Vol. 27 pp. 19-
22.

Resource Book, 2008. mpz//www.recycle.msu.edu/downloads/educational resource 4—4-
08.pdf

Selke, S.E., Joshi, S., Sidique, S., Dabholkar, G. (2006). Michigan State University White
Paper on Solid Waste.
http://www.ecofoot.msu.edu/documents/solid.waste.wp.pdf.

Sidique, S. 2008Analyses of Recycling Behavior, Recycling Demand, and Effectiveness
of Policies Promoting recycling. Ph.D diss., Michigan State University.

Tiller, K.M, Jakus, P.M., and Park, W.M. 1997. Household Willingness to for Drop-off
Recycling. Journal of Agricultural and Resource Economics. Vol 22. No. 2. pp.
310-320.

Wooldridge, J .M. 2002. Econometric Mlvsis of Cross Section and PaLnel Dat_a.
Cambridge, MA: MIT Press.

30

Fig. 2.1. Example page of choice experiment a respondent might have seen.

WHICH OF THESE TWO PROGRAMS IS BETTER?

Compare the characteristics of Program A and Program 8 and choose the program that you
prefer for recycling at MSU.

Michigan State University Recycling Program
Characteristics

 

 

 

 

 

 

 

 

 

 

 

 

Program

Characteristics
(Click hyperlink for A B
more information)
chceyfibadbb: flaetgflajs Paper, Non-deposit Paper, Non-deposit
MSU program containers containers

. . All non-paper materials Group similar materials
Material Separation . . . .
at collection site mixed together In one together Incollectron

container containers

Collection Locgtrons Remote collection Central collection
you bring materials to
Malacca . .
for your room / offi ce Single large basket Multiple smaller baskets
MSU Cost (per person)
(to build facility, collect
and process materials, S150 S150
and operate program)

 

 

 

32. Which of the above programs do you think is better? (Pick One Response)
4".‘Program A 0 Program B

4 “__V__‘ _________._J\.
1 Back Next ...__.

31

Chapter 3 - Essay 2

3.1 Introduction

Environmental economists use many varieties of survey methods and modeling
tools for the estimation of non-market values. “Despite differences in format, all survey-
based nonmarket valuation methods attempt to measure trade -offs between money and
environment,” thus allowing for estimates of a Marginal Rate of Substitution (MRS) for
the environmental quality (Siikamaki and Layton 2007). While average trade-offs for a
population are often of interest, in many cases it would be usefirl to identify segments of
the population that may hold different preferences, and hence, have different MRS.
Recent econometric methods offer new tools for accounting for such preference
heterogeneity. Utilizing the same choice experiment data set, we will compare and
contrast two approaches for identifying preference heterogeneity within a population.

The first method, interaction effects models, requires the researcher to have an a
priori idea about the nature of the heterogeneity in the population. The approach uses
demographic variables or other knowledge about the individuals to segment the
population into distinct groups with different preference structures. The assumptions
about the segments can then be tested. The second method, latent class models, uses the
information inherent in the data to segment the population into different classes. The
latent class models can also use demographic information about the respondents to help
estimate the segments. As such, both models can incorporate individual-specific socio-
demographic variables into the preference structures that would otherwise not be

identified in the standard variant of the choice experiment. While there are random

32

parameters models and heteroskedastic error structures that can also incorporate

heterogeneity, those approaches are beyond the scope of our current research.

3.2 Motivation and Literature

The use of latent class models for explaining consumer heterogeneity arose out of
marketing research. Marketers were addressing problems of “identifying segments of
households in a population, determining their associated sensitivity to market variables,
and investigating the possible bases of segmentation” (Gupta and Chintagunta 1994).
Marketers developed techniques for taking aggregate demographic variables in
combination with revealed preferences, and determining the latent class membership
probabilities (Gupta and Chintagunta 1994).

The idea underlying LCM proposes that individual behavior “depends on
observable attributes and on latent heterogeneity that varies with factors that are
unobserved by the analyst” (Greene and Hensher 2003). Therefore, individuals are
implicitly sorted into some number of classes or segments, the individual may or may not
know which class she or he falls into, and the analyst does not know which class the
particular person falls into. Since the LCM does not require an assumption about which
class any one person falls into, the analyst avoids making “possibly strong or
unwarranted distributional assumptions about individual heterogeneity” (Green and
Hensher 2003). This is a useful characteristic for situations when the analyst has no
reasonable a priori way to distinguish the exact segments of the population. In contrast to

the interaction effects model where the researcher must make that a priori assumption.

33

Kamakura and Russell (1989) developed a model for preference segmentation that
rested on the “assumption that consumers can be placed into a small number of segments
each characterized by a vector of mean preferences and a single price sensitivity
parameter” (Kamakura and Russell 1989). They developed a model that both estimated
the segment level parameters and determined the likely segment membership for the
individual. The Kamakura and Russell (1989) method made possible model building that
then led to homogenous segments rather than deconstructing it from aggregate numbers
as previous LCM work had done.

The use of latent class modeling is becoming more common in environmental and
resource economics. Often LCM has been used in eliciting recreation site preferences or
choices (Boxall and Adamowicz 2002, Scarpa and Thiene 2005). The use of attitudinal
data from preference scales has also been incorporated into modeling the environmental
quality preferences of anglers (Morey et al. 2006).

“Latent class modeling can be viewed as a means of modeling heterogeneity
across individuals in a random parameters fi'amework” (Greene 2003). This characteristic
is useful as it allows estimation of unbiased models and “forecasting demand by
including individual characteristics. The goal is that through understanding heterogeneity
the researcher “will provide information on the distributional effects of resource use
decisions or policy impacts” (Boxall and Adamowicz 2002).

Work has been done by a number of researchers examining the performance of
the latent class model in comparison to other models. Greene and Hensher (2003)
contrasted latent class and mixed logit choice analysis using a travel choice data set.

