Tran.s,m RCT:B Vol. 188, No. I, pp. 29-41. 1984 Ol9l-2615/84 $3.00 + .

Xl
Pnnted I” Great Bntam Pergamon Press Ltd.

THE IMPACT OF FARE AND GASOLINE PRICE

CHANGES ON MONTHLY TRANSIT RIDERSHIP:
EMPIRICAL EVIDENCE FROM SEVEN U.S.
TRANSIT AUTHORITIES?

GEORGE H. K. WANG and DAVID SKINNER

U.S. Department of Transportation, Transportation Systems Center, Kendall Square, Cambridge,
MA 02142, U.S.A.

(Received I February 1983)

Abstract-This paper presents eight empirical models of monthly ridership for seven U.S. Transit
Authorities. Within the framework of these models, the impacts upon monthly ridership from
changes in the real fare and gasoline prices are examined. Important findings are: (1) the elasticities
of monthly transit ridership with respect to the real fare are negative and ineiastic, ranging from
0.042 to 0.62; and (2) the elasticities of monthly transit ridership with respect to the real gasoline
price are positive and inelastic, ranging from 0.08 to 0.80. Such results have important policy
implications for decisions based on the relationships of price, revenue, and ridership; and for
assessing the impacts of changing gasoline prices upon urban modal choice.

1. INTRODUCTION

At present the current administration proposes a gradual phasing out of federal operating
subsidies to local public transportation authorities. As a result, the burden of financing
transit deficits will fall increasingly with the states, local governments, and the transit
authorities. It will probably be necessary for some transit authorities to increase fares and/or
reduce service to compensate for the federal subsidy loss. Such changes will affect ridership
levels and hence revenues. At the same time, ridership levels will be affected by the
continuing fluctuation (since 1979) of gasoline prices causing a substitution effect as
commuters switch from automobiles to public transit and vice versa. Thus, it is necessary
at this time for policy makers and transit planners to evaluate the impacts of changes in
fares, gasoline prices, and service levels on ridership and revenue.
It is not uncommon for many transit planners to use the Simpson and Curtin formulat
as a rule of thumb, or to use fare elasticities calculated from the shrinkage ratio procedure,
to predict the decrease (increase) in ridership due to an increase (decrease) in fares. This
common practice has the following limitations:
(1) There is empirical evidence that the size of the fare elasticity is inversely related to
city size. Therefore, the Simpson and Curtin Formula should be applied with caution when
used for specific cities.
(2) The shrinkage ratio procedure4 is a naive method for calculating fare elasticity
because it does not adjust for the effects of other variables which impact transit ridership at
the time of a fare change. For example, it should not be used to determine elasticity when
fare and gasoline prices change simultaneously. It definitely should not be used when
monthly transit data has a strong seasonal and/or working day variation which often is the
case.

tThis paper represents the views of the authors and not necessarily those of the U.S. Department of
Transportation.
$The Simpson and Curtin rule is a popular rule of thumb used by transit planners to predict the impacts
of fare changes on transit ridership. Simply stated it offers that transit ridership will decrease 3.3% for every 10%
increase in fares over the previous level (see Curtin, 1968).
$The shrinkage ratio has been widely used to calculate transit fare elasticities based on data before and after
a fare change. Formally it is calculated as:

Q,-Q, P,--P,
c
RS=PIt-- p,
where Q, and P, represent the initial level of ridership and fare; and Q2 and PI represents the level of ridership
and fare after the fare change has been made (see Kemp, 1973 and Lago et al., 1981).

29
30 G. H. K. WANG and D. SKINNER

Certainly, the above limitations have been recognized for a long time (see Simon, 1965,
and Driver, 1979); and a few previous studies have attempted to assess the impacts of
changes in real fares, real gasoline prices, and services within the framework of statistical
monthly transit demand models. Some related previous works in this area are Gaudry
(1975) Agthe and Billings (1978) Kemp (1981) Bates (1981), Wang et al. (1981a).
For this reason, the purpose of this paper is: (1) to construct and empirically estimate
monthly models for transit ridership in seven U.S. transit authorities; (2) and to estimate the
impacts of changes in real fares, real gasoline prices, and transit service on transit ridership.
The seven transit authorities are: (1) Albany, New York; (2) Baltimore, Maryland; (3)
Atlanta, Georgia; (4) Des Moines, Iowa; (5) Jacksonville, Florida; (6) Miami, Florida; and
(7) New York City, New York.
There are at least three reasons for using monthly models: (1) Policy decisions have
sometimes been influenced by the latest monthly data. More systematic and rigorous
analysis of this monthly data in the framework of a monthly model would reduce ad hoc
interpretations. (2) In a monthly model, there is less simultaneity involved in the system of
structural relationships than in an annual model. Hence the estimation of monthly models
by single equation estimation procedures would suffer less simultaneous equation bias (if
simultaneity exists) (see Liu, 1969). (3) Monthly models provide a detailed picture of
seasonal and working-day variation on the behavior of ridership levels. Such information
is crucial to short-run planning of manpower requirements, operating revenues, and equip-
ment utilization.
This paper contains three other sections. Section 2 gives the general form of the models
for monthly transit ridership. Section 3 discusses issues of econometric estimation for these
models. Section 4 gives the empirical results and interpretations. Section 5 presents a
summary and conclusion.

