Annals of Operations Research 136, 329–349, 2005

c 2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Location of Service Facilities for the Elderly

MICHAEL P. JOHNSON ∗ [email protected]
WILPEN L. GORR
STEPHEN ROEHRIG
H. John Heinz III School of Public Policy and Management, Carnegie Mellon University, Pittsburgh,
PA 15213-3890, USA

Abstract. Senior centers provide a variety of supportive services for independent elderly adults. In many
metropolitan areas, the elderly population is growing and redistributing from central cities to suburbs, where
accessibility to senior centers is limited. Policy analysts need to locate senior centers to best meet changing
demands for service. We present alternative hierarchical facility location models for senior centers applied
to Allegheny County, Pennsylvania. We find that a model that minimizes consumer disutility and unserved
demands is preferred to one that maximizes utility alone, and that the former model is well-behaved in
response to changes in structural parameters.

Keywords: facility location, elderly services, demand forecasting, integer programming

Introduction

The elderly population in the United States has increased in size and in proportion to the
overall population throughout the 20th century. These trends are expected to continue over
the next 50 years, as the “baby boom” generation enters retirement years. “Middle-series”
population projections estimate that the population of those 65 or older, which increased
from 9.9 percent of the U.S. population in 1970 to 12.6 percent in 1990, will increase
to about 20.3 percent in 2050 (Federal Interagency Forum on Aging Related Statistics
2000). Moreover, services for elderly persons—persons aged 60 or more—are in high
demand and are given prominence by advocates for the elderly. These services include
home-delivered means for the homebound, transit for those facing mobility obstacles,
and senior centers.
Senior centers are facilities for elderly adults who live independently that provide
services such as lunchtime congregate meals, socialization/entertainment, and referrals
to other service agencies. The focus of this paper is on the design of planning models for
the location of senior centers across large study areas, such as a county or metropolitan
area, and the application of a particular planning model to a portion of Allegheny County,
PA, a region with high demands for senior-related services, including senior centers.
From a policy perspective, this problem is significant. Increasing suburbanization,
and aging of suburban residents “in place” means that demand patterns for senior services
∗
Corresponding author.
330 JOHNSON, GORR AND ROEHRIG

have changed over time, decreasing emphasis on service providers in the central city that
are traditionally accessible by mass transit and which provide comprehensive coverage
of their service areas, and increasing emphasis on suburban service providers that are
less accessible by mass transit and which may leave considerable service gaps in areas
with newer infrastructure. In addition, limitations in government budgets have placed
restrictions on the number and type of senior services that can be offered, imposing new
requirements on service providers to measure and rationalize their services, and imposing
on coordinating agencies the need for proactive planning based on population data and
quantitative models.
Location of senior centers is interesting because it has not traditionally been a focus
of operations research/management science inquiry. The problem of forecasting demands
for senior centers requires close examination of service providers and use of geographic
information systems (GIS) applications. Design of a location planning model for senior
centers using these demand forecasts combines elements of a number of location mod-
eling strategies. This paper exploits an opportunity to demonstrate significant returns to
the use of data-driven planning that other, more mature areas of public-sector operations
research/management science may not enjoy.
The goals of this paper are: to design optimization-based location models for senior
centers using domain-specific demand forecasts, to implement these models using integer
programming modeling and solver packages, and to interpret model outputs so as to
provide non-trivial policy insight to public-sector planners. We hope that this paper will
enable others to extend our forecasting and optimization modeling work.
There are no papers known to us that present quantitative planning models for the
location of senior centers. Recent research results related to this paper include: point
demand forecasting models for home-delivered meals (HDM), which extend traditional
area-based forecasting models to identify specific demand points and levels of demands at
those points (Gorr, Johnson, and Roehrig, 2001a); an interactive GIS-based heuristic for
location of HDM kitchens, combining facility location with vehicle routing for catchment
area design (Gorr, Johnson, and Roehrig, 2001b), and an integer programming (IP)
formulation of the associated location-routing problem, aspects of which are solved
using standard optimization methods and results compared to current operating procedure
(Johnson, Gorr, and Roehrig, 2002).
The models in this paper are based on three insights. First, reflecting available data
on senior service demand, we design a demand forecasting technique to address a service
whose attractiveness is distance-based. Next, based on observed demand characteristics
for senior centers, we advocate an integrated integer programming modeling approach, as
opposed to spatial interaction models or strategies based on spatial decomposition. Last,
we develop two hierarchical facility location models that capture relevant policy char-
acteristics: one that maximizes facility utilization, and another that minimizes consumer
disutility for travel.
We find that the senior center facility location model that minimizes the sum of
demand-weighted distance (a measure of consumer disutility) plus the total number
of unserved demands generates a solution that is more efficient, in terms of facility
LOCATION OF SERVICE FACILITIES FOR THE ELDERLY 331

utilization, and more effective, in terms of total demands served, than one that maximizes
utilization alone. In addition, the solutions generated by the former model are much
more spatially coherent, in terms of the fraction of demand nodes served by two or more
facilities, and compact, in terms of the fraction of demand nodes served by the nearest
senior center, than those generated by the latter model. Finally, the former model is quite
stable in response to changes in values of a variety of dollar-based structural parameters.
The remainder of the paper is structured as follows. Section 1 examines the elderly
population and senior service centers in Allegheny County, PA (our study area) and es-
timates a statistical regression model to estimate demands for senior centers. Section 2
uses demand forecasting results to justify an integer programming-based facility loca-
tion modeling approach. Section 3 presents a facility location IP model to maximize
utilization of senior centers, estimates various model parameters and examines solution
characteristics. Section 4 presents an alternative IP formulation that minimizes a sum of
consumer disutility for travel and unsatisfied demand and examines solution character-
istics. Section 5 concludes and identifies research extensions.