Other research has been done to examine various forecasting capabilities of differing

34

models. For example, Provencher and Bishop (2004) examined the forecasting
efficiencies of the random parameters logit versus the latent class logit model for -
predicting the recreational fishing decisions of anglers. On a more theoretical level,
Provencher and Moore (2006) derived the likelihood functions of the latent class models
of Boxall and Adomawicz (2002) and Morey et al. (2006), where Boxall and Adamowicz
used a latent class model and Morey et al. a latent class model that included attitudinal
data, and determined that the two are “structurally similar.” Each of these papers that
contrast modeling methods used data sets related to some sort of site choice, be it
recreational or travel choice. Compared to these examples using revealed preference data,
our research is unique in that we use stated preference data and methods, and our
application is aimed at estimating recycling program preferences.

Some research has been done investigating the relationship between recycling
attitudes and actual recycling behaviors as a method of explaining recycling preference
heterogeneity outside the realm of environmental economics. Looking just at recycling
preferences, some research has suggested that attitudinal questions do a better job of
explaining general behaviors rather than specific actions, while other research has
suggested values can only be interpreted in the context of tradeoffs and possible tradeoffs
(Ewert and Baker 2001). Studies including a gender constraint have also had mixed
results. Ewert and Baker (2001) state that the results can be generalized to say that higher
income, higher educated, younger, less conservative people tend to show greater concern
for the environment. Lack of knowledge about environmental issues has been shown to
be an important factor determining environmental attitudes and behaviors (Gamba and

Oskamp 1994, Oskamp et al. 1991). The results of these research studies are not

35

uniformly consistent. Jennings (2004) found the most important determinants of
recycling behavior to be both logistical and attitudinal. Much of the above mentioned
work has been done in the areas of psychology and sociology.

Within recycling literature, there is an extensive body of work on municipal and
regional recycling programs, both on participation rates and the economics of recycling at
the household level. Though little work has been done examining why some people
choose to recycle, and some choose not to recycle. Some research has been done polling
various municipal leaders exploring what is effective at increasing participation rates, but
none to our knowledge that examines the consumer heterogeneity. One such study found
that “democratizing” the program can increase the participation rates (F 012 and Hazlett
1991). Others examined the economics of household recycling; they derived a model
whereby information about what policies should be implemented can be examined (Choe
and Fraser 1999). Narrowing the discussion further to looking at valuation studies of
recycling program attributes, there is research on drop-off recycling programs (Tiller,
Jakus, and Park 1997), and curbside attributes (Caplan, Grijalva, Jakus 2002; Aadland,
Capland 2003).

Using methods similar to these latter studies, in this research the Attribute Based
Method (ABM) was used to identify the relative importance that different attributes have
on people’s preferences for alternative recycling programs. Using ABM allowed us to
estimate the preferences people have for alternative programs (combinations of program
attributes) that the university could implement. We included a cost constraint to allow for

a Marginal Rate of Substitution analysis, which will be discussed further later in the

papen

36

Attribute based methods of analysis seek to understand consumer choices in terms
of the attributes (or characteristics) of the alternatives that are chosen by individuals.
That is, goods or programs are seen as consisting of bundles of attributes that combine to
produce consumers’ utility (Louviere et al. 2000). The Attribute Based Method estimates
economic values for a technically divisible set of attributes, and the inclusion of price as
an attribute in the choice sets provides a constraint to the choice. Eliciting and
quantifying these trade-offs is one way our study differs from what other campus
recycling literature offers where the examinations tend to address behavioral and

attitudinal research.

3.3 Theory

Most stated preference studies and ABM’s fall under the broader umbrella of
Random Utility Models (RUM). Starting with a theory that builds on utility theory, “with
utility a fimction of product attributes” (Cropper et al. 1993), we can begin building a
model of preferences. The goods, or in our case bundles, are of use either individually or
in combinations that combine to produce the consumer’s utility (Lourviere et al. 2000).

The basic formulation of a RUM is: U in = Vin + sin where U in is the utility that person n
associates with option i; Vin is the explainable portion of utility that is estimated from
observable behavior and choices. While sin is the random and unobservable portion of
utility.

In our choice experiment individual n faced a choice of bundle A or bundle B.

Each respondent faced three experiments, resulting in 3 binary choices. The probability

(7!) that the individual will choose alternative i over alternative k is equal or greater than

37

the probability that the utility gained is greater than the alternative. Hence, the probability
of choosing i is:
nn(i)=Prob,-{Vni+£ni2Vnk+£nk;i¢k,VkeC}. (1)

1f the random terms are assumed to be independently distributed Type-I extreme
value variates; then we can estimate these probabilities with the conditional logit model.
When we then substitute the attributes associated with our deterministic component of
utility (V), and following Boxall and Adamowicz (2002), select a linear functional form

allowing the choice probabilities to take the form:

exp(#/3X i)
ZkEC exp(,u,8Xk)

 

(2)

an“):

where ,u is a scale parameter assumed to be 1, and ,6 is a vector of parameters. In this
model, ,8 is common for all individuals (Boxall and Adamowicz 2002).

With a logit or probit model the individual’s characteristics cannot be measured
since they do not vary across the choices the individual makes. “However, individual
specific variables can interact with altemative-specific attributes to provide some
identification of attribute parameter differences in response to changes in individual
factors” (Holmes and Adamowicz 2003). This requires the researcher to make a possibly
strong assumption about segment membership, and have the means to control for that. In
our study we examined the effect that being a student, faculty, or staff member of the
campus community has on choice preferences, since we have information about each

respondent’s status on campus. This is modeled in equations 3.

38

eXIX/$110 )

 

 

 

- z , 3.1
701(1) ZkeC eXp(#/31X k) ( )
. = eXP(#.32Xi) 3 2
7TH“) Zkecexp(#fl2Xk), ( ' )
. z CXP(#,33Xi) 3 3
fin“) ZkeCeprJfl3Xk) 9 ( ' )

 

ZkecexmzflrdiX k wflzde k +flfl3d3X k) ’

The above equations with ,8] , ,62, [33, representing students (3.1), faculty (3.2), and staff

(3.2) respectively, are the models that will be used in our regression analysis. Equation

3.4 represents the interaction effects where d1,d2,d3, are dummy variables for the

respondent’s status on campus, coded in the standard way as a one if they are the
respective status, zero otherwise. This allows us to explicitly segment the campus
population using the status on campus as the controlling variable. This is possible since
we have a strong a priori assumption about what the segmentation of the campus is, and
because we can easily determine which status each respondent had in answering the
survey.