2. GENERAL MONTHLY MODELS FOR TRANSIT RIDERSHIP

In general, monthly demand for urban transit use is a function of real fares, cost of
competing modes, service characteristics of urban transit, promotion, seasonal variations
due to social and institutional habits, and working-day variation. The service characteristics
include: (1) areas served, (2) in-vehicle time, (3) comfort, (4) reliability, and (5) mesh
density.? Unfortunately, there is no comprehensive and consistent data available on these
measures for the transit systems in the cities of this study. Therefore, these variables were
excluded from the model.
Based on these considerations, an initial statistical model for monthly transit ridership
is postulated as follows:

i=, ,=5

where Y, is monthly transit ridership at time t; _Xir,S,, WD, are explanatory variables at
time t, and U, is an error term. The Box-Cox transformation is

y(i) _ 1
I where ;i # 0
y(i)= 1”
f
In Y, 1. = 0
X!“’ _ ]
L whereA#O, i=l,2,3
$4
II
= ;1
In Y, i =o.

Clearly, when ;1 = 1 eqn (1) becomes linear and when ;1 = 0 the equation becomes linear
in logarithms.

tMesh density refers to route density that affects access time to and from the transportation mode (see Howell
ef al., 1975, p. 43).
 The impact of fare and gasoline price changes on monthly transit ridership 31

The variable X,, denotes the real transit fare. It is approximated by the adult cash fare
deflated by the Consumer Price Index for all urban wage earners. Ideally, the average
transit fare for a system should be measured by a composite index of various kinds of
transit and pass fares. However, the data for a composite transit fare index for any of the
local transit authorities of this study were not available. Thus, we compromised by using
the adult cash fare as a proxy to the underlying behavior of the average transit system fare.
The variable A’,, represents the real gasoline price. X,, denotes monthly vehicle miles
which are used as an approximation for the change in the level of transit services supplied
by each transit authority. The variable X5, is equal to 0 before January 1978, and is equal
to 1 otherwise. It is used to measure the effects of a changing definition in the recording
of ridership from revenue passengers to unlinked passenger trips for some transit
authorities. X6, is a trend variable which represents secular change occurring over a period
of years. It is defined as A’,, = (t - T/2) if T is an even number, and as X6, = (t - T + l/2)
if T is odd. t is a time sequence of 1, 2, . . T where T equals the total number of monthly
observations. The dummy variable, A’,,, measures the impact of the oil embargo (October
19733February 1974) on ridership, being coded 1 during the embargo and 0, otherwise.
S, and WD, represent, respectively, a set of variables for seasonal and working-day
variation in monthly transit ridership. The rationale and specification of these variables
for modelling seasonal and working-day variations in (1) will be discussed in Section 3.
The prior restrictions on the signs of the parameters of main interest are: (1) B, < 0;
(2)/$ > 0 and (3) & > 0. These prior restrictions on the signs of the parameters represent
the following hypotheses that we are interested in testing in this empirical study: (1) The
negative value of p, will support the hypothesis that ridership will decrease (increase) as
the real fare increases (decreases). (2) The positive value of fiZ will confirm the hypothesis
that there is a substitution effect between automobile and transit trips. It is generally
believed that commuters will substitute urban transit for automobile trips as the operating
costs of automobiles increases. Since 1979, the real gasoline price has fluctuated sharply
as compared to the other components of automobile operating costs. Thus, the gasoline
price is a reasonable proxy for capturing the variations in automobile operating costs. (3)
The positive value of p3 will support the hypothesis that transit ridership will increase
(decrease) as transit services increase (decrease).
It is generally recognized that surface transit and rapid transit are non-homogeneous
transit services. For this reason, we constructed separate models for surface and rapid
transit ridership for transit systems which have both components.

3. PROBLEMS OF ECONOMETRIC ESTIMATION

In this section we will briefly discuss our strategies for dealing with the following
econometric problems: (1) missing observations and strike effects; (2) estimation and
adjustment of seasonal and working-day (trading-day) variations in the regression model;
(3) choice of the functional forms for monthly transit models; and (4) autocorrelation
present in the residuals.
In the analysis of monthly transit ridership data, frequently there are missing
observations and observations with low values due to transit strikes. Our strategies is to
preadjust the ridership data due to these effects. The preadjustment method is as follows;
(1) using Ordinary Least Squares, the following model was estimated:

where Y, and X6, are defined in Section 2; and S, and WD, are defined in this section.
The variable Hi,, i = 1,2, . . . m, is equal to 1 in the ith missing observation and 0
otherwise. The estimated coefficient of H,,, CY, is the negative value of the estimated missing
observation (see Fuller, 1980). ST,,, being a zero-one dummy variable used to represent
strike effects, becomes a onetime zero-one dummy if the length of the strike period is one
month, which is often the case. The coefficient of ST,,, O,, is the estimated difference
32 G. H. K. WANG and D. SKINNER

between the would-be ridership with strikes and the observed ridership during the strike
period. (2) The preadjusted data was then computed as:

Y,,= Y,- f d,Hi,- i tqsq,.

