You are on page 1of 10

~ronrpn. &,.-A Vol. 19A. No. I. Pp. 73-82. 1985 0191-2607185 53 oO+ .

@I
Rmed III the U.S.A. 0 1985 Pergamon FTcssLtd.

PERFORMANCE EVALUATION FOR BUS TRANSIT

GORDON J. FIELDING,
TIMLYNN T. BABKSKY and MARY E. BRENNER
School of Social Sciences and Institute of Transportation Studies, University of California,
b-vine, CA 92717, U.S.A.

(Received 4 May 1984; in revised form 20 September 1984)

Abstract-Performance analysis has gained increased attention now that transit organizations are emphasizing
efficiency objectives. Debates over both the number of indicators required and their method of selection are
reviewed. Research upon which this paper is based uses FY 1980 Section 15 data to first identify and then test
a set of performance indicators which are useful for evaluation of fixed route, motor bus transit. Four parallel
data sets, based on transformations of the original data, and several exploratory factor analyses were used to
detect the underlying structure of the data. Rigorous testing verified that the structure represented the most
salient performance dimensions. A small subset of seven performance indicators was identified and tested as
representative of these underlying dimensions. These indicators can be used together or individually to assess
transit performance for a single system or for cross-sectional comparisotis.

After two decades of expansion, the American transit efficiency is a statement about the achievements of an
industry has begun to consolidate its rejuvenation and agency in transforming a set of inputs into a set of out-
emphasize more efficient production of service. In part, puts. Others have used technological efficiency for
this has resulted from the desire of federal, state and this concept in public sector analysis in contrast to eco-
local governmental sponsors to evaluate relative per- nomic efficiency (Burkhead and Herrigan, 1978).
formance in order to account for public assistance, and Confusion over defining efficiency has reduced the
in part, it has come from a desire by managers to improve value of many studies. Public transit agencies cannot
performance within their agencies. focus on a single objective function as they must respond
Many conference sessions, workshops and profes- to the objectives of various publics. Advocacy of
sional papers have been devoted to the merits of different single measures of transit performance integrating effi-
methods of performance analysis and the merits of spe- ciency and effectiveness has not aided transit perform-
cific performance measures. Miller (1980) has summa- ance. Separate measures provide more useful results.
rized this literature and suggested the need for a simple These can be grouped into indicators of system efficiency
framework using three or four measures of efficiency and indicators of system effectiveness. Efficiency indi-
and effectiveness. This paper develops such a framework cators define the relationship between resource input and
and advances the thesis that there exists a consistent set produced output. They include indicators of overall cost
of performance concepts relevant to fixed route transit efficiency, labor utilization and vehicle utilization. Ef-
operations. And further, that a small, unique set of in- fectiveness indicators generally reflect the ability of transit
dicators can be selected from these concepts for the mea- operations to meet certain goals. These include service
surement of transit performance. utilization, service quality and accessibility of service.
Not all researchers agree that multiple measures of
transit performance are needed. Several authors have
PERFORMANCE EVALUATION USING SECTION 15 DATA
advanced claims that a single measure is sufficient. Tul-
ley and Andersen (1981) examine transit performance Section 1.5 of the Urban Mass Transportation Act of
from the perspective of a fii seeking to maximize rid- 1964, as amended, provides for the collection of a unique
ership within allowable deficit limits. They suggest that set of comparable transit statistics by requiring all transit
deficit per passenger is the appropriate governmental applicants for federal operating assistance to report a
objective for transit. Nash (1980) prefers cost per pas- uniform set of information about their system. Regula-
senger or passenger mile when analyzing alternative in- tions to implement the legislation were published in the
vestments for management. Patton (1983) has suggested Federal Register on September 17, 1975. The methods
that transportation statistics on performance can be in- used originated from Project FARE (Uniform Financial
tegrated about the indicator of cost per passenger. Each Accounting Reporting Elements) which the Urban Mass
assumes that the overarching goal for transit is trans- Transportation Administration (UMTA) and the transit
porting passengers for the minimum cost. Kneafsey (1975) industry had initiated in 1972. The reporting system is
refers to this as efficiency-in-the-small or cost min- designed to meet the needs of: (1) individual transit sys-
imization to the firm. tems for comparing their performance with other transit
Allocative efficiency-what Kneafsey calls effi- systems with similar characteristics; (2) transit industry
ciency-in-the-large, is more suitable for transit per- associations for monitoring industry performance and (3)
formance analysis. This defines efficiency in terms of Federal, State and local government agencies for transit
resources used to produce service. By this definition, industry analysis and for financial assistance program
73
74 G. J. FIELDINGef al.