The above equations allows the researcher to segment the population, one could
say this is an example of manually segmenting population. There are times however
where that might not be feasible, or the researcher may not wish to be quite so heavy

handed in his or her approach to segmenting the population. In such cases the latent class

39

model can be of more use as it allows the researcher to take a more hands-off approach to
population segmentation.

With that in mind, we turn to the latent class models where we assume that there
exists within our population a certain number of S (s=l . . .8) segments. Further, individual
n belongs to segment 5. This allows us to express the utility function

as Vnils = ,BSXm- + gnils . With this, we can then make the utility parameters segment

specific, and rewrite equation (2)

 

- : exPUlsflin) 4
”n|s(l) ZkeCeXPfflsfls/Yk) ()

with ,8, segment specific and ,u, a scale parameter (Boxall and Adomwicz 2002).

Following Boxall and Adomwicz (2002), we consider an unobservable or latent
membership function M* classifying individuals into one of the S segments For any one

individual n, this function can be described by the following set of equations:

# *
Mns zrpsPn +rsSn+§ns

,, (5)
P” = ,6an + gm,

where M* is the membership likelihood function for n and segment 5. P; is a vector of
latent psychometric constructs held by n. Sn is a vector of observed sociodemographic
characteristics of individual n. Pn is a vector of observed indicators of latent
psychometric constructs held by n. F and ,8 p represent parameter vectors to be estimated.

5 representing the error for both equations. Following the Boxall and Adamowicz (2002)

methodology, we can then relate this function to the classical latent variables approach

40

where observed variables are related to the latent variable, M* is then represented at the

individual level as:

M;S=25Zn+§ns, s=1,...,S (6)
where Z n is a vector of both the psychometric constructs (Pu) and the sociodemographic
characteristics (Sn) , and AS a vector of parameters.

Swait (1994), points out that these membership functions are random and one
must specify the distribution of their error terms in order to use them in practice. Given
that we followed Kamakura and Russel (1989), Gupta and Chintagunta (1993), Swait
(1994), and Boxall and Adamowicz (2002), all point out that the error terms are assumed
to be independently distributed across individuals and segments with Type I extreme
value distribution and scale factor a . Incorporating these assumptions allows the
probability of membership in segment s to be characterized by equation:

CXPfa’lsZn)
fins 2 S ,
25:]exp(a/tszn)

 

(7)

Schmidt and Strauss (1975) developed a multinomial logit model in which probabilities
are explained by individual-specific characteristics rather than attributes of choices. This

is important since in our latent class modeling the ,6]- changes, while the X,- does not

change across the respondents. There are other functional forms that could be used, but
regardless of form chosen, 25:17am must equal 1, andO s fins 51 . Defining the

probability fl'ns (i) as the joint probability that individual n belongs to segment 5 and

chooses altemative i, this can then be expressed as the product of the probabilities

41

defined in equations (3) and (6): fins (i) = ”n3”n|s (i). Thus, the probability that a

randomly chosen individual n chooses alternative i is given by:
S
fin“): Zflnsfl'nlsa) (8)
s=1
and substituting the equations for the choice equation (4) and class membership equation

(7) probabilities provides the expression:

”n (i) = % gxlxa/lszn) [ CXP(#s,Bin) ] (9).
s=1 Zs=lexp(a/lsZn) ZkecexPlflsflst)

This model, “allows choice attribute data and individual consumer characteristics to

simultaneously explain choice behavior” (Boxall and Adamowicz 2002).

As discussed in the next section, in our survey data, each individual responds to
more than one stated preference question. Since it is likely that there are unobserved
characteristics specific to that individual that induce correlation across his or her
responses, estimation uses the random effects logit model (W ooldridge 2002). In random
effects models, the error term is treated as two separate components. One component is
the unobservable portion that is unique to the individual. The second component is the
random shocks across all individuals and all responses (Boxall et al. 2003). This utility

model can be written as:

Auij=flsAXj+vi+eij, (10)
where V) is the individual specific error term, eij is the random disturbance term across

all individuals and observations, and ,BS is the parameter vector for market segments,

and AX 1 represents the change in the program attributes that a respondent sees. For our

42

first model, the interaction effects model, the random effects specification was used, but

our second method, the latent class models, did not use random effects.

3.4 Data Collection

As mentioned, a choice experiment survey was conducted that elicited
preferences over possible recycling programs. The survey contained several sections
including background information and questions about MSU’s plans to develop a new
recycling program. In the survey, each respondent was presented with two possible
recycling programs and asked to select their preferred program. The recycling programs
were described by several attributes and each program had a cost associated with it.
Adequately describing the program and the attributes was crucial to ensure that when
each respondent made their choices, each person has in mind the same type of item. It is
well-known in the literature that stated choice surveys should be as realistic as possible so
that the respondents are able to provide accurate, realistic answers (Carson et al. 2001).
As such, this section reviews the efforts that went into the attribute selection and the way
that attributes were conveyed to respondents.

The survey instrument was developed in three phases that centered on focus group
studies, survey design, and pretesting/revision of the survey (Kaplowitz et al. 2004). A
series of key informant interviews were conducted with university administrators,
consulting engineers, faculty active in university environmental affairs, as well as student
activists. These were conducted so that the survey targeted appropriate recycling program
attributes and levels as well as to design an instrument that adequately informed

respondents about the various recycling program attributes and the possible levels for

43

each attribute. The focus groups were qualitative studies conducted to gauge the
recycling knowledge, behaviors, and attitudes of MSU. The focus group results then
guided subsequent interviews and the design and implementation of the survey itself
(Hansen et al. 2007).

