-
I FEB
ALFRED
P.
WORKING PAPER
SLOAN SCHOOL OF MANAGEMENT
FINANCIAl INFLUFNCES
Nancy
L.
ON AIRLINE SAFETY
Rose
Assistant Professor of Applied Economics
May
Sloan School of
1987
Management Working Paper #1890-87
MASSACHUSETTS
INSTITUTE OF TECHNOLOGY
50 MEMORIAL DRIVE
CAMBRIDGE, MASSACHUSETTS 02139
•if
81988
j
FINANCIAI INFLUENCES
Nancy
L.
ON AIRLINE SAFETY
Rose
Assistant Professor of Applied Economics
May
Sloan School of
1987
Management Working Paper #1890-87
Prepared for the Northwestern University Transportation Center
Conference on Deregulation and Transportation Safety
June 23-25, 1987
This research was supported by a grant from the United Parcel Service
to the MIT Center for Transportation Studies. 1 am grateful to
Severin Borenstcin and James Poterba for helpful discussions and comments.
Vera Young collected much of the data used in this paper, and An-.Ien Tai and
Douglas Staiger provided excellent research assistance.
Foundation
^
FEBO
81988
f-'-;cciVEO
FINANCIAL INFLUENCES ON AIRLINE SAFETY
Nancy
Rose
L.
Assistant Professor of Applied Economics
Sloan School of Management, MIT, Cambridge, MA, 02139
May
Substantial attention
a
hn'^
1987
been focused on possible declines
safety as
in airline
consequence of economic deregulation during recent years. There has, however,
been relatively
little
analysis of the effect of firms' financial conditions on their
This paper uses two kinds of data to evaluate claims that
safety performance.
deregulation has reduced nir safety.
airline accident rates
after deregulation,
and
First,
profitability.
important variations across
and accident
carriers.
constructed, and differences
A
statistical
1
in
983.
series
These data suggest no decline
in
data on
in safety
accident rates
Aggregate data may, however, mask
To
investigate the relationship of financial
rates at the firm level,
over the period 19.SS through
analyzed.
examine aggregate time
and provide weak evidence of an improvement
relative to the trend undet regulation.
characteristics
I
I
use data on 31 air carriers
Accident rates for individual carriers arc
accident rates for different groups of carriers arc
model of the determinants of individual accident rates
is
developed and estimated. This model relates profitability and other financial
variables to accident rales, controlling for operating characteristics that also
affect accident probabilities.
profitability effects
on safety.
The
results provide
some evidence of
possible
may
1.
INTRODUCTION
Airline safety has attracted tremendous attention in recent months.
number
A
record
of international airline passenger fatalities in 1985, as well as an increase
in
the domestic airline accident rate, focused attention on possible declines in airline
safety as a consequence of economic deregulation.
in
These
issues
newspaper columns, network news broadcasts, business
other forums; sec, for example,
Main
periodicals,
and a host of
Feaver (1986), Gattuso (1986), Kahn and Kasper
(1986),
Magnuson
(1986).
Thayer's (1986) argument against deregulation
(1987).
have been debated
(198.5),
Nance
(1986), Nolan (1986),
is
and Thayer
typical:
...competition unleashed by the economic deregulation of any
industry forces many if not all companies to cheat and cut
corners in ways that are dangerous to their customers, society
at large and even themselves. This pattern, visible in any
deregulated industry, is becoming increasingly obvious in the
trucking and airline industries.
Despite substantial media attention, however, there has been relatively
scientific research
little
on possible links between deregulation and safety conditions
the airline industry.
financial conditions
Graham and Bowes
and
in
(1979) investigate the link between firms
their accident rates,
maintenance expenditures, and service
complaints; Golbe (1986) analyzes the relation between accidents and profitability.
Both studies use pre-deregulation data; neither finds much support for the argument
that reduced profitability lowers airlines' safety performance.
Advanced Technology
(1986) analyzes bivariate correlations between financial measures and carriers'
inspection ratings in the Federal Aviation Administration's
(FAA) 1984 National
Air
Transportation Inspection program.
and inspection
financial indicators
shortcomings, particularly
in
that
While the study finds some relation between
failures,
it
its
methodology has substantial
control for other factors that
fails to
WcKenzie and Shugart (1986) analyze
affect inspection ratings.
may
the pattern of
aggregate air fatalities and passcngcr-miles over the 1973 through 1984 period, and
conclude that
fatalities
have ncn been adversely affected by indirect effects of
deregulation on traffic or direct effects of deregulation on fatality rates.