administration. Data is sent to UMTA and edited and ferent units of analysis; e.g. financial data is fiscal year
compiled by the Transportation Systems Center. Copies based while service consumed and service supplied in-
are available in magnetic tape version or as hard copy formation is reported for an average weekday, Sat-
in annual editions of the National Urban Mass Trans- urday and Sunday. Further, revenue data and subsidy
portation Statistics. data are not disaggregated by mode; they are reported
The first year of Section 15 reported statistics (FY for the entire system, This requires the development of
1979) was used by Anderson and Fielding (1982) to test weighting strategies for single mode performance anal-
the performance concept model developed by Fielding, ysis. Data organization, verification and correction tech-
Glauthier and Lave (1978). The results were not satis- niques used in this research are described in more detail
fying. Serious questions were raised about the validity elsewhere (Fielding et al. 1984).
and completeness of the first years data, although the The first task of analysis entailed compilation and
method of using factor analysis to identify the underlying assessment of each of the basic variables that would later
structure in the data held promise. If the data set could be included as performance indicators. Some of the basic
be improved, mote rigorous factor-analytic solutions could variables available in the Section 15 data base are better
be applied on different versions of the data to test the than others; some present unique problems for multi-
validity of the performance model. variate statistical analysis. Revenue and expense data are
This paper analyzes data from the second year (FY fairly complete while operators wage, state and local
1980) of reported statistics (U.S. Department of Trans- subsidies, maintenance and passenger data present par-
portation, 1982). It replicates the methods and revises ticular problems. Systems differ as to how local versus
the results from analysis of the first year (FY 1979) state subsidy is defined; the same funding source may
statistics. Because the thesis that a highly consistent set be designated differently by several transit systems.
of performance concepts exists and that they can be Maintenance data is often suspect; systems using outside
represented by a small, unique set of performance in- contract maintenance labor often claim zero maintenance
dicators remains controversial, emphasis is given to ex- personnel and cost. Finally, passenger data is the weakest
plaining how: part of the Section 15 data base; between 13% to 25%
(i) variables were selected, verified and groomed; of the data is missing.
(ii) performance indicators were selected and calculated The problems identified here have not gone unnoticed.
in alternative ways to minimize bias; Suggestions from industry personnel and academic re-
(iii) different methods of factor analysis were used to searchers are continually solicited and evaluated to im-
explore the structure of performance concepts; prove the data base. Although agreement on definitions
(iv) tests were used to verify the structure of perform- and methods for refining the data collection process has
ance concepts. been slow in coming, the section 15 data base is becom-
ing increasingly sound. Both the quality of the data and
Categories of indicators the consistency with which it is reported are improving.
A wide variety of Section 15 statistics were evaluated
as possible performance indicators. Three categories of Missing data procedures
statistics-service inputs, service outputs and service Missing information in Section 15 data poses a unique
consumption-provided the framework to organize the analytical problem. Both valid zeros and no informa-
much larger set of data. Figure 1 portrays the organizing tion reported codes are represented by zeros. Whenever
framework originally advanced as a performance concept possible, other information available in the data base
model by Fielding et al. (1978). Cost-efficiency indi- was pieced together to provide for missing data or to
cators measure service inputs (labor, capital, fuel) to the distinguish between valid zeros and failure to report.
amount of service produced (service outputs: vehicle Missing values encountered at any point in the com-
hours, vehicle miles, capacity miles, service reliability). putation of basic and ratio variables and during the mul-
Cost-effectiveness indicators measure the level of service tivariate statistical procedures cause a snowball effect
consumption (passengers, passenger miles, operating of missing information. If any case is missing even one
revenue) against service inputs. Finally, service-effec- piece of information it is automatically deleted from the
tiveness indicators measure the extent to which service computations and subsequent analyses. The missing val-
outputs are consumed. ues problem has a cumulative effect as cases are dropped
from the analysis. Thus, from a total of 304 transit sys-
Data preparation tems running fixed route, motor bus service, only two-
The Section 15 data base includes information from thirds of the cases-198 systems-had enough infor-
15 separate required forms and three voluntary levels of mation available to use in the final analyses.
reporting. At the required level of reporting, information
is available for revenues, subsidies, expenses, wages and
SELECTING PERFORMANCE INDICATORS
benefits, service schedules, maintenance performance,
energy consumption, accidents, employee counts, ser- There were many potential performance indicator ra-
vice supplied and service consumed. For many purposes tios available from the Section 15 data base. In selecting
the level of detail is excessive and key pieces of infor- the set of performance indicators to be used, the data
mation are often fragmented into tiny units which must had to include variables that would relate to the con-
be aggregated before analysis. Many of the variables ceptual model, i.e. those that would best represent the
relevant to the performance concept model refer to dif- three categories of performance concepts-cost-effi-
Performance evaluation for bus transit 75

Service
Inputs
Labor
Capital
Fuel

Service- Effectiveness

Service Service
outputs Consumption
Vehicle Hours Passengers
Vehicle Miles Passenger Miles
Capacity Miles Operating Revenue
Fig. 1. Framework for a transit Performance concept model.