The result of our focus groups and interview efforts was the identification of five
program attributes with various possible levels for each attribute. The program attributes
consisted of the materials to be recycled (paper, cardboard, non-deposit containers, and
deposit containersz), the type of recycling container individuals would be provided, the
location where recyclables would be taken, the amount of waste stream separation
required by individuals, and the amount the program would cost the institution. (See

Table 3.1 and discussion below).

 

2 Michigan has a deposit on certain types of beverage containers.

44

Table 3.1 RecyclingProgram Attributes

 

Attributes Attribute Levels

 

Materials Paper, Non Deposit Containers, Deposit Containers, Cardboard,
(Baseline)
Paper, Non Deposit Containers, Deposit Containers
Paper, Cardboard, Non Deposit Containers
Paper, Non Deposit Containers
Paper, Deposit Containers, Cardboard
Paper, Deposit Containers
Paper, Cardboard

Location Floor/Classroom (Baseline)
Building
Drop-off
Central

Containers No Bin (Baseline)
One large bin
Bag
Multiple small bins

Preparation One stream (Baseline)
Multiple streams
Complete separation

Cost $50
$150
$300
$500

 

In the survey, respondents were told that MSU was designing a recycling program
and that their feedback was needed on the type of program to be implemented. After
informing the respondents about all of the possible program attributes and their levels, we
then presented the three pair-wise program choice experiment questions (see figure 3.1

for an example of one of the choice pairs).

45

Note that there are many possible combinations of all the attributes, yet each
respondent only saw three of the many possible pairs. The attributes pairs were then
varied across respondents following an experimental design. The experimental design
consisted of a mix of pairs from a main-effects design for these attribute levels and
random pairing of the attribute levels By varying the attribute pairs across respondents,
the statistical techniques outlined above can be used to identify the effect that each
attribute level has on the probability that an altemative is preferred.

Since the respondents were asked to make choices under a constraint, cost, it was
crucial that we explain how program costs would be paid. The program costs were costs
to the institution, and the construction costs would be paid for through a capital bond
mechanism, as was explained to the respondents. Further, respondents were told that
when the costs of the program were incurred, it would leave less money for the institution
to use in other areas. In the choice questions, the cost to the institution was expressed in
per-person terms. Hence, the cost associated with the program is not a direct cost to the
respondent, rather it is an opportunity cost to the campus — as more money is spent on the

recycling program, less money is available to spend on other campus items.

46

Fig 3.1 Example of a choice experiment question a respondent may have seen
WHICH OF THESE TWO PROGRAMS IS BETTER?

Compare the characteristics of Program A and Program B and choose the program that you
prefer for recycling at MSU.

Michigan State University Recycling Program
Characteristics

 

Program

 

Characteristics
(Click hyperlink for A B
more information)

 

Recyclable Materials

 

 

 

 

 

Paper, Non-deposit Paper, Non-deposit
Megappfggfzmhe containers containers
Material Se arati on All non-paper materials Group similar materials
[2 . . . .

at collection site mixed together In one together In collection

container containers
ggilgiiblgrmlafe‘iglgt‘og Remote collection Central collection
W Single large basket Multiple smaller baskets

for your room/office

 

MSU Cost (per person)
(to build facility. collect .

and process materials. $150 $150
and operate program)

 

 

 

 

 

32. Which of the above programs do you think is better? (Pick One Response)
(T'Program A C) Program B

 

Back Next t:::tf_‘;>

47

3.4 Interaction Effects Model

We tested whether we can pool data from the three populations. In examining the
pooled model with a simple random effects logit, we determined that three types of
person on campus: students, faculty, and staff have different utility functions. Performing

the log likelihood ratio tests allowed us to reject the pooled sample versus the restricted

samples. We tested this with a log likelihood ratio test which is: — 2(LL, - Z LL,)which
follows a 12 distribution with K(M-l) degrees of freedom (Wooldridge 2002). LL j is the

unrestricted pooled sample log likelihood value, and LLiare the log likelihood values for

the separate student, faculty and staff models. K is the number of restrictions in the
model, while M is the number of model treatments. For this test, we had 16 restrictions
on the models, with 3 models being tested, resulting in 32 degrees of freedom for the
pooled sample. We also tested each of the possible combinations of faculty, staff and
students against each other to see if we could pool together data for any of these
subsamples.

Table 3.2 also shows that we can with confidence reject the hypothesis that the
pooled sample is equal to the sub-samples. However, for the combined sub samples, we
can only say with 99% confidence that the students are different than staff or faculty. For
staff and faculty we cannot only reject the hypothesis that they are the same at the 90%
confidence level. In the results that follow, we report on the pooled model, and the
interaction effects model that separates students from the combined group of faculty and

staff.

48

Table 3.2 Tests on the Interaction Effects Models

 

 

Degrees
of LR Test
LL Freedom Value Significance Level

Pooled (LL!) 4479,52
Student (LL,) -2726.36
Staff(LL,.) -616.57
Faculty(LL.) -1100.30 32 72.58 0.00005
Faculty+StafflLle -1724.10
Faculty(LL,) -1 100.30
Stafflll.) -616.57 16 14.46 0.56449
Faculty+Student(LL,-l -3846.39
Faculty (L14) -1 100.30
Student(LL,.) -2726.36 16 39.46 0.000093
Student-l-Staff(LLj ) -3 364.76
Student(LL,) -272636
Staff(LL,.) -6l6.57 16 43.66 0.000022

 

LL=Log Likelihood, LR= Likelihood Ratio

Looking at the interaction effects coefficients in table 3.3, we find that almost all
of the attributes are significant except for the waste-stream separation variables for
students. For faculty and staff all the variables except for multiple stream separation were
significant. While at the combined, unrestricted population level, the coefficient and p-
value are so weak that preferences between the baseline of total separation and multiple
stream separation were not significantly distinguishable. The coefficients sign and size
also provide information about the ordinal ranking of the different attributes compared to

the baseline attribute comparison. Negative signs indicate that that attribute level is less

49

preferred, while a positive sign indicates that that level is more preferred to the baseline
level. The coefficients indicate that people want to recycle more materials, and that they
would like more decentralization of location, and some type of bin for recycling.
However, since we cannot compare coefficients across logit models, we need to
determine a marginal rate of substitution in order to gauge the magnitude of for one
attribute over another as revealed by the various models. This will be discussed in a latter

section of the paper.