,=I ,=I

The primary reasons for the initial adjustment of the data in this study are: (1) we want
to avoid the problem of estimating the regression models with autocorrelated errors and
missing observationqt and (2) the number of independent variables included in the model
have to be restricted in order to use a Box-Cox transformation to choose a linear vs
logarithmic functional form.
It should be mentioned that, for hypothesis testing, we multipled the estimated
standard error by a factor [N - P/N - P - (m + l)]“’ to take account of the loss of
(m + Z) degrees of freedom by preadjusting m missing observations and I strike effects in
the original data (see Seber, 1978). The variable P represents the number of parameters
estimated in model (1).
Visual inspection of monthly transit ridership data reveals that it exhibits strong
seasonal and working-day variation. Our approach is to model seasonal and working-day
variation along with other parameters in the regression model. The set of variables
representing fixed and changing seasonal variations are specified as:

S, = 2 (N,,~+ ch,K&l)sin(WK.X6,>+
K=l
t
K=l
(h,, + S2,,&>~0s (WK.&,I

where

21r.K
w,=-- (4)
12 .

We can view S,, as specified in (4), as being comprised of two parts:

&= i
K=l
[qK sin(WK.&J + hK cos(WK~&>l+ b,, ~0s (W6~X6,> (5)

representing fixed seasonal variation and:

s2t = Ki, a2,K ' ix& ' sin(wKx6t)l + 5 62,K[x6t ’ cos( wK x6t>l (6)
K=l

representing changing deterministic seasonal variation (see Stephenson and Farr, 1972).
Strong month-to-month variation in the ridership data is the result of working-day
variations. The reasons for the existence of these working-day variations in the transit
ridership data are as follows:
(1) The level of ridership on weekdays is higher than that of weekends; thus there exists
a weekly cycle in the daily ridership data.
(2) Monthly ridership data is an aggregate of daily data.
(3) Each calendar month has a different number of working days and holidays.
In this study, we adopted the specification of working-day variation variables for
transit series suggested by Wang (1981b). The set of working day variables are specified
as:

WD, = $, WW2, + $,WW6,

tThe usual Durbin-Watson statistic is less powerful in detecting first-order autocorrelations when the number
of missing observations are larger than three. An alternative, modified D-W statistic should be calculated (see
Savin and White, 1978). Further, the usual Cochrane and Orcutt (C-O) type of autoregressive transformation
is not strictly applied in this case because V(r&J*) # d21; where T is the C-O type autoregressive transformation
matrix and g* represents error terms with missing observations.
 The impact of fare and gasoline price changes on monthly transit ridership 33

where

WW2, = [TD, - 5.(TD,, + the number of holidays in the tth month)]

WW6, = [TD,, - (TD,, + the number of holidays in the tth month)].

$, is the coefficient of the weekday effect and $6 is the coefficient of the Saturday effect.
The coefficient for the Sunday effect can be derived as $,,, = - 5$, - tie, TD, denotes the
number of occurrences of the ith day of the weeks on month t. TD,, can take on only two
values, 4 and 5. TD, represents the total number of working days occurring in month t
(it should be noted that holidays occurring during the weekdays of month t have been
subtracted out).
The rationale for the inclusion of working-day variations in the model rather than the
preadjustment of working-day variation from the data are as follows: (1) Preadjustment
of working-day variation in the series by the Census X-l 1 procedure has an undesirable
property because the working-day variation coefficients estimated by the X-l 1 procedure
are biased (Cleveland and Devlin, 1982); on the other hand, the working-day variation
estimated by this procedure is unbiased. (2) Autocorrelation functions of residuals fail to
detect working-day variation omitted from the series and the inclusion of working-day
variables in the model provides us with a direct test on the existence of working-day
variation in the series as well as a procedure for directly estimating it from the data (see
Wang, 198lb).
In general, the application of economic theory and prior knowledge of the transit
industry can serve only to provide information about the signs of the variables and which
potential variables to include in the model. For this reason, researchers usually choose
arbitrarily between linear and logarithmic functional- forms of the model. In this paper,
the Box-Cox transformation technique (Box and Cox, 1960) is used as a test to select
alternative functional forms for the model (l).? The statistical procedures adopted in this
study to estimate the optimal values of 1, and the parameters of the model in (1) are as
follows: (1) first transform Yt and xixi,into Yy) and Xl:‘, i = 1,2,3. The notation for YI(1)
and X$ is defined in Section 2. The value of L is specified in the range of [- 1.5, 1.51 with
incremental intervals of 0.1. Following this transformation, ordinary least squares
estimation is performed on each set of transformed variables. The value of the log-
likelihood function of (1) was calculated and the estimated optimal value chosen was the
one which maximizes the log-likelihood function of (1). Then a 95% confidence interval
for the optimal value of f is constructed. Based on this confidence interval, a test for the
choice of the functional form was performed. If both linear and logarithmic functional
forms are rejected, then the generalized functional form according to the optimal value
of 3, is chosen.
When the functional form is chosen, the next step is to detect and model the structure
of autocorrelations in the estimated residuals by various statistical tests. These test
statistics include: (1) the Durbin and Watson statistics (D-W); (2) generalized higher order
D-W statistics (see Vinod, 1973); (3) autocorrelations of estimated residuals; and (4)
application of an overfitting procedure to the estimated residuals to consider the general
class of error structures (see Fuller, 1976) such as:

(1 - #L@‘2)li, = e,

(1 - /@)(I - P,2B’*)z&= e,.

