Transportation Research Part E 38 (2002) 1±18

www.elsevier.com/locate/tre

Cost structures of public transit systems: a panel data analysis

a,* b
Matthew G. Karlaftis , Patrick McCarthy
a
Department of Transportation Planning and Engineering, Faculty of Civil Engineering,
National Technical University of Athens, 5 Iroon Polytechnion Street, 157 93 Zografon, Athens, Greece
b
School of Economics, Georgia Institute of Technology, Atlanta, GA USA
Received 8 March 2000; received in revised form 22 January 2001; accepted 27 January 2001

Abstract
Results from numerous cost studies have generated con¯icting results on public transit production
technologies. Because prior studies have employed various sub-samples of public transit properties, the
diverse results may either re¯ect alternative technologies or sampling di€erences. Based on a panel of nearly
all transit systems reporting for Section 15 (US National Transit Data) purposes from 1986 to 1994, this
paper explores whether public transit production technology di€ers by size and operating characteristics of
the system. The results indicate that US transit properties are heterogeneous with di€erent production
technologies, which implies that transit cost analyses based upon a set of heterogeneous systems will
generate incorrect inferences on public transit cost and production structures. Ó 2001 Elsevier Science Ltd.
All rights reserved.

Keywords: Public transit; Cost structures; Translog; System size; Section 15; Cluster analysis

1. Introduction

Characteristics of the underlying production technology of ®rms in regulated industries have

attracted considerable interest in the literature due to the vast array of valuable information
provided by such analyses. Policy makers and governmental agencies may be interested in the
underlying production technology in order to set pricing policies. Firms on the other hand may be
interested in the cost e€ects of introducing new or discontinuing existing services. Further, the
examination of the production structure can shed light on the issue of privatization. While most
transit systems in the US are publicly owned and operated, privatization is widely debated with
both avid supporters and strong opponents. A ®nding of diseconomies of scale may imply that,

*
Corresponding author. Tel.: +30-1-671-2092; fax: +30-1-671-2335.
E-mail addresses: mgk@central.ntua.gr (M.G. Karlaftis), mccarthy@econ.gatech.edu (P. McCarthy).

1366-5545/01/$ - see front matter Ó 2001 Elsevier Science Ltd. All rights reserved.
PII: S 1 3 6 6 - 5 5 4 5 ( 0 1 ) 0 0 0 0 6 - 0
2 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18

for example, a city can have di€erent parts of its system operated by separate companies at a
lower unit cost of output. Information on scale economies, input factor and cost elasticities,
average and marginal cost, and optimal ¯eet and fare, can be obtained from the underlying
production structure of transit systems.
For reasons that will be examined later in the paper, most of the literature concerned with un-
covering the underlying production structure of transit systems has used the cost function approach.
While in the 1960s and 1970s most authors employed linear and Cobb±Douglas functional forms to
estimate transit system cost functions, in the 1980s and 1990s researchers almost exclusively used
¯exible functional forms, and most notably translog cost functions. Beginning with Viton (1981)
widely cited paper, many authors used parts of the readily available Section 15 US transit systems
data as the basis for a wide variety of ¯exible cost function models (a summary of many of the cost
function studies appears in Table 1). Due to di€erences in empirical speci®cations of the model,
di€erences in the transit properties analyzed, or di€erent time periods, these studies have often
reached con¯icting conclusions regarding the existence of scale economies, economies of density and
input factor elasticities. For example, researchers have reported results on economies of density and
scale that vary depending upon the size of the systems in the sample and upon the systems forming the
sample (Berechman, 1993, Karlaftis et al., 1999a). The easy availability of Section 15 data has led
most researchers to use these data in their estimation of transit system production characteristics.
Nevertheless, researchers have generally based their empirical work on di€erent time periods, cross-
sections, or both. As a result, the comparability of the results across studies is questionable. Further,
in many studies researchers have used cross-section time series data from largely heterogeneous
systems, introducing a size related bias in their results (Berechman, 1993).
In this paper, we generalize previous work in two important directions. First, we develop a
panel data set composed of transit systems that reported operating data for Section 15 purposes
(256 systems) over a 9-year time period (1986±1994 ). 1; 2 These data allow us to examine both the
cross-section and the temporal production characteristics of transit systems of all sizes. Second, as
noted above, a drawback in pooling data and analyzing systems of vastly di€erent sizes and
operating environments is that the results will likely be skewed towards the most ``in¯uential''
group of ®rms (Berechman, 1993; Braeutigam et al., 1984). Further, in pooled models, it is not
possible to account for di€erences in economies of density or scale between subgroups of ®rms. To
avoid these biases and accurately capture the production characteristics of di€erent-sized transit
systems, we introduce a grouping scheme that classi®es individual properties into homogeneous
subsets.
In this paper, we employ a two-step procedure to analyze the production technology of public
transit ®rms. First, based upon size and operating characteristics, we use cluster analysis to
classify transit systems into homogeneous groups. Second, we develop separate cost function
models for each group of transit systems. This two-step process minimizes the pitfalls of size
related biases and enables us to explicitly capture the production characteristics of transit systems

1
Section 15 data is now called the National Transit Database. We will use the term Section 15, since the data used in
the study came from Section 15 publications.
2
There are approximately 290 systems reporting for Section 15 purposes. The systems excluded did not report
operating data over all the years in the sample (for further discussion on the data see Section 3).
 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18 3