50

Rafi—Eco :momon— .. :oZ 1!. UQZ £35950 Eamon— " DO

 

:dd nod Ndvd ddd vwdd mod 23

dddd dvd dwvd odd modd wmd D «Sufi

dddd dd.m- dddd odd. dddd odd- $83 as .80
dddd 9d. ded d _ d- dddd nNd- cough—om oZ
Nam—d 2d- hmmd Ed Edd odd 332% 23:32
35338 =e 883%

539.30.:—
dddd end dodd 3d dddd odd 23:32
dddd tad dddd :d dddd Gd mam
dddd and dddd Nwd dddd and 0:0
Adam—03m: :3 oz

E05350
dddd ow. T dddd dv. T dddd mm. T 3.550
dddd mm. T dddd m _. T dddd mm. T modem
Edd mod- medd w_d- «mod Ed- wEEEm
AoE—omamv Became—380E

553.5
dddd ST dddd do. T dddd 8. T “Shoppes modem
dddd 2T dddd mad- dddd odd- 0Q conun—
dddd mud- dddd nmd- dddd mod- nceoadao .UQ comma
dddd and- dddd and- dddd Rd- UDZ 59$
mood med- owed Rd- dddd dad- demon—END .UQZ modem
dddd dud- dddd 3d- dddd mmd- 0Q .UDZ con—em
cease @826 do .82 8%;

Batons—2

 

51

83¢ .aou
.esm Ea page

033i «000
ficovam

0=_fl>rnm HOOD
.682 Base

 

£2.32 Book—m 5:228:— 05 59c 83:.th none—52.5 ad 935,—.

3.5 Latent Class Model

Given previous research regarding what influences choices regarding recycling,
both in the general population and on a college campus, a set of questions were
developed for the survey to gauge the attitudes and knowledge regarding recycling and
the environment on campus. These included questions related to attitudes about
recycling, the time involved with recycling, and general recycling knowledge
information. We then used responses to the questions to test for their effect in segmenting
the population.

Using the LIMDEP 9.0 software regression package, we constructed a latent class
model using 5 of the additional questions from the survey. The variable questions we
tested are in table 3.3. These questions were chosen since it was felt that they would
provide good indicators of how people feel regarding recycling and the amount of time or

effort required.

52

 

 

Aoflzcofio d . _n conEoE tau—sound Egan new user? x8830 the—sown
Adm—3.8 do . Us 82: S v S .8 o 23> >88: 8 8:585
. a d_ a bmbm «E. D Em Encamvgtfim
EwE 8 Bo. 638 tax: m 3 _ a me 388 amDmE swan—Bog
an @2982 on BE 35 30365 comma no omen. 2: Sons 30% Pa cinemas—305— ROI:
2 u 232 d u 0:253 8200 8250
Adno: 4.1.8.0 ..:o_§=oa BEEP—3:0 o «co
E395 8 wow: on 2:03 >258 of 55.3 mm? _ a 0885 b: .«o ten 03» 2:03 r. H D
800: . W85 aw==932 mo Scones 2: Sons 20E Emu. 8 8:— 2:03 r. Bmoeom
826030
80o: :qu43 :sz an 2982 8 note was 08: do do. a moo—S : 483% 5.. “EC. oEafionoQ
£532.82: 3.20 com 833:? ha $305 8: Ed 204 05 .«o cote—Etc.“ as; 252 333:3 oz
252
3:33:32 Eon—wow :o 3:25:— 05833: 8 com: 833...; .3 553.88: 833:3 Tue—z

 

2252 820 E03..— 2: E 688,—. mo_na_._a> .3 Bow 8...; 2:. 1m 03:.

53

Using the LIMDEP 9.0 software, we were able to arrive at the results in Table 3.5
examining the latent class tests. There is no set test for determining the exact number of
segments. Using the Akaike Information Criterion (AIC) and the Bayesian Information
Criterion (BIC) as guides provides some help, however “conventional rules for this
purpose do not exist and judgment and simplicity play a role in the final selection of the
size of S” (Boxall and Adamowicz 2002). So following the suggestion that choosing the
segment where we balance a decrease in the AIC begins with an incease in the BIC, we
suspect that the optimum number of segments that the campus can be divided into is 2.
This concurs with what our log likelihood ratio tests suggested when testing the
interaction effects. Though it could be that there are a different number of segments, for
our study we are going to examine more closely only the two segment case.

Table 3.10 in the appendix examines what effect being a faculty or staff member has
on segment membership for the two segment population for the LCM with Faculty and
Staff as the class membership variables. In the table, the First constant is the baseline
measurement, so everything is measured against that level. Hence, the average person is
less likely to fall into that category than the person into the second. This probability can
also be seen at the bottom of Table 3.5. From this we can learn the probability of
membership in the first instead of the second segment of the population is 0.21 for faculty

members, 0.11 for staff members and a 0.32 for students.

54

Table 3.5 AIC and BIC Results for Each of the LCM Models that Converged

 

 

LCM Latent Class Explanatory Variables Number Of AIC BIC
Segments Parameters
1 No Variables 15 1.028 1.038
1 Demographic Questions 15 1.028 1.038
1 Staff and Faculty Dummies 15 1.028 1.038
2 No Variables 31 1.016 1.035
2 Staff and Faculty Dummies 33 1.013 1.034
3 No Variables 47 1.005 1.035
3 Demographic Questions 65 1.006 1.046
4 Demographic Questions 87 0.998 1.052

 

Only those models that created a Hessian - Positive Definite solution are reported.