These various tests are employed complementarily to consider the large class of
autoregressive error structures in the residuals.

tThe Box-Cox transformation technique (see Box and Cox, 1963) was used by Zarembka (1968) to test the
functional .forms for the demand for money. Wang ef al. (198 1) applied this technique to choose functional forms
for air cargo demand. Their paper discusses the implications of choosing alternative functional forms for
long-term forecasts.

TR-B Vol 188. No IL4

34 G. H. K. WANG and D. SKINNER

According to the identified autoregressive structure of the estimated residuals, Gener-

alized Least Squares is used to estimate the parameters of the model, given the chosen
functional form.? Finally, the estimated residuals from the model estimated by GLS is
again examined by various autocorrelation tests. The model is considered to be adequate
when the residuals are distributed approximately as white noise.

4. EMPIRICAL RESULTS
4.1 Estimated monthly transit ridership models
Following the estimation strategies mentioned in Section 3, we estimated eight monthly
transit ridership models for the seven U.S. cities of this study. New York City has a
separate model for surface transit and rapid transit, respectively. The other six cities have
one equation for surface transit ridership because rapid transit is not present. The
estimated monthly models and associated statistics are reported in Table 1. Because of the
space limitation of Table 1, the estimated coefficients of seasonal and working-day
variation for the (corresponding) monthly models are reported in Table 2. The sources of
data used in this study are reported in the Appendix. Generally, the variables used in these
eight equations have the correct sign and are statistically significant for at least the 10%
level.
There are, however, some exceptions among the variables labeled X,, through X,,.
Neither the changing definition (X,,) nor the trend variable (X,,) are significant at the 5%
level in the Albany equation. The oil embargo variable (X,,) is not significant at the 5%
level in the Atlanta equation. For the Baltimore equation, both the real gasoline price (X,,)
and the oil embargo variable (X,,) are not significant at the 5% level, and the trend variable
(X,,) did not indicate the existence of any secular trend and was not included in the final
equation. The real cash fare variable (X,,) for Des Moines is not significant. Des Moines
did not have any nominal fare change during the estimation period and hence the only
variation from which to measure changes in ridership comes from the Consumer Price
Index. As well, for Des Moines, the trend variable (X,,) was omitted from the final equation
because there was no indication of significance. Neither the real fare (X,,) nor the real
gasoline (X,,) variables for Miami are significant at the 5% level. The trend variable (X,,,)
was omitted from the final Jacksonville equation because there was no indication of
significance. The real fare variable (X,,) for New York City rapid transit is not significant
at the 5% level. New York City, unlike other cities of this study has both a surface (bus)
and rapid transit system. The cash fare for surface transit was included in the rapid transit
equation; and the cash fare for rapid transit was included in the surface transit equation.
However, neither of the attempts to test for a substitution effect proved to be statistically
significant.
In estimating the equation for Atlanta transit, two improvements have been made to
Bates’ work (1981). First, a set of harmonic terms are used to represent seasonal variation
in linked trips. Bates used only a school-day variable to represent seasonal variation in
linked trips and thus made a specification error by omitting some of the seasonal variation.
Since the vehicle mile variable itself exhibits seasonal variations, the correlations between
the vehicle miles variable and omitted seasonal variation variables are not equal to zero.
Hence the estimated coefficients of Bates’ equation are biased. The direction of biases can
be analyzed following the framework of specification analysis suggested by Theil (1957)
(see Cassidy, 1981, Chap. 6, for a clear discussion for practical econometric users). On the
other hand, our Atlanta model has complete specification of seasonal variation in the
equation. Thus, we did not commit a specification error for seasonal variation in the
model. This is the reason why Bates’ estimate of the vehicle-mile variable is 0.45 and the
corresponding coefficient estimate is 0.22 in our model. Hence, his coefficient estimate of

tit should be mentioned that a more efficient procedure to simultaneously estimate 1 and autocorrelation
has been proposed by Savin and White (1978) and Gaudry and Dagenais (1979). However, due to the large
number of independent variables in our models, an efficient computer program for the simultaneous estimation
of 1 and p was not available.
 Table I. Estimatedmonthly transitridershipmodels for seven U.S. transit authorities

CITIES Fmn ii2 D-Y F-STAT Am(l)

5 Xlt X2t 5 X5t X6t X7t
--

(1) ALBANY LIMM 1442.92 -2767.99 588.59 27.97 6 ,937 2.04 89.86 -
(107.95) (525.72) (143.23) (16.37) ,::, t *IDt

llr perlod: I: 1973 to 12: 1980

(2) An.AnlA LlKfAR 2.54 - .014 .034 .609 .120 ts ,909 2.05 66.94 ,213
(:g, (.083)
t *t
l.45) (.001) (.ooeI l.2061

fir p&o& 1: 1970 to 12: 1979

(3) BMwxlllf LII(EM 11556.30 -17210.60 515.30 4j. .743 2.40 20.81 .643
t T
(1739.56) (5854.22) (FZ) D60.05)
Tlw perlod: I: 1973 to I: 1981

(4) 9fs w)IlES LOG 6.19 -.29 .I6 tit trio, .813 2.12 19.51 .307
I.911 (.47) ,::, l.07)

Tlr perlod: 9: 1976 to 12: 1980

(5) .MfxsomlLLf LOG 4.00 -.26 .19 .32 ls .916 1.97 35.06 -
t -t
(-32) (.031 t.031 (.W