Fromm
(Sec
(1968) for an historical perspective on airline safety, and Oster and Zorn
(1984) for an analysis of
In this study,
I
commuter
air carrier safety.)
investigate the relationship between accident rates
and
financial performance, controlling for operating characteristics of carriers that
The
affect their accident probabilities.
section 2,
I
structure of the study
is
as follows.
may
In
discuss economic models of firms' quality or safety choices, and assess
their implications for the airline accident-financial
performance relationship.
Section 3 analyzes the aggregate accident data for
all
U.S. certificated route
scheduled carriers over the 1955 through 1986 period. The use of a longer time
more
period than earlier studies permits
precise estimates of both the accident
trend rate under regulation and the effect of deregulation on the lc\cl or direction
of that trend.
profitability
These data
also arc used to estimate the effect of industry-wide
measures on accident
profitability effects
rates; the
data provide no evidence on aggregate
on safety. Section 4 explores the determinants of accident
rates for indi\idual carriers.
Data arc collected
through 1983 period, covering
a
broader range of carriers over
period than the data sets employed
carriers are constructed,
carriers are analyzed.
for 31 carriers o\er the
in earlier
and differences
in
Finally, a statistical
studies.
a
much
1955
longer time
Accident rates for individual
accident rates for different group*^ of
model of the determinants of individual
accident rates
is
The
estimated.
profitability effects
results provide
on safety. Section
.'>
weak evidence of
possible
contains a brief conclusion.
MODELS OF THE PiETERMINANTS OF AIRLINE SAFETY DECISIONS
2.
The
grounds.
A number
levels
is
ambiguous on
theoretical
of private incentives encourage safety independent of direct
government intervention.
liability
on safety
effect of deregulation
First,
insurance companies base the premium rates for
insurance on an assessment of
risk.
These companies have a strong
incentive to monitor safety performance of carriers, and to increase
may
increa se
taken into account.
premiums
not decrease, total costs once insurance
,
Second, firms have an important stake
tion for providing safe service in order to attract
develop a reputation for being
less safe will,
Some
passengers to safer competitors.
for
This effect suggests that reducing safety
firms that take on additional risk.
expenditures
premiums
and
in
are
maintaining a reputa-
retain business.
Airlines that
other things equal, be likely to lose
empirical research has analyzed the
strength of these reputation incentives; see Barnett and Lofaso (1983) on passenger
responses to the
DC- 10
crashes.
on aircraft manufacturers'
Borenstein and Martin
conference.
when
Chalk (1985, 1986) on the
profitability,
Zimmerman on
effects of fatal accidents
and the work being done by Severin
the profit effects of accidents for this
Third, employees have strong incentives to monitor safety, particularly
linked to maintenance of equipment or operating procedures.
Pilots
and
fiight
attendants, because of their extensive exposure to an airline's flights, are at highest
risk
from any decline
safety, or at
exposure.
in air safety.
minimum,
to require
(The strength of
employment opportunities
These groups are
likely to resist reductions in
compensating wage differentials
this effect
depends
critically
upon the
for these employees, for if their
for higher risk
alternative
wages are above
their
opportunity wage, reduced safety
may
reduce their economic rents rather than
leading to higher wages).
These private
In particular,
or
its
own
do not guarantee the optimal
may
asymmetric information about safety
socially desirable.
about
incentive*;
Asymmetric information
is
present
lead to less safety than
when
a firm has
of safety that insurance firms, employees, and customers cannot
lc\cl
less safety
than customers or employees desire.
firms.
effects of
In these
however; firms
models, firms do not provide unambiguously lower safety,
various parameters.
In
to
overprovide safety, depending on the values of
Shapiro's (1982) model, for example, a firm
higher quality (safety) to improve
profitability
its
One
of deregulation assume.
reduce safety proxisjon
b\-
Yet
is
limited liability laws, shareholders of
amount of
this
potential
the firm
may produce
reputation.
none of these models establishes
and safety choices.
limited to the
numerous others
asymmetric information on product quality or safety choices
may choose
In addition,
to
Models by Akerlof (1970),
Klein and Lcfficr (l^SI). Shapiro (I9R2. 1983). Allen (I98.S) and
examine the
is
knowledge
do not obtain. Linder these circumstances, the firm may have an incentive
provide
by
of safety, however.
level
is
links
between the
bankrupt firm have
their equity holdings.
critics
financial conditions
their
may
Because of
the possibility of bankruptcy.
n
many
precisely the relation that
mechanism by which
level of
"downside"
This creates an asymmetry
risk
in
shareholders' expected returns, with unlimited positi\e returns, but limited negative
returns.