Table 1
Economies of scale, density, and capacity utilization from ¯exible cost functions: empirical results
Author(s) Database Type of Measure of Scale Economies Economies
function output economies of density of capacity
utilization
Cross-section studies
Viton (1981) US Short-run veha -miles Increasing Increasing
Long-run Decreasing
(large
systems)
Williams and Hall (1981) US Long-run paxb -miles Increasing
Williams and Dalal (1981) US Long-run veh-miles Increasing
(large
systems)
Decreasing
(small
systems)
Obeng (1984) US Long-run pax-miles Decreasing
Short-run pax-miles Increasing
Button and O'Donnell (1986) UK Long-run pax-miles Decreasing
Obeng (1985) US Short-run pax-miles Increasing
Long-run pax-miles Decreasing
de Rus (1990) Spain Long-run veh-kilometers Decreasing/
constant
pax Increasing
Caves and Christensen (1988) US Short-run multi-output Constant Increasing Increasing
Long-run Increasing Increasing Increasing
Karlaftis et al. (1999a) US Short-run veh-miles Increasing Increasing
(small, (medium,
medium) large)
Decreasing Decreasing
(large) (small/
suburban)
pax Increasing
Time-series studies
Berechman (1993) Israel Long-run pax-trips Increasing
Berechman and Giuliano US Long-run veh-miles Decreasing
(1984)
pax Increasing
De Borger (1984) Belgium Short-run veh-miles Decreasing
Berechman (1993) Long-run pax Increasing
Applebaum and Berechman Israel Long-run multi-output Increasing Increasing
(1991)
Colburn and Talley (1992) US Long-run multi-output Increasing Increasing
Karlaftis et al. (1999b) US Short-run veh-miles Decreasing Constant
pax Increasing
a
Vehicle.
b
Passengers.
4 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18

that signi®cantly vary in size. The following section summarizes the methodology. Section 3
summarizes the data and identi®es the proposed grouping scheme and Section 4 presents the
estimation results. Section 5 discusses the policy implications of the results obtained from the
analysis.

2. The empirical model

The duality that exists between ®rm technology and costs (Varian, 1992) implies that a ®rm's
cost function summarizes all of the economically relevant information embodied in its production
function. Let CV ˆ CV p; y; Z; k; c† be a ®rm's total variable cost function, where y is a measure of
output, p is a vector of prices for the variable inputs, k is a ®xed factor of production, Z is a vector
of market related variables, and c re¯ects the existing state of technology. 3 Similar to other work
(see Table 1), we use a translog statistical function to approximate a transit system's variable
costs. In particular,
X
J X
J
ln CV ˆ a0 ‡ ay ln y ‡ ai ln pi ‡ aN ln N ‡ at t ‡ ak ln k ‡ giy ln pi ln y
iˆ1 iˆ1
X
J X
J X
J
‡ giN ln pi ln N ‡ git ln pi t ‡ gik ln pi k ‡ gyt t ln y ‡ gNt t ln N ‡ gyN ln pi ln N
iˆ1 iˆ1 iˆ1

1X J X J
1 2 1 2
‡ gyk ln y ln k ‡ gNk ln N ln k ‡ gij ln pi ln pj ‡ gyy ln y† ‡ gNN ln N †
2 iˆ1 jˆ1 2 2
1 2 1
‡ gkk ln k† ‡ gtt t2 ‡ e; i; j ˆ 1; . . . ; J ; 1†
2 2
where t is a time trend that is included to re¯ect changes in technology over the period and e is a
normally distributed disturbance term. The only market-size related variable in the cost function
is network size, denoted by N .
Each variable in (1) is normalized by its mean value. It is important to note that this expansion
is intended to be an approximation of a ®rm's cost structure around mean production level
(Friedlaender and Spady, 1981). The usual point of approximation is the sample arithmetic mean
of the vector py; Z; c. In practice, using the mean as the point of approximation is the same as
normalizing each variable to have a unit mean and using the normalized variables in estimating
Eq. (1). 4 The point of approximation corresponds to a ®rm whose input factor prices, output and

3
Oum and Waters (1996) provide an excellent summary of recent work in this area.
4
The essential observation is that the mean of each normalized variable is equal to one. To obtain this result, consider
the variable X , with mean and sample size n. Normalizing the variable X amounts to subtracting the mean of the
variable X † from each of the observations Xi †. Since the variable is in logarithmic form, the normalization yields:
ln Xi ln X ˆ ln Xi =X †. Finally, we can show that the mean l† of the normalized variable is:
P
1X n
1 Xi
lˆ Xi =X † ˆ  n ˆ 1:
n iˆ1 X n
 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18 5

technological conditions are those given by the sample arithmetic means (Friedlaender and Spady,
1981).
To be consistent with cost minimization, it is necessary that the cost function be homogeneous
of degree one and concave in variable factor prices. To ensure linear homogeneity in factor prices,
the following restrictions are imposed on the parameters:
X
J
ai ˆ 1; cij ˆ cji 8i; j;
iˆ1
2†
X
J X
J X
J X
J X
J
cij ˆ cji ˆ ciy ˆ ciN ˆ cit ˆ 0:
iˆ1 jˆ1 iˆ1 iˆ1 iˆ1

Global concavity must be checked ex-post because ex-ante restrictions the coecients of 1 cannot
ensure that this property is satis®ed.
Di€erentiating the cost function 1 with respect to factor prices and using Shephard's lemma
(Shephard, 1953), yields the following input share equations (Varian, 1992): 5
o ln C p; y; Z; c† oC p; y; Z; c† pi pi xi
ˆ ˆ ˆ si p; y; Z; c†
o ln pi opi C p; y; Z; c† C p; y; Z; c†
XJ
ˆ ai ‡ cij ln pj ‡ ciy ln y ‡ ciN N ‡ cit t ‡ ei ; i ˆ 1; . . . ; J : 3†
iˆ1

Combined with Eq. (1), the share equations in (3) form a system of seemingly unrelated regression
equations (SURE). Assuming classical additive disturbances, Zellners (1962) iterative technique is
used to obtain ecient estimates of the system's parameters. 6
The variable cost function provides short-run measures of returns to network size (RTSZ) and
capacity utilization (RTCU) where capacity is de®ned in terms of a public transit system's rolling
stock of capital. 7 Holding capital stock ®xed, short-run returns to density re¯ect the impact on
costs from a more intensive use of the mass transit network and are expressed as
o ln C
RTSZ ˆ 1 ; 4†
o ln y
whereas returns to capacity utilization, which re¯ect the impact on costs from a proportional
increase in output and network size, are 8