ciency, cost-effectiveness and service-effectiveness. to focus on the inflated variance due to the presence of
Particular attention was given to the availability and re- an outlier, rather than the more true-to-data variance
liability of the Section 15 reported data from which the present across the range of the other cases. Four cases
ratios would be calculated. As noted, some of the Section were dropped from the analysis because of the outlier
1.5 data variables were more complete or more reliable quality of their reported statistics.
than others. The next task was to check the univariate descriptive
Table 1 lists the initial set of 48 variables selected for statistics for each of the selected performance indicator
multivariate analysis. The variables are organized under ratios to evaluate the distribution of the case values across
the performance concept to which they relate. This set the variable range. Two descriptive statistics that provide
of 48 variables in most cases (other than passenger data) information on how far a variable deviates from a nor-
represents the most complete, generally reliable and non- mal-like distribution of values are skewness and kurtosis.
redundant performance indicators available in the Sec- For a normal distribution of data, both skewness and
tion 15 data set. kurtosis equal zero; for each statistic the further from
Variables based on revenue-capacity miles were not zero the value, the less normal-like is the data distri-
included because of the inconsistency in measurement bution. The less normal-like the distribution, then the
of that variable across systems. Ratios based on popu- more questionable are the statistical results.
lation data were not included because available popu- The skewness and kurtosis values for the list of 48
lation information reflected total urban population rather variables ranged from - 5.212 to 16.098 and from 1.373
than service area population. Otherwise, performance to 263.908 respectively, indicating that the distributions
indicator ratios comparable to the 1979 data analyses were far from normal. The proposed multivariate pro-
were selected for use. This facilitated comparison with cedures to be used on the performance indicator data set
previous results and identification of shifts due to the were considered relatively robust, i.e. valid even un-
better-collected, cleaner and more complete data. der deviations from normality. Robustness is of greatest
concern when using inferential statistical techniques.
However, even descriptive techniques, like the ones used
Distribution of data here, could be affected by highly skewed data. As the
One of the first tasks for exploring the data set was goal of this research was to provide a highly reliable set
to search for extreme outliers and to remove them from of consistent analytical findings that could serve as a
the analysis. Extreme outliers could force the analysis benchmark for cross-year comparisons, it was important
76 G. J. FIELDING PI ul.

Table 1. Performance indicators by concept

CUST EFFICIENCV -S SERVICEEFFECTIVENESS


MslJRt3
Labor Efficiency Utilization of Service

TVH/Dp Vehicle Hrs Per Employee TPAWRVH Passngr Trips Per Rev Veh Hrs
TPAS/RVM Passngr Trips per Rev Veh Ri
TPAS/W Passngr Trips per Peak Veh
PASWTPS Passngr Mi per Passngr

(beratina Safety

TVWACC l.DOO,OOD Veh Mi per Accident


RVH/KC Rev Veh Hrs per Accident
Vehicle Efficiency
Revenue Gerwation
TVWAVEH Vehicle Hrs Per Active Veh
WPVEH Vehicle Hrs per Peak Veh Req REWPVEH Passngr Rev per Peak Veh
TVWAVEH Vehicle Mi per Active Veh REV/RW Passngr Rev per Rev Veh Hr
TVIWVEH Vehicle Hi per peak Veh Req OREWRVH (per Rev per Rev Veh Hr
RVWTVM Revenue Veh Mi Per Veh Wi REV/TPAS Passngr Rev per Passngr

Fuel Efficiency Public Assistance

RVWFUEL Revenue Veh Ni per Cal Diesel RVlVLSUB Rev Veh Hrs Per Local Cap,
TVWFLEL Vehicle li (Bus) per Gal Diesel
RVWSSLB &?Lh*% per State Cap,
Mintenance Efficiency @er Asst
RWUOSUB Rev Veh Hrs Per Tot @er Asst
TVIVEXP Total Veh Mi per Naint Expense RVH/TSUB Rez;hA;g per Tot cap.
TV?l/!NT Veh Mi per Ralnt brployee
TVR/RCAL l.ODO,OOO Veh fli Per Roadcall TPA!YLOA Passngn Per Local Cper Asst
TPASJTSUI Passngrs per Tot @er. cap Asst
OutLNJt wr Dollar Cost REWTSUB Passngr Rev Per Tot cper.
Cap Asst
RVWDEXP Revenue Veh Hrs per REV/OSlB Passngr Rev per Tot @er Asst
Over Expense PAS/OSuB Passngrs per Tot @eer Asst
TVWOEXP Veh Mi Per Cper Expense
WlW Rev Veh Hrs Per lot Labor, COST EFFECTIVENESS MEASURES
Fringe Expense
RWU(wG Rev Veh Hrs per @er Labor, Service Ccmsmvtim LW Exwnse
Fringe Expense
RVHAWG Rev Veh Hrs Per Veh Mint Labor, PAUDEXP Passngrs pr LQeer Expense
Fringe Expense PASWOEX Passngr I41 r;orzrpense
RWUADU; Rev Veh Hrs per A&in Labor, PAS/W Passcgrs
Fringe Expense K nefits
PAS/FUEL Pa!xiE per Gal Diesel
PASWTEX Passngt rnl per Tot Expense