Firms that arc near insolvency might choose,
maintenance expenses and gamble on an increase
bankruptcy.
example,
accidents
in
Even
in this
its
effect
to
reduce
an effort to avoid
Bulow and Shoven (1978) and Golbe (1981) describe
bankruptcy decision and
behavior.
in
for
a
model of the
on stockholders' preferences for the firm's
risk
type of model, however, the attractiveness of increa.sed risk
through reductions
in
safety expenditures
is
ambiguous.
bankruptcy costs are
If
For
high, firms will try to avoid actions that raise the risk of bankruptcy.
airlines,
the tendency for failing carriers to exit the industry by mergers or acquisitions
rather than liquidation of assets
safety:
if
an airline increases
its
its
value to potential acquirers
its
goodwill or reputation).
by limited
distress
liability.
may
provide another argument against reducing
accident rate by reducing
likely to
its
safety expenditures,
be lower (the accidents
will
have eroded
This could tend to counteract the incentives provided
While more complex models of
carriers' responses to financial
provide some predictions for the direction of financial infiuences on
safety decisions, the issue
empirical investigation.
3.
is
may
is
one that seems
It is
likely to
be resolved only through
to this type of analysis that
1
now
turn.
AGGREGATE DATA ON A IRLINE SAFETY
An
analysis of the potential effect of deregulation on safety logically begins
with a comparison of Ihc safety performance of the industry before and after
deregulation.
If
deregulation materially affected airlines' safety decisions,
observe a change
in
the level of aggregate safety after 1978.
we should
In this section,
I
explore the behavior of industry-wide accident measures over time, and investigate
the relationship between profitability and safety at the industry level.
rationale for using accident rates as a
3.1; trends in
measure of safety
is
discussed
in
The
subsection
aggregate accident data over the 19.S5-1986 period are explored
subsection 3.2; and the relationship between aggregate accidents and industry
profitability
is
investigated in subsection 3.3.
in
3.1
Accident Rates as a Measure of Safe ty
While wc canncM (ibser\c safety
directly, a
are available, including accidents, "incidents,"
and
levels of safety inputs
number of
FAA
inspection results and citations,
Each has particular advantages
such as maintenance.
and disadvantages, and depending on the questions of
these measures might be incorporated
tradition
by focusing on accidents
a description of
safety;
in
in a statistical
observations.
analysis.
I
follow a long
the analysis below (see the data appendix for
because they arc such low frequency events,
in
more than one of
interest,
may
what constitutes an accident). Accidents
distinguish changes
proxies for safety levels
it
may
be a noisy proxy for
be difficult to
the safct\ distribution without a long time scries of
Howc\cr.
this
measure has
a
number of advantages
for the present
analysis.
First, the
numhci of
air carrier accidents (fatal
outcomes which seems of most concern
,
analysts:
what
fatal accident?
is
to passengers,
and nonfatal)
and possibly
reflect safely
to policy
the probability that this flight will be involved in a fatal or non-
FAA
inspections and citations arc
more
as are maintenance expenditures, training programs,
and the
transformation of safety inputs into safety performance
focusing on performance
may
provide
n
more
reflective of safety inputs ,
reliable
is
like.
itself
Because the
uncertain and noisy,
measure of the actual
level
of
safety.
Second, accident reporting and detection
more
is
quite accurate, particularly for
serious accidents and larger airlines and relative to reporting of incidents
detection of safety violations through
FAA
inspections.
Incident reporting
more
subjecti\c.
it
is
more
difficult to detect
non-reporting, and reporting
sensitive to "campaigns' that highlight the issue
in
less
is
consistent than accident reporting; judgment of what constitutes an incident
and
may
may
be
be
an attempt to improve reporting
rates.