1. Nature of elderly services in Allegheny County

1.1. Elderly population characteristics

The senior center location problem has its roots in the authors’ efforts to provide quan-
titative planning expertise to elderly client-serving agencies. These research initiatives
have focused on needs for elderly-related services in Allegheny County, PA. Recent field
research supervised by the authors1 has found that while Allegheny County had the high-
est number of elderly persons in Pennsylvania and is 13th in the U.S. in this category, the
numbers of persons over 65 had actually declined between 1990 and 2000. After 2000,
aging younger cohorts will probably result in a net increase in the elderly population.
Though the elderly population is well-represented in Census tracts throughout Al-
legheny County, the greatest concentrations of elderly population are found in the city of
Pittsburgh and surrounding first-ring suburbs. However, between 1990 and 2000, among
all age categories of persons 55 and older, suburban Allegheny County showed either
smaller population decreases or larger population increases than the city of Pittsburgh.
Long-term secular patterns of suburbanization in Allegheny County are likely to require
that elderly services, such as senior centers, respond creatively to the changing spatial
distribution of elderly clients.

1.2. Senior center characteristics

Based on a survey of 68 senior-serving agencies in Allegheny County, we have classified

the 42 senior centers that responded to the survey in terms of programming, personnel,
space utilization and transportation and fees. Senior centers in this small sample have
been divided into three classifications: 8 “type 1” centers that serve 50 or more clients
332 JOHNSON, GORR AND ROEHRIG

daily and are open for about seven hours; 18 “type 2” centers that serve 30–50 clients
daily and are open for about seven hours, and 16 “type 3” centers that serve fewer than
30 clients daily and are open for six or fewer hours. Considering five categories of senior
center services—meals and nutrition, social activities, health and wellness, information
and referral, and education programs, type 1 and 2 senior centers tend to be strong across
all service categories, while type 3 senior centers tend to be significantly weaker in the
areas of health and wellness and education.
Senior centers tend to have more employees overall, and more full-time and better-
paid staffers as the service level increases, with type 1 senior centers significantly more
advantaged than other centers, especially regarding use of volunteers. Most senior centers
lease their space and share their space with other community organizations; estimates of
actual space used range so widely as to be unreliable for planning purposes. Interestingly,
there was little difference between city and suburban senior centers regarding client transit
modes; most senior centers report that a majority of their clients drive to the site, and
such centers are roughly equally divided between city and suburbs. Finally, most senior
centers charge only nominal fees for their services.
Defining services provided by senior centers in terms of “demand units” (e.g.
meals), defining clients as those persons who have consumed a certain number of service
units annually from specific senior centers, and using the GIS methods of geocoding and
buffering, we are able to define the usage rate (fraction of all elderly persons who are
clients) by distance from senior centers. Our unit of analysis is the Census block group—
a standard areal unit defined to have a population of about 1,000 persons within which
demographic characteristics show limited variation. Our primary outcome measures are
average use rate—the number of clients residing in block groups within specified dis-
tances (e.g. 0.25 miles or less) of any senior center, divided by the number of persons
aged 60 or older within the same block groups—and marginal use rate—the number of
clients residing in block groups within a given distance increment (e.g. between 0.125
and 0.25 miles) of a senior center divided by the number of persons aged 60 or older
within the same block groups. Figure 1 (below) shows that average use rates decline from
10.6 percent for block groups less than 1/8 mile from the nearest senior center to 5.4
percent for block groups 2 or more miles away. Marginal use rates decline more steeply,
from 10.6 percent to 2.0 percent over the same distance.
Use rates differ substantially by urban/non-urban classification: average use rates
in urban areas range from 14.3 percent for block groups less than 1/8 mile away from a
senior center to 9.1 percent for those two miles away, while use rates in non-urban areas
range from 5.6 percent to 0.5 percent for the same distance standard.
As is the case for many public services, especially those provided with significant
volunteer support, in a decentralized manner and to lower-income clients, the financial
performance of senior centers is difficult to measure. Relying on the same survey of
elderly client-serving agencies described above, three primary agency-level measures
are defined: percentage of total operating costs that support administration, cost per
service unit, and cost per consumer. These measures vary quite widely and indicate that
agencies suffer diseconomies of scope while not enjoying economies of scale. The fact
LOCATION OF SERVICE FACILITIES FOR THE ELDERLY 333

Figure 1. Senior center average and marginal use rates versus distance.

that these cost data are self-reported make them merely suggestive of the actual cost
structure of senior centers in Allegheny County; a more reliable estimate will require
data collection by an independent body. In our planning model to follow, we will use a
different, and better-behaved, estimate of senior center fixed and variable costs.