Revenue Generation per Exoense

OREVKEXP Dper Rev tooper Expense


TREWTEX Tot Rev to Tot Expense

to begin with a set of data that had a minimum of dis- reported data; (4) logs of the weighted data variables.
tributional problems. The purpose for developing these four sets was to ensure
that when final results from multivariate analyses were
Logarithmic and weighted data sets reported, most contingencies for possible bias in the data
To counter any possible bias in the analyses and to had been addressed. Consistent results across the four
provide a comparable set of more normally distributed data sets would provide evidence that, indeed, a stable
performance indicator variables, the base 10 logarithms performance concept structure had been found in the
of the 48 performance indicators were calculated to shift data.
the distribution of the data to a less skewed curve. This
provided two sets of comparable data--the 48 perfotm-
ante indicator ratios calculated from the Section 15 re- EXPLORATORY ANALYSIS

ported data and a set of 48 logarithm variables. Multivariate analyses were used to search for a highly
A third and fourth set of indicators were also devel- consistent set of performance concepts relevant to fixed
oped. As mentioned above, revenue data is reported as route transit and for a small, unique subset of perform-
a total for the whole system; it is not broken down by ance indicators. Factor analysis is ideal for detecting the
mode when more than one mode exists. Capital and structure within a set of data and for determining those
operating assistance are also not consistently separated few key variables with which a whole range of infor-
by mode. Therefore, a third set of performance ratios mation can be represented. The prime objective in this
was developed using basic variable data and revenue research was to search for the minimum amount of data
statistics that were weighted to eliminate revenue and necessary to convey the maximum amount of perform-
subsidies from modes other than bus transit. Then, a full ance information. Parsimony and consistency were the
set of 48 base 10 logarithms was calculated on the weighted key criteria; factor analysis was the most efficient means.
data: again, to provide more normally-distributed data.
As a result of the cleaning, verifying and grooming,
four somewhat different sets of performance indicator Factor analysis
data were available: (1) ratios from reported data; (2) The most distinctive characteristic of factor analysis
logs of reported data variables; (3) ratios from the weighted is its ability to reduce a large set of data to a smaller set
Performance evaluation for bus transit 77

of components or factors which portray the un- for motor bus operations it was difficult to validly com-
derlying structure of relationships among a set of vari- pare fuel efficiency across systems. Local and state sub-
ables. Based upon the correlation patterns of a large sidy related variables (e.g. RVH/LSUB, RVH/SSUB)
number of variables, the objective of the factor analytic were dropped because definitions of local versus state
technique is to group together those variables which are subsidies were inconsistent. Capital subsidy variables
highly correlated with each other. The analyst then in- were dropped because they can greatly shift from year
terprets each factor according to the variables belonging to year.
to the group. The idea is to summarize many variables The passenger miles (PASM) variable was missing
by using a few representative factors. from almost 20% of the cases. To increase the number
Factor analysis not only provides information on the of cases entering into the analysis, variables based on
number of factors underlying the data, it also determines PASM (e.g. PASM/OEX, PASM/TPS) were eliminated
which variables grouped on a particular factor are most from the data set.
highly related or representative of the identified factor. Variables related to active vehicle counts were also
The factor loading of each variable on the respective deleted. Distinctions were not always made between school
factors can be interpreted as the correlation of the var- buses, charter buses and other motor buses. Some agen-
iable with the factor; high factor loadings represent high cies listed more active vehicles than total vehicles while
correlations. for others, vehicle inventories were incomplete. Varia-
In performing any factor analysis, Comrey (1973) sug- bles using peak vehicle requirements (PVEH) were
gests that there are three problem areas that could exist retained.
in the data, and obscure the underlying data structure: The variable RVMlTVM was eliminated because 65
(1) two variables carry highly redundant information of the cases had revenue vehicle miles equal to total
(collinearity); (2) a variable loads across several factors vehicle miles, a strong indication of a definitional prob-
equally well (poorly defined structure in the variable) lem, which greatly inflated the kurtosis value of the
and (3) one factor has all or most of the variables weight- variable. The roadcall related variable, TVM/RCAL,
ing heavily on it (poorlydefined structure in the data was dropped because the definitions for what makes a
set). true roadcall were unreliable. The variables related to
The first exploratory factor analysis was begun with total expense (e.g. PASMITEX, TREVITEX) were de-
the most complete set of performance indicator ratios leted because total expense is not truly comparable across
available in the Section 15 data. It remained necessary systems; there are no set parameters for depreciating
to assess how well these variables measured the target capital costs. Finally. REV/RVH was so highly corre-
information and how relevant the indicators were for lated with OREVlRVH that it was eliminated.
cross-system analysis. The next task involved determin- With each exploratory factor-analytic pass through the
ing from the set of 48 variables which subset of variables data sets, the variables were checked against the factor
provided the best cross-system measures, and best de- structure to determine if remaining variables presented
fined the structure in the data while testing the data for any of the structural problems mentioned above. With
the three possible contaminating problems listed above. each pass through the data, the underlying structure be-
As four parallel sets of performance indicators were came more clearly defined. The number of cases entering
available, the same type of exploratory factor analysis into the analysis had increased from 128 to 198 and the
was carried out on each set. After each exploratory factor same general solution appeared across the four different
analysis was performed, the resulting factor loading ma- sets of data.
trix was evaluated. Finding similar results across the four The final set of 30 performance indicators that emerged
data sets would signal detection of the consistency in the after the fourth pass through the data reflected a strong
data which would point to the true underlying struc- set of performance indicator variables. These portrayed
ture in the variables. such highly consistent factor loadings across all data sets
that it was evident that the most salient features of the
Variable elimination performance concept model had been identified.
In the first exploratory analysis on the full set of 48 Table 2 lists the 48 performance indicator variables
performance indicators, a total of 128 cases were in- selected for analysis from the Section 15 data base. They
cluded in the analysis. As mentioned earlier, factor anal- are portrayed within the framework of the conceptual
ysis will drop from the analysis every case missing any model. Those variables eliminated prior to the final anal-
piece of information. Because the missing values were ysis are marked with an asterisk to offset them from the
scattered throughout the 48 variable set, the snowball final set of 30 performance indicators used in subsequent
effect had eliminated nearly two-thirds of the cases from analyses.
the analysis. Thus, in the next exploratory pass through
the data, it was decided to eliminate those variables that
FINAL FACTOR ANALYSIS ON 30 PERFORMANCE
compounded the missing data problem and those that
INDICATOR RATIO VARIABLES
were still somewhat questionable as to the quality and
comparability of reported information. This increased The final factor analysis was carried out on the cleaned
the number of cases to approximately 198. set of 30 performance indicator ratio variables. After all
Fuel related variables (RVMIFUEL, TVMIFUEL) were the data cleaning and verifying strategies, after all the
eliminated because with four different types of fuel listed exploratory passes through the data and after all the
78 G. J. FIELDINGet al.