Similarly,
violations,
all
FAA
and the
inspection results and safety citations are unlikely to detect
level
FAA's enforcement
the
of citations
activity,
which
may depend
is
critically
upon the
signal
by the
airlines, or
they
may
weaker enforcement, not necessarily better
may
may
reflect
FAA
by the
refiect a decision
"seriousness" about enforcing safety violations (they also
its
For
unlikely to be constant over time.
example, the FAA's recent record-setting fines of major trunk carriers
lax safety procedures
intensity of
to
reflect
Reported
safety, in earlier periods).
accidents are likely to be more consistently measured through time.
Third, because this study focuses on measuring the effects of economic
deregulation on safety, the selected safety measure should abstract from changes
from the Professional Air Traffic Controllers Organization
resulting
in
(PATCO)
strike
1981 and the subsequent dismissal of the participating air traffic controllers.
"Incidents," such as near misses, are likely to include a higher proportion of events
that
the
may
be partially or wholly attributable to air
number of
accidents
seem somewhat
may
traffic controller errors.
While
be affected by air traffic control conditions, accidents
than an incidents index.
less sensitive
Finally, aggregate accident rates will yield fairly precise estimates of accident
probabilities, particularly given the large
major
number of
certificated air carriers ("part 121" carriers,
taxi operators) currently operate well over
aircraft miles, per year.
in
to identify the probability of
accident
U.S.
air
million flights, logging 3 billion
air carriers,
combined
the range of 15 to 25 during recent years, allows us
an accident with a great deal of precision.
feature of the aggregate statistics
is
each year.
which excludes commuter and
This substantial exposure of U.S.
with annual total accidents
carrier
.S
flights in
is
frequently overlooked.
When
This
an individual
small relative to the frequency of accidents, one more or one fewer
may
substantially change
its
accident rate (although this will average out
across carriers).
This scnsiti\it\ often
i'>
argument
cited as an
Industry-wide total accident rates
accident statistics.
will
for disregarding
have
a
much
smaller
variance, however, and therefore will provide a reasonably powerful test of changes
through time.
(Note that
this
argument
out of every 5 or
account for only
I
accidents will be
less precisely
be
will
more
which
less true for fatal accidents,
total accidents; the rate for fatal
estimated than
be the corresponding rate for
will
total accidents).
Given these considerations, the analysis below uses accidents and accident
rates
per thousand dcpnrfurcs to measure airline safety.
number of
I
measure accidents by the
accidents rather than passenger-based measures of accident incidence
(such as passenger fatalities).
we have data on
the
This choice
is
number of passengers
dictated by two considerations.
affected by accidents only for fatal
Using passenger-based measures would
accidents.
one-quarter of total accidents, throwing out
safety performance.
First,
much
Second, one accident that
limit the analysis to less
than
of the available information on
kills all
200 people on board
may
have different implications about the safety of the system than 20 accidents with
the
same
total fatalities.
The
number of
decision to focus on the
implicitly judges the second scenario to reflect lower safety levels,
3.2
Aggregate Acciden
Table
3.1
t
accidents
all else
equal.
Trends
number of
presents information on the total
accidents, aircraft miles,
revenue departures, and accident rates per 100,000 departures for U.S. certificated
air carriers'
Both
fatal
scheduled passenger and cargo operations, from I9.S5 through 1986.
and nonfatal accidents are included
accidents arc also tabulated separately
in
in total
column
2.
both fatal and nonfatal accidents over time, which
is
accidents (column
There
is
a
I);
fatal
steady decline
sustained after economic
in
TABIF
ACXinrNI RATES,
I'.S.
3
1
CERTIFICATED AIR CARRIERS,
SCHEDUIED PASSENGER AND CARGO OPERATIONS
(14CER
121
OPERATIONS)
10
deregulation of the industry
the
number
^
and
improvement
aircraft
declines even though
of flights and aircraft miles increase sharply through time, implying
even larger declines
columns
The number of accidents
1978.
in
(^
in
in
(and
safc>
the accident rates per 100,000 departures, as reported in
accident rates per million miles, not shown).
in
attributable primarily to the substantial improvements
is
and aviation technology over the
of radar technology: development of the
and operating
This
last
jet
40 years. These include the diffusion
aircraft (with
enchanced power, range,
and materials advances; the introduction and
altitude): metallurgical
continued improvement of navigational and landing aids (such as automatic
electronic glide slopes,
simulators for
pilot training,
and the
like.
rates continue to fall through the deregulated period,
possible that deregulation has caused the rate of decline to deviate from
To
estimate the effect of deregulation on accident trends,
accidents as following
pilot,
ground proximity warning systems); more sophisticated
Although accident
term trend.