Tlr perlod: 1: 1976 to 3: 1960

(6) ltllv(I LOG 8.52 -.oo@ .12 .002 - .Ol 6

t
.825 1.61 27.34 .762
*t
f.13) (.052) (.W (.ooOS) I.02)

llr period: I: 1973 to 12: 1980

(7) NW YORK LOG 4.Bo -.lS .I7 .07 .04 -.002 45 .900 2.02 49.25 .381
(SURFAtE t.071 (.02) (.0004) t -t
(.42) (.07) (.07)
TRAIlSIT)
llae perlod: I: 1972to 10: 1960
^
WY WIRK LOG 9.77 -.Os .21 -.OOl .868 2.09 47.13 .207
(8) ,::, (G, ‘St *t
(RAfw 1.33) (.M) (.04) (JO031
1MNsl1)
llr perlod: I: I972 to 10: 1960
 Table 2. Estimated coefficients for seasonal and working-day variation for the monthly transit ridership models reported in Table I

st
CllltS FORM Xl 551 SC2 ss2 SC3 553 SC4 SS4 SC5 SSS SC6 m w6

ALBANY lItEM 56.16 70.24 -9.10 -68.4? -11.70 18.03 36.78 30.72 -24.06 -13.59
(6.69) (6.64) (6.34) (6.25) (6.41) (6.@N (5.82) (6.22) (5.56) 0.99) (K, g::, 0
I
ATLAurA LllCM -.Ol c::, -26 - .03 .O? -.06 .12 -.02 .Ol -.004 -.M .03 -.05
l.09 l.03) t.021 l.02) (.OZ) (.OZ) 7;
(.02) i.02) (.a?) t.011 i.01) t.031
P
OUfllME lII(LM 273.24 231.41 - 309.84 -783.19 -191.22 374.32 256.22 166.55 -179.05 54.99 -71.05 $
(126.56) f 124.06) (78.16) (I7.26) (60.551 (51.22) (48.S3) (51.09) (rn.31) 11.45) (69.10) 0

MS mms LOG .I2 -.M -.M .02 -.a309 -.Ol -.004

s
*
l.02) .r:% t.021 (.02) t.011 (.OlO) (.Ol) (:z, (.016)
P
JACKSONVILLE LOG .axJl .03 -.004 -.005 .OO? .02 -.009 .llO9 -.005 .ooT -.OZ %
c:%, (.003) I.0031 (.003) (.Jm4) i.003) (.@34) (.004) (.004) (:Ey (.OO? (.a311 (.fJw 2
2.
nlN4l Lo6 -.Ol &e, - .003 -.02 -.07 .OZ .O? -.07 .07 -.ss7 B
-.Ol
(.OtW I.0041 (.oosl t.0031 (.a)31 (.003) (.aJ3) (.W3) dE, (.002 (%, (.OW

)ItY YORK LO6 -.02 .03 .04 -.03 -.oz -.03 .a2
(SIRfACE (.OW &?6, (.ooa) (.005) (.~5) I.0051 (.MK) &E, (.Ool) (:z
TWNSII

MY YORK 106 -.Ol .02 .03 .03 -.0006 -.a3 .01 -.a% .007 -.009
(RAPID (.cQ5) (.001) (.rn4) (.ooO t.0031 i.003) (.003) (.OO? (.ooo71 t.0041
lRNl5IT)
 The impact of fare and gasoline price changes on monthly transit ridership 31
the vehicle-mile variable suffers upward bias. Second, Bates (1981) used ordinary least
squares for estimation throughout his study. None of the D-W statistics of his log models
(Bates, 1981, p. 94) indicate that the residuals are free from a first-order autoregressive
process. As a consequence, the standard errors of the estimated parameters are biased and
so too are the corresponding t-statistics. In our estimation procedure, generalized least
squares was used when the error terms were found to have a first-order autoregressive
representation.
The Box-Cox transformation technique was used to test the choice of functional forms.
In this study, it is found that linear forms are preferred over log forms for Albany, Atlanta,
Baltimore; and log forms, for the rest of the monthly models.?
Six out of eight models have autocorrelation in the residuals and the first order
autoregressive process is adequate to describe autocorrelation properties for these esti-
mated residuals.

4.2 An overview of empirical results

In this subsection, we first discuss the similarities and differences of the empirical results
of these seven cities. Some policy implications based on these empirical results are then
discussed. Finally, a brief comparison of our empirical results with those of previous
studies is presented.
For convenience of our discussion, the estimated elasticities with respect to real fares
and real gasoline prices for all seven cities are presented in Table 3. The table shows several
interesting findings. Transit elasticities with respect to real fares are negative and inelastic
for all seven cities of this study, ranging from -0.42 to -0.62. It is believed that the
underlying reason for the differences in the real fare elasticities among these cities is due
to the difference in the substitutability of transit trips by other modes of transportation.
It is apparent that transit is mainly used for working trips during the weekday at which
time traffic is generally congested in the central business district of larger cities and
commuters usually have to encounter parking problems and pay higher parking fees. It
is obvious that, under such conditions, the automobile is a costly substitute for transit.
Therefore, it is expected that the real fare elasticity for cities with congested roads and high
parking fees is low relative to cities with less traffic congestion and parking problems. This
result also suggests that transit users consider the payment of fare as a cost to reduce the
time spent traveling to and from work. We also observe that the elasticity of demand for
bus service with respect to the real fare is higher than the elasticity of demand for rapid
transit. This implies that automobiles and bus service are closer substitutes for each other
than automobiles and rapid transit. There are two possible reasons for this: (1) buses, autos
and taxis usually share the same road conditions, and (2) bus trips are usually shorter and
more causual, with taxi service, or even a walk, being a close substitute.
The findings of real fare elasticities have some interesting policy implications: (a) In
general, transit demand with respect to the real fare is inelastic. Hence an increase in the
fare will increase total revenue. (b) The fact that the real fare elasticities of surface transit
is more elastic than that of rapid transit has two implications: (1) the impact of the same

tThe empirical results of testing linear vs logarithmic forms by the Box-Cox transformation technique are
summarized in the following tables:

