Travel Cost Method

.
Chapter 9
THE TRAVEL COST MODEL
George R. Parsons
University of Delaware
1.
INTRODUCTION
The travel cost model is used to value recreational uses of the environment.
For example, it may be used to value the recreation loss associated with a beach
closure due an oil spill or to value the recreation gain associated with improved
water quality on a river. The model is commonly applied in benefit-cost
analyses and in natural resource damage assessments where recreation values
play a role. Since the model is based on observed behavior it is used to estimate
use values only.
The travel cost model is a demand-based model for use of a recreation site
or sites. A site might be a river for fishing, a trail for hiking, a park for wildlife
viewing, a beach for swimming, or some other area where outdoor recreation
takes place. It is useful to separate travel cost models by those that estimate
demand for a single site and those that estimate demand for many sites.
Single site models work like conventional downward sloping demand
functions. The quantity demanded for a person is the number of trips taken to
a site in a season and the price is the trip cost of reaching the site. Variation
in price is generated by observing people living at different distances from the
site. Price is low for people near the site and high for those living further away.
The demand function slopes downward if trips decline with distance to the site.
Single site models are useful when the goal is to estimate the total use or
access value of a site. The elimination of a site is the usual application. The
1
Chapter 9
lost value is the total consumer surplus under the single site demand function -the difference between a persons total willingness to pay for trips and the actual
trip cost incurred over a season. Some valuation examples are
a beach closure due to an oil spill,

a fish consumption advisory that closes a lake for fishing, or
a development that eliminates a natural area for wildlife viewing.
It is also possible to use a single site model to estimate the value associated
with a change in the cost of access to a site. For example, an increase in an
entry fee or the opening of a new entrance could be evaluated using a single site
model. More difficult, and really calling for the transfer of a single site model
from one site to another, is valuing the addition of a new site such as reservoir
created by a dam.
There are some variations of the single site model for valuing changes in site
characteristics such as improved water quality on a lake or an increase in the
number of hiking trails in an wilderness area. However, this is not the strength
of the model. When the goal is to value changes in site characteristics at one or
more sites or to value the access to more than one site simultaneously, a
multiple site model is preferred.
The random utility maximization (RUM) model is the most widely used
multiple site model. A RUM model considers an individuals discrete choice of
one recreation site from a set of many possible sites on a single choice occasion
in a season. The choice of site is assumed to depend on the characteristics of the
sites. For example, an individual making a fishing trip may consider trip cost,
catch rate of fish, and site amenities. The choice of site implicitly reveals how
an individual trades off one site characteristic for another. Since trip cost is
always included as one of the characteristics, the model implicitly captures
trade-offs between money and the other characteristics.
Some examples of valuing changes in site characteristics (at one or more
sites) using the RUM model are
C
C
C
C
C
improvements in water quality on lakes and rivers

increases in the catch rate of fish on lakes and rivers
improving the conditions of access to several local parks
increases in moose populations in several hunting areas
increasing the number of mountain biking trails in a state park
The RUM model may also be used to value access to one or more sites
simultaneously. For instance, it may be used to value the loss of several beaches
closed due swimming advisories or to value several ski slopes opened in a
development project.
This chapter is organized into two sections. The first covers the single site
model. The second covers the RUM model.1 Both sections are organized
around a table listing the steps required to estimate a basic version of the model.
The emphasis is placed how one applies a basic modern version of the model
and not on the frontiers of modeling. For more on recent developments see
Herriges and Kling (1999) or Phaneuf and Smith (2002). For a good historical
account of the model which was first applied over 50 years ago, see Ward and
Beal (2000).
1.
THE SINGLE SITE MODEL
This section is divided into four parts: basic model, steps in estimation, an
example, and variations. I will focus on using the single site model to value site
access since that is its strength and most common application.
2.1 Basic Model
The single site model is a demand model for trips to a recreation site by a
person over a season. The quantity demanded is the number of trips a person
takes to the site. The price is the trip cost of reaching the site which includes
a persons travel expenses and time cost necessary to make the trip possible. In
its simplest form the single site model is
(1) r = f ( tcr )
where r is the number of trips taken by a person in a season to the site and tc r
is the trip cost of reaching the site. Like any demand function, one expects a
negative relationship between quantity demanded (trips r) and price (trip cost
tcr). People living closer to the site face a lower cost of reaching the site and,
all else constant, probably take more trips.
Trip costs alone will not explain an individuals demand for recreation trips.
Chapter 9
It will also depend on things like income, age, experience in the recreation
activities available at the site, and proximity to other recreation sites.
Accounting for these factors gives a more realistic demand function with a set
of shifters
(2) r = f ( tcr , tcs , y , z )
where tcs is a vector of trip costs to other recreation sites, y is income, and z is
a vector of demographic variables believed to influence the number of trips.
By incorporating trip cost to other sites, the model now accounts for
substitutes. The tcss are the prices of trips to substitute sites. If a person lives
near a substitute site, the number of trips (r) is likely to decline as the person
substitutes trips to that site and away from the site of interest in the analysis. A
positive coefficient on tcs would pick up this substitution effect. The other
shifters work much as one would expect. A positive coefficient on income, for
example, implies that the number of trips taken increases with income and so
forth.
Figure 1 shows a linear version of equation (2) corresponding to
(3) r = tcr tcr + tcs tcs + y y + z z .
If a person faces a trip cost of tcr0 in this model, he or she takes r0 trips. The
area A is his or her total consumer surplus for trips to the site during the season
-- the difference between total willingness to pay for trips (area A+B) and total
trip cost (area B). This is also called the individuals access value for the site.
If the site were closed for a season, the individual would lose access to the site
and consequently the area A. Notice that access values increase the closer a
person lives to the site.
Figure 1. Access Value in a Linear Single Site Model
A more general expression for consumer surplus or access value, which

applies to any functional form, is the area under the demand curve between an
individuals current price and the choke price. The choke price is the price at
which trips fall to zero in the model. Mathematically the consumer surplus is
(4) w =
tc r choke
f ( tc r , tc s , y , z ) dtc r
tc r 0
where tcr choke is the choke price and tcr 0 is the individuals trip cost.
In application then, one seeks to estimate an equation like (3) using
visitation data to a site. Table 1 shows the form of a typical single site data set -trip count, trip cost, and demographic data for a sample of individuals. The data
are gathered by survey. Once assembled, equation (3) is estimated by
regressing trips (r) on the relevant explanatory variables in the data set (tcr , tcs ,
y, z1, z2). After the model is estimated, the parameters are used to compute
access value (area A) for each individual in the sample. Means are computed
and an estimate is extrapolated to the population. (The model is usually not
linear but the principle is the same regardless of functional form.) The next
section describes these steps in detail. For a derivation of the basic model from
utility theory or from a household production function see Freeman (1993) or
Bockstael (1995).
Chapter 9
Table 1. Typical Form of a Data Set for Estimating a Single Site Model
Number of
Trips to the
Site for the
Season
Trip
Cost to
Site $
Trip Cost
to
Substitute
Site $
Annual
Household
Income
$000
Number
of
Children
(y)
Years
Engaged in
this Form
of
Recreation
(z1)
(r)
(tcr)
(tcs)
45
200
45
17
150
20
100
21
12
65
77
22
98
55
ID
25
(z2)
2.2 Steps in Estimation

The steps are shown in Table 2. I will discuss each of these in a separate
subsection. I will assume that a model is being estimated for the purpose of
valuing site access, that the analyst is using day trip data only, and that the data
are individual-based. This is the most common application of the model.
Valuing quality changes at a site, as already noted, is not the strength of the
single site model and incorporating overnight trips, while possible, is less
common and presents special complications.
Following this section, I will discuss an actual single site application by
Brent Sohngen (2000). I will describe how he handled each step of the analysis.
He estimated a model of beach use for two Lake Erie beaches.
2.2.1 Step 1: Define the Site to be Valued
The analysis begins with a definition of the site to be valued. The site
might be a park, lake, beach, wilderness area, river segment, or some other area
used for outdoor recreation. The boundaries are often easy to delineate. This
is true in the case of a park, lake, or wildlife reserve. In other cases the
delineation is not so clear. A river for fishing or white-water rafting, for

example, calls for the analyst to define the length of the segment to be analyzed.
A site for hunting or ocean fishing usually requires researcher defined
boundaries as well. In these cases, one seeks to define a site broadly enough to
include most, if not all, of the area affected by the policy being analyzed.
Government agencies, park services, and tourist bureaus often have
literature and maps which help in defining a site. Sometimes government
agencies charged with managing natural resource use within a state have their
own definitions for sites, such as wildlife management units. These site
definitions often work for a travel cost model.
Table 2: Steps in Estimating a Single Site Model
Step 1
Define the Site to be Valued
Step 2
Define the Recreation Uses and the Season
Step 3
Develop a Sampling Strategy
Step 4
Specify the Model
Step 5
Decide on the Treatment of Multiple Purpose Trips
Step 6
Design and Implement the Survey
Step 7
Measure Trip Cost
Step 8
Estimate the Model
Step 9
Calculate Access Value
2.2.2 Step 2: Define the Recreation Uses and the Season

In some cases the site will have a single or dominate recreation use. In other
cases there will be multiple uses such as fishing, swimming, boating, and
viewing on a lake. Ideally one would like to include all recreation types and
estimate separate demand functions for each. Sometimes, policies call for
focusing on a single recreation type. Again, government agencies, park
services, and tourist bureaus often have information and data to help identify
major uses.
If recreation types are similar enough, they may be aggregated. For example,
if there are many types of boating, one might treat these as a single group. The
Chapter 9
more similar the recreation types, the less problematic the aggregation.
Aggregating sail-boating and motor-boating is less problematic than aggregating
motor-boating and swimming. Most studies have some aggregation. Beach use,
which can include sunbathing, swimming, surfing, jogging, and even surf
fishing, is often treated as a single recreation type. Aggregation simplifies data
collection and analysis. It means less information from each respondent, fewer
observations, and less modeling. Again, one must be careful not to bundle
recreation types that are too dissimilar.
Since individuals are sometimes observed engaging in more that one type
of recreation on a single visit, a common practice is to identify the primary
purpose of the recreation trip and classify the use accordingly. For example,
one might ask respondents in a survey to report the number of trips taken
primarily for the purpose of fishing, and then primarily for the purpose of
boating. And, so on.
Along with defining the uses of the site one must also define the season
for each use. Hunting and fishing may have a season defined by law. Skiing,
rock climbing, and beach use will have a season defined by periods of favorable
weather. Others uses, such a viewing and hiking may be year round.
2.2.3 Step 3: Develop a Sampling Strategy
Next, a strategy is developed to sample the users and potential users of the
site. There are essentially two approaches: on-site and off-site sampling.
In on-site sampling recreationists are intercepted at the site and asked to
complete an oral or written survey. The survey may be completed on the spot
or handed to people to mail to a specified address at a later time. It is even
conceivable to recruit respondents on-site and later mail a survey to their home.
In this case, the recreationist only reports an address at the site.
On-site samples have the advantage of hitting the target population directly.
Every person interviewed has visited the site. Compare this to a random survey
of the general population where the percent of people who have visited the site
in the current season is likely to be quite small for most types of outdoor
recreation. In this case, the number of contacts or interviews required to get a
reasonable sample of users can be rather large. For this reason, on-site samples
are popular for single site studies.
There are a number of issues to be aware of when using on-site samples.
First, people who do not visit the site are missed. This implies a sample with no
observations taking zero trips. This compromises the accuracy of the estimated
intercept, the choke price, for the demand function. Think of a scatter of data
points being used to estimate the demand function in Figure 1. With on-site
sampling the data are truncated at one trip. We have no scatter of data points
at zero trips and are forced to estimate the intercept using trip data over
individuals having taken one or more trips. This is extrapolating outside the
range of the observed data.
Second, on-site samples can be difficult to conduct in such a way that a
random sample of users is obtained. Think about randomly drawing a user on
a beach. When and where do you sample? A strategy must be devised. For
example, randomly drawing several week days and weekends during the season
for interviewing and then interviewing every tenth person is a strategy that
attempts to approximate a random sample. Clear entry points, such as gates,
help in on-site random sampling. Consideration must also be given to how one
conducts a survey on-site. Interrupting someone as they sleep on the beach or
as they put a boat in the water is hardly advisable. Catching respondents at an
opportune time with minimum disruption will help response rates and extend
common courtesy. Consideration must also be given as to whether or not one
samples a person as they arrive or depart. The latter has the advantage that
respondents know more about the actual recreation experience catch rate of
fish, time spent on the site, activities, and so forth.
Third, in estimation one must correct for selection bias inherent in on-site
samples. The error term implicit in equation (3) in an empirical analysis will be
truncated such that no observations less than one are observable. This will
cause the estimated demand function to be too steep giving biased parameter
and welfare estimates unless corrected in estimation. See Creel and Loomis
(1990). On-site sampling will also over sample more frequent users. This is
called endogenous stratification. A person that visits a site 10 times during the
season is ten times more likely to be sampled than someone visiting the site only
once. This also introduces bias which must be corrected. See Shaw (1988) or
Haab and McConnell (2002, p. 174-181).
The alternative sampling strategy is off-site sampling. In off-site sampling
individuals from the general population are contacted, usually by mail or phone,
and asked to complete a survey. Unlike an on-site sample, an off-site sample
will include people who take trips (participants) and people who do not take
trips (nonparticipants). This gives the information needed to estimate the
intercept, avoids the selection biases, and is simpler to design for random
10
Chapter 9
response. It also possible to model the individuals decision of whether or not