Table 2. Forty-eight performance indicator variables used in The patterns of factor loadings were so similar be-
analyses tween the reported data, weighted data and the two sets
COSTEFFICIEKY Kk3JRE.s of logs that it appeared very convincing that the under-
lying structure in the data set had, indeed, been found.
Seven factors, accounted for approximately 83% of the
variance. Table 3 shows the pattern of factor loadings
for the final weighted data set. Factors one, two and
three represent output per dollar cost, utilization of ser-
vice and revenue generation per expense, respectively.
These first three factors directly relate to the three major
categories of the performance concept model-cost ef-
SERVICEEFFICIENCYIEASlREs ficiency, service effectiveness and cost-effectiveness.
Factors four, five and six represent labor efficiency,
vehicle efficiency and maintenance efficiency, respec-
tively; each a refinement of cost-efficiency. Factor seven
is clearly related to safety.

VERIFYING THE FINAL 1980 FACTOR ANALYSIS

COSTEFFECTIVENESS
MSLRES The adequacy and strength of the final solution were
determined by Thurstones five criteria for detecting sim-
ple structure solutions in factor analysis results (Harmon,
1967). The rotated factor loading structure was compared
against Thurstones criteria for evaluating structure for
its simpleness and met each of the qualifying con-
Wariable anitted prior to final analysis ditions. This was convincing evidence that a clear, un-
derlying structure in the data had been found.
considerations for data quality, these 30 variables were In interpreting and portraying the factor loading pat-
chosen to represent the best possible information on per- tern, i.e. in Table 3, an arbitrary cutoff of .5 was used
formance currently available in the Section 15 data base. to determine the representative variables for each factor.
Principal component factor analysis with varimax or- It was felt that a high cut off value would make for easier
thogonal rotation was carried out on the four different and clearer interpretation of the factors.
sets of 30 performance indicator variables. Two different The next point investigated was how much of the
computer routines were used-SPSS-PA1 and BMDP- variance of the final factor solution was not being ac-
P4M. The latter was used to compare as closely as pos- counted for by those identified high-loading varia-
sible the current analyses with the previous work. bles. The data were tested and it was found that for each

Table 3. Rotated orthogonal factor loading matrix

FACTOR 1 FACTOR 2 FACTOR 3 FACTOR 4


OmpT PER DOLLAR COST UTILIZATION OF SERVICE REVENUE GENERATICNI USOREFFICIENCY
EXPENSE

TvMiOEXP .90 PAS/OEXP .m REWOSUS .s2 lVH/EMP .91*


RVMlTWG .67 PAMWAG .66 OREWOEXF -91. RVHiOEMP .66
RVH/OEXP .67 TPASlRVH .66 OREWRV .M TVH/PVE 53
RVH/OWAG .63 TPASIPVH .a4 REWPVEH .n TVM/EMP .Sl
TVMIMEXP .,I TPASlRVM .83 REWTPAS .70
RVH/OSUS .61 PASOSUS .67 RVHlOSUB .66
RVWVMWG .s6 PAS/OSUS 61

Percnlo,Vtianc8Explalrad: Percent of Vananca E@aind: P,,ro,VananceEx~d: Percemofvarianw Elqlsined:


26.7 16.6 12.6 9.1

FACTOR 5 FACTOR 6 FACTOR?