5
Shephard's lemma indicates that the vector of derivatives of the cost function with respect to the input prices results
in the conditional optimal demand for the input vector xi p; y; Z; c†.
6
Zellner's iterative technique yields estimators which, in large samples, are more ecient than equation-by-equation
least squares (Berndt, 1991). This is a result of both the non-diagonality of the disturbance covariance matrix formed by
Eqs. (1) and (3) and the di€erence in the regressors in Eqs. (1) and (3).
7
When network size is included in the model, Berechman (1993) distinguishes economies of size from economies of
trac density depending upon how the output variable is de®ned. If output is `produced' output (e.g. vehicle miles) then
a `network held constant' cost elasticity measures economies of size; if output is `consumed' output (e.g. passenger
miles) then a network held constant cost elasticity measures economies of trac density.
8
A positive value for RTSZ and RTCU re¯ects economies of scale, a negative value re¯ects diseconomies of scale and
a 0 re¯ects constant returns to scale.
6 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18
 
o ln C o ln C
RTCU ˆ 1 ‡ : 5†
o ln y o ln N
In addition, short-run average and marginal cost (SRAC and SRMC) functions are
CV exp ln CV†
SRAC ˆ ˆ ;
y y
" # 6†
oCV o ln CV CV XJ
SRMC ˆ ˆ ˆ AC ay ‡ cyy ln y ‡ ciy ln pi ‡ cky ln k ‡ cyN N :
oy o ln y y iˆ1

3. The data

A problem that often characterizes cost studies of transit systems is the heterogeneity of the
sample of agencies used in the estimation sample. This leads to several problems. First, analyzing
systems of vastly di€erent sizes and operating environments will likely result in parameter esti-
mates that are not representative of the ®rms in the sample, or are skewed toward the most in-
¯uential group of ®rms (Berechman, 1993). Examples include Viton (1981), Williams and Dalal,
1981), Williams and Hall (1981), Obeng (1984, 1985, 1987), and de Rus (1990), who include transit
systems that operate in urban areas ranging from very small cities with low population densities to
very large and dense metropolitan areas. Second, most studies based upon panel data include
systems of di€erent sizes and con®gurations that fail to account for the potential di€erences in
scale economies between sub-groups of ®rms (Braeutigam et al., 1984). Third, it is dicult to
comparatively evaluate the results from alternative studies that are based upon di€erent cross-
sections of public transit systems and over di€erent time periods.
In this study, we make an explicit e€ort to overcome these problems. Initially, we developed a
data set comprising most US transit systems reporting for Section 15 purposes over a 9-year time
span (1986±1994). This yields a set of 256 systems for a 9-year time period resulting in a total of
2304 observations. 9 These panel data enable one to control for both cross-section and temporal
characteristics of the included systems. 10 It is important to note that other sources of useful
operational data, besides Section 15 data, exist; but, an explicit e€ort was made to avoid the use of

9
Because some ®xed-schedule ®xed-route systems, that are the focus of this paper, also provide demand responsive
services and/or urban rail services (they are multi-mode rather than single-mode operators), the analysis performed
re¯ects only the transit services provided. While this may not, in a few cases, capture the overall performance of multi-
mode operators, it is not expected to signi®cantly a€ect the results since demand responsive services generally account
for only a small portion of total services provided (less than 5%). Further, Section 15 reporting allows the separation of
the operating characteristics of the transit, paratransit, and rail aspects of a system's operation.
10
Despite the availability of Section 15 data, there have been no extensive cross-section-time series analyses of these
data. Most studies have been based on large cross-section but short time-series (2±3 years) data sets. In general, most
existing cross-section studies of mass transit in the US have used a subset of Section 15 that falls within the 9-year
period of this study. From a purely time-series perspective, there have been three main studies based on US data.
Berechman and Giuliano (1984) analysis using California data, Colburn and Talley (1992) study employing Virginia
data, and Karlaftis et al.'s (1999b) analysis of the Indianapolis' transit system.
 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18 7

diverse sources since one of the goals of this paper is to form a large sample of transit properties
whose data are easy and straightforward to acquire to make the comparability of di€erent
methodologies possible.
The pooled data used in this study have several advantages over purely cross-section or time-
series data (Hsiao, 1993). From a statistical perspective, by increasing the number of observations,
panel data have higher degrees of freedom and less collinearity (particularly in comparison with
time series data), thus improving the eciency of the parameter estimates. Second, panel data
allow researchers to analyze questions that cannot be adequately addressed in time-series or cross-
section data. For example, suppose that for a cross-section of transit agencies, labor expenses
raise operating costs 60%. If we have a homogeneous population of public transit systems, then
each system's operating costs increase by 60%. This could also be an implication for a sample of
heterogeneous ®rms. However, an alternative explanation in a sample of heterogeneous public
transit ®rms is that the expenses on labor have no e€ect on (4/5) of the ®rms and raise operating
costs 300% on (1/5) of the ®rms. Although we cannot distinguish these hypotheses in a cross-
section sample, it is possible to discriminate between them by identifying the e€ects of labor
expenses on a cross-section of time-series for the di€erent transit systems. Third, panel data allow
researchers to test whether simpler speci®cations are appropriate. With panel data, it is possible to
control for cross-section homogeneity by introducing additional parameters in the model. Testing
for homogeneity is equivalent to testing the null hypothesis that these additional parameters equal
0. Fourth, the availability of panel data enables researchers to explore whether di€erences in
parameter estimates re¯ect sampling di€erences or di€erences in the underlying production
technology. And ®fth, this data set will allow us to better examine, without being constrained by
degrees of freedom concerns, the di€erences in parameter estimates between systems of varying
sizes and operating attributes.

3.1. System groupings

From the national data set, we used cluster analysis to separate transit systems into homoge-
neous groups. As is well known, cluster analysis is a statistical technique that groups items to-
gether on the basis of similarities or distances (dissimilarities). That is, cluster analysis is a
technique for grouping items (transit systems in this case) into clusters of similar elements. In
cluster analysis, an e€ort is made to identify similar elements by their characteristics and to form
groups (clusters) that are homogeneous and di€erent from other groups. Correlations and
functions of distances between elements are used in de®ning the clusters. Clustering is a multi-
variate technique that can use any number of variables, but where variables must be numeric and
for which numerical di€erences ``make sense''. The common use for cluster analysis is when data
are not scattered evenly through n-dimensional space, but rather form locally dense areas, modes,
clusters.
In general, there are two di€erent clustering methods: hierarchical clustering and K-means
clustering. Hierarchical clustering is also called agglomerative clustering because it is a combining
process. The method starts by considering each point as its own cluster. At each step it calculates
the distances between each cluster, and combines the two clusters that are closest together. This
combining continues until all the points are in one ®nal cluster, so the user has to select the
number of clusters that make sense. It is interesting to note that there are a number of rules for
8 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18