in
I
its
it
is
long-
model
logarithmic decline over time and allow for the possibility
a
that the accident trend shifts
up or down with deregulation. This suggests an
equation of the form:
(3.1)
where
TIME
In
(ACCIDHNT RATE) =
is
a linear
the years 1978-86.
time trend and
otherwise.
of the fatal accident rate
is
accidents
is
a
DEREG
is
*
a
TIME
dummy
+
B2
*
DEREG
variable equal to
I
for
1980. fatal accidents are zero, and the logarithm
undefined.
equal to zero and introducing
number of
In
Bq + B|
dummy
I
treat this
\ariable.
by setting
ln(fatal accidents)
ZERODUM,
equal to one when the
zero and equal to zero otherwise (see Pakes and Griliches
11
Hausman,
(1980) and
Hall,
and Griliches (1984)
for a similar econometric treatment
model of patents).
in a
I
estimate equation (3.1) by ordinary least squares regression, using three
different
measures of accident
(TOTACC),
number
the
reported
the
in table 3.2.
total
The
and
(1.4) percent per year
number
total accidents per
of fatal accidents
number of
the
number of
rates: the
in a
decline
given year
in
of total accidents
accidents through time
fatal accidents declining
in
quite strong, with
1
and
The
accident rates over time.
error, .6) to 5.0
total accident rate per
The equations
6.4 (.6) percent per year.
the two measures based on total accidents (columns
percent of the variance
results are
by 4.4 (standard
throughout the period, and the
hundred thousand departures declining
is
given year
(TACCDEP), and
100,000 departures
(FATACC). These
in a
2) explain
for
roughly 90
third equation, for fatal
accidents, explains 60 percent of the variation in fatal accidents over the period.
Moreover,
although
in
-
three equations, accident levels are below trend after deregulation,
this effect
column
exp(B3)
in all
1
1).
is
statistically significant
only for the total accident equation
(with total accidents roughly 30 percent below trend, as measured
This
may
suggest that improvements in safety technology or aircraft
operation (such as cockpit
management techniques adopted by many
these and subsequent equations are qualitatively the
is
used
in
is
if
its
alternative approach to evaluating pre-
B
for a discussion
implications for
OLS
regressions).
and post-deregulation performance
to estimate the accident equations over the regulated period only,
how
results for
the square root of
place of the log of accidents (see appendix
of the probability distribution of accidents and
An
same
airlines in the
The
1980s) have advanced safety even faster than the long-term trend.
accidents
by
and examine
well the regulation experience predicts accidents over the deregulated period.
Figures 3.1 through 3.3 compare predicted accident rates based on this approach
12
lABI r
RI C^RISSinN
A\M
>SIS n]
.3.2
AfiGRFGATr ACCinrNI RATF*; OVI R
19.';V1')H6
inftotal
Dependent Variable:
Constant (Bq)
TIMF
((^,)
Df-RFG
(B-,)
accidents
I
IMI
Actual V. Predicted Totai Accidents
Fig.
Fig.
3.1
Actual
V.
Predicted Fatal Accidents
Actual
V.
Predicted Total Accidents/
3.2
uc,ccc
Departures
n
u
m
Fig.
3.3
b
e
r
eo
65
70
75
85
13
with actual accident rates for the 1955 through 1985 period.
curves are generated from equations of the form of
and
3.2), the predicted levels
years, as expected from the negative
Although
results.
DEREG
fatal accident rates
in
is
to actual
deregulation
in the
in
and
the full time period
vary considerably from year to year, figure
3.3 suggests that the deregulation experience
the regulatory era; there
estimates
rate
total accidents
conform reasonably closely
Actual rates tend to be lower than predicted levels
levels.
DEREG
excluding
(1),
For rates based on
estimated over the 1955 through 1977 period.
(figures 3.1
The predicted
is
consistent with predictions based on
no evidence that actual
fatal accident rates are higher
the deregulated environment.
Aggregate Acciden t s and Industry Profitability
3.3
A
final possibility to
be explored with the aggregate data
between accidents and profitability
profits induce airlines to
at the industry level.
is
the relationship
The notion
that lower
shade on maintenance and other safety expenses
may
suggest that lower profit periods should be associated with higher accident rates.
To
test this hypothesis.