to visit the site (see Step 8). Unfortunately, as noted above, off-site samples can
be costly to assemble because the participation rates from the general population
for a given site tend to be low.
Off-site sampling also raises the issue of determining the extent of the
market. Is the full day trip population located within 20 minutes of the site, 3
hours of the site, or further? A local fishing pond may have a small geographic
market. A popular ocean beach may have a large geographic market. In
principle one would want to sample the entire market randomly.
In a day trip model of recreation use, a safe bet for the extent of the market
is a maximum days drive to reach the site -- perhaps 3 to 4 hours. However, the
further one gets away from the site, the lower the rate of participation and the
higher the cost of assembling a sample with participants. Prior knowledge of
where users come from may help establish the extent of the market. For
example, parks may keep records of where their visitors come from. Stratified
sampling is also possible whereby residents living closer to the site are sampled
more heavily than those living further away. In this case one must correct for the
sampling in estimation and welfare computation.
Using an off-site targeted population is a common and effective way to
circumvent the participation rate problem. For many types of recreation, users
must obtain a licence or pay a registration fee. There are, for example, fishing,
boating, and hunting licenses. Participation rates among the set of license
holders is significantly higher than the general population. Hence, if one can
sample from the set of license holders, the number of people one must contact
will be considerably lower than over the general population. Often times
agencies that issue fishing licenses or boating permits will give an analyst
access to addresses of license or permit holders. If so, one can randomly sample
from this set.
For a good example of studies using on-site surveys see Sohngen (2000) and
Siderelis and Moore (1995). See Boyle ( ) for a discussion and analysis using
a targeted population.
2.2.4 Step 4: Specify the Model
Before data collection begins the right hand side variables for equation (2)
must be idntified. This establishes the information the analyst needs to gather
in the survey.
11
Every model includes an individuals trip cost to the site (tcr ). This serves
as own price in the demand model. Step 7 addresses measurement of trip cost
at length. Most models also include a measure of trip cost to substitute sites
(tcs). There are usually three or fewer included. The analyst looks for sites
frequently visited by the same population of users, sites similar in character to
the site of interest in the analysis, and sites nearby. From this mix, a set is
formed. Occasionally, proxies for substitute sites are used, such as the number
of lakes or acreage of water within a certain distance of ones home.
As shown in equation (2) most models also include a measure of income (y)
and a set of demand shifters (z) or demographic variables. The shifters are
factors other than trip cost believed to influence the number of trips taken over
a season. Some common shifters are
family size
age
gender
urban/rural residence
occupation
level of education
club membership
equipment ownership
attitudinal information
experience in activity
Family size or family composition may matter for many types of recreation
such as beach use or hiking. Families with young children are more likely to use
a beach. Club membership might include a fishing or hunting club,
environmental group, or some other such association. One might use
subscriptions to specialized recreation magazines as a explanatory variable as
well. These variables pick up unobserved measures of intensity of interest in
the recreation activity.
Urban/rural, occupation, and education are sometimes used as shifters as
well. Occupation and education are usually categorical variables such as
unemployed (yes/no), student (yes/no), retired (yes/no), and so forth. Education
might be high school completed (yes/no) and college completed (yes/no).
Attitudinal information is response data based on questions like Would you
consider yourself an advocate for the environment? Experience includes things
like number of years a person has been rock climbing or a self evaluation of
level of expertise. This might come from the answer to a question like How
would you rate you level of rock climbing expertise? Novice, intermediate, or
expert?
Although the list is long, most analysts are parsimonious in their selection
12
Chapter 9
of shifters. Studies typically include anywhere from one to five variables.

Sometimes analysts will report two or three model specifications such as a short
model that includes only trip cost and income and a longer model that adds a set
of shifters.
2.2.5 Step 5: Decide on the Treatment of Multiple Purpose Trips

Recreation trips may be single or multiple purpose. On a single purpose trip
an individual engages only in recreation at the site. On a multiple purpose trip
a person does more. For example, they may visit family or friends along the
way, take side trips for business, go shopping, go site seeing, visit other
recreation sites and so forth.
Single purpose trips fit the travel cost model well. An individual leaves
home, travels to a recreation site, engages in some type of recreation, and
returns home. All travel expenses can, more or less, be attributed to creating the
recreation experience.
Multiple purpose trips are more complicated. Trip expenses no longer
purchase a recreation experience alone. Instead, they purchase a package of
experiences. The logic of treating trip cost as the price of a recreation trip or
recreation experience becomes tenuous. Attempts to apportion trip cost across
the different purposes meet with little success. There is no logical way to
identify the marginal cost of the recreation portion of the trip unless some
restrictions are placed on the model (see Parsons and Wilson (1997)). How then
does one handle multiple purpose trips in the analysis?
With day trip data the most common approach is to assume all trips are
single purpose. If a case can be made that individuals engage in few other
purposes along the way and that these other purposes are mostly incidental, then
all trips may be treated as single purpose with little objection. With day trip data
this is often a reasonable assumption. With overnight data it is not. Indeed, this
is one reason many analysts confine their attention to day trips.
Another approach is to drop multiple purpose trips from the analysis. In this
case, the analyst ask respondents to report multiple and single purpose trips
separately or to report only single purpose trips. The model is then confined to
single purpose trips.
A last approach is to amend to the basic model to accommodate multiple
purpose trips. Mendelsohn et. al. (1992) and Parsons and Wilson (1997) are
13
recent examples. Mendelsohns framework broadens the definition of a site to

include multiple purposes and proceeds with the same logic as the basic travel
cost model. Their application is to a multiple site model where the other
purposes are recreation at other sites but the reasoning would apply to a single
site model. Parsons and Wilson (1997) present an empirical model, with
theoretical support, that introduces a simple demand shifter for multiple purpose
trips in a single site model in which all trips are analyzed together. They also
offer an argument for why access values for a site may embody the consumer
surplus associated with the other purposes of the trip.
2.2.6 Step 6: Design and Implement the Survey
The next step is to design and implement the survey. It is useful to start
with a survey from a past study as a guide. The survey is usually divided into
four parts
Introductory material
Trip count questions
Last trip questions
Demographic/Household Characteristic questions
The introductory material introduces the interviewer, identifies his or her

affiliation, explains the purpose of the study, and provides assurances that keep
respondents interested (eg., the survey is short and the responses are
confidential). There are usually opening questions as well that are easy to
answer, familiarize the respondent with the material, set the stage for the trip
count questions, and help define the site and the season.
The next three parts of the survey are used to form a data set like that shown
in Table 1. The trip count questions ask the respondent to report the number of
trips taken to the site over a designated time period. These questions may be
divided by recreation type (number of fishing trips, number of boating trips and
so forth), by day and overnight trips, and/or by multiple and single purpose
trips.
The last trip questions pertain to the most recent trip taken by the respondent
and typically include information such as time spent on-site, number of people
sharing travel expenses, other expenses incurred, and information on the trip
experience such as number of fish caught. These data are used to construct trip
14
Chapter 9
cost and sometimes to create other explanatory variables in the demand model.
These are gathered for the last trip only because gathering them for each trip
over the season can lengthen a survey considerably and are difficult for
respondents to recall for every trip.
The demographic/household characteristic questions correspond to the
vector z in equation (3). These questions close out the survey and include the
respondents income and location of the respondents hometown which is
required to estimate trip cost. Income is used as a shifter and usually to calculate
trip cost. In principle, one needs to know how trip cost will be calculated in this
step to know what data to gather. For this reason steps 6 and 7 overlap
somewhat.
It is essential to use good survey research methodology in developing the
survey. Champs Chapter 3 is a good reference on survey research methods.
They cover issues pertaining to effective question writing, keeping response
rates high, sampling, and more. While there is no need or room to repeat the
principles of good survey research here, there are a couple of issues of special
concern in travel cost surveys that deserve some attention: trip recall and trip
categorization.
When people are asked to report their number of trips, the analyst assumes
they will remember how many were taken. Since the model calls for a count of
trips over a season, respondents are often asked to recall the number of trips
taken over many months or even a year. This raises obvious questions. How
accurate is an individuals recall of past trips? Will approximations be valid?
There is no evidence I am aware of from controlled experiments as to how
serious recall error may be, but it is hard to deny the potential. One approach to
improve recall is do the survey at several intervals over the season asking
respondents to report trips taken only over the proceeding month or so. While
sensible, this approach can raise survey costs considerably and lower responses
rates through respondent attrition over the season.
Off-site surveys are usually conducted immediately following a season
while trip recall is good. On-site surveys are done within a season, which is also
likely to lessen recall problems. However, with on-site surveys the seasonal
data are truncated because the respondent can only report the number of trips
taken to date. The balance of the season is unknown. Two approaches are used
to fill in the data for the end of the season: ask respondents to estimate the
number of trips or predict the number trips based on current trip behavior. The
later assumes an individuals rate of trips is constant throughout the season.
Note that on-site samples that gather addresses only for an end-of-the-season
15
mail survey have a complete trip count but probably have a greater recall
problem.
As mentioned in step 2, there is often more than one type of recreation at the
site. If so, the survey may be designed for multiple recreation types. The most
common survey strategy is to proceed use-by-use in the questioning. For
example, in valuing recreation at a local lake one might begin by asking How
many day trips did you take to the site for the primary purpose of fishing?
Then, following the fishing questions, there would be a similar block of
questions about swimming, and then boating, and so on.
There are short-cuts that reduce the length of the survey. One might simply
ask people to report How many trips did you take to the site for purposes of
recreation? And then, What is your primary recreation use of the lake? This
avoids questioning about each recreation use. People are then classified by their
primary use. There is a model of lake visitation for people that use the site
primarily for fishing, one for people that use the site primarily for boating, and
so on. There is some mixing of trips within models in this case. However, if
people tend to have a single dominate use for a site, this strategy can be
effective.
It is also useful to isolate side trips and trips originating form somewhere
other than the residents hometown. For example, if a person is on a business
trip and takes a trip to the beach or goes on a whale watching tour while on the
trip, this is a side trip. The trip cost from the individuals home to the site
would clearly overstate the real marginal cost of the trip. It is easiest to isolate
and delete side trips from the analysis by clearly defining a day trip as being to
and from ones primary residence when the trip count is being made. See
Parsons and Wilson (1997) for an approach that adjusts trip cost and then
incorporates side trips into an analysis.
A similar issue arises when an individual owns a cabin or cottage near the
recreation-site of interest. Suppose the person spends most of the summer at the
cottage and makes frequent day trips to the site from that second residence.
How should these trips be handled in the analysis? As one 3 month long
overnight trip from their permanent residence or as it many day trips to the site
from their cottage? While the former choice should, in principle, embody all the
individuals consumer surplus it is simply not amenable to the single site model.
There are two strategies that one might consider. Either drop the observation
from the sample and lose a heavy recreation user or include the observation and
count the trips as day trips from the cottage. The latter approach understates the
full the surplus, but avoids deleting important observations. To identify these
16
Chapter 9
persons is it necessary to include a question that identifies whether or not the

individual owns property near the recreation-site and makes trips to the site
from there.
2.2.7 Step 7: Measure Trip Cost
Once the raw data are assembled and organized, the trip costs or prices to
the site and any substitutes are computed. Trip cost is the sum of the expenses
required to make a trip possible. Typical costs for a day trip include
travel cost
access fees
equipment cost
time cost
Travel cost includes all transit expenses. In a model of day trips where most
or all of the trips are made by car, travel cost is measured using the U.S.
Department of Transportations or the American Automobile Associations
estimate of the average cost of operating a vehicle per mile. Current studies are
using about 35 cents per mile. These costs include fuel and upkeep. The round
trip distance to the sites is usually calculated using a software package such a
PC Miler. The per mile cost is multiplied by the round trip distance to arrive at
trip cost. Tolls, if any, are added as well.
Since travel costs may be shared by several people, efforts are sometimes
made to apportion the costs. For example, one might ask respondents to report
the number of people sharing travel cost on a last trip and divide the cost
equally. Or, one might ask directly for an individuals share of the cost. In
either case, this is an example of a last trip question mentioned in Step 6. Since
it difficult to pose these questions for each trip and since there is no logic way
to handle separate trips within a season differently in a single site model, the
analyst typically relies on last trip data.
If the site or any of its substitutes have an access fee, that fee is included in
the trip cost. Sometimes sites will have annual or weekly passes, senior
discounts, or admit children for free. Making adjustments for seniors and
children is easy, but accounting for discounts is more difficult and usually
ignored. Typically the daily fee is used.
Equipment costs vary by type of recreation. In fishing one needs bait,
tackle, a rod, and sometimes use of a boat. For beach use there may be chairs,
17
umbrellas, surf boards and so on. For bird watching, there are binoculars and
film. For something like bait the cost is simply the market price of bait for a
day of fishing. For durable goods an imputed rent is needed. If one rents or
charters a boat for fishing the cost is simply the fee. If one owns a boat the rent
or cost of its service flow needs to be imputed. One approach for imputing such
costs is to use the rental fee for comparable services. This is, no doubt, an
overstatement of the cost, but is usually easy to obtain. Often times equipment
cost is excluded from the trip cost estimate. It is difficult to estimate and, in
some cases, is a negligible portion of trip cost. If incorporated, equipment cost
is usually obtained using a last trip question.
An alternative strategy for estimating trip cost, access fees, and equipment
cost is simply to ask individuals to report their expenses on the last trip to the
site. This is usually done by expense category. The estimates are then summed
to arrive at trip cost. The advantage of this approach is that it uses perceived
cost information and the researcher need not construct the cost estimates. Since
individuals base trip decisions on perceptions of cost, which may diverge from
actual costs, the respondent reported estimate is compelling. However,
objective estimates based on researcher computation as described above is most
common. The data are cleaner in the sense that they are uniform across
individuals and there are no missing or otherwise peculiar numbers.
The most difficult issue in computing trip cost, and certainly the one that
has received the most attention in the literature, is estimating the time cost of the
trip. The time lost traveling to and from the site as well as the time spent on the
site is time that could have been devoted to other endeavors. The value of those
lost opportunities is the time cost of the trip. Time cost often accounts for a
sizable portion of the total trip cost and deserves careful attention.
In most applications the estimate of time cost is related to a persons wage
in some way. This has a theoretical basis so long as the individual has a flexible
working arrangement and can substitute work time for leisure time at the
margin. Under such conditions, in theory, an individual increases the number
of hours worked until the wage at the margin is equal to the value of an hour in
leisure. Multiplying the hourly wage times travel and on-site time, in this case,
is a fair estimate of time cost. Unfortunately, this simple model breaks down for
many individuals. The simple leisure/work trade off does not apply to
individuals working a fixed 40 hour a week job for a salary. These individuals
do not have the flexibility to shift time in and out of work in exchange for
leisure. The tradeoff is also implausible for retired folks, homemakers, students,
and unemployed persons.
18
Chapter 9
Despite the difficulty of extrapolating the simple flexible leisure/work model