de®ning distances between clusters (average linkage, centroid method, Ward's method, single
linkage, complete linkage, and others), and each rule can generate a di€erent sequence of clusters.
K-means clustering is an iterative follow-the-leader strategy. First, the user speci®es a starting
number of clusters, k. Then a search algorithm ®nds k points in the data, called seeds, which are
not close to each other. Each seed is then treated as a cluster center. The routine then goes through
the points and assigns each point to the cluster that it is closest to. For each cluster, a new cluster
center is formed as the means (centroid) of the points currently in the cluster. This process
continues as an alteration between assigning points to clusters, and recalculating cluster centers
until the clusters become stable.
In this paper we used hierarchical clustering and Ward's minimum distance algorithm as the
distance de®ning rule (this method was used because it is a typical hierarchical cluster analysis
technique for grouping transit systems (Hartgen and Segedy, 1997)). The choice of how many
clusters to use is largely up to the analyst and should be chosen after a thorough search. In this
analysis we considered between 5 and 15 clusters. For practical and computational ease, we
clustered each year's observations separately. While this approach can be criticized because it does
not group systems into fully homogeneous groups (it is possible, for example, that a transit system
is statistically more similar to another transit system than it is to itself in a di€erent year), the
e€ort was made to keep systems in the same group over the years to facilitate the comparison
between the systems and the transferability of results. It has to be noted that this approach makes
the ®rst year clustering relevant since it is the year that is used to form the clusters. The subsequent
years' clusters are used to assess the stability of the ®rst years' clusters. 11
Systems were clustered on a variety of variables (such as number of vehicles, wage rate, number
of employees (total and per category), total vehicle miles, passengers, etc.), but the results re-
mained very stable throughout this process. Regardless of the variables used for clustering, most
systems remained in the same (or neighboring) clusters. The ®nal clustering was based on total
operating cost for the system, number of vehicles, total employees, and vehicle miles traveled. The
analysis yielding the most robust results gave us six general groups (clusters) of transit systems
(Table 2 shows the operating characteristics of the di€erent system categories). Group 1 contains
some of the largest US transit systems, while group 6 contains some of the smallest suburban/rural
systems. 12 Combining cluster one, which contains only ®ve systems, with cluster two was con-
sidered; but, both the statistically signi®cant distance between the data in these two clusters and
the distinctly di€erent production technologies by their systems (Section 4.2), suggested the need
for di€erent clusters.
For the 256 US transit systems examined the six clusters may seem low, especially when
compared to the 12 clusters (for 120 systems) selected by Fielding et al. (1985) and the 11 clusters
(for 121 systems) selected by Hartgen and Segedy (1997). But, an explicit e€ort was made to select
a number of clusters that would be large enough to create homogeneous groups of systems yet low
enough to permit the applicability of the results to a suciently large number of systems. The six
cluster case provides a reasonable compromise between these con¯icting objectives.

11
Over the 9-year period, approximately 10% of the public transit ®rms changed groupings.
12
While, for space considerations, we do not present the full list of systems belonging to each group, this list is
available upon request from the authors.
Table 2
Transit system operating characteristics
Operating cost Number of vehicles Gallons of fuel Total employees Vehicle miles Passengers ('000s)
('000s $) traveled ('000s)
Mean S.D. Mean S.D. Mean S.D. Mean S.D. Mean S.D. Mean S.D.
All systems n ˆ 256† 22,621 72,547 129 296 1444 3479 391 1140 5100 11,164 14,781 53,932
Group 1 n ˆ 5† 430,524 82,875 1816 264 21,796 4041 6821 1242 67,869 14,859 315,110 111,390
Group2 n ˆ 23† 98,326 33,183 557 174 6388 1850 1678 549 22,791 6087 61,660 21,045
Group 3 n ˆ 47† 21,140 10,712 150 64 1717 312 404 177 6030 2380 12,191 6909
Group 4 n ˆ 69† 5495 1882 46 15 463 149 111 34 1825 876 3381 1342
Group 5 n ˆ 53† 2453 730 23 6 210 64 57 14 862 207 1487 554
Group 6 n ˆ 59† 1015 431 10 4 92 45 23 9 408 170 583 336

Table 3
Descriptive analysis of the variables
Price of labor ($) Price of fuel ($) Price of material Share of labor ($) Share of fuel ($) Share of material
($) ($)
Mean S.D. Mean S.D. Mean S.D. Mean S.D. Mean S.D. Mean S.D.
All systems n ˆ 256† 16.51 4.2 0.77 0.23 8321 3527 0.7 0.08 0.22 0.08 0.07 0.03
Group 1 n ˆ 5† 23.41 3.5 0.95 0.1 9357 1901 0.8 0.07 0.15 0.07 0.05 0.01
Group2 n ˆ 23† 20.58 4.4 0.87 0.11 9125 3008 0.76 0.06 0.18 0.07 0.06 0.02
Group 3 n ˆ 47† 17.74 3.9 0.82 0.13 8589 2722 0.73 0.08 0.2 0.07 0.07 0.02
Group 4 n ˆ 69† 16.13 3.9 0.75 0.19 8143 2998 0.7 0.07 0.22 0.07 0.07 0.02
M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18

Group 5 n ˆ 53† 13.6 3.7 0.8 0.17 8008 2986 0.68 0.07 0.24 0.08 0.08 0.03
Group 6 n ˆ 59† 12.68 3.3 0.9 0.13 7600 2624 0.66 0.08 0.25 0.08 0.08 0.03
9
10 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18

4. Model estimation

4.1. Variable formulation

The dependent variable in the translog cost function formulation is the total annual system
operating cost, excluding depreciation and amortization of intangibles. Total vehicle miles
(TVM), including deadhead as well as in-service mileage, is the measure of output used in this
study, since bus operation is the primary determinant of cost in a transit system. 13
The model includes three variable inputs, labor, fuel and materials. The price of labor was
estimated by dividing the total labor expenses (including wages, fringe bene®ts, and pension
payments to operators, maintenance and administration employees) by the paid labor hours to
operators, maintenance, and administrative employees. Fuel price was computed by dividing total
fuel cost by total fuel gallons. Measuring fuel price by this procedure has an important limitation
in that the prices of diesel and gasoline cannot be directly determined and the computed fuel prices
are rather approximate. 14
Materials cost is computed as total operating cost minus labor and fuel costs. 15 The materials
expenses include maintenance, parts, casualty and liability, purchased transportation, taxes, etc.
The materials expenses category does not refer to cost for materials per se; rather, it refers to all
other expenses besides labor and fuel (an interesting alternative may be to collect the CPI or
another index for each city as the measure of materials price). A factor price was obtained by
dividing the total materials cost by peak-hour vehicle requirement, based on the realization that
peak vehicle requirement is the major contributor to material expenses. Finally, the empirical
model assumes that public transit systems are minimizing costs conditional on the amount of
existing capital. That is, we estimate a short-run translog cost function where the ®xed factor of
production is the number of buses operated by the system. Table 3 gives some summary statistics
for the variables included in the estimation process.