1.3. A demand forecasting model for senior centers

Observed use rates have been shown to decline nonlinearly with respect to distance from
a senior center, but not to converge to zero within a reasonable distance from a senior
center. For the purposes of a location planning model, it is useful to estimate demand
levels for current and proposed senior centers that account for variation in spatial and
demographic characteristics of areas in which clients live and are consistent with known
variations in use rates. We have estimated such a model. We assume that income, size of
elderly population and race affect use rate, in addition to distance from senior centers. In
our regression model, observations were the 1,107 block groups in Allegheny County;
the dependent variable was the measured use rate according to data provided by elderly
client-serving agencies; and independent variables were the inverse of distance between
the block group and the nearest senior center, population density, the fraction of males
in their 60s and the fraction of females in their 70s. Regression results are contained in
Table 1, below.
This model explains only 24 percent of all variance in the values of the dependent
variable, though all of the independent variables have the expected sign and are statis-
tically significant. Despite the deficiencies of this forecasting model, we believe that it
represents a plausible alternative to a modeling approach that relies entirely on distance
between demand points and facilities as the primary measure of user utility/disutility.
Refining this model to reduce the amount of unexplained variance is a subject of ongoing
research.
334 JOHNSON, GORR AND ROEHRIG

Table 1
Regression results for senior center use rate model.
Regression
Variable Definition Avg. value coefficients
USE RATE Fraction of all elderly residents in a Census 0.049 –
block group who are clients of a senior center
Constant – – −0.0890∗∗
INV DIST Inverse of the distance, in miles, between the 1.249 0.01∗∗
block group centroid and the nearest
senior center
POP DENS Number of persons in the block group per 5587 0.0000012∗∗
square mile
PCT M60S Fraction of males in the block group aged 0.178 0.1786∗∗
60 or older
PCT F70S Fraction of females in the block group aged 0.231 0.3154∗∗
70 or older
Adjusted R-squared 0.243
F-statistic 89.809
†
Observations represent 1,107 Census block groups in Allegheny County, PA.
∗
Coefficient significant at 5% level.
∗∗
Coefficient significant at 1% level.

We estimate total usage by multiplying a block group’s estimated use rate by the
number of persons aged 60 or older in each block group. Note that in contrast to Figure 1,
which indicates a sharp decline in facility utilization with respect to proximity to senior
centers, the model estimated in Table 1 implies a much smaller impact of distance upon
estimated demand levels. For a given block group, the only independent variable in
the estimated demand model that varies is inverse distance to the nearest senior center.
Estimating demands between block groups and all senior centers, even those that may not
currently serve residents of a particular block group, we find that the contribution of the
inverse distance term to total estimated demand (0.01∗ INV DIST times the number of
elderly residents) is small and highly skewed: the average contribution to total estimated
demand is 0.77 demands, with a standard deviation of 1.58, a minimum value of zero
and a maximum value of 51.33.
The fact that estimated demand levels by block group show relatively little variation
over potential senior center locations means that the friction of distance, or “distance
decay” effect, in terms of spatial interaction models (Fotheringham and O’Kelly, 1989)
is relatively small. This implies that there is significant uncertainty as to the facility
that a particular user might patronize, given known characteristics about the user and
the service environment. As a result, the conventional assumption of location-allocation
models that consumers travel to the nearest center to receive services would appear to
be supported only modestly. Thus, this facility location problem may require the use of
spatial interaction models to best capture observed user behavior. However, for simplicity,
LOCATION OF SERVICE FACILITIES FOR THE ELDERLY 335

the facility location models used this paper will not integrate spatial interaction models.
Doing so is a subject for further research.

2. Application of demand forecasting model to facility location modeling

There are two contrasting strategies one might use to locate senior centers using the
demand estimation model from the previous section: an incremental strategy and a com-
prehensive strategy. The incremental approach is inspired by Brandeau and Chiu (1994),
who locate public facilities on a tree network by identifying a user-equilibrium configura-
tion. Given an existing network of senior centers and well-defined service areas for each
center, a similar strategy begins by defining a senior center’s catchment area as all block
groups within three miles of a senior center outside the city of Pittsburgh, and for sites
within the city, all block groups within one mile. Rivers and other natural barriers define
other boundaries to the catchment area. The next step estimates demands for proposed
senior centers, and chooses the senior center to add to the network that optimizes some
measure of interest, say total utilization. As the configuration of senior centers changes,
demand levels for each block group would be recomputed using the regression model
in Table 1, where the choice of the closest facility to the block group must account for
capacity restrictions. This process would repeat until all demands are assigned or budget
and/or capacity restrictions hold.
An alternative approach uses a partial-equilibrium assumption common in integer
programming-based facility location models (ReVelle, 1987; Daskin, 1995). In this view,
alternative configurations of facilities do not affect the markets for goods relevant to the
planning model. In particular, creation of a new configuration of service facilities does not
affect markets for activities provided by these centers. Thus, the planning problem may
be solved in a single step, using mathematical programming. The demand forecasting
model of the previous section may be adapted for use in this planning framework as
follows. For a given existing or potential facility, estimate the number of potential clients
associated with every demand node in the network. By repeating this process for every
facility, we estimate the total number of potential clients at each demand node that could
patronize all facilities as well as the total number of potential clients at each demand
node that could patronize any particular facility.
For reasonably-sized problem instances, it is likely that the integrated optimization
modeling approach based on partial-equilibrium analysis is easier to implement than the
full-equilibrium analysis. It is also more useful from a strategic planning perspective,
when configurations of facilities over a large study area could change substantially.
However, the fundamental assumption of the demand forecasting strategy used, that
demands at one location are independent of demands at another location, is problematic.
We venture that for relatively low demand levels, and minimal perceptions of service
externalities, this approximation is not unreasonable.
By computing utilization rates for all demand nodes and all current facilities as
described above, we find that the average estimated utilization rate is 0.176, with a
336 JOHNSON, GORR AND ROEHRIG

Figure 2. Study area for elderly facility location model—Eastern Allegheny County, PA.