to many individuals in a recreation data set, the most commonly used approach
to value time is still wage-based. For people with fixed work schedules most
studies impute an hourly wage using annual income. Reported annual income
is divided by the number of hours worked in a year -- a number in the range
2000 to 2080. Another approach is to impute an individuals wage using a
simple wage regression over the subset of individuals in the sample earning an
hourly wage (Smith, et. al. (1983)). In this case, wage is regressed on income
and a vector of individual characteristics such as age, gender, and education.
The fitted regression is then simulated over non-wage earners to impute a wage.
For a recent application see McConnell and Strand (1994, p.100).
In wage-based applications, it is also common to see some fraction of the
imputed wage used to value time, anywhere from 1/3 of the wage to the full
wage, as the value of time. According to Feather and Shaw (1999), this practice
stems from early transportation literature wherein analysts had imputed the time
cost in empirical travel studies in this range. The recreation literature has more
or less accepted 1/3 as the lower bound and the full wage as the upper bound,
but neither is really on firm footing. For example, Feather and Shaw (1999)
show that in theory for those on a fixed work week it is possible for the value
of time to be greater than the wage. Finally, there are approaches for inferring
values of time from market data in the recreation context. See McConnell and
Strand (1981), Bockstael et. al. (1988), and Feather and Shaw (1999).
Time traveling to the site as well as time spent on-site should be included
in any calculation of time cost. While the time of getting to and from the site
is more or less fixed, time at the site is chosen by each individual and may vary
across the sample. Nevertheless, on-site time is typically assumed to be
constant across individuals and valued the same as travel time. Sometimes
analysts use the sample average length of stay on the last trip as an estimate of
the fixed on-site time. Others allow the on-site time to vary across the sample
using last trip data as each persons on-site time estimate for each site.
It should be evident that measuring trip cost calls for considerable researcher
judgement. Moreover, many of the trip cost components are endogenous or
chosen by individuals. For example, the purchase of equipment, number of
people a person fishes with, mode and route of travel, time on-site, choice of
residence (starting point for trip), and so forth are all chosen. Yet, the model
assumes that individuals take price as given. This creates the potential for bias
parameter estimates. There are a few attempts in the literature to deal with this
endogenity. McConnell (1992) suggests an approach that allows on-site time to
19
be endogenous, and Parsons (1991) suggests an approach for purging trip cost
of its endogeneity due to choice of residence. See Randall (1994) for a criticism
of the trip cost model fueled chiefly by the issue of trip cost endogeneity.
2.2.8 Step 8: Estimate the Model
The next step is to estimate the model specified in Step 4. In most modern
single site applications, the model is estimated as a count data model. The
dependent variable (number of trips) is a nonnegative integer, and the frequency
of zero and small numbers of trips typically make up a sizable fraction of the
data set. Count models are well suited for these data. See Hellerstein (1999),
Creel and Loomis (1990), and Greene (1997, p. 931-946) for details.
The basic count data travel cost model is a Poisson regression. The number
of trips taken by a person to a site in a given season is assumed to be generated
by a Poisson process. The probability of observing an individual take r trips in
a season is
(5)
Pr (r) =
exp( ) r
.
r!
The parameter is the expected number of trips and is assumed to be a function

of the variables specified in the demand model. To ensure nonnegative
probabilities, usually takes a log-linear form
(6) ln( ) = tcr tcr + tcs tcs + y y + d z .

Substituting equation (6) into (5) gives an expression for the probability of
observing an individual take r trips as a function of trip cost, income, and
individual characteristics. Equation (6) is the Poisson form of the recreation
demand specified in equation (2).
The parameters in equation (6) are estimated by maximum likelihood. For
each person in the sample the analyst knows r, tcr, tcs, y, and z. Using these data
and equations (5) and (6) the probability of observing the number of trips
actually taken is constructed for each person in the sample. The likelihood of
observing the actual pattern of visits then is the product of these probabilities
20
Chapter 9
(7) L =
n =1
exp( n ) nrn
rn !
An individual is denoted by n = 1, ... , N, so r n is the number of trips taken

by person n. In estimation, the parameters , on which depends according to
equation (6), are chosen to maximum L. There are many software packages
available for estimating by maximum likelihood. GAUSS and LIMDEP are
popular for Poisson forms.
Consumer surplus, or access value, for each person in the sample (area A
in Figure 1) has an explicit form in the Poisson model. For individual n the
surplus is
(8) S n = n / tcr .
where n is the expected number of trips from equation (6). Once the parameters
of the model are estimated, equation (8) is used to calculate the surplus value for
each individual in the sample and then aggregated over the population of users
to arrive at a total access value. This is discussed in detail in step 9.
The actual form of the probability used in estimation varies somewhat from
equation (5) depending on whether one uses on-site or off-site sampling. Recall
from step 3 that on-site random samples are truncated at one trip and over
sample more frequent users. Either of these complications will bias parameter
estimates unless corrected statistically.
The corrected probably for an on-site sample is a slight variation on the
basic Poisson probability in equation (5). It takes the form
rn 1
exp( n )n
(9) pr ( rn | rn > 0) =
( rn 1)!
This corrects for truncation at one trip and endogenous stratification and differs
from the basic Poisson regression only by r n-1 replacing rn. See Haab and
McConnell (2002, p. 174-81) for the derivation. With on-site sampling then
equation (9), instead of (5), enters the likelihood function for each individual.
Consumer surplus is still measured as shown in equation (8). See Shaw (1988)
and Greene (1997, p. 936-7) for more on truncated regressions in the Poisson
21
model.
Off-site random samples avoid the problem of truncation and endogenous
stratification and have the advantage of including non-participants. This allows
the researcher to model the decision to participate in recreation at the site or not.
For example, in the simple Poisson model of equations (5) and (6) the
probability of not participating (r = 0) can be modeled along with the decision
of how many trips to take (r = 1,2,3 ,....). The same Poisson process is assumed
to generate the outcome for any count of trips, including zero. For individuals
taking zero trips, Pr(rn=0) = exp(-n) enters the likelihood function. Again,
modeling participation is important because it helps pin-down the choke price
or intercept on the demand function by using observed choices to estimate
Pr(rn=0).
A more complicated version of the model is the hurdled Poisson which
assumes that the decision of whether or not to take a trip and the decision of
how many trips to take are generated from different Poisson processes. The
probability of taking zero trips is assumed to be exp(- n), where n is some
function of individual characteristics that governs whether or not a person takes
trips at all. This may include some of the same variables used in the trip
frequency portion of the model as well as some new variables. The individuals
probability of taking one or more trips then is just (1-exp(- n)). The model
becomes
(10)
Pr( rn = 0) = exp( n )
Pr( rn | rn > 0) = (1 exp( n ))
exp( n )nrn
rn !(1 exp( nrn ))
The term 1-exp(-n) scales the all the non-zero probabilities so that they sum to
one. The likelihood function is then constructed using the probabilities in
equation (10).
In some cases the hurdled model is estimated with a double hurdle. For
example, the population of nonparticipants in a given season may be divide by
those who never take trips to the site of interest and those who do but for one
reason or another have not in this particular season. For example, in a study of
fishing on a river, there may be two types of nonparticipants: those who never
fish and those who fish but not at this site in this season. The analyst may wish
to treat these groups differently -- the latter group may suffer from a site loss
while that former will not (at least in terms of use values). See Shonkweiler and
22
Chapter 9
Shaw (1996) and Haab and McConnell (1996) for more on the double hurdled
model.
Finally, it is common to see a variation of the Poisson Model know as the
Negative Binomial Model to estimate travel cost models. In the Poisson Model
the mean and variance of r n are constrained to be equal. To the extent that this
constraint is unreasonable, the model is in error. The Negative Binomial Model
is an approach for relaxing this constraint. The basic reasoning and structure
above still applies, but the estimation is somewhat more complex. See Greene
(1997, pp. 939-40) and Haab and McConnell (2002).
2.2.9 Step 9: Calculate Access Values
In the final step, access value for the site is computed using the estimated
model. Access value may be reported as
a mean seasonal value per person,

a total seasonal value for the population,
a per trip value per person, and/or
a total discounted present value of the site.
The seasonal per person estimate for any observation in the sample is the
area A in Figure 1. In the Poisson Model this is S n in equation (8) . The
estimated seasonal access value for the nth individual in the sample then is
exp( tcr tcrn + tcs tcsn + y yn + d zn )
,
(11) S$ n = n =
tcr
tcr
where denotes an estimated value using the results of the Poisson regression,
and the subscript n on an explanatory variable denotes the value of that variable
for individual n.
If one has estimated a Poisson Model using a randomly drawn off-site
sample, the sample mean access value is
23
(12) Soff =
n=1
where Soff is an unbiased estimate of the mean access value over the population
of participants and nonparticipants in the market area sampled. A reasonable
estimate of aggregate seasonal access value is
(13) AS = Soff POPoff
where POPoff is the total number of people in the relevant geographic market.
POPoff is typically from census data. For example, if one sampled the entire
population over 16 years old within a days drive, that population is the relevant
value for POPoff. If one is using a targeted population such as all individuals
holding a fishing license, POPoff is the total number of people holding a license
in the relevant time period.
If one is estimating a model with an on-site sample, the sample mean access
value is a biased estimate of the population mean because is over samples more
frequent visitors to the site. A corrected sample mean is
(14) Sonc =
R n
1 N Sn
, where N * = j .
*
N n =1 rn
j =1 j
where dividing the individual surpluses by r n and the summation by N* are

corrective weights. The term n j is the number of persons in the sample taking j
trips, and R is the largest number of trips taken by a person in the sample. A
reasonable estimate of aggregate seasonal surplus in this case is
(15) AS = Sonc POPon
where POPon is the total number of participants at the site over the season.
POPon may be gathered from an independent data source on participation rates
24
Chapter 9
at the site or may be estimated from survey data. In the latter case one must
again adjust for over sampling frequent visitors.
In principle, an off-site study using equation (13) and an on-site study using
equation (15) estimate the same aggregate value AS. S onc is larger than Soff
because it excludes nonparticipants. But, POP on is smaller than POPoff for the
same reason. One way of thinking about Son POPon is that it implicitly
assumes that anyone who has not taken a trip has zero surplus and is excluded
in the calculation of AS.
Soff POPoff , on the other hand, includes
nonparticipants and each will have a S$ n > 0 since everyone has some positive
probability of taking one or more trips in a Poisson model. And, the
nonparticipants contribute many low surplus values to the overall mean giving
the off-site model its lower seasonal mean.
An alternative method of estimating aggregate surplus is to compute an
average per trip per person value and then multiple this by an estimate of the
total number of trips taken to the site. Since average per trip values are the per
person seasonal value of the site divided by the number of trips taken, the
average per trip value in a Poisson Model is
(16) t =
( n / tcr )
1
=
.
n
tcr
This applies to on-site and off-site models alike. To arrive at an aggregate value
for the site multiply the average per trip value by the number of trips taken to
the site during the relevant season. This gives
(17 ) A S = t TR IP S
where TRIPS is the total number of day trips to the site over the relevant season.
Many parks and major recreation sites collect data or at least estimate TRIPS
which makes this a popular method of measuring aggregate surplus.
Finally, it is common to see the discounted present value of a site computed
using the seasonal aggregate estimate. The easiest approach is to assume no
25
change in the use of the site over time, no change in the change in the character
of the site, and a constant rate of discount. Then, using the conventional formula
for a the value of a perpetuity, the discounted present value of the site is
(18) PV = AS / i.
where i is the real rate of discount, usually a number in the range of .01 to .05.
See Krutilla and Smith ( ) for a discussion of examples of measuring PV when
use and values change over time.
In reporting site values, be it S, AS, t, or PV, it is important to be clear what
is and is not included in the value. For example, it may be a value of day trips
for rocking climbing in a park. If so, it should be noted that this excludes
overnight and side trips, other types of recreation taking place in the park, and
nonuse values.
2.3 A Single Site Application
Sohngens (2000) model of beach recreation on Lake Erie in 1997 is good
example of a modern application of the single site model. He estimated two
models -- one for Maumee Bay State Park and the other for Headlands State
Park. Muamee Bay is in the western Ohio. Headlands is in eastern Ohio.
Maumee Bay offers opportunities for recreation beyond beach use including
golfing, camping and so forth. Headlands is more natural.
Shongen defines the sites as the beaches located withing these two parks
(step 1). The recreation users are people visiting a beach within either park for
recreation or pleasure (step 2). All forms of beach recreation are aggregated
into a single type. The data were gathered on-site (step 3).
Both models were specified with own price (tcr), income (y), one or two
substitute prices (tcs1 and tcs2 ), and a variety of explanatory variables (z) (step
4). The Maumee Bay model used one substitute site; the Headlands model used
two. In both cases the substitutes were nearby beaches similar in character to
the beaches under study. The Headlands substitutes were located on either site
of Headlands Park. Income was measured as annual household income divided
by 10,000. The demand shifters were the scaled responses to five attitudinal
questions. These questions asked individuals to rank (from 1 to 5) how
important certain issues were in the choice of making a visit. The issues
included water quality, maintenance, cleanliness, congestion, and facilities.
26
Chapter 9
Higher rankings indicate the issue is more important to the individual visitor.
Finally, the model included a dummy variable for whether or not the primary
purpose of the last trip was for beach use.
Multiple purpose trips were included in the demand model along with single
purpose trips (step 5). Since all trips were single day visits by people living
within 150 miles of the site, other purposes are likely to be incidental side trips
of little consequence. As noted above, they included a demand shifter in the
model for trips with the sole purpose of visiting the beach. Sole = 1 if the sole
purpose was to visit the beach and Sole = 0 otherwise.
The survey was conducted on-site. Random beach users were handed a
survey and asked to return it by mail (step 6). Respondents reported their
number of day as well as overnight trips to the beach over the entire season.
Over 90% of the trips to Headlands were day trips; over 66% to Maumee Bay
were day trips. Only day trips were considered in the analysis. Their hometown
(and zip code) was given when respondents were asked for their home address.
They also gathered data on other activities while on a typical beach trip and the
attitudinal data on factors that affected a decision to visit the beach. The latter
were gathered to constructed the demand shifters specified in step 4. The
response rate was 52% for Headlands and 62% for Maumee Bay.
Trip cost was measured as the sum of travel expenses and time cost (step 7).
Distances to the site were measured as the linear distance from the center of an
individuals home zip code to the beach using latitude and longitude
coordinates. That distance was doubled (round trip) and then multiplied by 33
cents per mile giving total transit cost. The average distance traveled to
Headlands was 26 miles and to Maumee Bay was 35 miles. Time cost was
measured as an imputed wage times travel time. Travel time was calculated
using the estimated round trip distance and assuming people traveled at 40 miles
per hour. The imputed wage was 30% of annual household income divided by
2040. On-site was ignored.
Four different functional forms were estimated (step 8). Two continuous
(linear and log-linear) and two count (Poisson and Negative Binomial).
Truncation at zero trips due to on site sampling was accounted for in each
model. Endogenous stratification (over sampling of more frequent users in on
site samples) was not.2 The parameter estimates for the Poisson models for
Maumee Bay and Headlands are shown in Table 3.
The coefficient on own trip cost (tcr) is negative and significant in both
regressions, so the demand function is downward sloping. The coefficient on
income is positive in both regressions but significant in only the Headlands
27
regression. The effect of substitute sites on trips works as expected in the