4.2. Estimation results

Table 4 reports the translog estimation results for the pooled sample and for each of the six
groups of homogeneous transit systems. Recall from Table 2 that Group 1 corresponds to the
largest of the public transit systems with an annual operating cost that averages $430 million,
a network that averages 67 million vehicle miles traveled and 315 million passengers, and an

13
Deadhead miles are miles traveled by revenue vehicles when not in revenue service (not available for passengers).
This includes miles traveled to and from storage facilities and other non-revenue service mileage such as maintenance or
training mileage.
14
Nevertheless, it should be noted that in most transit systems the amount of consumed diesel fuel is much higher
gasoline. In many systems, for example, diesel accounts for over 90% of fuel used. This suggests that the calculated fuel
prices primarily re¯ect the price per gallon for diesel fuel that the transit systems pay.
15
Several studies have combined all other expenses besides labor in the fuel expenses category (Viton, 1981; Obeng,
1984, 1985). As a result, the authors reported own-price and substitution elasticities involving fuel that did not re¯ect
only fuel. To overcome this problem, we utilize a third variable factor of production, materials factor, that explicitly
accounts for expenses other than labor and fuel.
Table 4
Short-run translog cost functions dependent variable: short-run costs 106 †a
Variable Model
Pooled Group 1 Group 2 Group 3 Group 4 Group 5 Group 6
Constant term 9.27 12.53 11.30 9.77 8.47 7.72 6.85
Price of labor ($/h) 0.71 0.79 0.76 0.73 0.70 0.66 0.65
Price of maintenance ($/h) 0.15 0.15 0.18 0.20 0.21 0.23 0.28
Number of buses 0.29 1.06 0.13(ns) 0.44 0.33 0.47 0.59
Output (vehicle-miles, units of 105 ) 0.78 0.09 1.01 0.55 0.59 0.34 0.26
Network Size (miles) )0.03 )0.04 )0.05(ns) )0.002(ns) )0.03 0.13 )0.04
(Price of labor)  (Price of labor) 0.08 0.002(ns) 0.03 0.07 0.04 0.08 0.11
(Price of maintenance)  (Price of maintenance) )0.02 0.02 )0.05 )0.02(ns) )0.06 )0.09 )0.01
(Number of buses)  (Number of buses) )0.51 )0.59(ns) 2.40 0.49 0.11(ns) 0.53 )1.11
(Output coecient)  (Output coecient) 0.20 )2.93(ns) )2.55 0.02(ns) 0.53 0.04(ns) )0.84
(Network Size)  (Network size) 0.0071 1.62 0.01(ns) 0.02(ns) )0.13 0.01(ns) )0.02(ns)
(Price of labor)  (Price of maintenance) )0.01 0.01(ns) 0.06 )0.01 0.04(ns) )0.01 )0.05
(Price of labor)  (Number of buses) 0.03 0.39 0.08 0.06 0.05 0.01(ns) )0.02
(Price of labor)  (Output ) )0.01 )0.44 )0.02(ns) 0.005(ns) )0.01 0.03 0.02(ns)
(Price of labor)  (Network size) )0.01 )0.13 )0.02 0.003(ns) )0.008(ns) )0.01 )0.02
(Price of maintenance)  (Number of buses) )0.17 )0.39 )0.07 )0.11 )0.15 )0.10 )0.14
(Price of maintenance)  (Output) )0.07 0.47 )0.10 0.005(ns) 0.02(ns) 0.003(ns) )0.0009(ns)
(Price of labor)  (Network size ) 0.28 0.17 0.05 0.009 0.03 0.01 0.08
(Number of buses)  (Output) 0.17 2.04(ns) )1.17 )0.24 )0.43 )0.10(ns) 0.84
(Number of Buses)  (Network size) 0.81 )1.21(ns) )1.04 0.84 1.13 1.07 0.50
(Output)  (Network size) )0.84 0.79(ns) 2.35 )0.97 )0.70 )1.01 )0.59
Dummy for 1987 0.13 0.02(ns) 0.01(ns) 0.02(ns) )0.05 )0.09 )0.10
Dummy for 1988 0.16 0.04(ns) 0.11 0.02(ns) 0.03(ns) 0.02(ns) )0.09
Dummy for 1989 0.11 )0.03(ns) )0.04(ns) )0.04(ns) )0.08 )0.15 )0.02(ns)
Dummy for 1990 0.15 0.07 0.06(ns) 0.09 0.06 0.07 0.07
Dummy for 1991 0.07 0.09 0.05(ns) 0.07 0.03(ns) 0.04 0.08
Dummy for 1992 0.10 0.07 0.04(ns) 0.07 0.003(ns) 0.007(ns) 0.03(ns)
Dummy for 1993 0.15 0.02(ns) 0.02(ns) 0.08 0.05 0.03(ns) )0.01(ns)
M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18