standard deviation of 0.133, minimum value of 0 and a maximum value of 0.86. We

believe that these estimated utilization rates, combined with observed utilization rates of
0.15 or less as shown in figures 1 and 2, justify a partial-equilibrium analysis. However,
for services that command significantly larger market shares, or which exhibit strong
user externalities, or when a small number of new facilities are to be sited, an iterative,
full-equilibrium approach may be preferable.
The partial-equilibrium modeling approach is incorporated into two alternative
facility location modeling frameworks: covering models (Toregas, et al., 1971; Church
and ReVelle, 1974) and median models (Hakimi, 1964, 1965). Set covering and p-center
models assume that benefits per unit of demand for services is roughly constant over a
coverage distance S between the facility and the demand node, then drop quickly to zero
for distances greater than S. p-median and fixed-charge models assume that benefits per
unit of demand decrease in the distance between demand points and facilities. Though
results of the previous section indicate that usage rates computed using a forecasting
model decrease more slowly than observed usage rates computed directly from senior
center client records, it is appears that the nature of traveled-to services provided at senior
centers, and the likelihood that elderly clients face constraints on transit accessibility
imply that median models are more appropriate than center models for senior center
location planning.
p-median models minimize demand-weighted distance, or average distance, be-
tween demand points and facilities. Thus, they optimize a measure of efficiency. ReVelle
LOCATION OF SERVICE FACILITIES FOR THE ELDERLY 337

(1987) addresses the need for incorporating distance-dependent demands for services by
presenting a median-based model for maximizing consumer utilization, which is assumed
to decrease linearly over a coverage distance S as a function of the distance between de-
mand points and the nearest facility. This model optimizes a measure of effectiveness.
Further, ReVelle quotes earlier work (ReVelle, Church, and Schilling, 1975) to show that,
if all demands can be met by facilities located within a distance S, then the maximum
utilization model and the minimum average distance models are equivalent.
This analysis is not wholly applicable to the senior center location problem for two
reasons. First, the analysis of demands for senior centers presented above does not indicate
that average use rate declines to zero for a reasonably large travel distance.2 Second, the
problem we solve in this paper incorporates capacity constraints, which imply that, in an
optimal solution, clients may not patronize the facility nearest to them. However, it may
be that for elderly services with the special demand characteristics described earlier,
effectiveness criteria (maximizing utilization) are closely related to efficiency criteria
(minimizing average distance). Since maximizing utilization is the primary policy goal
of the elderly-serving agencies in Allegheny County surveyed recently, this will be the
primary objective we pursue initially.

3. A utilization-maximizing senior center location model

3.1. Formulation

In the previous section, we argued that an integrated optimization modeling approach is

preferred to an iterative, full-equilibrium modeling approach. We also argued that, under
certain conditions, maximizing utilization—the goal of elderly client-serving agencies—
is preferable to minimizing travel costs, but that the two are closely related. Interviews
by these authors with agency administrators did not indicate that equity, or the perceived
fairness, from one stakeholder group’s perspective, of alternative potential facility con-
figurations, was a dominant concern. Thus, the senior center location model that follows
has a single objective—maximizing utilization.
Another important consideration is the notion of differential service levels presented
earlier. Define k as an index of services, k = 1, 2, . . . , p, where k increases in the quality
of the service provided. Redefine the senior center service levels presented in Section II
so that k = 1 denotes “type-3” senior centers, k = 2 denotes “type-2” centers and k = 3
denotes “type-1” centers. Using definitions of services provided by hierarchical networks
(Tien, El-Tell, and Simons, 1983; Narula, 1986) and details of senior centers presented
above, we argue that:

• The facility hierarchy is successively inclusive, i.e. a type k facility provides all the
services of a type k-1 facility plus at least one additional service, and
• The service hierarchy is globally inclusive, i.e. a type k facility provides k and lower
levels of service to customers at all nodes.
338 JOHNSON, GORR AND ROEHRIG

Assume that a potential type k facility at node j has capacity C jk < ∞, a fixed cost
f jk and a per-client variable cost v jk . The fixed cost f jk is composed of a capital cost
(for new facilities) and an operating cost (for all facilities). Finally, assume that there is
a total budget amount B for building and operating new facilities. Thus:

• The maximum number of facilities to be located is implied by the budget constraint

and does not have to be defined explicitly;
• Some demands may not be served in an optimal solution to the elderly service center
location problem, and
• A demand node may be served by two or more facilities, i.e. the nearest facility to a
given demand node may not serve all demands.

The demand forecasting model in the previous section computed the total potential
demand, at each demand point, for senior services provided throughout the study area. We
believe that these demands are mildly distance-dependent, i.e. that the number of seniors
that would visit various centers will decrease as a function of the distance from the
centers. However, to reflect expressed concerns of senior center administrators, we seek
to locate senior centers in such a way as to maximize total utilization. Thus, our model
does not optimize a function of distance traveled, though we impose limits on patronage
of specific facilities that reflect the mildly distance-dependent nature of demands. The
model definition follows.