City numbers* (1) (2) (3) (4) (5) (6) (7) (8)

L-max 357.4 620.2 215.1 - 167.5 - 153.7 -487.4 -749.7 - 678.2

Optimal /I 0.7 1.1 0.9 -0.03 -0.2 - 0.06 0.07 0.2
95%
Confidence (1.1,0.3) (1.3,0.9) (1.5,0.4) (0.2, -0.19) (0.9, -1.1) (0.6, -0.61) (0.3,-0.3) (0.5, -0.1)
interval
Functional
form Linear Linear Linear Log Log Log Log Log

*Note: see Table 1

38 G. H. K. WANG and D. SKINNER

Table 3. Estimated elasticities

CITIES TIME PERIOD FUNCTIONAL REAL RGAL GASOLINE

FOW4 FARE

1. Albany, NY 1973: 1 to 1980:12 LINEAR -. 62* .22.

(0.05) (0.05)

2. Atlanta, GA 1970: 1 to 1979:12 LINEAR -.042” .25*

f .013) f .06)

3. Baltimore. MD 1973: 1 to 1981:l LINEAR -. 38* ,12

(0.13) C.10)

4. Des mines, 10 1976: 9 to 1980:12 LOG -.2& .80*

1.47) 1.31)

5. Jsckaonville. FLA 1976: 3 to 1980:3 LOG -.26+ .18’

C.03) C.03)

6. Niami. FL 1973: 1 to 1980:12 LOG -.009 .12

c.052 C.09)

7. New York 1972: 1 to 198O:lO LOG -. 15’ .17*

(Stirface Transit) (.07) C.07)

8. New York 1972: 1 to 1980:%0 LOG -*OS . oa*

(Rapid Transit) (.010 (.04)

NOTE: 1. Estimated elasticities evaluated at sample mean for linear

fmccional form.

2. The standard errora for the estimated elasticities are

contained in parentheses.

3. *Denotes variable is stastisticslly significant at 5X level.

percentage increase in real fares will cause relatively less of a percentage change in rapid
transit ridership than in bus transit ridership; and (2) an increase of the same percentage
in fares will increase revenue for rapid transit more than that for bus transit.
Transit demand elasticities with respect to the real gasoline price are positive and
inelastic for all the cities of this study. The estimated real gasoline price elasticities lie in
the range of 0.08-0.80. It is found that the cross elasticities of bus demand with respect
to the real gasoline price is greater than those of rapid transit demand with respect to the
real gasoline price.t This is expected because the automobile is a closer substitute for bus
transit than is the automobile for rapid transit. These findings are interesting in terms of
policy implications. For instance, an increase in the price of gasoline will have substantially
more of an impact on ridership and revenue of surface transit than on rapid transit. In
general, the cross elasticities of transit demand with respect to the real gasoline price is
in the range of 0.08-0.80. Thus, as the real price of gasoline increases (decreases) so will
ridership and revenue, but the increase (decreases) is small.
The vehicle mile variable was incorporated in the transit demand models of Atlanta,
Georgia, New York City, and Jacksonville, Florida. We did not incorporate this variable
in the transit models of other cities because monthly data for this variable was not available

?We should mention the limitation of this conclusion because in this study only one rapid transit equation
(New York City) was estimated; the other equations are all for surface transit. However, some previous studies
which used both annual and cross-section data obtained similar results (see Lago et al., 1981).
 The impact of fare and gasoline price changes on monthly transit ridership 39

at the time this study was undertaken. The elasticities of transit demand which respect to
vehicle miles is reported in Table 4.
Based on this limited empirical evidence, the elasticities of demand for surface transit
with respect to vehicle miles are positive and statistically significant at the 5% level. It is
also very interesting to observe in the analysis that the elasticities of surface transit with
respect to vehicle miles are higher than those of rapid transit.
The monthly transit ridership for all seven cities exhibits strong seasonal variation.
Generally, the pattern of seasonal variation is for low ridership in the summer and winter
months. In the case of Miami, Florida, as was expected, there is peak ridership in the
winter months because of the tourist season. As well, it should be stressed that the
empirical evidence shows specific seasonal ridership patterns among the cities due to, it
is felt, differing social and institutional habits such as tourist seasons, popular vacation
seasons, and school vacations.
There is also strong working-day variation for all cities in the transit ridership data.
The level of ridership is high during the weekday and low during weekends (see Table 2)
because transit is mainly used for trip-to-work purposes. On weekends the auto becomes
a more desirable mode of transportation because the transit system may not serve areas
commuters wish to visit.
Before comparing our study to previous related works, it is worth mentioning that the
estimated elasticities depend on model specification, the time frame used for observations
(such as monthly, yearly, or cross sectional), the length of sample period, and the
procedure used to calculate elasticities (such as point or arc elasticity as well as the level
at which the elasticity is estimated such as the mean). Therefore, it is very difficult to make
direct comparisons between elasticities of this study and other studies. We thus compro-
mise by comparing only the magnitudes of elasticities.
In a survey paper on fare elasticities, Kemp (1973) found that fare elasticities computed
from demonstration studies and direct demand models estimated with annual time-series
data were in the range of -0.1 to -0.7. The range of variation is because of city size,
central city area, and time of day. Recently, Lago et al. (1981) presented a more updated
survey of fare elasticities estimated prior to 1980. They found that transit fare elasticities
were also inelastic and varied from -0.04 to -0.87. Further, they observed that transit
fare elasticities vary by city size, mode, and by time of day. The empirical results of our