Dummy for 1994 0.06 0.08 0.08 0.07 0.08 0.03(ns) )0.004(ns)
Degrees of freedom N † 2304 45 207 423 621 477 531
% of points satisfying second-order conditions 68 97 86 89 81 78 81
system R2 0.998 0.999 0.999 0.999 0.999 0.999 0.998
a
Unless indicated by (ns), all variables are signi®cant at the 0.05 level.
11
12 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18

average ¯eet size of 1816. In contrast, the smallest group of ®rms, Group 6, averages just over $1
million in operating costs, has an average of 10 vehicles and annually serves 583,000 passengers.
Looking at the group of results in Table 4, we see that the system R2 for each model is exceedingly
high at 0.998 or above. However, we also see one of the advantages of estimating separate models for
groups of homogeneous systems. For the pooled model, only 68% of the systems satisfy the second-
order conditions. This rises considerably for the separate models, varying between 78% for Group 5
to 97% for the largest set of public transit systems. This suggests that the estimated models su€er
from some speci®cation errors but that these errors have been signi®cantly reduced by classifying
systems into homogeneous sub-groups and estimating separate equations for each.
An examination of the ®rst-order coecients in Table 4 reveals some interesting di€erences
between the larger and smaller systems. For the average ®rm in the largest group of public transit
properties, labor constitutes the largest share of operating costs at 79%, which falls continuously
as average ®rm size falls. For the average ®rm in Group 6, labor represents 65% of operating
costs. Maintenance expenditures, on the other hand, increase with decreasing ®rm size, 15% of
operating costs for average ®rm in the largest group, rising to 28% for the average ®rm in the
smallest group. Not surprising, fuel is more important for the largest group of ®rms, representing
14% of operating costs for the average ®rm. For smaller ®rms, this falls to around 6%, except for
properties in Group 5 where fuel is 11% of operating costs for the average ®rm. It is also inter-
esting to note the coecient estimates for the time trend. The time trend considered in this es-
timation is not the usual linear time trend (a time trend whose value is 0 for the ®rst year, 1 for the
next, 2 for the following year, and so on) frequently used. Rather, a series of dummy variables was
created, one for each year in the sample. This way, the time trend is not restricted to being of equal
slope over the years, but it can have varying magnitudes, allowing for much greater ¯exibility.
Interestingly, for the panel data model each of these time variables is positive, indicating an in-
crease in operating cost over time. However, when the data are disaggregated into homogeneous
sub-groups, the results are much di€erent. With some exceptions, operating cost increased little
over the nine year period. Group 1 and Group 3 experienced increasing costs through most of this
period and for all but the two smallest groups, operating costs increased in 1994.
We also see that, in both the pooled and separate samples, the ®xed factor of production is
positive and, for all but Group 2, signi®cantly di€erent from 0. This is inconsistent with long-run
cost minimization where ®rms are expected to invest in capital up to the point at which the savings
in operating costs equals the increase in annualized capital costs. Oum and Waters (1996) note
that a positive sign for the ®xed factor of production is a common ®nding in transportation cost
function analyses and Oum and Zhang (1991) attribute this to a kink in the variable cost function
due to a discrete shift in the user cost of capital beyond the point at which capital is currently
utilized. Thus, using the ®xed factor of production mis-speci®es the operating cost function. To
correct this, Oum and Zhang suggest using a measure of capital's service ¯ow. 16 Unfortunately,

16
Oum and Zhang (1991) argue that the user cost of capital is the interest opportunity cost plus the rate of
depreciation for utilization levels at or below current usage rates. Beyond this point, the user cost of capital is simply the
interest opportunity cost since the depreciation cost falls to 0. This induces a kink in the variable cost function at which
point derivatives are unde®ned. Oum and Zhang obtained correct (negative) signs on the capital variable when the stock
of capital was replaced with a measure of capital's service ¯ow.
 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18 13

for this study there is not sucient information in the Section 15 data to generate a measure of
service ¯ow.

4.3. Properties of the production technologies

To further examine the structure of the underlying production technology, the hypothesis of
linear input factor separability and homotheticity can be statistically tested from the estimated
cost function. Factor separability implies that the marginal rate of technical substitution between
two inputs is independent of the prices of other inputs. And a sucient condition for separability
in input factors is that the cross terms of the input price variables are 0, that is, cij ˆ 0 for all i; j.
Also, the restrictions for linear homogeneity in prices (Eq. (2)) combined with a restriction that
the parameters of the cross terms between input price and output variables are 0 (i.e. ciy ˆ 0 ) for
all i, provides a necessary and sucient condition for an underlying production technology to
be homothetic. 17 At the 0.05 level, separability was rejected in all cases with the exception of
Group 4.
With respect to homotheticity, the test results indicate a relationship between size and ho-
mothetic production. In particular, homotheticity was rejected in the two largest groups of
public transit ®rms, Groups 1 and 2 whereas for Groups 3, 4, and 6, homotheticity was ac-
cepted. The anomaly was Group 5, a group of smaller systems characterized by non-homo-
thetic production technology. 18 The ®nding of non-homotheticity implies that scale economies
are not independent of the input factor prices used in the production process; that is, changes in
the price of an input factor will a€ect the cost elasticity besides a€ecting the demand for that
factor.
For each of the inputs we calculate the own price elasticities of demand and, to explore sub-
stitution possibilities among inputs, we use the Allen±Uzawa partial elasticities of substitution
(Uzawa, 1962). For the translog function (Berndt, 1991), the partial elasticities of substitution are
de®ned as 19

cij ‡ si sj
rij ˆ ; i; j ˆ 1; . . . ; J ; for i 6ˆ j;
si sj
7†
c ‡ s2i si
rii ˆ ii ; i ˆ 1; . . . ; J ;
s2i

17
The translog model (Eq. (3)) reduces to a Cobb±Douglas cost function when each of the cross-product and second-
order terms is set to equal to 0; this can be expressed as: cij ˆ cyy ˆ ciy ˆ 08i; j (Berndt, 1991).
18
In addition to separability and homotheticity, it is straightforward to check for simpler production technologies,
the most popular of which is a Cobb±Douglas function which allows for economies of scale but assumes a unit
elasticity of substitution between inputs. Consistent with most studies, Cobb±Douglas technology was strongly
rejected in all cases.
19
The symmetric Allen±Uzawa substitution elasticities are de®ned as rij ˆ C oC=opi opj †± oC=opI oC=opj †. The own
price elasticity of the demand for input i is given as rii sI where sI is the share of input i in costs.
14 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18

Table 5
Short-run own factor demand elasticities and elasticities of substitution
Model Price elasticitya
ell eff Emm rlf b rlm rmf
Pooled )0.17 )0.45 )0.58 )0.55 0.38 1.13
Group 1 )0.19 )0.17 )0.64 0.63 1.10 )1.63
Group 2 )0.19 )0.19 )0.69 )0.53 0.44 )0.65
Group 3 )0.16 )0.45 )0.69 )0.16 0.86 1.22
Group 4 )0.24 )0.39 )0.88 0.02 0.82 1.08
Group 5 )0.20 )0.36 )0.44 )0.38 0.89 1.14
Group 6 )0.16 )0.31 )0.77 )0.13 0.66 1.68
a
e: own-price elasticity of demand, r: Allen partial elasticity of substitution; where: l ˆ labor, f ˆ fuel, m ˆ maintenance.
b
Elasticities of substitution are symmetric.