Sets:

i = 1, 2, . . . , n: set of demand nodes (Census block group centroids); I = {i}

j = 1, 2, . . . , m: set of potential facility locations; J = { j} and J ⊆ I
k = 1, 2, . . . , p: set of facility service level categories; K = {k}

Data:

λik = total demand, for type k senior center services at demand node i
ei jk = upper bound on the number of demands at node i for type k services that may be
provided at facility location j
f jk = fixed annual cost of locating a type k facility at candidate location j
v jk = variable per-client annual cost of locating a type k facility at candidate location j
B = total budget available for locating all facilities
C jk = capacity of type k facility at candidate location j

Decision Variables:


1, if a facility of type k is located at candidate site j
X jk =
0, otherwise
LOCATION OF SERVICE FACILITIES FOR THE ELDERLY 339

Yi jk = number of demands at node i for type k services that are assigned to a facility at
candidate site j
Model SCL (Senior Center Location):

n 
m 
p
Maximize Yi jk (1)
i=1 j=1 k=1

s.t.:

p
Yi jk ≤ λik · X jh ∀i, j, k (2)
h=k

p
X jk ≤ 1 ∀ j (3)
k=1
m
Yi jk ≤ λik ∀i, k (4)
j=1
 

m 
p 
n
f jk · X jk + ν jk · Yi jk ≤B (5)
j=1 k=1 i=1
n
Yi jk ≤ C jk ∀ j, k (6)
i=1
X jk ∈ {0, 1} ∀ j, k (7)
0 ≤ Yi jk ≤ ei jk and integer∀i, j, k (8)

Objective (1) maximizes the total number of demands for all types of services
that are provided. We have previously argued that this effectiveness objective is closely
related to an efficiency objective.
Constraints (2) require that demands for type k service at demand node i cannot
be provided by a type k facility unless there is a facility sited at potential location j that
provides type k or higher service. The multiplier λik on the right-hand-side ensures that
the right-hand-side will always be greater than the left-hand-side yet provides a tight
constraint upper bound for the linear programming relaxation.
Constraints (3) enforces that only a single facility (of any service type) may be
sited at potential facility location j. Constraints (4) enforce that the number of demands
at node i for type k services actually served by all facilities cannot exceed the total
number of demands at node i for type k services (the left-hand-side may be strictly less
than λik due to budgetary and capacity constraints that could result in unmet demand).
Constraint (5) ensures that the sum of fixed and variable costs associated with facility
location decisions cannot exceed the total budget B. Constraints (6) ensures that the total
demands for type k service assigned to potential location j cannot exceed the capacity
C jk . Constraints (7) and (8) define integrality and bounds for decision variables.
340 JOHNSON, GORR AND ROEHRIG

Model (1)–(7) has o(n . m . p) decision variables and constraints. A model instance
consisting of 1000 Census blocks, 100 potential facility locations and 3 service levels
results in o(105 ) decision variables and constraints. Since this paper focuses on examining
solution characteristics and sensitivity analysis, we apply model SCL to a significantly
smaller dataset and solve the model using a commercially available software package.

3.2. Data

We apply model SCL to a subset of Allegheny County, the portion bordered on the south
and north by the Monongahela and Allegheny rivers, and on the east by the Allegheny
County border. This region was chosen because of the relatively high density of existing
senior center services and the natural borders formed by the rivers. This region contains
374 Census block groups and 28 currently operating senior centers. In addition, we have
defined eight proposed senior centers, most located in the relatively less-served eastern
portion of the study area. Figure 2, below, shows existing and proposed senior centers,
plus indications of the number of elderly persons in each Census block group.
We assume that existing centers, if chosen in an optimal solution to model EFL,
have the same service level as they do currently. That is, for existing facilities there
is the option of “opening” or “closing” but not one of “resizing”. However, proposed
senior centers, if sited, may provide either “basic” or “comprehensive” service. We have
assumed that the proposed senior centers have high-quality management and thus enjoy
low fixed and variable costs relative to existing senior centers. However, unlike existing
centers, these centers must be built anew or renovated from current uses; thus a capital
cost is incurred.
To account for uncertainty in facility capacities and costs, we have made “educated
guesses” regarding these parameters, and have refined the definitions of service levels
from three to two, consistent with the judgment of the agency that manages elderly
services delivery in Allegheny County. Service level 1 now denotes “basic” services,
and service level 2 denotes “comprehensive” services. We assume that annual capital
costs are zero for existing facilities and are $80,000 and $100,000, respectively, for
proposed “basic” and “comprehensive” facilities, respectively. Annual operating costs
are $30,000, $50,000 or $80,000 for existing facilities and $40,000 for proposed “basic”
facilities and $50,000 for proposed “comprehensive” facilities. Variable costs are $500
per client per year for existing “basic” facilities, $1,000 per client per year for existing
“comprehensive” facilities and $500 per client per year for all proposed facilities.
The total capacity of existing senior centers is 4,161 clients. Total annual fixed costs
for existing centers are $1,270,000. Assuming that all centers are at 90% of capacity,
total annual variable costs are $2,833,200, resulting in a total current annual budget of
$4,103,200. We assume that our planning problem is to use a 10% increase in budget
to select a configuration of new and existing senior centers to maximize utilization of
services, allowing perhaps two or three new facilities to be built, or more if existing
senior centers are closed.
LOCATION OF SERVICE FACILITIES FOR THE ELDERLY 341

By applying the demand forecasting model of Section III to the study area, we are
able to estimate two key demand quantities: total demand for senior centers by block
group (the number of clients that would patronize the closest facility, new or existing,
providing basic or comprehensive services) and the maximum demand, by block group,
for a given facility (not necessarily the closest one), were it included in the optimal
solution, and independent of demands at any other block group for any other facilities.
The average demand, by Census block group, for senior centers, is 60.69 clients; the
minimum demand is 0 and the maximum demand is 812.65.