Table 4. Estimated transit demand elasticities with respect to vehicle miles

CITIES TIt4E PERIOD FUNCTIONAL

FORM

Atlanta, CA 1970: 1 to 1979:12 LINEAR .27*

C.09)

Jacksonville, FL 1976: 1 to 1980:x LOG FORM .32*

(0.06)

New York City

Stirface Transit 1972: 1 to 1980:10 LOG .07*

(0.07)

Rapid Transit 1972: 1 to 1980:10 LOG .21*

(0.04)

*Denotes variable is statistically significantly at 5 percent level.

The standard errors for the estimated elasticities are contained

in parentheses.
40 G. H. K. WANG and D. SKINNER

study are comparable to the two surveys mentioned above. With regards to the gasoline
price elasticity, Wang and McCarthy (1978) found that the elasticity of national quarterly
transit demand with respect to the gasoline price is 0.25. At the city level, Agthe and
Billings (1978) found that the elasticity with respect to real gasoline price is 0.42 for
Tucson, Arizona; Bates (1980) found that the elasticity of monthly transit demand with
respect to real gasoline prices was 0.23 for Atlanta, Georgia; while Kemp (198 1) observed
that the elasticity of monthly surface transit demand with respect to real gasoline price was
0.29 for San Diego. Our results for this elasticity for the seven cities of our study range
from 0.08 to 0.80, a range which also includes all of the author’s previous estimation of
this elasticity. Thus again, our empirical results are reinforced by those of previous studies.

SUMMARY AND CONCLUSIONS

In this paper we have constructed and estimated eight empirical models of monthly
transit ridership for seven U.S. transit authorities. The merits of this type of approach lie
in the ease of implementation and the abundant availability of data from local transit
agencies. Within the framework of these models, we examined the impacts of changes in real
fares and gasoline prices upon ridership for each transit authority with the following
important findings:
(1) Elasticities of monthly transit ridership with respect to the real fare are negative and
inelastic, ranging from 0.042 to 0.62. Furthermore, the elasticity of demand for surface
transit, with respect to the real fare, is greater than the elasticity of demand for rapid
transit.7 These findings have some interesting policy implications: (1) The percentage
decrease in ridership is less than the percentage increase in the real fare. Hence, an increase
in the real fare will increase total revenue. (2) An impact of the same percentage increase in
real fares will cause relatively less of a percentage change in rapid transit ridership than in
bus transit ridership. Thus an increase of the same percentage in fares will increase revenue
for rapid transit more than for surface transit.
(2) The elasticities of monthly transit ridership with respect to the real gasoline price are
positive and inelastic, ranging from 0.08 to 0.80. These elasticities indicate that as the
gasoline price increases (decreases) so will ridership and revenue, but by a small amount.
(3) The monthly transit ridership for all seven cities exhibits strong seasonal and
working-day variation. Generally the pattern of seasonal variation is low for ridership in the
summer and winter months. However, in the case of Miami, Florida, as is expected, there
is peak ridership in the winter months. For working-day variation patterns, ridership is high
during weekdays and low during weekends because transit is mainly used for working-trip
purposes.
In conclusion, the positive contributions of this paper are in providing new empirical
evidence on fare and gasoline price elasticities from models of ridership from seven U.S.
transit authorities. These elasticities, obtained from empirical models estimated from
monthly time series data, are superior to elasticities obtained from transit ridership data
before and after a fare change (the Simpson and Curtin Formula for shrinkage ratio
procedures) because our methodology controls for the effects of other variables. Our
findings also suggest that the general Simpson and Curtin formula which predicts a ridership
decline of 3.3% for each 10% increase in fare should be used with extreme caution in
predicting at the disaggregated city level and for predicting by specific transit mode.

REFERENCES
Agthe D. E. and Billings R. B. (1978) The impact of gasoline price on urban ridership. Annuls Reg. Sci. 9&96.
Bates J. W. (1981) A study of demand for transit use. Ph.D. Dissertation, College of Business, Georgia State
University.
Box G. E. P. and Cox C. R. (1964) An analysis of transformation. J. Roy. Stat. Sot. B26, 211-243.
Cassidy H. J. (1981) Using Econometrics: A Beginner’s Guide. Reston, Virginia.
Cleveland W. S. and Devlin S. J. (1982) Calendar effects in monthly time series: modelling and adjustment. J.
Am. Stat. Assoc. 77, 52C528.
Curtin J. F. (1968) Effect of fares on transit riding. Highwa_r Res. Rec. 213, 8820.
Driver J. C. (1979) Price elasticity estimates by quasi-experiment. Appl. Economics 11, 147-155.
 The impact of fare and gasoline price changes on monthly transit ridership 41

Fuller W. A. (1976) Introduction to Statistical Time Series. Wiley, New York.