where si is the estimated factor share for input i. Further, rij < >†0 implies that factors i and j are
complements (substitutes).
For the pooled sample and each sub-group, Table 5 reports the price elasticities of demand and
the substitution elasticities. For all estimation groups, the demand for labor, fuel and materials
are price inelastic. But di€erences are revealed among the groups. Groups 1 and 2 have similar
price elasticities of demand for each of the inputs. There is considerably more variation, however,
among Groups 3±6. Relative to Groups 1 and 2, the demand for labor in Group 4 ®rms is more
price sensitive. Groups 3 and 6, however, have less wage sensitive demands for labor. The demand
for fuel shows more uniformity between larger and smaller ®rms. For the two largest ®rms, a 10%
increase in the relative price of fuel reduces quantity demanded 1.7±1.9%. In Groups 3±6, how-
ever, a similar fuel price increase has a much larger impact, reducing the quantity of fuel de-
manded 3.1%±4.5%. Although the di€erence is not quite as great, we also see that smaller ®rms
(excepting those in Group 5) are more price sensitive to materials' prices than are the larger
systems (Groups 1 and 2).
The second part of Table 5 reports the substitution elasticities for each of the groups. Here we
see some interesting di€erences between the systems. For the largest set of transit properties in
Group 1, labor is a substitute for fuel and materials whereas fuel and materials are complements
in production. In addition, the substitution possibilities for material inputs are relatively elastic. A
1% increase in the relative price of materials increases the labor-materials ratio 1.1% but decreases
the fuel-materials ratio 1.63%. On the other hand, for the smaller groups (speci®cally Groups 3±
6), we see that labor and fuel are complements in production (excepting Group 4 in which the
substitution elasticity is quite small) whereas materials is a substitute for both labor and fuel.
Overall, the results suggest that labor and materials are substitutes in production regardless of
size. For the smallest (largest) systems, fuel is a substitute (complement) for materials and a
complement (substitute) for labor. And consistently, materials and fuel are most sensitive to
changes in relative prices.
Table 6 provides estimates for short-run returns to network density (Eq. (4)), capacity utili-
zation (Eq. (5)), SRAC, and SRMC. For the pooled model, we see that a representative ®rm
producing at the mean experiences short-run returns to capacity utilization. Holding network size
 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18 15

Table 6
Short-run returns to scale, returns to network density and costs
Model RTCUa RTSZb SRACc SRMCd
Pooled 0.22 0.25 2.963 2.31
Group 1 0.91 0.95 5.559 0.50
Group 2 )0.01 )0.01e 4.027 4.07
Group 3 0.45 0.45e 3.307 1.83
Group 4 0.41 0.44 2.913 1.72
Group 5 0.66 0.53 2.647 0.90
Group 6 0.74 0.78 2.391 0.62
a
Returns to capacity utilization.
b
Returns to network size.
c
Short-run average cost.
d
Short-run marginal cost.
e
Network size coecient not signi®cantly di€erent from 0.

and rolling stock constant, a 1% increase in vehicle miles reduces the average cost of production.
22%. Combined with a 1% increase in network size, average cost falls 25%. Consistent with short-
run economies, estimated SRAC is $2.96,$0.65above estimated SRMC of an additional vehicle
mile.
However, if we look at the size speci®c results we see fairly large di€erences in the elasticity
measures, particularly for the largest ®rms. For ®rms in Group 1, returns to capacity utili-
zation and network density, respectively, are 0.91 and 0.95. That is, a 1% increase in vehicle
miles (or vehicle miles plus network size) reduces the average cost of production by nearly 1%.
With the exception of ®rms in Group 2, which produce at close to constant returns to scale,
the short-run scale economy results are qualitatively similar but the cost bene®ts are not great,
about. 4% for ®rms in Groups 3 and 4 and ranging between 0.53% and 0./78% for the
smallest ®rms.
The cost estimates reported in the last two columns of Table 6 reinforce these economy of scale
results. With the exception of ®rms in Group 2 which exhibit very slight decreasing returns to scale,
the short-run operations of ®rms in Group 1 and Groups 3±6 exhibit increasing returns to scale
which implies an average cost of production that exceed the marginal cost of production. The last
two columns con®rm this relationship where the largest di€erence between average and marginal
cost of production occurs for those systems that experience the largest scale economies. There is a 10-
fold di€erence between the average cost and marginal cost of production for the largest ®rms,
$5.56versus $0.50. For ®rms in Groups 3 and 4, there is less than a 2-fold di€erence and for the
smallest ®rms there is nearly a 4-fold di€erence, $2.39 versus $0.62. These di€erences seem somewhat
high, particularly for the smallest systems, but not necessarily so. Consider the set of largest ®rms.
One can easily expect that a public transit system with a rolling capital stock of more than 1800
vehicles (which characterizes Group 1) will have a high average cost of production combined with a
low marginal cost of production. This is precisely the situation reported in Table 6. Moreover, in
comparison with the smallest set of systems, Group 6, the average cost of production in Group 1 is
more than twice that of Group 6 in comparison with a 20% di€erence in the marginal cost of pro-
duction.
16 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18