3.3. Model solution

We model SCL using AMPL Version 8.0 (ILOG-CPLEX Division 2003) and solve SCL
using CPLEX 8.0 for AMPL (ILOG-CPLEX Division 2003). The PC used to solve SCL
has 256 MB of RAM, and a total of 384 MB of virtual memory, using a 2 GHz Pentium 4
processor. The problem instance examined here has 15,101 variables, 15,828 constraints
and 93,264 nonzeros. To minimize model solution times, we specify non-default values
for a variety of CPLEX run-time options.
Solving SCL required 1,707,573 MIP simplex iterations and 1,706 branch-and-
bound nodes, resulting in a solution time of 14,401 seconds, the predefined limit. 4,412
of the possible 8,967 demands were served in this solution. Demands at the block group
level are allocated to multiple facilities; Table 2 below indicates that demand nodes were
served by a minimum of 0 and a maximum of 4 facilities; most demand nodes that were
served at all were served by 2 facilities.
In addition, the budget constraint was binding. 21 of the possible 36 facilities
were sited, 11 of which were “basic” facilities and 10 of which were “comprehensive”.
Eighteen of these facilities were used to capacity.
Facility location and demand allocation variables were processed and displayed
graphically using the ArcView 3.2 geographic information system (Environmental Sci-
ences Research Institute, 1999). Figure 3 indicates that demands at specific nodes are
often served by “basic” and “comprehensive” senior centers simultaneously. Only 2.3% of
all demand nodes have demands served by the closest facility, and the demand-weighted
distance associated with this solution is 24,113.66. Thus, model SCL does not meet one
priority of senior centers, that demands be served by the closest facility, if at all possible.

Table 2
Demand splits in solution to model SCL.
Number of demand nodes Number of facilities assigned
226 0
39 1
78 2
21 3
10 4
342 JOHNSON, GORR AND ROEHRIG

Figure 3. Location of senior centers and allocation of demands to senior centers, model SCL.

The next section examines an alternative senior center location model that attempts
to address effectiveness (utilization) and efficiency (average distance) criteria concur-
rently.

4. A consumer disutility-minimizing senior center location model

To minimize consumer disutility, we formulate an objective function that minimizes

average distance traveled between demand points and facility locations.
Define the following data:

di j = distance, in miles, between demand node i and candidate location j

Then the desired objective function becomes:


n 
m 
p
Minimize di j · Yi jk (9)
i=1 j=1 k=1

Objective (9) is similar to the traditional p-median objective, except that the al-
location decision variables Yi jk are not 0-1 variables, but general integer variables that
allow demands to be split between facilities.
In the absence of a constraint specifying a certain number of facilities to be sited,
minimizing total distance traveled could result in no facilities being opened. To disallow
this outcome we modify our model to minimize the sum of demand-weighted distance
and total unsatisfied demands. Define decision variable δik = unsatisfied demand for
LOCATION OF SERVICE FACILITIES FOR THE ELDERLY 343

service level k at block group i 

and let M be a large positive constant. Then the total
n p
amount of unsatisfied demand is i=1 k=1 δik , and for each demand at node i for type-k
service, we must have δik ≥ λik − mj=1 yi jk .
Thus, we define a new model called SCL2 that minimizes the sum of average travel
distance and unsatisfied demands, subject to funds availability:
Model SCL2

n 
m 
p 
n 
p
Minimize di j · Yi jk + M δik (10)
i=1 j=1 k=1 i=1 k=1

s.t.:

p
Yi jk ≥ λik · X jh ∀i, j, k (11)
h=k

p
X jk ≤ 1 ∀ j (12)
k=1

m
λik − yi jk ≤ δik ∀i, k (13)
j=1
 

m 
p 
n
f jk · X jk + ν jk · Yi jk ≤B (14)
j=1 k=1 i=1
n
Yi jk ≤ C jk ∀ j, k (15)
i=1
X jk ∈ {0, 1} ∀ j, k (16)
0 ≤ Yi jk ≤ ei jk and integer ∀i, j, k (17)
δik ≥ 0 (18)

with data, variables, objective and constraints as defined above. Model SCL2 results in
o(n . m . p) total decision variables, and o(n . m . p) constraints, as before.

4.1. Model solution

Solving model SCL2 required 1,760,401 MIP simplex iterations and 2,147 branch-and-
bound nodes, resulting in a solution time of 14,440 seconds, the limit set in CPLEX
run-time parameters. 4,457 of the possible 47,447 demands are served in this solution,
slightly more than in the solution to the utilization-maximizing model SCL; however, the
demand-weighted distance component of the objective function has a value of 1,539.16,
a reduction by a factor of 10 from the demand-weighted distance associated with model
SCL. Table 3 below indicates that many fewer demand splits occur in this model as
344 JOHNSON, GORR AND ROEHRIG

Table 3
Demand splits in solution to model SCL2.
Number of demand nodes Number of facilities assigned
305 0
56 1
13 2

compared to SCL. In fact, only 18.8% of all demand nodes served by facilities are served
by more than one facility.
Twenty-two of the 36 facilities are sited, and of these, 21 are at capacity. 12 of the
22 higher-utilization facilities sited are “basic” facilities; the remaining 10 are “compre-
hensive”.
As expected, demands served in the solution to model SCL2 are far more spatially
concentrated than those for the original model. Moreover, nearly every facility sited is in
or very near to the block group which contains the demands it serves: nearly 94% of all
demand nodes are served by the facility closest to it. Thus, to the extent that demands for
senior centers conform to the conventional partial-equilibrium facility location modeling
assumption of low friction of demand, model SCL2 appears to represent a solution to
the senior center siting problem that is preferable to that of the original model SCL.
We now determine the response of model SCL2 to changes in cost components.
These changes include: a 10% reduction in the overall budget, a 10% increase in fixed