Maumee Bay model, the higher the cost of reaching a substitute all else
constant, the more trips taken to the site. The results is a bit weaker in the
Headlands model. One of the substitute prices has a negative but insignificant
coefficient and the other is positive and significant.
The dummy variable for whether or not the primary purpose of the trip is for
beach use works in opposite directions in the two models. The attitudinal
variables also produced variable results. Recall that these variable are
characteristics of the respondents, not sites. One would expected the coefficients
on these variables, however, to be correlate with the characteristics of the site.
For example, if a site is congested, one would expected a negative and
significant coefficient on congestion. This would indicate that people who
consider congestion important in making a trip decision are less likely to make
a trip to a congested site.
28
Chapter 9
Table 3: Single Site Poisson Models for Headlands and Maumee Bay
Maumee Bay
Variable
Headlands
Parameter Estimate
Parameter Estimate
tcr
-.040***
-.026***
Income
.018
.040***
Sole
-.016
.292***
tcs1
.004***
.005
tcs2
--
-.004
Water Quality
-.053
-.139***
Maintenance
-.270***
.033
Cleanliness
.176**
.028
Congestion
-.065*
-.066***
Facilities
.098**
-.004
Constant
2.648***
2.433***
R2
.38
.29
Sample Size
230
345
Notes: Three asterisks denotes that a coefficient that is statistically significant at

the 99% level of confidence; two denotes significance at 95%; and one
significance at 90%. Water quality, maintenance, cleanliness, congestion, and
facilities are individual, not site characteristics, based on respondents ranking of
the importance of each beach characteristic in making a trip decision. The ranking
ran from 1 to 5, where 1 = strongly disagree it is important and 5 = strongly agree
it is important. Source: Sohngen (2000).
The results are then used to estimate access values (step 9). A per trip, an
annual aggregate, and a discounted present value estimate was computed for
each park. The per person per trip values using the model above are $25 ( =
1/.04) for Maumee Bay and $38 (= 1/.026) for Headlands. The range of per trip
values from all models (not reported here) was $14 to $33 for Maumee Bay with
a midpoint of $23.50, and $11 to $39 for Headlands with a midpoint of $25.
The Ohio Department of Natural Resources reported the total number of
29
trips taken to the beaches during the 1997 season at 224,000 for Maumee Bay
and 238,000 for Headlands. The total access value of the beaches for day trips
was estimated by multiplying total trips to each beach by its per trip value. This
is my equation (18). Sohongen reported these using the midpoint of the per trip
values across the models, so I report the same. These were $5.6 million ( =
238,000$23.50) for Maumee Bay and $5.6 million ( = 224,000$25) for
Headlands. Last, assuming a real discount rate of 3% and no future change in
the use of the beaches, the total discounted present value of each beach is $187
million ( = $5.6/.03). This accounts for day trip beach use only, excluding all
other uses as well as non-use value.
2.4 Variations
While the single site travel cost model using individual data to estimate the
access value of a site is the most defensible and widely used application, there
are some variations on this theme. I will briefly mention a few of these here and
direct the reader to the latest literature.
First, there are single site zonal travel cost models. These are estimated
using aggregate visitation rate data and average trip costs from predefined zones
near a recreation site. The zonal model has fallen out of favor due to its lack of
consistency with basic theory. Nevertheless, when data are limited, the zonal
model can provide a useful approximation. See Hackett (2000) for a recent
application and Loomis and Walsh (1997) for more detail.
Second, there have some efforts at valuing quality changes with single site
models. These all seek to estimate a shift in a single site demand function due
a change in quality and to then use the area between the functions as an estimate
of the value of the quality change. There are two versions of these models:
pooled and the varying parameter. These can be estimated with cross section
data on many sites (see Smith and Desvouges (1985) and Loomis (1988)), using
time series data on a single site (see Brown et. al. (1983)), or using data that
combines actual trips to a site with hypothetical visits to a site (see Layman et.
al. (1996) or McConnell (1986)).
Third, there are systems of single site demand equations. These are stacked
single site models for a group of substitute sites. See Bockstael, McConnell, and
Strand (1991, pp.254-6) for a discussion and Burt and Brewer (1971) for the
first application to a system of demands. Morey (1981) estimated a system of
share equations and Ozuna and Gomez (1994) and Shonkwiler (1999) have
estimated systems of count data models. These models account for substitutes
30
Chapter 9
and allow one to estimate access value for more than one site simultaneously.
However, as Bockstael, McConnell, and Strand (1991) explain, it is impossible
to value quality changes without placing rather stringent restrictions on the
model. Furthermore, it is difficult to estimate such models when the number of
sites rise above more than a half dozen or so. The RUM model, on the other
hand, can hundreds or even thousands of substitute sites and works well in
valuing quality changes.
3. THE RANDOM UTILITY MODEL

This section is divided into four parts like the previous section: basic model,
steps in estimation, an application, and variations. I will discuss the use of the
RUM model for valuing site access and changes in site quality. As a general rule
the RUM model is preferred to a single site model. It does a better job of
capturing site substitutes and it is better for valuing quality changes.
3.1 Basic Model
The RUM model considers a persons choice of a site for a recreation trip.
Instead of a quantity demanded as in the single site model, there is a site
chosen. In choosing a site a person is assumed to consider its price and its
characteristics. The price is the trip cost. The characteristics are things about
the site that matter to people such as ease of access and environmental quality.
While the time frame for a single site model is a season, the time frame for the
RUM model is a choice occasion. When analyzing day trips a choice occasion
is simply a day.
The theory works as follows. On a given choice occasion, a person considers
visiting one of S sites denoted as i = 1,2,....,S. Each site is assumed to give the
person a site utility vi. The utilities are assumed to be a function of trip cost and
site characteristics. The utility for site i assuming a linear form is
(19) vi = tc tci + q qi + ei
where tci is the trip cost of reaching site i, qi is a vector of site characteristics,
ei is a random error term, and the s are parameters. One expects site utility to
decline with trip cost ( tc < 0), to increase with desirable characteristics such a
31
easy access and good environmental quality, and to decrease with undesirable
characteristics. The random error term accounts for unobserved factors.
In theory, a person chooses the site with highest utility. Site k, for example,
is chosen if
(20) tc tck + q qk + ek tc tci + q qi + ei for all i S .
A useful way of expressing this result is in terms of trip utility. An individuals

trip utility is
(21) u = max( v1 , v2 ,..., v S ) .
Trip utility is the maximum attainable site utility on a given choice occasion
assuming a person visits a site. If site k gives the highest utility, the person visits
site k and attains trip utility u = v k .
Since people may choose not to take a trip on a given choice occasion, it is
common to see a no-trip utility included the choice set. Let a persons no-trip
utility be
(22) v0 = 0 + e0 .
No-trip utility is the highest utility a person can attain in any activity other than
visiting one of the S sites. A person now chooses from S + 1 alternatives -- S
sites and no-trip. Now, it is useful to define a choice occasion utility as
(23) u* = max{v0 , v1 ,..., v S } .
Choice occasion utility is the maximum attainable utility on a given choice

occasion. If no-trip gives the highest utility, choice occasion utility is u* = v 0.
If visiting site k gives the highest utility, choice occasion utility is u* = vk . And,
so forth.
Individual characteristics similar to the vector z in the single site model (age,
32
Chapter 9
family size, years of experience in a recreation activity, and so on) may also
enter the model and do so in one of two ways. First, they may be used to capture
differences in participation in recreation across the sample. Some people like
to fish, and some do not. Some people like to go to the beach, and some do not.
And, so forth. To capture differences in participation characteristics are entered
as shifters in the no-trip utility function
(24) v0 = 0 + 1 z + ei
where z is a vector of characteristics believed to influence a persons propensity

for recreation. In this way, no-trip utility varies across the population. Consider
a model of recreational fishing. If gender is a dummy variable in the vector z (
zg = 1 if female) and men like to fish more than women, a positive coefficient
on zg allows women to have a higher no-trip utility than men and captures this
effect. Or, consider rock climbing. If interest in climbing decreases with age, a
positive coefficient on age in the vector z would capture this effect.
Individual characteristics may also be used to capture differences in
preferences across the population for different sites. Some people like natural
beaches, while others like more developed beaches. Some people trailer a boat
and prefer a site with a boat ramp, while others look to rent a boat or moor a
boat at some preselected site. To captures differences in preferences for different
sites, individual characteristics are interacted with the relevant site
characteristics.
Assume there are m site characteristics and let qi = (q1i,q2i,...qmi). Suppose
in an analysis of beach use that there are people for whom surf fishing at a
beach is an important part of their recreation experience and for others of whom
it is not. Suppose I have data on whether or not a person owns a surf fishing
license. I might specify site utility as
33
(25) vi = tc tci + q1 ( q1i z1 ) + q2 q2 i + ... + qm qmi + ei
where q1i is a measure of the quality of surf fishing and z1 = 1 if the person owns
a surf fishing license and = 0 if not. In this way q1q 1i affects site utility only for
individuals who own a surf fishing license. Sites with good surf fishing give
higher utility for the fishing population but not for the general population. It is
possible to concoct numerous interactions of this sort.
Now, lets turn to how the RUM theory can be used to value site access at
one of the S sites. Suppose site 1 is closed due to an oil spill. Using RUM
theory, I can express a persons choice occasion utility with and without a spill.
The utility without a spill, the baseline, is
(26) u *(baseline) = max{v0 , v1, v2 ..., vS } .

The utility with a spill is
(27) u *( spill ) = max{v0 , v2 , v3..., vS } .

The utility u*(spill) excludes site 1 since it is closed. The decline in utility or
welfare loss due to the spill is u*(spill) - u*(baseline). This is the difference in
the utility a person can attain when the site is and is not available. The change
in welfare is
(28) u* = max{v0 , v2 ..., v S } max{v0 , v1 , v2 ..., v S } .
Using equation (28), suppose a person visits site 1 if there is no closure. It

must be true for that person that v1 > vi for all i. If site 1 where to close, the
person would choose the alternative with the second highest utility. Suppose
that is site k. If so, the person visits k instead of 1 and utility declines from v1
to vk. By the same reasoning, if no-trip has the next highest utility, the person
34
Chapter 9
no longer takes a trip and utility declines from v 1 to v0. The closer the second
highest utility is to site 1 utility, the smaller the decline in welfare.
If a person visits a site other than site 1 if there is no closure, then there is
no loss in utility. For example, if a person visits site k when site 1 is open, then
u*(spill) = u*(baseline) = v k. The same alternative that maximized utility
without closure maximizes it with the closure. Note that there is no accounting
here for congestion at site k that may follow from the closure of site 1.
To convert the decline in utility in equation (28) into monetary terms, divide
by an individuals marginal utility of income. In the RUM model the negative
of the coefficient on trip cost, - tc, is a measure of the marginal utility of
income. It tells us how much an individuals site utility would increase if trip
cost were to decline or, what is the same, if income were to rise for that trip.3
The welfare loss in monetary terms then is
(29) w = [ max{v0 , v2 ..., v S } max{v0 , v1 , v2 ..., v S }] / tc
w is a per choice occasion value. In a day trip model it gives the welfare loss
for a day due to the oil spill. To convert it to a seasonal value comparable to the
single site model, multiply by the number of choice occasions in the season. The
seasonal value for the loss of site 1 is
(30) W = T w
where T is the number of choice occasions in a season.

The value of access to more than one site is computed by dropping many
sites from equation (27). For example, if sites 1 through 5 are closed due the oil
spill, sites 1 through 5 are dropped in equation (27). It also possible to value
new sites by adding them to the choice set. In this case, the trip cost and site
characteristics for the new site or sites must be specified as well.
The model is used in a similar way to value changes in site characteristics.
Instead of dropping a site from the choice set, a site characteristic is altered
which, in turn, changes site utility. For example, if water quality is a site
35
characteristic in the vector q i, a decline in water quality at site 1 is captured by

writing q1 as q1#, where the element in qi measuring water quality is smaller in
q1# than in q1 . The utility at site 1 then declines from v1 = tc tc1 + q q1 + e1 to
v1# = tc tc1 + q q1# + e1 .

In this case an individuals choice occasion utility without the decline in
water quality is u*(baseline) in equation (26). The choice occasion utility with
the decline in water quality is
(31) u * ( dirty ) = max{v0 , v1# , v2 ..., v S } ,

where v1# is the utility at site 1 with the now poorer water quality. Following the
same reasoning as a site closure, the welfare loss is
(32) w = max{v0 , v1# , v2 ..., v S } max{v0 , v1 , v2 ..., v S } / tc .
Equation (32), like equation (29), captures an array of possible behavioral

responses. If an individual does not take a trip or visits a site other than site 1
without the decline in water quality, there is no change in utility. The same
alternative maximizes utility with and without the change in water quality, so
choice occasion utility is unchanged. If an individual visits site 1 without the
change, he or she may either continue to visit site 1, visit another site, or take
no trip with the change. In each case, the choice occasion utility declines.
If the individual continues to goes to site 1, utility drops from v 1to v 1#. The
individual maximizes utility by making the same trip but the recreation
experience is diminished due to the decline in the water quality. If the individual
goes to another site (say k) or no longer takes a trip, then utility declines from
v1 to vk or from v1 to v0 . All of these manifestations are captured in equation (32).
And again, this gives a per choice occasion value. A seasonal value is
computed as before using equation (30). And again, as with site access value,
quality changes can be evaluated at many sites simultaneously by altering the
36
Chapter 9
relevant attribute for each affected site in equation (31).