VEHCLE EFFICIENCY MAlNTENANCEEFFlClENCY SAFEI-Y

TVM/PVEH .92 lvM/MNT .w* TVMIACC .93


PVEH/OP -.77 PVEH/MNT .9Q RVHIACC .63
NIPVEH .77

penen, of Varianca Ex$4med: PercantdVdEd: pacwndvwianceEap4ained:


7.2 6.6 6.6

TOTAL AMOUNT OF VARlANCE EXFIAINED: 63%


Performance evaluation for bus transit 79
factor, approximately 95% of the information was still A final set of factor analyses was carried out on the
being represented while overall, 86% of the variance of four sets of performance indicators where missing values
the original factor structure was represented in the subset had been replaced with estimates. The number of cases
of high-loading variables. then being analyzed increased from 194 to 280.
It was plausible that an increase in the number of cases
Reliability being analyzed could shift a weak or unstable factor
A third question regarding the set of high-loading var- solution to a different factor structure. The final set of
iables defining the factor structure centered on the re- factor-analytic solutions carried out from the data sets,
liability-in a statistical sense-of the grouped varia- which included estimated values, were entirely consist-
bles. Cronbachs Alpha was calculated for each group ent with the earlier results.
of variables gathered together on a particular factor. Thus, after rigorous testing of the final 1980 factor
Cronbachs Alpha can be used to evaluate the internal analysis. It was found that: (1) the same general under-
consistency of a group of variables to see if they essen- lying structure had consistently appeared across all
tially target the same underlying information (Carmines checking routines; (2) not only the same factors ap-
& Zeller, 1979). Alpha values range from zero to one peared, but they also appeared in the same order and (3)
with a value equal to one representing perfect reliability, with minor fluctuations, the factor loading patterns were
or internal consistency in this case. An alpha value of generally the same. Therefore, it was concluded that
.8 is considered very reliable. a stable, consistent and reliable simple structure had
A standardized item alpha was calculated for each been detected out of the larger group of performance
group of high-loading variables on each factor, across indicators.
each of the four sets of slightly different data. The alpha
values hovered around the .8 criterion on the weighted
SELECTING REPRESENTATIVE MARKER VARIABLES
data set, and were all well above .8 on the log set of
the weighted data. This was true on all Factors except The objective for this research was the establishment
Factor 5 which produced an uninterpretable alpha value. of a small, unique subset of performance indicators that
Factor 5 measures the positive and negative poles of the is particularly useful for performance evaluation by sys-
vehicle efficiency concept as shown in the negative and tems of all sizes. The goal was to identify the minimum
positive factor loadings. Thus, it confounds the calcu- amount of data necessary to convey the maximum amount
lation of a standardized item alpha. of performance information. To accomplish this, the fac-
tor loading data in the rotated factor-structure solutions
Factor structure stability on the final variable sets were used. High factor loadings
Two final questions were raised regarding the 1980 represent a high correlation of a particular variable with
final factor analysis. They both focused on a single con- a particular factor. When a variable has a high factor
cem-how globally relevant was the final factor struc- loading on only one factor, it can be said to represent
ture? Would the underlying structure of the data remain that factor both statistically and conceptually.
stable over different theoretical assumptions or an in- To select a small subset of easily accessible perform-
crease in data cases? ance indicators from the final factor structure, five cri-
Classical inferential factor analyses were carried out teria were used: (1) Representativeness of a variable vis-
on the four performance indicator variable sets. This type a-vis a factor was reflected in a high factor loading on
of analysis assumes that the data comes from a sample only one factor; (2) the distribution of values in the
of cases from a larger population. All solutions and re- variable had to be as close to normal-like as possible;
ported statistics are mathematically adjusted to predict (3) ease of collection of the variable was assessed via
values as they would exist in a larger population. Thus, the percentage of data missing; (4) the variable had to
it is conceivable that, if a factor structure is somewhat have been well captured by the factor structure in general
weakly defined, a different structure could emerge from (high communality) and (5) the variable selected had to
an inferential solution than from a principal components be easily understood by transit managers.
analysis. However, results from both the inferential and Seven representative or marker variables were se-
principal components analyses were consistent across the lected from the final factor structure-one variable rep-
four data sets. resenting each factor. Seven alternate markers were
To test whether the final structure in the analyses would also identified. These alternates can be used equally well
remain stable over an increase in data cases, an esti- for assessing performance: it is the performance dimen-
mation procedure for missing data was used. The BMDP sion that is important rather than the marker variable.
statistical computing package includes a program whereby Table 4 lists the selected marker variables and their
missing data values can be estimated. Multiple regres- alternates.
sion on the variables with data is used to predict a most The markers and the alternate set of markers are highly
likely estimate for any case missing data on some subset reliable (alpha range is from .802 to .937). Thus, with
of the variables in the analysis. When no prediction can a maximum of seven variables from a much larger data
be made from other available data, the mean of the set, performance of a transit system can be assessed. To
variable of interest is used to replace the missing value. assess the three major categories represented in the con-
When any case is missing too much of its data, it is not ceptual model, the first three marker variables would
used in the estimation procedure. be sufficient. Further, any one of the seven factor con-
80 G. J. FIELDINGet al.