Figure 4. Location of senior centers and allocation of demands to senior centers, model SCL2.
 Table 4
Complete model results.
Branch Number # Opened % Demands % Demands
Solve MIP -and- demands Demand- Fixed Variable Total faclities served by served by
Model name and Objective time simplex bound served weighted cost cost cost # Facilities below two or more nearest
run number value (sec.) iteration nodes (% of Total) distance ($1,000) ($1,000) ($1,000) opened capacity facilities facility
Maximize 021230b 14,401 1,707,573 1,706 4,412 (9.3) 24,113.66 $1,740 $2,773.5 $4,513.5 21 3 71.7 2.3
utilization [SCL]
Minimize 031113 14,400 1,760,401 2,147 4,457 (9.4) 1539.16 $1,820 $2,693.5 $4,513.5 22 1 18.8 93.9
demand-
weighted
distance +
unserved
demands [SCL2]
Minimize demand- 031114 14,403 2,150,918 8,023 4079 (8.6) 1479.06 $1,520 $2,542 $4,062 20 1 14.7 93.4
weighted distance +
unserved
demands; 10%
LOCATION OF SERVICE FACILITIES FOR THE ELDERLY

reduction in
overall
budget [SCL2]
Minimize demand- 031113b 14,401 1,619,231 3,650 4303 (9.1) 1577.91 $1,760 $2,753 $4,513 21 1 12.9 93.7
weighted distance +
unserved
demands; 10%
increase in
fixed
costs [SCL2]
Minimize demand- 031113c 14,401 3,336,471 21,909 4290 (9.0) 1450.64 $1,760 $2,753.3 $4,513.3 21 1 17.4 95.1
weighted distance +
unserved
demands; 10%
increase in
variable costs [SCL2]

Model SCL maximizes utilization. Model SCL2 minimizes the sum of demand-weighted distance and unserved demands.
345
346 JOHNSON, GORR AND ROEHRIG

facility siting costs, and a 10% increase in variable facility operating costs. Details of
these run results, called 03114, 03113b and 031113c, respectively, as well as the base-case
results for models SCL and SCL2, are contained in Table 4, below.
These results indicate that model SCL2 is quite stable in response to changes in
cost parameters. Although the number of sited facilities and the number of demands
served by sited facilities decreases as compared to base case SCL2 results, as expected,
the spatial concentration and coherence of demands served remains high. The decrease
in the fraction of demand nodes served by two or more facilities ranges from 7.4% (Run
031113c) to 31.4% (Run 031113b), and the fraction of demand nodes served by the
nearest facility remains within 1% of base case. Budget utilization remains high: for all
three variants of SCL2, total costs are within 1% of the budget. The decrease in total
fixed costs ranges from 3.3% (Runs 031113b and 031113c) to 16.5% (Run 031114), and
changes in total variable costs ranges from 2.2% increase (Runs 031113b and 031113c)
to a 5.6% decrease (Run 031114). These sensitivity analyses indicate that model SCL2
appears to meet two desirable goals: maximizing facility utilization and minimizing
consumer disutility.
Note, however, that all models investigated in this paper require substantial com-
putational resources: run times are limited only by a solver parameter. This is an indica-
tion that solutions for larger problem instances will require alternative methods such as
optimization-based heuristics or meta-heuristics.

5. Conclusion

This paper has presented an optimization-based planning model for the location of senior
service centers throughout a large study area. This model represents a contribution to
the literature because it integrates three distinct areas of location modeling research:
median-related models, hierarchical location models and tradeoffs between utilization-
oriented and consumer disutility-oriented objectives. The first area is relevant because
of the gradually decreasing level of consumer benefits as a function of distance that
distinguishes median models from coverage models. The second area is relevant because
of the need for public agencies to provide different levels of senior services to diverse
local communities. The third area is relevant because some public organizations may
judge maximization of utilization of services (an effectiveness criterion) to be more
important than alternative efficiency or equity criteria, even though minimization of
demand-weighted distance, a measure of consumer disutility (an efficiency criterion)
may result in more satisfied clients. In terms of applications, this work represents a novel
application of location modeling methodology to a domain, elderly services, which will
assume increasing importance in the United States in coming years. Finally, our models
use GIS-based demand forecasting methods that are not common in practice.
Computational results indicate that the senior center location model variant that
minimizes the sum of consumer disutility, measured by demand-weighted distance, and
total unserved demand (Model SCL2) is more desirable than the location model variant
LOCATION OF SERVICE FACILITIES FOR THE ELDERLY 347

that maximizes utilization (Model SCL) along a variety of metrics: number of demands
served, demand-weighted distance, facility utilization, demand splits, and proximity of
demands to facilities that serve them. Moreover, model SCL2 is robust to changes in
a number of dollar-valued parameters: as the total budget decreases by 10%, or cost
components increase by 10%, facility and budget utilization remain high and demands
continue to be served primarily by most-proximate, single facilities. These results indicate
that this senior center location model may provide substantial benefits to senior-serving
agencies.
However, computational requirements for both models are substantial. Run times
are at the pre-defined solver limit of 4 hours, an indication that solution of more realistic-
sized problem instances, such as all of Allegheny County, which has about three times
the number of block groups as the problem set examined in this paper, will require the
use of alternative solution methods to generate policy recommendations in a reasonable
amount of time.
Subsequent work in this area should address a number of additional issues. First,
these models could be enriched by an explicit consideration of equity impacts of alter-
native siting strategies for senior service centers. That is, given that a relatively small
fraction of all potential demands for senior service centers is met in solutions for either of
the models presented in this paper, it may be useful to address the impacts, with respect
to a number of standard facility location model equity measures, of locating facilities
especially in areas where the very elderly, who may be approaching the limits of their
ability to drive safely, are expected to increase in numbers.
Second, it may be useful to apply spatial interaction models to observed senior center
utilization data to clarify the extent of a distance-decay effect for senior center services.
If these spatial interaction models provide little evidence for a distance-decay effect, then
it may be necessary to incorporate spatial interaction models into the prescriptive facility
location modeling framework.
Third, estimates of fixed and variable costs of senior centers in Allegheny County,
based on self-reports of senior center administrators, proved to be of insufficient quality
to use directly in a planning model. Acquiring such data would add greater credibility to
our solutions and provide more detailed guidance to decisionmakers regarding sensitivity
analyses.
Finally, we argued that the interactive heuristic approach to solving the senior center
siting problem is likely to be less efficient than the integrated optimization modeling
approach represented by model SCL. However, this assumption needs to be tested in
practice. Both solution methods should be evaluated via spatial decision support systems
to determine the relative value of each to unsophisticated end-users.