So far, I have treated the RUM model as deterministic -- assuming all
parameters and the error terms are known. In application, however, the
parameters are estimated and the error terms are unknown. The error terms are
assumed to come from some know random distribution. The choice occasion
utility in equation (23) then is also random. This is easy to see by substituting
the random site and no-trip utilities into equation (23) giving
(33) u *(baseline) =
max{0 + e0 , tctc1 + qq1 + e1,..., tctcS + qqS + eS }
If the error terms in equation (33) are random, choice occasion utility u* is
random. For this reason, in application, one uses the expected value of choice
occasion utility, instead of a deterministic value. Expected choice occasion
utility, the applied counterpart to equation (23), is
(34) eu *( w / o) =
E[max{0 + e0 , tctc1 + q q1 + e1,..., tctcS + q qS + eS }]
Equation (29) for access value and (32) for a quality change now take the forms
(35) w = [ E (max{v0 , v 2 ,...v S }) E (max{v0 , v1 , v2 ,...v S }) ] / tc
(36) w = E (max{v0 , v1# , v2 ,...v S }) E (max{v0 , v1 , v2 ,...v S }) / tc
In estimation, the form of the distribution for the error terms determines the
form of the expected value of the choice occasion utility. Each of the
behavioral responses to the a site closure and a quality change described above
still apply. However, now, each response occurs with some probability. Annual
37
values, once again, are calculated by multiplying the per choice occasion values
in equations (35) and (36) by the number choice occasions in the season.
In estimation then, one seeks to estimate the parameters of the site and notrip utilities in equations (19) and (24) using visitation data. Table 3 shows the
form of a typical data set -- a count of trips to each site in the choice set, trip
cost to each site, detailed site characteristics for each site, and demographic data
across a sample of individuals. For simplicity the table pertains a simple model
with three sites and one site characteristic (water quality). In most applications
the number of sites and site characteristics is much larger. Once the data are
assembled, the parameters of site and no-trip utility are estimated using some
form of a discrete choice multinomial logit model (more on this in the following
sections). The estimated parameters are then used to value site access or quality
changes at specified sites for each individual in the sample using equation (36)
and (37). Seasonal measures and means across the sample are computed. And
finally, the estimates are extrapolated to the population. The next section covers
these steps in detail.
38
Chapter 9
Table 3: Typical Data Set for Estimating a RUM Model (3 site choice set)
Number of Trips to
Site
I
D
Trip Cost to Site

$
Water Quality
Index at Site
1-10
Income
$000
Age
yrs
Site
1
(r1)
Site
2
(r2)
Site
3
(r3)
Site
1
(tc1)
Site
2
(tc2)
Site
3
(tc3)
Site
1
(q1)
Site
2
(q2)
Site
3
(q3)
(y)
(d1)
17
45
158
15
10
67
43
111
201
35
22
28
29
33
345
109
39
12
12
66
123
78
51
3.2 Steps in Estimation

The steps in estimating the basic RUM model are listed in Table 4 and are
similar to the steps for a single site model. To avoid repetition I will focus on
aspects of the steps in the RUM model that are different from the single site
model. It will be useful to refer the single site counterpart for most steps.
I will assume that a RUM model is being estimated for the purpose of
valuing site access and quality change at several sites, that the analyst is using
day trip data only, and that the data are individual-based.
Again, I will follow this section with a brief discussion of a RUM
application by Matt Massey and myself. We estimated a model of Mid-Atlantic
beach use by Delaware residents to value beach closures and loss due to erosion.
I will describe how we handled each step of the analysis.
Some other good applications to read if you are just beginning to learn about
39
the use of trip cost RUM models are Montgomery and Needelman (1997) and
Murray, Sohngen, and Pendelton (2001). I also recommend two publications
which go into considerable depth in laying out the development of their model
including basic theory, survey design, data collection, computer programs, and
so forth. These are McConnell and Strand (1994) and Hoehn et. al. (1996).
Table 4: Steps in Estimating a RUM Model

Step 1
Identify the Impacts to be Valued
Step 2
Define the Population of Users to be Analyzed
Step 3
Define the Choice Set
Step 4
Develop a Sampling Strategy
Step 5
Specify the Model
Step 6
Gather Site Characteristic Data
Step 7
Decide on the Treatment of Multiple Purpose Trips
Step 8
Design and Implement the Survey
Step 9
Measure Trip Cost
Step 10
Estimate Model
Step 11
Calculate Access and/or Quality Change Values
3.2.1 Step 1: Identify the Impacts to be Valued

The analysis begins by identifying the impacts to be valued. These will take
the form of site closures, openings, or quality changes at one or more sites.
Some examples of impacts are the closure of many lakes and rivers due to a fish
consumption advisory, the opening of a new park or skiing area, the expansion
of several hunting areas, a change in water quality across several urban beaches,
and so forth. Identifying the impacts to be evaluated at the outset defines the
research problem and gives direction for the succeeding steps.
40
Chapter 9
The impacts may be hypothetical. For example, a site is currently open and
you want to analyze the loss associated with its possible closure, or a site
presently has no hiking trials and you want to consider the gain associated with
adding trails. The impacts may also be actual events. For example, several lakes
are closed due to contamination and you want to analyze the welfare loss.
With a site closure or opening one identifies the areas and, in turn, the sites
that are affected. With a quality change one is identifying the affected sites and
the changes that will occur there. In this case, it is useful to begin thinking early
in the analysis about how quality will be measured for each site. For example,
will objective measures of quality be used -- like levels of dissolved oxygen for
water quality? Or, will perceived measures of quality be used -- like a rating
based on the reporting from respondents in a survey ? Also, are there published
measures or will you need to construct new measures as part of the analysis?
And, how will these quality changes map into the policy being evaluated? It is
also necessary to establish early on that there is sufficient variation in quality
across sites to measure its effect on site choice. If quality is more or less
uniform across the sites for the specific characteristic of interest, it is impossible
to measure its effect on site choice. See Parsons and Kealy (1992) for an
example using objective measures and Adamowicz et. al. (1997) for an analysis
using objective and perceived measures.
In some cases the impact may be analyzed as either a change in access or a
change in quality. For example, a fish consumption advisory might be analyzed
as a quality change instead of a closure, especially if people continue to use the
site. In this case one includes a dummy variable for fish consumption advisories
in the characteristic vector q. For this approach to work some sites in the choice
set must have advisories currently.
3.2.2 Step 2: Define the Population of Users to be Analyzed
The next step is to define the population of users for whom values will be
estimated. In principle, this includes all users and potential users of the affected
sites -- persons who use the sites without the changes and who might use the site
with the changes. A geographic area encompassing all users is the market for
the changes under consideration.
41
One way to capture the market in a day trip analysis is to define the
population as all individuals residing within a days drive of the affected sites.
In practice one approximates this market using boundaries that are convenient
for sampling. The most common market definition is the set of residents in one
or more states within a days drive of the sites. In some cases analysts will focus
on a particular state. This may be called for by the decision makers funding the
study or simply due to limited resources. As with the single site model, one also
needs to consider what types of recreation to include and whether or not to
aggregate recreation uses.
Here are some examples of day trip market definitions: Parsons and Kealy
(1992) define the market for water based recreation on Wisconsin lakes as all
Wisconsin residents. Adamowicz et. al. (1997) define the market for hunting
in Alberta as Alberta residences holding provincial moose licenses. Bockstael,
Hanemann, and Kling (1987) define the market for Boston beaches as the
Boston metropolitan area. See step 4 for a discussion of defining the market
before sampling begins.
3.2.3 Step 3: Define the Choice Set
The next step is to define the choice set S in equation (22). This includes
defining and determining which sites belong. In principle, one wants to include
all the sites with impacts identified in step 1 plus all other sites that may serve
as substitutes for these sites for the population of users defined in step 2. In
practice, one always approximates this set. Again, political boundaries play a
role in constructing the choice set and defining the sites.
It is easiest to see how analysts go about defining sites and determining what
to include in the choice set by giving some examples. In Parsons and Kealy
(1992), we analyze lake recreation in Wisconsin. Individual lakes in the state
over a certain size are defined as the sites. There are over 1000 such lakes. If a
lake is within 150 miles of a persons home, it is included in his or her choice set.
This is based on the fact that few people traveled over 150 miles to make a day
trip in the data.
Andrews (1996) analyzes trout fishing in eastern Pennsylvania. He defines
sites as the management units used by the state fish and wildlife service. These
42
Chapter 9
are stream segments and lakes known to have trout. Each is different enough
in character to be managed separately and data are organized by these units by
the state. There are over 2000 sites. The choice set includes any site within
185 miles of a persons home. The longest day trip in the data set is 183 miles.
Parsons, Jakus, and Tomasi (1999) study lake recreation at the reservoirs
operated by the Tennessee Valley Authority in the southeastern United States.
Each persons choice set is defined as 14 specific reservoirs in the TVA system.
Shaw and Jakus (1996) study rock climbing in northeastern United States. Their
sites are defined as the four major climbing areas in the region. Each person has
all four sites in his or her choice set. One major climbing area is excluded on
the grounds that it is distinctly different from these four. Morey, Watson, and
Rowe (1993) study salmon fishing in Maine and Canada. Nine rivers in the
region known for salmon fishing are included each persons choice set.
McConnell and Strand (1994) study marine recreational fishing on the east
coast of the United States and define coastal counties as the sites. Again, if the
site is within 150 miles of the persons home it is included in the choice set.
Hausman, Leonard, and McFadden (1995) study recreational fishing in Alaska
and defined sites as one of 17 large fishing regions in the state. These regions
serve as each persons choice set.
As you see, choice sets vary in size from 3 or 4 sites to over a thousand
sites. Site definitions can vary from highly aggregated regions or counties to
rather narrowly defined units such as beaches or small segments of a river. In
choosing the number of sites and the degree of aggregation of sites it is best to
error on the side of too many sites and narrow site definitions. Modern
econometric software packages can handle a large number of sites and the
feasibility of randomly drawing sites in estimation to approximate larger choice
sets is also an option (see Parsons and Kealy (1992)).
If you must aggregate sites into regions or counties, the general rule is to
group similar sites together. The sites should be similar in all characteristics
including trip cost. The less similar the sites, the more bias one is likely to
encounter (see Parsons and Needelman (1992)). For approaches that mix
aggregated and non-aggregated sites together see Lupi and Feather (1998) and
Parsons, Plantinga, and Boyle (2000).
Finally, there has some concern about using researcher versus individual
43
defined choice set in RUM models. The definitions I have described above are
all researcher defined. Peters et. al. (1995) and Hicks and Strand (2000) have
argued that people cannot perceive this many sites in making a choice and
suggest an approach using choice sets determined by people in the survey.
Individuals identify sites they consider in site choice and this makes up the
choice set. For a counter argument to this approach see Parsons, Massey, and
Tomasi (1999). Researcher defined choice sets still dominate the RUM
literature.
3.2.4 Step 4: Develop a Sampling Strategy
Next, a sampling strategy is developed. Virtually all published studies use
some form of off-site random sampling. The users identified in step 2 are
contacted by phone or mail and asked to report trips taken to the sites defined
in the choice set in step 3. See Champs chapter 3 or Dilman (1999) for random
sampling with mail and phone surveys. In some cases a targeted population of
users, such as people with boating licenses, will be contacted randomly. In most
applications participants and non-participants will be sampled. This enables the
analyst to model the decision of whether or not to participate in the recreation
activity.
Like the single site model, sampling for a RUM model runs into the issue
of low participation rates from the general population. The fraction of people
participating in most forms or outdoor recreation is low enough that random
samples from the general population can yield a small sample of actual users.
The conventional way of dealing with this problem is to stratify the sample -sampling counties or communities nearer the sites more heavily. Participation
rates are higher in the nearby communities. In some cases stratification is the
only way to get nearby users in the sample if local communities are relatively
sparsely populated. This not only increases the number of participants but also
allows for a more even geographic distribution across the sample of users which
can sharpen parameter estimates by ensuring variation of trip cost across the set
of sites. As with the single site model, a target population of license holders
circumvents the low participation rate problem.
Unlike the single site model, on-site sampling is usually not an option. On-
44
Chapter 9
site sampling or what is known as choice-based sampling is covered

extensively in the RUM literature. To estimate parameters without bias, site
weights are needed. The weights are based on the actual distribution of day
trips by the population of users being studied and are used to adjust the count
of trips to the different sites to reflect their true occurrence. I know of no RUM
applications of the trip cost model that estimated using a choice based sample.
See Train (1986, pp. 48-9), Ben-Akiva and Lerman (1985), or Laitila (1999) for
more on working with on-site samples.
3.2.5 Step 5: Specify the Model
Before data collection begins one specifies the elements that will be
included in the model in equations (19) and (24).
Site utility in equation (19) includes trip cost (tc i) and a vector of site
characteristics (qi) for to each site. In some cases it will include interactions with
individual characteristics. Trip cost is the same as in the single site model, the
sum of travel and time cost. The vector qi includes characteristics that matter to
people in making a site choice. This will vary depending on the type of
recreation being studied. Fishing will include catch rate of fish, rock climbing
will include difficulty of the climb, boating will include presence of boat ramps
and so forth. It is difficult to make a generic list, but here are some common
attributes one sees in a RUM model
amenities
size
access
environmental quality
park (yes/no)
opportunities on-site
sport success
remote location (yes/no)
character of nearby area
special features
Amenities are measures such as acreage of tree coverage. Size might be

total acres in the site, length of beach, or number of trails. Access might be a
variable indicating the presence of a boat ramp or one indicating that the site is
only accessible by 4 wheel vehicle. Environmental quality includes measures
of water quality or a variable for the presence of a fish consumption advisory.
A variable designating whether or not it is a park is a common site
45
characteristic. By opportunities on-site, I mean things like camping available