Table 4. Performance indicators best representing the underlying performance concept

1 output per $ cost RVlVCfZXP TVWCEXP


2 Utilization of TPM/RVH TPAWRVR
service

3 Revenue Generation (*(EV/ou(P REV/ON3


per Expense
4 Labor Efficiency TMEW RVwoEw
5 Vehicle Efficiency TVIwJVEH wwVEH
6 Mintenance Efficiency mvm1 PVEIWWT
7 safety TVWACC RVHIAaz

cepts could be assessed via the relevant marker most revealing types of analysis comes about when per-
variable. formance achievements for an agency are compared not
only against those of previous years, but also against
those achieved by other agencies which are regarded as
KEY INDICATORS FOR PERFORMANCE EVALUATIONS
its peers. Through performance comparisons within a
The seven performance indicators identified in the peer group, the achievement of any one system can be
previous section represent the key underlying dimensions demonstrated in sharp relief (Fielding et al., 1985).
of performance as found in the Section 15 data base. Figure 2 demonstrates the effectiveness of conducting
These performance variables can be validly used for single system performance comparisons to a peer group.
cross-sectional comparisons of transit agencies. Jacksonville Transit Authority (JTA) was compared against
The first three indicators-Revenue vehicle hours per its peers across the seven performance indicators iden-
operating expense (RVH/OEXP), unlinked passenger trips tified in this research for fiscal years 1980, 1981 and
per revenue vehicle hour (TPAS/RVH) and operating 1982. The graphic display of Jacksonvilles performance
revenue per operating expense (OREV/OEXP)-are the information reveals that JTA is above its peer group
most informative global measures available in the data average on nearly all performance indicators over the
base. They represent the first three factors which account three years of data.
for most of the variance in the underlying data. Further, For each year, safety (TVMIACC) is the strongest
they each relate to one of the performance model con- area and maintenance efficiency (TVMIMNT) is the
cepts noted in Fig. l-cost efficiency, service effec- weakest area of system performance. Labor efficiency
tiveness and cost effectiveness. The four additional in- (TVHIEMP) is consistent and appears to be improving,
dicators can be used to evaluate more specific areas of while vehicle efficiency (TVM/PVEH) shows a sharp
cost and service which have important impacts on overall decline between 1980 and 1981. The drop in utilization
performance. of service (TPAS/RVH) over the three years is somewhat
The set of seven alternate marker performance in- reflected in the dip in revenue generation per expense
dicators can be thought of as the second best list for (OREVIOEXP) but seems to have had little effect on
conducting performance evaluations. They also represent JTAs output per dollar cost (RVHIOEXP). The low
the underlying dimensions of performance. These have performance on maintenance efficiency and the decline
been identified to assist transit analysts in selecting equally over time in vehicle efficiency are performance trends
valid cross-sectional measures when local or state re- Nhich may require managerial intervention.
quirements request reporting of particular statistics. For Using Section 15 data and the performance indicators
example, vehicle miles rather than vehicle hours are identified and tested above, transit managers can monitor
preferred by some governmental agencies as a measure system performance. Unusual performance vis-a-vis other
of service produced. The dimensions identified are more members of the peer group, or evidence of trends within
important than the indicators chosen to represent the one system over time, can alert the manager to the need
structure. for corrective action. Thus, Section 15 data and the set
of performance indicators listed here can provide the
The evaluation process basis for insightful managerial control over the transit
One of the most important functions of any perform- operations. The indicators provide vital signs for the
ance evaluation is to identify the major strengths and major dimensions of transit performance: they may not
weaknesses of the object being evaluated. In transit, this indicate the cause of change, but they should alert man-
involves measuring the amount of output produced by agement to the need for analysis.
an agency per unit of input and whether the output is Research that has led to the development of the per-
consumed. Evaluations can be static snapshots of the formance evaluation methodology has been based on the
agency at one point in time, or they can be trend analyses Section 15 data tape available through Transportation
of the same system over months or years. One of the Systems Center. However, for individual agency per-
Performance evaluation for bus transit