OPMARG,
is
a
calculate the industry's average operating margin,
defined as
OPMARG
(3.2)
This
I
=
I
-
(operating expenses)/(operating revenues))
measure of the industry's
profit
and tax payments. Because
OPMARG
capital leases-- equivalently,
it
may
margin, calculated before interest, lease,
treats returns to all capital-debt, equity,
be more comparable across carriers than are
return on equity or net income measures.
To
insure that
contaminated by the costs of current accidents, and to
OPMARG
refiect the
is
not
assumption that
and
pronts arc likely to influence nccideni rates with a lag
from the precccding period
(if at all),
I
OPMARG
use
the regression estimates (see section 4.3 for further
in
discussion).
Tabic
y.? presents the estimated coefficients for
TIME,
using the three measures of accident rates that were used
These equations arc estimated o\er the
variables.
series
DEREG
table 3.2 as
in
OPMARG,
and
dependent
i9.'>6-l984 period, as the data
used to construct operating margin were unavailable after 1983. The results
provide no support for the hypothesis that low profits reduce safety,
The
aggregate.
TIMF
coefficients are essentially
the accident lc\el equations, although
from zero
in
any of the equations.
it
DEREG
unchanged, and the
coefficients arc substantially similar to the earlier results.
for only
at least in the
DEREG
now
is
negative
not statistically distinguishable
is
OPMARG
In all three regressions.
has
a
counterintuitive positive sign, implying that lower profit margins correspond to
lower accident rates.
OPMARG
is
In all the equations,
however, the standard errors on
are substantial, and the point estimate of the operating margin coefficient
statistically indistinguishable
To
from zero.
test
whether
this
is
a
function of
pooling the regulated and deregulated samples into a single regression,
estimated equations that allow
OPMARG
to differ pre-
I
also
and post-deregulation.
Unfortunately, the small number of post- 978 obser\ations leads to enormous
1
standard errors on the deregulation coefficients, which make
either the hypothesis that
OPMARG
The
operating margins.
limit the
impossible to reject
has no effect on accidents
period or the hypothesis that the coefficients are the
deregulation.
it
in
same before and
results also are robust to using current, rather
While the imprecision of the estimated
power of these
tests,
there
is
no evidence
in
the deregulated
after
than lagged,
OPMARG
coefficients
the aggregate accident data
15
TABLE
3.3
RBGRFSSinN ANAI YSIS OF AnORFGATK ACCIDRNT RATHS OVER TIME:
THE INEIUENCE OE LAGGED PROEITABILITY
1956-1984
ln(total
Dependent Variahlc:
16
that safety hn^ dctcrinrntcd since deregulation, nor
is
any evidence of
there
negative effects of average profitability on the level of aggregate accidents.
3.4
Conclusions from the A<tf re aate Accident Analysis
This analysis of aggregate accident data for the major domestic
pro\idcs no support
the \icu that safety levels have deteriorated in the
foi
aftermath of deregulation, nor docs
profitability with aggregate safety.
McKenzie and Shugart's (I9R6)
on aggregate
it
indicate any correlation of aggregate
The
results in section 3.2 are consistent with
findings that deregulation had no discernible effect
airline fatalities, cither directly or
through
passenger miles, based on data from 1973-1984.
consistent with Oolbc's
on aggregate accident
air carriers
(
l98fS)
rates.
The
its
effects
on increased
air
results in section 3.3 arc
time series analysis of the influence of profitability
Using data from 1952-1972 and controlling
number of departures and general economic
conditions,
Golbe finds that
for the
profitability
tends to be associated with positive but statistically insignificant effects on accident
rates.
While the aggregate data indicate no adverse change
1978. they prcnidc
a
Intcrfirm differences
weak
in
test
in
accident rates after
of the hypothesis that deregulation reduced safety.
accident rates permit
of deregulation's effects on accidents.
The
more
precise
and more powerful
tests
next section analyzes firm-specific
accident data.
4.
ANALYSIS O F INPIVIDUAL AIR CARRIER ACCIDENT DATA
The argument
that safetx has declined since deregulation typically
is
an implicit link between the financial health or profitability of a carrier and
investment
in safety.
Advocates of
this
based on
its
view claim that more profitable carriers
17
choose to spend more on maintenance and safety-enhancing procedures, while low
profit carriers "cut corners"
on safety expenditures (Thayer (1986), Nance (1986)).
Deregulation, they argue, reduces overall profits, and therefore safety.
argument
is
correct, then safety should be linked to observable
profitability or financial health.