or a museum is present. Sport success is catch rate of fish, bag rate for hunting,
sightings for birding watchers, and so forth. The character of the nearby area
might be indicating if a beach is developed or natural. And finally, special
features are unusual characteristics that may make difference in site choice such
as a dam on a river. Finally, if a site or group of sites is distinguished from
others sites in the choice set in some more general way, it is common to see an
alternative specific constant assigned to that site. An alternative specific
constant is a dummy variable the site utility for one site or more sites.
The no-trip utility in equation (24) is a vector of individual characteristics
that govern how often and whether or not a person makes a trip. These play the
same role as the vector z in the single site model. Some common characteristics
are age, occupation, and urban/rural residence essentially the same list shown
in single site step 4.
In some cases individual characteristics enter the site choice model as
interaction terms. As explained in earlier, this happens when one believes that
a certain characteristic affects people differently. For example, the presence a
boat ramp may matter to people who trailer a boat but may not to people who
keep a boat at a marina. Or, the difficulty of a rock climb may matter to expert
climbers but not to novice climbers or at least the degree of its importance may
be different. An interaction is entered into the model as shown in equation (25).
Specification always depends on available data. For this reason steps 5 and
6, model specification and collection of site characteristic data, work in parallel.
The analyst has a certain specification in mind but amends it to accommodate
available data. For some good of examples of differences in specification by
recreation types see Sidderelis (1995) for a boating, Shaw and Jakus (1996) for
rock climbing, Karou (1995) for fishing, Morey (1981) for skiing, Parsons and
Massey (1999) for beach recreation, and Adamowicz (1997) for hunting.
3.2.6 Step 6: Gather Site Characteristic Data
The next step is to gather site characteristic data. As noted, step 5 works in
tandem with step 6. The typical sources for characteristic data are the state and
federal agencies responsible for managing the resource. Environmental
46
Chapter 9
protection, natural resource management, and fish and wildlife agencies are
often good sources. These agencies may have data on indicators of
environmental quality and physical measures such as size and elevation
Furthermore, the agencies often have ready made site definitions. A fish and
game agency, for example, may have management units for which data are
gathered for their own purposes. In most cases these agencies are a good
starting point for data collection.
Other sources of data are tourist bureaus, clubs and associations,
universities, scientists, and newspapers. In some cases fieldwork where the
analyst constructs the primary data him or herself or interviews knowledgeable
people is necessary. Even in cases where variables are not constructed by field
observation, such visits can confirm or amend existing site and variable
definitions.
In some circumstances the site characteristic data are gathered in the survey.
For example, individuals may be asked to rate the quality of hunting at each site
visited in the current season. Using these responses the analyst constructs a
hunting quality index for each site. One could do the same with catch rate of
fish, view amenities and so on.
There are several problems with this approach. First, the data for each site
are usually confined to those who have visited the site in the current season. 4
This is likely to bias quality measures upward at each site. People who visit a
site are those who find the site desirable. Second, popular sites have more data
than less popular sites. Many sites may a single or no visitor. This causes an
asymmetry in the quality of variable measurement across sites. Third, perceived
variables are more difficult to map into actual policy changes.
An alternative method for working with survey data is to preform an
auxiliary regression using the reported measure for the characteristic (such as
number of fish caught) as a dependent variable and observable site
characteristics such as lake size, depth, elevation, and presence of regulations
as explanatory variables. The unit of observation is a site and the number of
observations is the number of respondents times the number of different sites
visited by each. The fitted regression is then used to predict catch at each site.
For some examples see McConnell, Strand, and Blake-Hedges (1995).
47
3.2.7 Step 7: Decide on the Treatment of Multiple Purpose Trips

In a day trip RUM model multiple purpose trips are handled in much the
same way as in the single site model. Either all trips are assumed to be single
purpose or multiple purpose trips are identified in the survey and dropped from
the analysis. Again, this is a larger issue when overnight trips are being
analyzed.
Another way of dealing with multiple purpose trips is to define site
characteristics in such a way that they account for the potential of other
activities while visiting the site. For example, a beach characteristic in a RUM
model might be a dummy variable for the presence of shopping nearby. This
method at least recognizes that an experience may be broader than recreation at
the site alone.
It also possible to imagine a variant of Mendelsohn, et. al.s (1992) approach
for the RUM model. Alternatives would be defined as portfolios of sites. Each
site would be represented by a dummy variable in the RUM model (like a site
characteristic) and trip costs would be measured as the cost visiting the entire
portfolio. People would choose portfolios. Welfare losses could be computed
for losses of one or more sites from the portfolios.
3.2.8 Step 8: Design and Implement Survey

Next is the design and implementation of the survey. It is organized in the
same four parts as a single site survey
Introductory material
Trip count questions
Last trip questions
Demographic/Household Characteristic questions
The introductory material, last trip questions, and demographic questions are
essentially the same as in a single site survey. Unlike the single site survey, one
gathers data on trips to every site in the choice set. The responses are used to
48
Chapter 9
form a data set like that in Table 3. RUM surveys face the same recall and trips
categorization issues faced with single site models. See single site step 6 for a
discussion. Recall is perhaps a larger problem in a RUM survey, because we are
concerned about many sites.
There are essentially three approaches for counting trips to the sites. One
provides the respondent a preestablished list of sites. The respondent reviews
the list and records the trips over the relevant time period. This approach is
easiest for data handling. The sites are defined and respondents count fits
neatly into the definition. For sites off the list, there is usually a catch-all
category such as some other site. This approach is also somewhat less
burdensome. The site names serve as a reminder to people and recalling exact
site names is not necessary.
The second approach is open-ended. People are asked to list all the sites they
have visited over the relevant time period along with a count of trips to each.
Once gathered one must rework the data such that site definitions are consistent
across respondents. This can be time consuming. Sites often have more then
one name, people using nearby town names, and sites with similar names can
lead to confusion. When the number of potential sites runs into the hundreds or
thousands, this approach may be hard to avoid. Even in these cases an insert,
separate from the survey, listing (and numbering) the sites is worth considering.
One useful innovation to use in conjunction with the open-ended approach
is a map-and-sticker. Respondents are asked to place a sticker on a map for
each site visited. The sticker is numbered to correspond to a table in which the
respondent records the number of trips taken to the site. This eases the final
assembly of the data and avoids any confusion about the name and location of
the site. It also has the advantage of aiding in finer site definitions depending on
individuals use of sites. For example if the there is heavy visitation to a site
which seems divided by trips to the northern and southern portion of the site,
one might opt for dividing the site along these lines.
Mail surveys or mail-phone surveys favor the approach where a list of sites
is provided to the respondents. Phone surveys call for an open-ended approach
unless the number of sites is less than a dozen or so. However, one can
approximate the preestablished list approach by being interactive. The caller
has a list of sites. During the survey respondents are asked to count trips region
49
by region. For example, Did you visit any beaches on the Outer Banks? If so,
which beaches? When the respondent gives a beach, the caller verifies that the
beach fits the list and proceeds. If the respondent has trouble recalling the
name, the caller can ask about the general area and provides some clues till the
correct beach is identified. If the respondent gives and unknown name, it is even
possible to find the closest beach. This can be an extremely valuable way to
construct the trip data file and can eliminate the post survey site definitions.
A third approach is to work with last trip data alone. In this case one gathers
information only on the last site visited what site was it, how many trips were
taken to that site over the season, and how many were taken to all sites in total?
A basic RUM model can be estimated with these data. The survey is less
complex and recall is a smaller issue. This comes at the expense of far fewer
trips and hence less data for the final analysis.
3.2.9 Step 9: Measure Trip Cost
This step is essentially the same a single site step 7 except that trip cost is
measured to every site in a persons choice set, not just the visited sites. This
invariably calls for a software package to compute travel time and distances
between respondents hometowns and the sites. PCMiler is a popular one.
Respondents hometowns and the site names must defined so they map into the
chosen software package. Small communities are sometimes excluded.
Spellings must be exact. Some packages require state abbreviations as part of
the title. Most software packages will compute time and distances between zip
codes as well as an approximation.
For more discussion of the issues surrounding the measurement of trip cost
such as the value of time and using individual versus household measures see
single site step 7.
3.2.10 Step 10: Estimate Model
Next, the analyst seeks to estimate the parameters and in equations (19)
and (24) specified in step 5. This is done in a probabilistic framework. An
expression for the probability of visiting a site is formed. The probability that
50
Chapter 9
a person visits site k is

(37) pr ( tc tck + q qk + ek tc tci + q qi + ei for all i S
and 0 + 1 z + e0 )
Under different assumptions about the distribution of the error terms e i, different
forms for equation (37) are derived. The simplest is the multinomial logit (ML)
for which the probability that an individual visits site k is
(38) pr ( k ) =
exp( tc tck + q qk )
S
exp( 0 + 1 z ) + exp( tc tci + q qi )

i =1
Notice that the probability of visiting site k depends the characteristics of site
k (in the numerator and denominator of equation (38)) and on the characteristics
of other sites (in the denominator). Also notice that each person has a
probability for each site and no trip. The no-trip probability has exp( 0 + 1 z )
in its numerator, and its denominator is the same as the denominator in equation
(38). This ML probability assumes that error terms in equation (19) and (24)
are independent and identically distributed with Weibull distribution. See
Greene (1997, p. 913).
The parameters are estimated by maximum likelihood using the probabilities
in equation (38). If one has data on N persons each visiting one of the S sites
or taking no trip for a given choice occasion, the likelihood of observing that
pattern visits in the data using the ML probability model is
N
(39) L = pr (i ) rin
n =1 i = 0
51
where rin= 1 if individual n visited site i and = 0 otherwise. The probability pr(i)
is the logit form from equation (38). The parameters are estimated by choosing
values of and to maximize L. These are the maximum likelihood estimates
for the model. This is the form of the model that is most likely to generate the
patterns of visits actually observed in the data.
In circumstances where individuals are observed taking multiple trips to
sites over the season the same likelihood function is used with rin now equal to
the number of trips taken to site i by individual n.5 Many software packages are
available for estimating logit models. LIMDEP, GAUSS, and TSP are common.
This ML model is criticized for a restriction know as the independence of
irrelevant alternatives (iia). This restriction implies that the relative odds of
choosing between any two alternatives is independent of changes that may
occur in other alternatives in the choice set. For example, the iia property
implies the following. If an improvement in a characteristic at site k causes a
10% increase in the probability of visiting that site, then the percent change in
the probability of visiting each of the remaining sites in the choice set must
decrease by 10% (a proportional reduction in all other probabilities). If some
sites are better substitutes for the site experiencing the improvement, this result
is unrealistic. One expects good substitutes to have a larger percentage
reduction in probability than poorer substitutes.
There are essentially two methods for relaxing iia: the nested logit (NL) and
the mixed logit (MX). Both introduce correlation among the site and no-trip
utility error terms which allows for more general patterns of substitution in the
model. The NL model nests alternatives in groups believed to be close
substitutes. For example, in an application of beach use, where sites include
ocean and bay beaches, nesting the ocean and bay beaches into separate groups
may make sense. This has the effect of allowing beaches within each nest to be
better substitutes for one another. The iia assumption still holds for sites within
a nest, but is relaxed for sites from different nests. In this way the model allows
for a richer pattern of substitution. The error terms within each nest are viewed
as having shared unobserved characteristics which leads the correlation of the
error terms within a nest. (eg. all ocean beaches may have large waves and
similar view amenities). A common application of the NL model is to nest sites
and no-trip in separate nests. The reasoning here is that sites are more likely to
52
Chapter 9
serve as a better substitutes for one another than no-trip. Or, put differently,
sites are likely to share unobserved characteristics. See Morey (1999) for more
on NL models in recreation demand, Greene (1997, p.121) for an introductory
discussion, and Morey, Rowe and Watson (1993) or Parsons and Hauber (1998)
for applications.
The second method for relaxing the iia assumption is a mixed (MX) or
random parameters logit. MX models are estimated used simulated probability
techniques and are fast becoming the standard for estimating trip cost RUM
models. Like the nested model, the mixed model is a generalization of a basic
multinomial logit model that allows the parameters and to be random. The
variation in each parameter is interpreted as a component of the error term in the
site utilities which leads to correlation among the utilities and a more general
pattern of substitution. See Train (1999) for an example. NL and MX models
may be estimated using GAUSS or LIMDEP.
3.2.11 Step 11: Calculate Access and Quality Change Values
In the final step, access or quality change values are computed. Value may
be reported as
a mean per choice occasion value per person,

a mean seasonal value per person,
a total seasonal value for the population,
a per trip value per person, and/or
a total discounted present value of the site.
The parameters estimated in the previous step are used to calculate welfare
changes using equations (35) and (36). The form of the expected maximum
utility depends on the assumed distribution for the error terms in the model.
With the Weibull distribution in the ML model, the expect maximum utility of
a choice occasion is
53
(40) eu = ln exp( 0 + 1 z ) + exp( tc tci + q qi )
i =1
This is the log of the denominator in equation (38). It is a preference-weighted

utility index for a choice occasion in the sense that all alternatives are included
and the higher the alternatives utility the larger its roll in the expression. Again,
its form follows directly from the assumed distribution for the error terms. With
equation (40) then, one can value per choice occasion site access and quality
changes for one or more sites using equations (35) and (36).
Per choice occasion site loss for person n is
S
(41) s n = ln exp(0 + 1 zn ) + exp( tc tcin + q qin )
i =6
S
ln exp(0 + 1 zn ) + exp( tc tcin + q qin )
i =1
where the first 5 sites are lost. A denotes an estimated value using the
estimation results and the subscript n on an explanatory variable that denotes the
value of that variable for individual n.
For a quality change the per choice occasion value is
S
(42) sn = ln exp(0 + 1 zn ) + exp( tc tcin + q qin* )
i =1
S
ln exp(0 + 1 zn ) + exp( tc tcin + q qin )
i =1
where q*i is a vector indicating a quality change at some or all of the S sites.
Again, as the distribution of the error term changes, the form of the expected
maximum utilities in equation (42) change. If the sample is randomly drawn a
54
Chapter 9
simple mean of sn is presented for the per choice occasion value. If the sample
is stratified, the sample mean is adjusted to represent the population.
The seasonal value for each individual is the total number of choice
occasions times his or her per choice occasion value. So, the mean seasonal per
person value is
(43) S = T s
where s is the sample mean per choice occasion value (adjusted for
stratification if necessary) and T is the total number of choice occasions in the
seasons. In a day trip model, T is the number of days in the season. The
aggregate seasonal value over the population is
(44) AS = S POP
where POP is population of users and potential users. This might be all
residents within driving distance of the site, all people owning a fishing license,
or whatever defined the population in sampling. AS is occasionally converted
to a discounted present value assuming a constant following of recreation
services from the site or sites. As shown with the single-site model in section
3.2.9 this is
(45) PV = AS / i
where i is rate of discount. Equtions (43) through (45) work for site access and
quality changes alike.
Last, it is not unusual to see per trip access or quality change values
presented is a RUM analysis. A per trip per person value is
55
(46) t = AS / trips
where trips is the total number of day trips by the relevant population. One can
use an external estimate of the total number of day trips taken to the site over
the relevant season or estimate the number of trips using the RUM model. In
either case the per trip value applies to the same scenario consider for site access
or quality change above.
An alternative approach for estimating per trip values arises when one has
estimated a model that excludes no-trip form the choice set. In this case, the one
estimates per trip, instead of per choice occasion values, in the basic RUM
model and per trip values flow naturally from the results. See Parsons and
Massey (2002) for an example. Although perhaps convenient, one must keep in
mind that these results come from a restricted model that disallows no-trip as
and alternative.
3.3 A RUM Application