Acknowledgment

Presented in slightly different form at the Ninth International Society of Locational De-
cisions Conference (ISOLDE IX), June 14, 2002, Fredericton, New Brunswick, Canada
348 JOHNSON, GORR AND ROEHRIG

and INFORMS Fall Conference, November 17, 2002, San Jose, California. Portions
of this research were supported by the National Science Foundation Faculty Early Ca-
reer Development (CAREER) Program, “CAREER: Public-Sector Decision Modeling
for Facility Location and Service Delivery” and by the Jewish Healthcare Foundation,
Pittsburgh, PA. The insights and recommendations of three anonymous reviewers of this
manuscript are gratefully acknowledged.

Notes

1. All references to field data collection are contained in printed reports available from the authors upon
request.
2. In the city of Pittsburgh, no block group centroid is more than a mile from a senior center, so it is not
possible to measure decreases in usage rates that approach zero. In the Pittsburgh suburbs, marginal usage
rates decline to 1% at a distance of 4 miles from a senior center. Therefore, at least in suburban areas, a
focus on maximizing utilization may be appropriate.

References

Brandeau, M.L. and S.S. Chiu. (1994). “Facility Location in a User-Optimizing Environment with Market
Externalities: Analysis of Customer Equilibria and Optimal Public Facility Location.” Location Science
2(3), 129–147.
Church, R.L. and C. ReVelle. (1974). “The Maximal Covering Location Problem.” Papers of the Regional
Science Association 32, 101–118.
Daskin, M.S. (1995). Network and Discrete Location: Models, Algorithms, and Applications. Wiley-
Interscience, New York.
Environmental Sciences Research Institute. (1999). ArcView GIS Version 3.2. Redlands, CA.
Federal Interagency Forum on Aging Related Statistics. (2000). “Older Americans 2000: Key Indica-
tors of Well-Being,” World Wide Web: http://www.agingstats.gov/chartbook2000/slides.html. (Accessed
11/17/03)
Fotheringham, A.S. and M.E. O’Kelly. (1989). Spatial Interaction Models: Formulations and Applications.
Kluwer Academic, Amsterdam.
Gorr, W., M.P. Johnson, and S. Roehrig. (2001a). “Point Demand Forecasting,” Heinz School Working Paper
Series 2001–02. Pittsburgh, PA: H. John Heinz III School of Public Policy and Management, Carnegie
Mellon University. World Wide Web: http://www.heinz.cmu.edu/wpapers/detail.jsp?id=183 (Accessed
11/17/03).
Gorr, W., M.P. Johnson, and S. Roehrig. (2001b). “Facility Location Model for Home-Delivered Services:
Application to the Meals on Wheels Program,” Journal of Geographic Systems 3(2), 181–197.
Hakimi, S.L. (1964). “Optimum Location of Switching Centers and the Absolute Centers and Medians of a
Graph,” Operations Research 12, 450–459.
Hakimi, S.L. (1965). “Optimum Distribution of Switching Centers in a Communication Network and Some
Related Graph Theoretic Problems,” Operations Research 13, 462–475.
ILOG-CPLEX Division. (2003). AMPL Version 8.0. Incline Village, NV.
ILOG-CPLEX Division. (2003). CPLEX 8.0 for AMPL. Incline Village, NV.
Johnson, M.P., G. Wilpen, and S. Roehrig. (2002). “Location/Allocation/Routing for Home-Delivered Meals
Provision: Models & Solution Approaches.” International Journal of Industrial Engineering 9(1), 45–56.
Narula, S.C. (1986). “Minisum Hierarchical Location-Allocation Problems on a Network: A Survey,” Lo-
cation Decisions: Methodology and Applications, Annals of Operations Research 6, 257–272.
LOCATION OF SERVICE FACILITIES FOR THE ELDERLY 349

ReVelle, C. (1987). “Urban Public Facility Location.” In E.S. Mills (ed.), Handbook of Regional and Urban
Economics, Volume II: Urban Economics. Amsterdam: North-Holland.
ReVelle, C., Richard C., and D. Schilling. (1975). “A Note on the Location Model of Holmes, Williams and
Brown.” Geographical Analysis 7, 457.
Tien, J.M., K. El-Tell, and G.R. Simons. (1983). “Improved Formulations to the Hierarchical Health Facility
Location-Allocation Problem.” IEEE Transactions on Systems, Man, and Cybernetics SMC-13, 1128–
1132.
Toregas, C., R. Swain, C. ReVelle, and L. Bergman. (1971). “The Location of Emergency Service Facilities.”
Operations Research 19, 1363–1373.