Fuller W. A. (1980) The use of indicator variables in computing predictions. J. Econometrics 12, 231-243.
Gaudry M. (1975) An aggregate time-series analysis of urban transit demand, the Montreal case. Transpn Res.
9, 249-258.
Gaudry M. and Dagenais M. G. (1979) Heteroscedasticity and the use of Box-Cox transformation. Economic
Lett. 2(3), 225-229.
Hovel1 P. J., Jones W. K. and Moran A. J. (1975) The Management of Urban Public Transport. Saxon House,
Great Britain.
Kemp M. A. (1973) Some evidence of transit demand elasticities. Transpn 2, 25-52.
Kemp M. A. (1981) Bus services in San Diego-a study of patronage growth in mid-1970’s. Working paper, The
Urban Institute.
Lago A. M., Mayworm P. D. and Mcenore J. M. (1981) Transit ridership responsiveness to fare changes. Trajic
Q. 35, 117-142.
Liu T. C. (1969) A monthly recursive econometric model of the United States: a test of feasibility. Rev. Economics
and Stat. 51, 1-13.
McCallum B. T. (1972) Relative asymptotic bias from errors of omission and measurement. Econometrica 40,
757-758.
Savin N. E. and White K. J. (1977) The Durbin-Watson test for serial correlation with extreme sample sizes
or many regressors. Econometrica 45, 198991996.
Savin N. W. and White K. J. (1978) Estimation and testing for functional form and autocorrelations: a
simultaneous approach. J. Econometrics 8, l-12.
Savin N. E. and White K. J. (1978) Testing for autocorrelation with missing observations. Econometrica 46,
59-67.
Seber G. A. F. (1977) Linear Regression Analysis. Wiley, New York.
Simon J. L. (1966) The elasticity of Liquire in the U.S. and a simple method of determination. Econometrica
34, 193-205.
Stephenson J. A. and Farr H. T. (1972) Seasonal adjustment of economic data by application of the general linear
statistical model. J. Am. Sfatist. Assoc. 67, 37-45.
Theil H. (1957) Specification errors and the estimation of economic relationships. Rev. fnt. Statist. Instit. 25,
41-51.
Vinod H. D. (1973) Generalization of the Durbin-Watson statistics for higher order autoregressive process.
Commun. Statist. 2, 115-144.
Wang H. G. and McCarthy E. (1978) Quarterly forecasting models of transportation aggregates. Transportation
Systems Center Staff, Study Rep. No. SS-221-U9-7(4).
Wang H. G., Maling W. and McCarthy E. (1981) Functional forms and aggregate U.S. domestic air cargo
demand: 195&1977. Transpn Res. 15A, 249-256.
Wang H. G., Ward D. and Hassler F. (1981a) Possible effects of fare increases on New York ridership and
revenue. Transportation Systems Center Staff, Study Rep. SS-24-U3-199.
Wang H. G. (1981b) An intervention analysis of interrupted urban transit time series: two case studies. Proc.
Business and Economic Statistics Section, American Statistical Association, pp. 424429.
Zarembka P. (1968) Functional form in the demand for money. J. Am. Statist. Assoc. 63, 502-511.

APPENDIX
This Appendix discusses the measurement and sources of data used in this study:
Y,: monthly ridership, usually measured in thousands. Data for Atlanta was obtained from the Appendix
contained in Bates (1981); for Jacksonville, which is measured in 10,000’s, from the unpublished records of the
Jacksonville Authority; and for Miami, from the unpublished transit operating statistics of the Miami Transit
Authortiy. Data for the other cities was obtained from Monthly Transit Ridership published by the APTA
Statistical Department. APTA recorded monthly ridership as revenue passengers up to and including December
1977. From January 1978, APTA has measured ridership as unlinked transit passenger trips, The Atlanta data
is entirely measured as passenger linked trips while Jacksonville and Miami is entirely measured in revenue
passengers. New York City data was obtained from published records of the New York City Transit Authority.
X,,: real adult cash fare. Data for Atlanta was obtained from the Appendix in Bates (1981). New York City
data was obtained from published records of the New York City Transit Authority. All other fare data was
obtained from the APTA Fare Summaries, published by the APTA Statistical Department, and deflated by the
national Consumer Price Index for all urban wage earners, published by the Department of Commerce and
obtained from the Survey of Current Business. The adult cash fare for New York City, however, was deflated
by the Consumer Price Index for all wage earners in New York City, and the adult cash fare for Miami by the
Consumer Price Index for wage earners in Miami.
X2,: real gasoline price per gallon including taxes. Monthly gasoline prices which are an average of the weekly
pump level prices were obtained from the Oil and Gas Journal. The retail gasoline prices were deflated by the
Consumer Price Index for all urban wage earners except for Miami and New York City which were deflated
by the local Consumer Price Index.
X,,: vehicle miles supplied, which is used in the equations of three cities. The Atlanta data was obtained from
the Appendix contained in Bates (198 1). The Jacksonville data from unpublished records of the Jacksonville
Transit Authority. New York City from the Transit Record published by the New York City Transit Authority.