Matt Massey and I have estimate several RUM models of beach use in the
Mid-Atlantic region of the United States using a 1997 beach use survey. See
Parsons and Massey (1999) or Parsons and Massey (2002). Here, I present
another version of that model following the steps outlined in the previous
section. The Mid-Atlantic beaches in our analysis include all of New Jersey,
Delaware, and Maryland ocean beaches -- 62 beaches.
We identify two impacts for two analyses in our study (step 1). The first is
the potential closure of beaches in the state of Delaware due to oil spills, water
pollution, other environmental episode. These are analyzed as lost site access.
The second is beach erosion or narrowing of beaches in the state. This is
analyzed as a quality change. Our quality measure is beach width. Beach
width data is provided by the states and there is sufficient variation across the
beaches in this characteristic.
We define our market as all residents in the state of Delaware (step 2). We
56
Chapter 9
recognize that there were a large number of out of state users. However, budget
limitations and our key interest in exploring some methodological issues in the
model result in our using this narrow market definition. We also treat all uses
of the beach as a single recreation type. We aggregated sunbathing, swimming,
surf fishing and so on.
We define our choice set as all ocean beaches within a days drive of
Delaware residents (step 3). This includes 62 beaches in four states. Beaches
(our sites) were defined using the political boundaries of beach communities
which is consistent with how people identify beaches in this area. For example,
Ocean City MD, Rehoboth DE and Cape May NJ were all separate beaches in
our analysis. Every person had all 62 sites in their choice set.
Delaware is a small state with 3 counties only. The most populated county
is located in the north. The ocean beaches are along the southern most county.
We randomly sampled an equal number of residents over the age of 16 from
each of the three counties (step 4). We stratified in this way to avoid a
population dominated by residents from the northern most county. We were not
too concerned about low participation rates. Historic data convinced us that half
or more of the population used the beaches in a typical year.
There are a number of things we thought would influence day trips to the
beaches in this region (step 5). Among these were:
trip cost
natural vs developed beach
private or limited access
presence of boardwalk
availability of parking
availability of bathhouses
and other facilities
beach width
These, more or less, were our targets as we set out to gather the data in the next
step. For individual characteristic data we thought occupation, education,
family composition, and flexibility in work would be important.
The site characteristic data were gathered from various sources: state
departments of natural resources (including interviews with experts in those
agencies), field trips, interviews with scientist working on a data on the physical
characteristics of the New Jersey beaches, tourist guides, maps, newpapers, and
locate web sites (step 6).
57
We analyzed day trips only and assumed all trips were single purpose or at
least that side trips were incidental and easily ignored without introducing error
(step 7).
We used a random mail survey of 1000 Delaware residents in the Fall of
1997 (step 8). An initial mailing of the survey was followed by a reminder
postcard one week later and a second mailing of the survey after three weeks.
Our response rate was 55%. Individuals were asked to report day, short
overnight, long overnight, extended stay, and side trips separately. The survey
was eight pages long. Respondents were asked to complete a 2-page table of
trips to 62 beaches. A map insert was provided to help identify beaches.
Trip cost was measured as the sum of travel expense, time expense, and
beach fees (step 9). Travel cost was 35 cents times round trip distance plus tolls
and parking fees. Many trips to New Jersey beaches are via toll roads and on
some routes a ferry is used to cross the mouth of the Delaware Bay. After the
shortest route was determined using PC Miler, the toll routes were identified and
their cost computed. Many New Jersey beaches have fees. We used a per day
fee which was published by beach for the 1997 season in a local newspaper.
Time costs were estimated a wage proxy times round trip travel time. The wage
proxy was annual household income divided by 2080. PC Miler was used to
compute round travel times.
We estimated a nested logit model. The results are in Table 4. The site
characteristics are variables appearing in site utility. The individual
characteristics are variables appearing in no-trip utility.
As shown, the site characteristics that increase a sites day trip utility are
boardwalk, amusements, good surfing, having a park, and good parking. The
characteristics that decrease a sites day trip utility are private, being too wide
or too narrow, and having high rises nearby. None of these are surprising.
Length of the beach and park (state or federal) were insignificant and facilities
had the wrong sign. The alternative specific constant for Atlantic City and for
New Jersey were positive and significant.
The inclusive value coefficients are parameters which characterize the
degree of substitutability of among alternatives within a nest. To be consistent
utility maximization, these coefficients should fall between 0 and 1. The closer
the coefficient is to 0 the greater the degree of substitutability. A coefficient
58
Chapter 9
equal 1 is the same as not nesting. See Morey (1999) for an excellent discussion
of NL models and interpreting inclusive values. Out results indicate a higher
degree of substitution among sites within the same state. The coefficient on IV
beaches suggests that the no-trip/sites nest is a mis-specification.
The individual characteristics that decrease no-trip utility and hence increase
the probability of taking a beach trip are number of kids under 10 years old in
the household, flexible time, owning a beach cottage in Delaware or New
Jersey, student, working part time, or volunteer. No-trip utility increases with
age, retired, and working at home.
For site access values, we considered the closure of each of the 62 beaches
separately and the closure of groups of beaches (step 11). For the groups of
beaches we considered the 6 northern most Delaware beaches and 8 southern
most beaches. For the beach erosion scenarios we considered a narrowling of
all developed beaches in Delaware to less than 75 feet wide. We report seasonal
per person, seasonal aggregate (for Delaware residents), and discounted present
values (for Delaware residents) in each case.
The mean seasonal per person value, equation (43), for the loss a single
beach ranged from about $5 for the northern most beaches in New Jersey to
about $135 for the most popular beaches in Delaware and Maryland. These were
Rehoboth (DE), Cape Henlopen (DE), and Ocean City (MD). Recall that our
sample considers only Delaware residents and the scenario assumes all other
beaches remain open. These relative sizes make sense. The northern New
Jersey beaches are far away and have many nearby substitute. The higher
valued beaches are close to population centers and have many of the desirable
characteristic noted in the results above.
These values translate into mean seasonal aggregate losses, using equuation
(43), that range from $2.9 million for the least valued northern New Jersey
beaches to $77.8 million for the highest valued Delaware beaches. The relevant
population is all residents of Delaware over the age of 16 in 1997. That is
consistent with the sample frame. It is important to keep in mind that this value
excludes overnight trips, people from other states, and non-use values.
We are particularly interested the loss of groups of beaches. For example,
if an oil spill should occur, it is likely that more than a single a beach would be
lost. The northern beaches in Delaware are particularly vulnerable. Losing these
59
simultaneously gives a mean seasonal per person loss of $698 which translates
into an aggregate seasonal loss of $402 million. Losing the southern beaches
gives a mean seasonal per person loss of $554, an aggregate loss of $319
million. Again, overnight trip, out of state residents, and non use value is
excluded.
Loss of beach width due to erosion is a major issue on beaches in the MidAtlantic. So, we considered a scenario where all developed beaches with a width
of 75 feet or greater are narrowed to less than 75 feet. If a beach is not
developed it not likely to erode and so was excluded. To calculate an
individuals expected maximum utility with erosion, all developed beaches have
site utilities computed with wide = 0 and narrow = 1. The mean seasonal per
person loss for a narrowing of Delawares beaches is about $76. The
corresponding aggregate loss is $44 million.
60
Chapter 9
3.4 Variations
There are a number of variations on the basic RUM model and an active
research agenda advancing the technique (see Herriges and Kling (1999)). I will
mention a couple variations on the model I consider particularly important.
First, there is the application of Kuhn-Tucker models to recreation demand.
These models have features of both RUM models and demand systems. They
can be used to value quality changes and access and, it has been shown recently
that these models may be used in settings with a large number of sites (see
Phaneuf et. al. (2000) and Phaneuf et al. (2002)). Kuhn-Tucker models are
noted for their utility theoretic link between participation and site choice.
Second, there is increasing use of revealed and stated preference data in
combination in the context of RUM models. This enables an analyst to consider
behavior beyond the range of observable data and hence a much wider range of
policy options (see Adamowicz (1994)).
Third, a time dimension is being introduced into RUM models. This allows
trips to have interdependence over time and allows for time periods (eg.,
weekend versus weekday) to be treated differently. See Adamowicz (1994).
Fourth, the mixed logit (MX) mentioned in section 2.2.10 has been widely
adopted as a means of introducing complex patterns of substitution and
unobserved heterogeneity into the model. This model is almost certain to see
even wider use. See Train (1999).
Fifth, RUM models are often extended to include more than site choice.
They may incorporate choice target species for fishing, choice of boat or shore
in fishing, and even type of recreation. See Parsons and Hauber (1998) or
McConnell and Strand (1994).
Finally, the basic model presented here may be extended to overnight trips
but care must be taken to account for multiple purpose trips and to calculate
lodging cost accurately. Neither is easy. See Shaw and Ozog ( ) and Hoehn
et. al. (19 96).
61
Table 4: RUM Model for Mid-Atlantic Beaches

Site Characteristics:
Variable
Definition
tc
travel plus time cost
Length
Length of beach in miles
.13 (8.3)
Boardwalk
Boardwalk present = 1
.41 (6.3)
Amusement
Amusements nearby = 1
.48 (15.4)
Private
Private or limited access beach = 1
-.17 (6.3)
Park
State or federal park
.04 (0.6)
Wide
Wide beach (= 1 if > 200 feet)
-.33 (12.8)
Narrow
Narrow beach (= 1 if < 75 feet)
-.20 (5.4)
AC
Atlantic City = 1
.42 (7.2)
Surf
Good surfing = 1
.40 (15.5)
HighRise
High rises present on beach = 1
-.30 (9.4)
ParkWithin
Park located within the beach = 1
.25 (5.1)
Facilities
Bathhouse, restroom facilities

present = 1
-.05 (1.1)
Parking
Parking available at beach = 1
.13 (1.8)
New Jersey
New Jersey beach = 1
.51 (33.9)
IV(NJ)
Inclusive Value on New Jersey

Beaches
.49 (36.9)
IV(DE)
Inclusive Value on Delaware

Beaches
.99 (38.7)
IV (Beaches)
Inclusive Value on Beaches
2.06 (11.0)
Individual Characteristics:
Parameter Estimate (t-stat)

-.04 (63.9)
62
Chapter 9
Variable
Definition
Variable
Definition
Constant
.25 (5.3)
ln(age)
Log of age
.20 (7.0)
Kidsu10
Number of kids under 10 in