5. Conclusions

Relative to virtually all other US studies of public transit costs based upon Section 15 data, the
analysis presented in this paper is unique for two basic reasons. First, rather than focusing upon a
relatively small subset of public transit ®rms, this study includes 256 public transit systems. In
addition, the data comprise a cross-section of these 256 ®rms, covering a nine year period from
1986±1994, which provides an extensive set of panel data with 2304 observations. Second, the
present study explicitly focuses upon the role that system size has upon a public transit ®rm's
production technology.
The primary conclusion from this study is that size matters if public transit policies are based upon
analyses of transit costs using a heterogeneous collection of public transit ®rms. Whether one ex-
amines own price elasticities for inputs, elasticities of substitution between inputs in the production
of public transit, or scale economies, there will be di€erences in these measures by size of public
transit ®rm. Moreover, controlling for size di€erences, using the same time frame, and employing an
identical empirical speci®cation suggests that the observed di€erences among systems of varying
sizes are not the result of sampling problems but rather re¯ect di€erences in production technology.
Consistent with expectations, the largest systems operated with the greatest returns to capacity
utilization and have the highest estimated average cost of production. For Groups 2±6, both the
average cost of production and the marginal cost of production were directly related to system
size. The larger the system, the higher the respective cost. However, this was not true for Group 1
where the marginal cost of production was found to be lower than any other system and sub-
stantially lower in some cases (e.g. Group 1 versus Group 2). A further di€erence was the ®nding
that the smallest set of ®rms experienced greater short-run economies than mid-size ®rms. For
example, a 1% increase in vehicle miles reduced the average cost for ®rms in Group 6 by 0.78%
compared with a 45% decrease for ®rms in Group 3 and a 0.01% decrease for ®rms in Group 2.
Di€erences in substitution elasticities were also observed when ®rms were separated into homo-
geneous groups. In particular, only labor and materials were found to have the same (substitute)
production relationship for all groups. Whereas in the pooled data set, fuel was a complement for
labor and a substitute for materials, in the disaggregated analyses this was not uniformly true.
The research reported in this analysis suggests a number of interesting avenues for further study.
The focus of this work is on the short-run. Do we observe similar size-related di€erences for long run
scale economies and substitution elasticities? In addition, we have estimated a model which uses
vehicle-miles as a measure of output. What di€erences would occur if a `consumed' measure of
output were used rather than a produced measure? And related to this, how would the results change
if we de®ned a multi-product production function which included both consumed and produced
measures of output? Finally, we have assumed that all variables on the right hand side of the cost
function are exogenous. Although this may be a reasonable assumption for a short time period, it is
clearly restrictive over a longer time span.

Acknowledgements

A previous version of this paper was presented at the 1999 annual Transportation Re-
search Board meeting in Washington DC. The authors would like to thank anonymous
 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18 17

referees for their helpful comments. All remaining errors are the responsibility of the
authors.

References

Applebaum, E., Berechman, J., 1991. Demand conditions, regulation and the measurement of productivity. Journal of
Econometrics 47, 379±400.
Berechman, J., 1993. Public Transit Economics and Deregulation Policy. North-Holland, Amsterdam.
Berechman, J., Giuliano, G., 1984. Analysis of the cost structure of an urban bus transit property. Transportation
Research B 18B (4/5), 277±287.
Berndt, E.R., 1991. The Practice of Econometrics: Classic and Contemporary. Addison-Wesley, Reading, MA.
Braeutigam, R.R., Daughety, A.F., Turnquist, M.A., 1984. A ®rm speci®c analysis of economies of density in the US
railroad industry. The Journal of Industrial Economics XXXIII (1), 3±17.
Button, K., O'Donnell, K.J., 1986. The costs of urban bus provision in Great Britain. Transportation Planning and
Technology 10, 293±303.
Caves, D.W., Christensen, L.R., 1988. The importance of economies of scale, capacity utilization, and density in
explaining interindustry di€erences in productivity growth. The Logistics and Transportation Review 24 (1),
958±976.
Colburn, B.C., Talley, W.K., 1992. A ®rm speci®c analysis of economies of size in the US urban multiservice transit
industry. Transportation Research 26B (3), 195±206.
De Borger, B., 1984. Cost and productivity in regional bus transportation: the Belgium case study. Journal of Industrial
Economics 37 (1), 35±54.
de Rus, G., 1990. Public transport demand in Spain. Journal of Transport Economics and Policy, May, 189±201.
Fielding, G.J., Brenner, M.E., Faust, K., 1985. Typology for bus transit. Transportation Research 19A (3),
269±278.
Friedlaender, A.F., Spady, R.H., 1981. Freight Transport Regulation. MIT Press, Cambridge, MA.
Hartgen, D.T., Segedy, J.A., 1997. Peer groups for transit system performance, Working paper, Center for
Interdisciplinary Transportation Studies, The University of North Carolina at Charlotte.
Hsiao, C., 1993. Analysis of Panel Data. Cambridge University Press, Cambridge, UK.
Karlaftis, M.G., McCarthy, P.S., Sinha, K.C., 1999a. System size and cost structure of transit industry. ASCE Journal
of Transportation Engineering 125 (3), 208±215.
Karlaftis, M.G., McCarthy, P.S., Sinha, K.C., 1999b. The structure of public transit costs in the presence of multiple
serial correlation. Journal of Transportation Statistics 2, 113±121.
Obeng, K., 1984. The economics of bus transit operation. The Logistics and Transportation Review 20 (1), 45±65.
Obeng, K., 1985. Bus cost productivity and factor substitution. Journal of Transport Economics and Policy 19 (2),
183±203.
Obeng, K., 1987. Classi®cation of the bus transit policy variables. Transportation Planning and Technology 11,
257±272.
Oum, T.H., Zhang, Y., 1991. Utilization of quasi-®xed inputs and estimation of cost functions. Journal of Transport
Economics and Policy 25 (2), 121±134.
Oum, T.H., Waters, W.G., 1996. A survey of recent developments in transportation cost function research. The
Logistics and Transportation Review 32 (4), 423±463.
Shephard, R.W., 1953. Cost and production functions. Princeton University Press, Princeton, NJ.
Varian, H.R., 1992. Microeconomic Analysis. W.W. Norton, New York.
Viton, P.A., 1981. On competition and product di€erentiation in urban transportation: the San Francisco Bay area. Bell
Journal of Economics 12, 362±379.
Uzawa, H., 1962. Production functions with constant elasticity of substitution. Review of Economic Studies 29 (4), 81,
291±299.
Williams, M., Dalal, A., 1981. Estimation of the elasticity of factor substitution in urban bus transportation: a cost
function approach. Journal of Regional Science 21 (2), 263±275.
18 M.G. Karlaftis, P. McCarthy / Transportation Research Part E 38 (2002) 1±18

Williams, M., Hall, C., 1981. Returns to scale in the united states intercity bus industry. Regional Science and Urban
Economics 11, 573±584.
Zellner, A., 1962. An ecient method of estimating seemingly unrelated regressions and tests for aggregation bias.
Journal of the American Statistical Association 58 (2), 348±368.