If this
measures of firms'
This section reports on a preliminary exploration
of the determinants of accident rates for individual air carriers, focusing on the
relationship of accident rates to firms' financial characteristics.
(This
work
is
presently being extended to apply the econometric models of count data developed
in
Hausman,
Hall,
and Griliches (1984) and Cameron and Trivedi (1986)
airline accident rates.)
If a
The
analysis
(described in further detail
model
profit-accident relationship exists, then the effect of
deregulation on safety can be calculated from the effect of deregulation on
characteristics.
to
in
is
restricted to
major scheduled
the data section below),
firm
air carriers
and covers the period 1955-
1983.
The
carriers
section
is
structured as follows: Section 4.1 describes the sample of air
and the data collected
for each firm.
Section 4.2 estimates the accident
probabilities for each carrier across five-year time periods,
and
tests the similarity
of accident rates through time and across different types of carriers (such as large
versus small, high profit versus low profit).
In section 4.3,
I
parameterize these
accident probabilities, allowing them to vary continuously over time, and modelling
them
as a function of firms' operating
and service experience. These
and route
results are
used to
financial condition on their safety records.
characteristics, financial condition,
test the effects of firms'
18
4.1
Individual Air Carr ier Data
The datn
set
used fn evaluate individual air carrier safety performance consists
of information on 31 scheduled air passenger carriers certificated by the
the period
19.s,s
through
1*»R3.
The
trunks (12). local service carriers
territorial carriers (3),
(6).
over
and international
intra-Alaskan and intra-Hawaiian airlines
(5),
and intrastate carriers that expanded into interstate service
after deregulation in ]^79 (S).
availability arc listed in
Table
A number
entire period.
carriers include domestic
CAB,
The sample
4.1.
carriers
Note that not
of carriers
and
all
is
data
carriers are observed over the
exit the industry,
acquisitions, .nnd datn for intrastate carriers
their dates of
primarily through mergers or
not reported by the C.AkB prior to
1979.
The sample omits
taxis, intrastate
(charters),
four major classes of air carriers:
scheduled passenger
and cargo
carriers.
The
airlines
and
air
nonschedulcd passenger carriers
air carriers,
first
commuter
three groups were excluded primarily
becau.se of the unavailability or non-comparability of published data on their
operations and nnancial conditions.
carriers also
is
Omission of commuter and nonschedulcd
air
supported on theoretical grounds: most of these carriers differ
substantially from major scheduled air carriers
terms of their services,
in
technology, and scale and scope of operations, suggesting that statistical inferences
The
based on pooled samples ma> he quite misleading.
carriers
is
A
deletion of all-cargo
dictated by our focus on air passenger safety.
variety of operating and financial information
publications for each of the 31 sample carriers.
The
consists of system-wide information on accidents
(DEPART),
revenue aircraft miles (MILES),
total operating
was
collected from
basic data for each carrier
(TOTACC),
total
CAB
revenue departures
maintenance expenses
expenses (OPEXP). total operating revenues
in
(MAINT),
(OPREV). A number
of
:
19
TABLE
4
1
SAMPI F AIR CARRIERS
Basic Data
Availability
Carrier
Trunk
COMPIJSTAT
Avaiilability
Carriers:
American Airlines
Braniff Airways
Continental Air
Delta Air
I
I
ines
1955-1983
ines
Eastern Air lines
1955-1983
National Airlines
Northeast Airlines
1955-1979
1955-1971
Northwest Airlines
1955-1983
Pan American Airlifics
Trans World Airiines
1955-1983
United Airlines
Western Air 1 ,incs
1955-1983
1955-1983
Local Service Carriers
Y
Y
Y
Y
Y
1955-1983
Y
Y
:
Frontier Airiincs
Ozark Air I ines
Piedmont Aviation
Southern Airways
Texas International
USAir/Allegheny
,
1955-1983
1955-1983
1955-1983
1955-1979
Y
Y
1955-1983
Y
Y
Alaska Airlines
1955-1983
Y
Aloha Airlines
Hawaiian Airlines
1958-1974
Northern C^onsolidatcd/Wien Consol.'^
1955-1983
1955-1967
Intra- Alaska
and
1955-1982
Intra- Ha waiian
1955-1983
Wien Air
Territorial Carriers
Intrastate Carriers
1955-1972
1955-1966
1955-1966
:
Air California
1979-1983
Air Florida
1979-1983
Pacific
Southwest Airlines
Republic
Southwest
^Data
Y
:
Caribbean Airiincs
Pacific Northern
Pan American-Grace
Former
1955-1983
1955-1981
1955-1983
reflects
1979-1983
Y
1979-1983
1979-1983
Y
Northern Consolidated operations prior to
Air in 1968, and combined operations thereafter.