55

E05380 :monoa l :02 u UQZ £05850 5039 N GO

 

 

56

is N; eNd NNs £83862 Epsom .8 sea
N oN.NN_- N 8.x:- N 2:. N _N.o- cos-sea oz
N 8.2- N 8.8- d NNNN _ SNN Sauce. 63:32
_ _ N N 95388 as osaaaom
secs-Ego...—
_ oNNNN _ 3N8 N EN N _ SN 2832
N 8.va N NE: N 2.2 N :.NN 8m
N 2.3m N SEN _ :.NN _ 8.? 6.5
a v v v 95338 :3 oz
Pagan—SD
a 8.82- 4 N32:- 4 3:- 4 2.8- .850
N 3.22. N $.82- N 8.3-- N 3:.- E685
N NE _- N N2: _- N 2% _- N SN. 8:55
_ _ — fl Aoczogmv ECO-amm—oioo—m
ficzaocfl
N answ- N NSNN- N. N58- N NNS- 28928 .63..
e Names- e NNNON- v 3.9.. 4 EN? on .693
a NNNS. - 8.2:.- N 2. d T N 2.2- eaofiau .8 59$
N SN:- N 3.4%- N NNNY N 8.2.. 82 5%..
N N32- N Nan:- N NNNN- N NNNN- 63828 .82 59$
N NENN- N :.NNN- N. 2 . d N- e :.NN- do .82 .693
_ _ _ _ cease encased do .82 see
$3.532
seam mm: seam mm: #52 mm: 622 mm:
BEE-=3 oZ 3:555 £38m a. tSm 33253 02 BEE—=9 Egan an team

 

 

EoEmom vacuum EoEuom SEE

 

ease» assets,
2.. 2: E: 8E9» assess new .5- £55... 2: .8 ass: :20 2.35 ESPN-65. 5 mm: .3. 63.;

38%-ED u mu £35850 :Noaom - :02 n 092 £558.80 :Nomon u on

 

 

N £51 N NE:- N NN. a- Sage-N oz
_ NNNN N NNNN- N NN. _- Nashua 2332
N N _ 35338 =N 388%
5:33on
_ NNNNN N SNNN N 3N3 29:32
N 322 N N32 N N3“: Nam
N NoéN _ NNSN _ Nfl-NN 2.0
v v v 65338 Em oz
gunman-EU
v N.NNN- N NN.N$. v _N.N:.. 550
N _ 2: N- N NNNNN- N 83 N- @985
N NN.:- N NNN- N NN. :1 N523
_ _ _ Auczommmv Eco-amu—Utoo—m
5.3603
N NNNNN- N NNNNN- N NNNNN- N6 59$
N NSNNN- N NN.NN N- N :.NNN- on .59».
v NNeNT N N38- v 3.2:- mu do 59$
N 3+2- N NNNoN- N NodNN- 092 59$
N NNNN- N NEE- N :.3. mo .82 89$
N NNNNT N N_.NNN- N 31-2- 8 .82 5%;
N _ _ saga mu 8.82 59$
23.852
xfim Na: NEE mm: gm mm:
35me £33 USN tSm 330N—

 

NESS 88.5— 32328:— 2: 8.5 5.52 32:53- una 553535 no 829— _a=_u..a2 Nun «Ea-H

57

985850 zmomom I :02 n UDZ nNNNNNNNSNNNNU :8an U DO

 

N NN.NNN- N 2.2:. N NN.NNN- 8:888 02
N NN.NN- N NNN N- N NN.NN- N983 29%:
N N N NNNNNNNNNNNN NNN 988%
.3333on
N NN.NNN N NN.NNN. N NN.NNN. 29:92
N NN.NNN N NN.NNN N NN.NNN NNNN
N NN.NNN N NN.NNN- N NN.NNN- 25
N N. N. NNNNNNNNNNN NNNNN oz
gunman-EU
N. NN.NNN- N. NN.NNN- N. 3. NN-N- N888
N NN.NNNw N NN.NNN- N NN.NNN- No-95
N NN.NN- N NN.NN- N 2.5. NNNNNNNNN
N N _ 35335 8005929805
Home—39H
N N NNNN- N NN.NNN- N NN.NNN- 289:8 59$
N NN.NNN- N NN.NNN- N NN.NNN- on 59$
N- NN.NNN- N. NN.NNN- N. NN.NNN- 28550 do 59$
N NN. :N. N NN.NNN- N NN.NNN- 092 59$
N NN.NNN- N NN.NNN- N 8.9;- 22528 .82 59$
N NN.NNN- N NN.NNN- N NN.NNN- on .82 59$
N N N 99.3.va 28550 8.092 59$
23.532

 

viny— mmz owNNNo>< v.5 mmE omfio><
NNNNNNNN> oz 23 9955 23

x53 mm: owfiu><
flout-m 532285

 

NN.—«Ewom mmo-na tows-33‘ acme—Nzumaam .3 more REM-ES .w.m «Bah.

58

For each population segment, the individual Marginal Rates of Substitutions (MRS)
of an attribute level for program costs were calculated by dividing the coefficient on the

attribute level by the negative of the cost coefficient (a /— ,8). The averaged MRS

measures were calculated by multiplying the probability of being in class i by that

attributes individual marginal rate of substitution plus the other segments probability and
marginal rate of substitution, i.e., Average MRS = Z (Pr Class _i* induvidual _i MRS).

Tables 3.6 and 3.7 show the ordinal ranks of the various attributes and the individual
MRS measures controlling for staff and faculty through dummy variables. What we can
see in table 3.5 is that there is a segment of the campus that is very price sensitive, and a
segment that is not. The first segment is highly sensitive to the opportunity cost to the
campus of this new program. While the second segment of the campus is more willing to
spend the universities money. The magnitude of the values helps to provide an idea of
how much each segment prefers one attribute compared to another level. The MRS
values help us to get a sense of how much more an attribute is preferred. For instance, we
can see with location that not only do campus members prefer having a decentralized
location, but that they would prefer a building location at more than three times the
amount compared to a drop-off location option.