Standard Scorns

1981

1982
2

fotmance evaluations or for peer group comparison, the and in the final results. Rigorous cleaning. verifying and
Section 15 Annual Report would be more useful. The grooming procedures carried out before analysis insured
Annual Report is a compilation of financial and operating that the input data was as complete as possible. Careful
statistics which have been reported under Section 15 decisions regarding which variables to keep andor drop
requirements. Data is both aggregated by size and mode from the analysis provided the best possible set of per-
categories, and reported for individual transit systems. formance indicators available in the Section 15 data. The
To facilitate single agency performance evaluations, an use of four parallel data sets and several exploratory
instructional manual has been prepared and is available factor analyses detected the simple underlying structure
through the National Technical Information Service of of the data. Finally, the rigorous testing and validation
Springfield, Virginia (de la Rocha et al., 1984). The of that underlying factor structure was convincing that
performance instructional manual is a self-contained sys- the most salient performance indicator concepts had been
tem for performance evaluation based on use of the Sec- identified. The consistent and stable structure in the data
tion 15 Annual Report and the seven key performance enabled the selection of seven marker variables for
indicators identified in this article. evaluation. These too measured up to testing and veri-
fying procedures. Given the quality and content of the
Section 15 data, it is felt that the most salient features
CONCLUSION
for performance evaluation have been determined, and
The FY 1980 Section 1.5data has been used to identify that the marker variables can be used with confidence
and test the most easily accessible and parsimonious set by transit managers and governmental officials.
of performance indicators for fixed-route transit. The
research had two objectives: (1) to find the minimum Acknow/edgemenr-Research for this paper has been supported
amount of data necessary to provide solid and stable by the Urban Mass Transportation Administration. U.S. De-
performance evaluation capability, and (2) to test the partment of Transportation, Research Contract No. CA-I I-0026.
The U.S. Government assumes no liability for the contents or
validity of the methodology developed from the previous
use thereof.
analysis of FY 1979 data.
The use of factor analysis on a large set of performance
indicator ratios gleaned from the data the structure of REFERENCES
the underlying performance concepts. From the factor
Anderson S. C. and Fielding G. J. (1982) Comparative Analysis
structure, a subset of seven variables was identified and of Transit Performance. University of California. h-vine. In-
tested against the larger data structure. These seven var- stitute of Transportation Studies. Final Report, UMTA-CA-
iables are the most salient performance indicators cur- 11-0020-1.
Carmines E. G. and Zeller R. A. (1979) Reliobilit) and Validity
rently available in the Section 15 data base. They can
Assessmen?. Sage Publications, Beverly Hills.
be used together or individually to assess fixed-route Comrey A. L. (1973) A First Course in Factor Andyis. Ac-
transit performance. ademic Press, New York.
There is a great deal of confidence in the data used de la Rocha 0.. Brenner M. E. and Fielding G. J. (1984) The
82 Cl. J. FIELDING ef al.
Irvine Evaluation Method: An Instructional Manual for Use Miller J. H. (1980) The use of performance-based methodol-
in Conducting Internal Evaluations of Transit Performance. ogies for the allocation of transit operating funds. Transpn
University of California, Irvine. Institute of Transportation Quarterly 34, 574.
Studies. Final Report, UMTA-CA-11-0027-l. Nash C. A. (1980) Management objectives, fares and service
Fielding G. J., Glauthier R. E. and Lave C. A. (1978) Per- levels in bus transport. /. Transpn Econ. Policy 12, 70-85.
formance indicators for transit management. Transporration Patton T. A. (1983) Transit Performance Indicators. U.S. De-
7, 365-379. partment of Transportation, Transportation Systems Center,
Fielding G. I., Brenner M. E., de la Rocha 0.. Babitsky T. Cambridge. Staff Study (SS-67-0.3-01).
T. and Faust K. (1984) Using Section I5 Data for Transit Talley W. K. and Becker A. J. (1982) A single measure for
Performance Analysis. University of California, Irvine. In- evaluating public transit systems. Transpn Quarterly 36, 423-
stitute of Transportation Studies, FiiaI Report, UMTA-CA- 431.
I I-0026-2. U.S. Department of Transportation, Urban Mass Transportation
Fielding G. J., Brenner M. E. and Faust K. (1985) Typology Administration (1982) National Urban Mass Transportation
for bus transit. Trans Res. (forthcoming). Statistics: Second Annual Report, Section 15 Reporting Sys-
Harmon H. H. ( 1967) Properries of Different Types of Factor tem. Report No. UMTA-MA-O6-0107-82-1. Washington, D.C.
Solutions: Modern Factor Analysis. University of Chicago Vaziri M. and Deacon J. A. (1983) Application of Section 15
Press, Chicago. and Census Data to Transportation Decision Making. Uni-
Kneafsey J. T. (1975) Transport Economic Analysis. Lexington versity of Kentucky, Lexington. Dept of Civil Engineering,
Books, Lexington, MA. Final Report. UMTA-KY- 1 I-0002-83.

You might also like