household
-.26 (9.4)
Flextime
Flexible time available in work

schedule
-.14 (3.4)
Cottage (DE)
Own beach property in Delaware =

1
-1.3 (25.5)
Cottage (NJ)
Own beach property in New Jersey

=1
-.80 (16.4)
Retired
Retired = 1
.53 (10.5)
Student
Student = 1
-.90 (19.5)
Parttime
Work part time =1
-.56 (13.3)
Workhome
Work at home = 1
.94 (12.0)
Volunteer
Work as a volunteer = 1
-.16 (2.6)
Sample Size
565
Mean Log- Like
-94.05
4. CONCLUSION
The traditional single site model and contemporary RUM model are the two
most widely used travel cost models in recreation demand. The RUM model is
the modern workhorse. It is able to account to a broad array of substitutes. It
is possible to value changes in site access as well changes in quality at one or
more sites. It tells a convincing and defensible story. And, software is readily
available for estimation. Most advances in travel cost modeling are taking place
63
in the context of RUM models. The single site model requires less data and is
easier to apply. In circumstances where one is interested in access value at only
one site and the number of substitute sites is not large, the single site model is
often used and is defensible.
NOTES
1.This overlooks other multiple site approaches such as conventional demand systems and the
hedonic travel cost model. Neither have been as popular as the RUM model. See Ward and
Beal (2000) for a discussion of conventional demand systems and Brown and Mendelsohn
(1984) for the hedonic travel cost. See Bockstael, McConnell, and Strand (1991) for a critique
of the hedonic travel cost model.
2.For this reason, their probabilities corresponded to Greenes truncated model (1997, p. 937)
instead of my equation (9).
3.Sometimes site utilities are written as v i = tc(y-tc i) + qq i, and no-trip utility is written as v0i
= tcy + where y is a measure of per trip income. In this way trip cost is seen quite
explicitly as reducing income available for other uses and hence lowering welfare. The
coefficient tc (now positive) is also readily interpreted as a marginal utility of income. In
estimation y is constant across sites and no-trip utility, provides no explanatory power for
choice of alternative, and hence drops out in estimation.
REFERENCES
Adamowicz, Wiktor. L. Habit Formation and Variety Seeking in a Discrete Choice Model
of Recreation Demand, Journal of Agricultural and Resource Economics 19(1): 19-31.
1994.
Adamowicz, W., J. Louviere, and M. Williams. Combining Revealed and Stated
Preference Methods for Valuing Environmental Amenities, Journal of Environmental
Economics and Management 26:271-92. 1994.
Adamowicz, Wiktor L. Joffre Swait, Peter Boxall, and Michael Williams. Perception
Versus Objective Measures of Environmental Quality in Combined Revealed and
Stated Preference Models of Environmental Evaluation, Journal of Environmental
Economics and Management 32: 65-84. 1997.
Andrews, Tom. 1996.
64
Chapter 9
Ben-Akiva, M. and S. Lerman. Discrete Choice Analysis. Cambridge MA: MIT Press, 1985.
Bockstael, Nancy E. "Travel Cost Models," Handbook of Environmental Economics.
Blackwell Publishers: Cambridge, MA. 1995.
Bockstael, Nancy E. and Ivar E. Strand. The Effect of Common Sources of Regression
Error on Benefit Estimates, Land Economics 63(1): 11-20. 1987.
Bockstael, Nancy, W. Michael Hanemann, and Catherine L. Kling. Estimating the Value
of Water Quality Improvements in a Recreational Demand Framework, Water
Resources Research 23(5): 951-960. 1987.
Bockstael, Nancy E., W. Michael Hanemann, and Ivar E. Strand, Jr. Measuring the Benefits
of Water Quality Improvements Using Recreation Demand Models. Report to the U.S.
Environmental Protection Agency. College Park, MD: University of Maryland. 1984.
Bockstael, Nancy E., Kenneth E. McConnell, and Ivar E. Strand. Benefits from
Improvements in Chesapeake Bay Water Quality. Report to the U.S. Environmental
Protection Agency. College Park, MD: University of Maryland. 1988.
Bockstael, N.E., K.E. McConnell, and I.E. Strand. A Random Utility Model for
Sportfishing: Some Preliminary Results for Florida, Marine Resource Economics 6:
245-260. 1989.
Bockstael, Nancy E., Kenneth E. McConnell, and Ivar E. Strand, Jr. "Recreation"
Measuring the Demand for Environmental Quality, North Holland: New York: 1991.
Bockstael, Nancy E., Ivar E. Strand, and W. Michael Hanemann. Time and the
Recreational Demand Model, American Journal of Agricultural Economics 293-302.
May 1987.
Boyle, Kevin. Marine Resource Economics.
Brown, G. and R. Mendelsohn. The Hedonic Travel Cost Method, Review of Economics
and Statistics 66: 427-433. 1984.
Chen, H.Z. and S.R. Cosslett. Environmental Quality Preference and Benefit Estimation
in Multinomial Probit Models: A Simulation Approach, American Journal of
Agricultural Economics 78(3). 1998.
Cicchetti, Charles J., Anthony C. Fisher, and V. Kerry Smith. An Econometric Evaluation
of a Generalized Consumer Surplus Measure: The Mineral King Controversy,
Econometrica 44(6): 1259-1276. 1976.
Creel, Michael D. and John B. Loomis. Theoretical and Empirical Advantages of
Truncated Count Data Estimators for Analysis of Deer Hunting in California, Amer.
J. Agr. Econ. 434-441. May 1990.
Feather, Peter and W. Douglas Shaw. Estimating the Cost of Leisure Time for Recreation
Demand Models, Journal of Environmental Economics and Management 38:49-65.
65
1999.
Freeman, A. Myrick III. The Measurement of Environmental and Resource Values:
Theory and Methods. Resources for the Future: Washington, D.C. 1993.
Greene, William H. Econometric Analysis (3rd Ed.). Prentice Hall: New Jersey. 1997.
Haab, Timothy C. and Kenneth E. McConnell. Count Data Models and the Problem of
Zeros in Recreation Demand Analysis, American Journal of Agricultural Economics
78: 89-102. 1996.
Haab, Timothy C. and Kenneth E. McConnell. Valuing Environmental and Natural
Resources. Edward Elgar: Cheltenham, UK 2002.
Hanemann, W. Michael. Welfare Analysis with Discrete Choice Models. Valuing
Recreation and the Environment: Revealed Preference Methods in Theory and
Practice, Edward Elgar: Cheltenham, UK. 1999.
Hauber, Albert H. IV and G.R. Parsons. The Effect of nesting Structure on Welfare
Estimation in Random Utility Models: An Application to the Demand for Recreational
Fishing, American Journal of Agricultural Economics 82: 501-514. 2000.
Hausman, Jerry A., Gregory K. Leonard, and Daniel McFadden. A Utility-Consistent,
Combined Discrete Choice and Count Data Model: Assessing Recreational Use Losses
Due to Natural Resource Damage, Journal of Public Economics 56:1-30. 1995.
Hellerstein, Daniel. Can We Count on Count Models? Valuing Recreation and the
Environment: Revealed Preference Methods in Theory and Practice, Edward Elgar:
Cheltenham, UK. 1999.
Hellerstein, D. and R. Mendelsohn. A Theoretical Foundation for Count Data Models,
Amer. J. Agr. Econ. 75: 604-11. 1993.
Herriges, J.A. and C.L. Kling, (eds.). Valuing Recreation and the Environment: Revealed
Preference Methods in Theory and Practice. Edward Elgar: Aldershot U.K. 1999.
Herriges, Joseph A., Catherine L. Kling, and Daniel J. Phaneuf. Corner Solution Models
of Recreation Demand: A Comparison of Competing Frameworks, Chapter 6 in
Herriges and Kling (1999). Edward Elgar: Cheltenham, UK. 1999.
Hicks, R.L. and I. Strand. The Extent of Information: Its Relevance for Random Utility
Models, Land Economics 76 (2). 2000.
Hoehn, J.P., Theodore Tomasi, Frank Lupi, and Heng Z. Chen. An Economic Model for
Valuing Recreational Angling Resources in Michigan, Michigan State University,
Report to the Michigan Department of Environmental Quality. 1996.
Kolstad, Charles D. Environmental Economics. Oxford University Press: New York, 2000.
Krutilla and Smith (1975)
66
Chapter 9
Laitila, Thomas. Estimation of Combined Site-Choice and Trip-Frequency Models of

Recreational Demand Using Choice-Based and On site Samples, Economic Letters 64:
17-23. 1999.
Layman, R. Craig, John R. Boyce, and Keith R. Criddle. Economic Valuation of the
Chinook Salmon Sport Fishery of the Gulkana River, Alaska, under Current and
Alternate Management Plans, Land Economics 72(1): 113-28. 1996.
Lupi, F. and P.M. Feather, Using Partial Aggregation to Reduce Bias in Random Utility
Travel Cost Models, Water Resources Research 34(12): 3595-603. 1998.
McConnell, K.E. The Damages to Recreational Activities from PCBs in New Bedford
Harbor. Prepared for the Ocean Assessment Division, National Oceanic and
Atmospheric Administration, December 1986.
McConnell, K.E. On site Time in the Demand for Recreation, American Journal of
Agricultural Economics 74: 918-25. 1992.
McConnell, K E., I. Strand, and L. Blake-Hedges. Random Utility Models of Recreational
Fishing: Catching Fish Using a Poisson Process, Marine Resource Economics 10: 247261. 1995.
McConnell, K.E. and I.E. Strand. The Economic Value of Mid and South Atlantic
Sportfishing. University of Maryland, Report to the USEPA and NOAA. 1994.
McFadden, D. Economic Choices, American Economic Review 91(3): 351-378. 2001.
Mendelsohn, R., J. Hof, G. Peterson, and R. Johnson. Measuring Recreation Values with
Multiple Destination Trips, American Journal of Agricultural Economics 74: 926-33.
1992.
Morey, Edward. The Demand for Site-Specific Recreational Activities: A Characteristics
Approach, Journal of Environmental Economics and Management 8:345-371. 1981.
Morey, Edward. Two Rums Uncloaked: Nested-Logit Models of Site Choice and NestedLogit Models of Participation and Site Choice, Chapter 3 in Herriges and Kling
(1999). Edward Elgar: Aldershot U.K. 1999.
Morey, Edward, Robert D. Rowe, and Michael Watson. A Repeated Nested-Logit Model
of Atlantic Salmon Fishing, American Journal of Agricultural Economics 75: 578-592.
1993.
Ozuna, T. and I.A. Gomez, Estimating a System of Recreation Demand Function Using a
Semmingly Unrelated Poisson Regression Approach, Review of Economics and
Statistics 76: 356-60. 1994.
Parsons, George R. A Note on Choice of Residential Location in Travel Cost Demand
Models, Land Economics 67(3): 360-364. 1991.
Parsons, George R. and Mary Jo Kealy. Randomly Drawn Opportunity Sets in a Random
67
Utility Model of Lake Recreation, Land Economics 68(1): 93-106. 1992.

Parsons, George R. and Michael Needelman. Site Aggregation in a Random Utility Model
of Recreation, Land Economics 68(4): 418-433. 1992.
Parsons, George R. and Aaron J. Wilson. Incidental and Joint Consumption in Recreation
Demand, Agricultural and Resource Economics Review. 1997.
Parsons, George R. and A. Brett Hauber. Spatial Boundaries and Choice Set Definition in
a Random Utility Model of Recreation Demand, Land Economics 74(1): 32-48. 1998.
Parsons, George R., Paul M. Jakus, and Ted Tomasi. A Comparison of Welfare Estimates
from Four Models for Linking Seasonal Recreational Trips to Multinomial Logit
Models of Site Choice, Journal of Environmental Economics and Management 38:
143-157. 1999.
Parsons, George R., D. Matthew Massey, and Ted Tomasi. Familiar and Favorite Sites in
a Random Utility Model of Beach Recreation, Marine Resource Economics 14: 299315. 1999.
Parsons, George R., Andrew J. Plantinga, and Kevin J. Boyle. Narrow Choice Sets in a
Random Utility Model of Recreation Demand, Land Economics 76(1): 86-99. 2000.
Parsons, George R. and D. Matthew Massey. A RUM Model of Beach Recreation, in The
New Economics of Outdoor Recreation, ed. Nick Hanely and Douglass Shaw,
forthcoming 2002.
Peters, T., W.L. Adamowica, and Peter C. Boxall. Influence of Choice Set Considerations
in Modeling the Benefits from Improved Water Quality, Water Resources Research
31(7): 1781-1787. 1995.
Phaneuf, Daniel J., C. L. Kling, and J. Herriges . Estimation and Welfare Calculation in
a Generalized Corner Solution Model With an Application to Recreation Demand,
Review of Economics and Statistics 82: 83-92. 2000.
Phaneuf, Daniel J. and V. Kerry Smith. Recreation Demand Models Manuscript, North
Carolina State University, 2002.
Randall, Alan. A Difficulty with the Travel Cost Method, Land Economics 70(1): 88-96.
1994.
Shaw, D. On site Samples Regression Problems of Nonnegative Integers, Truncation, and
Enodgenous Stratification, Journal of Econometrics 37: 211-223. 1988.
Shaw, W. Douglass and Paul Jakus. 1996.
Shaw and Ozog, 1999.
Shonkwiler, J.S. Recreation Demand Systems for Multiple Site Count Data Travel Cost
Models, Chapter 9 in Herriges & Kling (1999). Edward Elgar: Cheltenham, U.K.
68
Chapter 9
1999.
Shonkwiler, J.S. and W. D. Shaw. Hurdle Count-Data Models in recreation Demand
Analysis, Journal of Agricultural and Resource Economics 21(2): 210-219. 1996.
Siderelis, Christos, Gene Brothers, and Phil Rea. A Boating Choice Model for the
Valuation of Lake Access, Journal of Leisure Research 27(3): 264-282. 1995.
Siderelis, Christos and Roger Moore. Outdoor Recreation Net Benefits of Rail-Trails,
Journal of Leisure Research 27 (4): 344-359. 1995.
Smith, V.K., W.H. Desvousges, and M.P. McGivney. The Opportunity Cost of Travel Time
in Recreation Demand Models, Land Econom. 59(3): 259-277. 1983.
Smith, V. Kerry and William H. Desvousges. The Generalized Travel Cost Model and
Water Quality Benefits: A Reconsideration, Southern Economics Journal 52(2): 371381. 1985.
Smith, V. Kerry, William H. Desvousges, and Matthew P. McGivney. Estimating Water
Quality Benefits: An Econometric Analysis, Southern Economic Journal 50(2), 422437. 1983.
Sohngen, Brent. The Value of Day Trips to Lake Erie Beaches, Dept. of Agricultural,
Environmental, and Development Economics, Ohio State University. 2000.
Train, Kenneth E. Mixed Logit Models for Recreation Demand, Chapter 4 in Herriges
and Kling (1999). Edward Elgar: Aldershot U.K. 1999.
Ward, Frank A and D. Beal. Valuing Nature with Travel Cost Models: A Manuel. Edward
Elgar: Cheltenham, UK. 2000.

Travel Cost Method

Uploaded by

Copyright:

Available Formats

Travel Cost Method

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Travel Cost Method

Uploaded by

Copyright:

Available Formats

.

a beach closure due to an oil spill,

improvements in water quality on lakes and rivers

THE TRAVEL COST MODEL

THE SINGLE SITE MODEL

THE TRAVEL COST MODEL

Figure 1. Access Value in a Linear Single Site Model

A more general expression for consumer surplus or access value, which

2.2 Steps in Estimation

THE TRAVEL COST MODEL

delineation is not so clear. A river for fishing or white-water rafting, for

Define the Site to be Valued

Define the Recreation Uses and the Season

Develop a Sampling Strategy

Specify the Model

Decide on the Treatment of Multiple Purpose Trips

Design and Implement the Survey

Measure Trip Cost

Estimate the Model

Calculate Access Value

2.2.2 Step 2: Define the Recreation Uses and the Season

THE TRAVEL COST MODEL

response. It also possible to model the individuals decision of whether or not

THE TRAVEL COST MODEL

of shifters. Studies typically include anywhere from one to five variables.

2.2.5 Step 5: Decide on the Treatment of Multiple Purpose Trips

THE TRAVEL COST MODEL

recent examples. Mendelsohns framework broadens the definition of a site to

The introductory material introduces the interviewer, identifies his or her

THE TRAVEL COST MODEL

persons is it necessary to include a question that identifies whether or not the

THE TRAVEL COST MODEL

Despite the difficulty of extrapolating the simple flexible leisure/work model

THE TRAVEL COST MODEL

The parameter is the expected number of trips and is assumed to be a function

(6) ln( ) = tcr tcr + tcs tcs + y y + d z .

An individual is denoted by n = 1, ... , N, so r n is the number of trips taken

THE TRAVEL COST MODEL

a mean seasonal value per person,

exp( tcr tcrn + tcs tcsn + y yn + d zn )

THE TRAVEL COST MODEL

(13) AS = Soff POPoff

where dividing the individual surpluses by r n and the summation by N* are

Soff POPoff , on the other hand, includes

THE TRAVEL COST MODEL

THE TRAVEL COST MODEL

regression. The effect of substitute sites on trips works as expected in the

Notes: Three asterisks denotes that a coefficient that is statistically significant at

THE TRAVEL COST MODEL

3. THE RANDOM UTILITY MODEL

THE TRAVEL COST MODEL

(20) tc tck + q qk + ek tc tci + q qi + ei for all i S .

A useful way of expressing this result is in terms of trip utility. An individuals

Choice occasion utility is the maximum attainable utility on a given choice

where z is a vector of characteristics believed to influence a persons propensity

THE TRAVEL COST MODEL

(25) vi = tc tci + q1 ( q1i z1 ) + q2 q2 i + ... + qm qmi + ei

(26) u *(baseline) = max{v0 , v1, v2 ..., vS } .