In addition, the Marginal Rates of Substitution (MRS) measures for the Interaction
Effects Models for Students, and the Faculty plus Staff were closer together than with the
two segments identified by the LCM estimates. Since the modeling approaches all use the
same choice data, we might expect that the MRS measures be somewhat similar, or at

least not as dissimilar those provided in the LCM estimates. This occurred for both the

59

model with faculty and staff dummy variable covariates, and for the model with no
covariates Afier combining the model estimates (Table 3.7) we see that the preference
rankings again are consistent, but the LCM again provides higher estimates of MRS than
the interaction effects model provides.

3.5 Discussion

An implication of this research is that there is strong support for the campus
population having at least two preference segments. This is made clear from the Log
Likelihood tests and from the results of the LCM. The interaction effects model suggests
that the two segments are students and a combination of faculty and staff. The latent class
model also suggests at least two segments, one segment that is highly price sensitive and
one that is less price sensitive.

The interaction effects and latent class models both provided identical preference
orderings when averaged across the segments. Both predicted that people largely want to
be able to recycle more things, at more decentralized places. A bag or bin is preferred to
none. Cost was always negative and highly significant. People consider deposit
containers the first item to be given up from the materials accepted, which makes
intuitive sense as they retain value after use.

We had a strong a priori idea of how the segments would break down. Interaction
effects also allowed for a Log Likelihood test of our hypothesis of different preferences
between students, faculty and staff, which confirmed out a priori assumptions. While the
LCM did provide some useful information regarding population segmentation, there is
yet to be developed a test of what is the correct number of classes as determined by the

LCM which is a weakness.

6O

The also found that in our case use of the attitudinal and other demographic data did
little to advance what we already knew regarding recycling preferences and MRS. If
anything, attitudinal questions performed worse than controlling for the segment

membership directly through faculty and staff dummy variables

3.6 Conclusion

This paper was motivated by the desire to examine the relative performance
capabilities of the latent class model and the interaction model. The advantage of the
LCM is that when the researcher does not have strong a priori knowledge of preference
heterogeneity, the model does impose potentially unwarranted assumptions onto the data.
However, when the researcher does have a good understanding of probable sources of
heterogeneity, then it may be worthwhile for the researcher to use that information and
use the simpler interaction effects model.

Other research using the LCM has shown that it does have strong potential for
understanding of preference segments. Our results were less clear cut. All the models we
examined provided the same averaged rankings of the attributes being studies, though the
rankings did differ within some subpopulations. It may be that LCM is of more use in
cases where the preferences are more heterogeneous whereas the findings from this study
indicated strong uniformity in rankings. Even though in our case both methods would
provide the policy maker with same ordinal preferences, the magnitude of the MRS
differed, and perhaps that is the cautionary tale of the two methods.

Future research could examine the out of sample predictive qualities that the two

models provide as our sample size is large enough to examine that question. Another area

61

would be to further explore what the three and four segment models look like and could
imply. Another area of furthering the research would be to rework the models so that the
latent class model is also a random effects model, allowing a closer comparison of the
two models. Finally, the possibility of examining a random parameters regression on the

data set might also provide insights.

62

£05850 :Nogoa - co Z n UQZ £05850 :NoN—oa n 00

 

 

oood oo.T oood oodm- oood oo.T oood Zimm- ANooS N5 .80
:od _Nd- Nmod de- mood vmd- mwod 5.? 5:838 oZ
Svd mod- vmod mwd mwud wod- hood mod N885 oat—=2
9:23.va =N SEQ—om

.8333on
oood 5d Nood ood oood owd oood :.o 22:32
oood mud on _ .o de oood mod wmod de mam
oood mwd oood no; oood Nwd oood mo; 0:0
NNNNNNNNNNV :5 oz

€055.50
oood hm. T oood vw. T oood hm. T oood ow. T 35:00
oood on. T oNod m:- oood Nm. T mood :.T woo-acum—
omod Ed- mmmd ovd- Nmod 2d- _o_d mmd- mama—Em
95—on9 EooNNNBQNNooE

5:325
oood o:- oood omT oood ST oood ooT Emonoso 59$
oood ST wood voT oood m _.T mood NoT OD comma
oood :.d- momd omd- oood ood- ommd mvd- annEaU .UQ comma
oood ond- oood NN. T oood whd- oood on UQZ comma
o5d omd- NNod owd- wood hmd- omod end. 283an .82 dogma
oood wed- oood mm. T oood wvd- oood va OD .UQZ comma—
NNNNNNNNNNN NaNNNao do .82 5%;

flaw—8a:

o2m>d .toou 023d .boou 023d E80 3.3-N— .hoou
N Eofiwom _ EoEwom N :58me _ Eofiwom
204 Nozntm> oz 204 NBEENNQ 338-.— EB tSm

 

NEE—z man—U 2.8a.— EoEwom-NEH 2: .3.— Nouafitmm E39530 dd 03:.

.< 5259?.

63

Table 3.10 Coefficients on Probability of Membership in Segment One
fi'om the Two Segment Staff and Faculty Dummies LCM

 

 

Coefficient P-Value
Constant -0.76 0.000
Faculty_Dummy -0.58 0.000
Staff_Dummy -l .34 0.000

 

64

References

Aadland, Capland. 2003. Willingness to Pay for Curbside Recycling with Detection and
Mitigation of Hypothetical Bias. American Journal Agricultural Economics, Vol
85 No (2) pp. 492-502.

Adamowicz, W.L., and Boxall, RC, 2001. Future directions of stated choice methods for
environment valuation. Paper read at Choice Experiments: A New Approach to
Environmental Valuation, April 10, 2001, at London, England.

Boxall, RC, and Adamowicz, W.L., 2002. Understanding Heterogeneous Preferences in
Random Utility Models: A Latent Class Approach. Environmental and Resource
Economics. Vol.23. pp. 421-446.

Boxall, P.C., Englin, J ., & Adamowiczc, W.L. 2003. Valuing aboriginal artifacts: a
Combined revealed-stated preference approach Journal of Environmental
Economics and Management. Vol. 45, pp. 213-230.

Caplan R.J., Grijalva, T.C., Jakus, RM. 2002. Waste not or want not? A Contingent
ranking analysis of curbside waste disposal options. Ecological Economics.Vol
43, No. 2-3 pp. 185-197.

Carson, R.T., N.E. Flores and NP. Meade. 2001. Contingent valuation: Controversies
and Evidence. Environmental and Resource Economics. vol. 19 pp.l73-210.

Choc, C. and Fraser, Iain. 1999. An Economic Analysis of Household Waste
ManagementJournal of Environmental Economics and Management.Vol 38. No
2. pp 234-246.

Cropper, M. l., Deck, L., Kishor, N., and McConnel, K.E., 1993.Valuing Product
Attributes Using Single Market Data: A Comparison of Hedonic and Discrete
Choice Approaches. The Review of Economics and Statistics, Vol. 75, No 2. pp.
225-232.

Ewert, A. and D. Baker. 2001. Standing for Where You Sit: An Exploratory Analysis of
the Relationship between Academic Major and Environment Beliefs. Environment
and Behavior. Vol. 33. No.5.pp. 687-707.

Folz, D.H., Hazlett, J .M. 1991. Public Participation and Recycling Performance:
Explaining Program Success. Public Administration Review. Vol 5 1. No.6. pp
526-533.

Gamba, R., & Oskamp, S. 1994. Factors influencing community residents ’ participation
In Commingled curbside recycling programs. Environment and Behavior. Vol. 6.
Pp. 587- 612.

65

Greene, W.H. 2003. Econometric Analysis Fifth Edition. Upper Saddle River, New
Jersey. Pretince Hall.

Greene, W.H., & Hensher DA. 2003. A latent class model for discrete choice analysis:
Contrasts with mixed logit. Transportation Research Part B 37, 681-698.

Gupta, S., & Chintagunta, PK. 1994. On Using Demographic Variables to Determine
Segment Membership in Logit Mixture Models. Journal of Marketing Research.
Vol.31. No.1 pp. 128-136.

Hansen, L.T. McMellen, C. Olson, L. 2007.Recycling Attitudes andBehaviors on a
college campus: Use of Qualitative Methodology in a Mixed-Methods Stuafv.
Working Paper, MSU.

Holmes, T.P., and Adomowicz, W.L. 2003. Attribute-Based Methods. In: Champ, P.A.,
Boyle,K.J., Brown, T.C., (eds.) A Primer on Nonmarket Valuation. Kluwer
Academic Publishers: Dordrect/Boston/London pp. 171-220.

Jennings, M. E. 2004. An Instrument to Measure the Recycling Attitudes and Beliefs of
Undergraduate Students at a Large Northeastern University. Online Submission,
Education Resources Information Center: 22.

Kamakura, W.A., & Russell, GJ. 1989. A Probabilistic Choice Model for Market
Segmentation and Elasticity Structure. Journal of Marketing Research. Vol.26
No.4. pp. 379-390.

Kaplowitz, M.D., Hadlock, T.D., Levine, R. 2004. A Comparision of Web and Mail
Survey Response Rates. Public Opinion Quarterly. Vol. 68 No.1 pp. 94-101..

Kelly, T. C., I. G. Mason, M. W. Leiss and S. Ganesh. 2006.University community
responses to on-campus resource recycling. Resources, Conservation and
Recycling. Vol. 47 . No. 1. pp. 42-55.

Louviere, J .J ., D.A. Hensher, and J .D. Swait. 2000. Stated Choice Methods: Analysis and
Application. Cambridge: Cambridge University Press.

Louviere, J ., Street, D., Carson, R., Ainslie, A., Deshazo, J .R., Cameron, T., Henseher,
D., Kohn, R., & Marley, T. 2002. Dissecting the Random Component of Utility.
Marketing Letters. Vol.13 No.3. pp. 177-193.

Morey, E., Thacher, J ., & Breffle, W. 2006. Using Angler Characteristics and Attitudinal
Data to Identify Environmental Preference Classes: A Latent-C lass Model.
Environmental & Resource Economics. Vol.34. pp. 91-1 15.

Oskamp, S., Harrington, M., Edwards, J ., Sherwood, D., Okuda, S., & Swanson, D. 1991.
Factors influencing household recycling behavior. Environment and Behavior,

66

v0]. 239 pp- 494'519.

Provencher, B., and Bishop, RC. 2004. Does accounting for preference heterogeneity
improvethe forecasting of a random utility model? A case study. Journal of
Environmental Economics and Management. Vol. 48. pp. 793-810.

Provencher, B. and Moore, M. 2006. A Discussion of “Using Angler Characteristics and
Attitudinal Data to Identify Environmental Preference Classes: A Latent-C lass
Model. ” Environmental & Resource Economics. Vol. 34. pp. 117-124.

Schmidt, P. and R. Strauss 1975. The Predictions of Occupation Using Multinomial Logit
Models International Economic Review Vol.16, pp 471—486.

Scarpa, R., & Thiene, M., 2005. Destination Choice Models for Rock Climbing in the
NortheasternAlps: A Latent-C lass Approach Based on Intensity of Preferences.
Land Economics. Vol.8]. No. 3. pp. 426-444.

Siikamaki, J ., & Layton, D. 2007. Discrete choice survey experiments: A comparison
using flexible methods. Journal of Environmental Economics and Managemem.
Vol. 53. No. 1. pp. 123-139.

Swait, J. R. 1994. A Structural Equation Model of Latent Segmentation and Product
Choice For Cross-Sectional Revealed Preference Choice Data, Journal of
Retailing and Consumer Services. vol.1, pp. 77—89.

Tiller, K.M, Jakus, P.M., and Park, W.M. 1997. Household Willingness to for Drop-off
Recycling. Journal of A gricultural and Resource Economics. Vol 22. No. 2. pp.
310-320.

Wooldridge, J .M. 2002. Econometric Anjalysis of Cross Section fiand Panel Dela.
Cambridge, MA: MIT Press.

67

   

ll"flillillflijlllflllifllfllIlflliiI