its
merger with Wien
20
additional measures arc constructed from these variables, including:
international operation'-
(
DINTL,
equal to one
otherwise), operating margins
operations,
expenditures pet mile
(AVSTAGE,
(RWAINTPM,
cumulative aircraft miles flown hy the
experience,
LNEXPFR). and TIME
the carrier has international
if
(OPMARG),
maintenance
real
millions of 1982 dollars), average stage length
in
thousands of miles) and
in
exposure to
airline operating experience
airline,
(a linear
EXPER, and
time trend).
(measured
in
the natural log of
Several data series are
missing information or arc incompatible for part of the I9.'>5-I983 time period
end of the period);
(typically either the beginning or the
available sample size for any given application.
a
subsamplc of the
COMPUSTAT
column
on overall corporate
and income statements are available from
For these
for the I967-I98.S period.
carriers,
denoted by
3 of table 4.1. a richer set of financial indicators could
including:
return on equity
the share of equity
in total
(ROE),
refiect
the
common
capitalization
book, values), and the current ratio
These indicators
reduce the
Appendix A.
in
carriers, detailed information
capital structure, balance sheet,
may
Further details on the sources,
construction, and a\'ailnhilitv of the data provided
For
this
Y
(rURRENT.
in
be constructed,
equity market to book ratio
(EQRATIO.
\arious aspects of
a
(MTB),
based on market, rather than
current assets/ current
a firm's profitability,
liabilities).
solvency, and
liquidity.
The
accidents.
statistical analysis focuses exclusively
For
total accidents, rather
indi\ idual carriers, fatal accidents arc
during the early part of the sample period.
fatal accidents,
on
it
will
than
fatal
extremely rare events. e\cn
Given the scarcity of observations on
be quite difficult to identify the parameters of accident-
operating characteristic iclntionships.
Moreover,
it
is
not clear that fatal accidents
21
are a priori fundamentally different signals of safety than are nonfatal accidents,
even though they are ex pest substantially worse outcomes.
4.2
Accident ProbablHtiBs across Carrier Groups
Before estimating models of the relationship bewteen accidents and firms'
operating characteristics,
to
it
is
useful to ask
have different accident probabilities.
whether different types of firms appear
In this section,
first
I
construct accident
rates for individual carriers over five-year time periods, beginning with the 1955-
1959 period and ending with the (four-year) 1980-1983 period.
rates vary substantially across carriers
and through time.
I
These accident
next explore whether
these variations are correlated with broad categories of operating or financial
characteristics (such as high or low profits, large or small size).
Table 4.2 reports the accident rates per thousand departures for each sample
carrier over each of the six time periods in the sample.
calculated as Nj^/Dj^, the
divided by the total
The
last three
number of
accidents for firm
number of departures
rows of the table report the
(in
Accident rates are
i
during time period
thousands) for firm
i
during period
total accidents, total departures,
estimated accident rate for the entire sample of carriers
in
t
t.
and
each period. The
accident rate for the entire sample declines continuously through time, with the
1980-83 accident rate
less
than one-fifth of the
This refiects the trends observed
in
initial
accident rate in 1955-59.
the aggregate data.
Likelihood ratio tests of
the equality of accident probabilities across time periods reject the hypotheses that
sample aggregate accident rates
The accident
and
any two adjacent five-year periods are the same.
rates for individual carriers exhibit considerable variance, although
some general patterns emerge:
PanAm
in
TWA,
for
example, the two international trunk carriers,
tend to have above average accident rates throughout the period,
22
TABI.n4.2
A(
Carrier
American
BranifT
Continental
r)elta
Faslem
National
Northeast
Northwest
Pan American
1 rans World
United
Western
Frontier
Ozark
Piedmont
Southern
lexas Intl.
IJSAir
Alaska
Aloha
Hawaiian
North. Consol.
Wicn Air
Caribbean Air
Pacific Northern
Pan Amcr-Cirace
Air Cal
Air Florida
PSA
Republic
Southwest
Total Accidents
Total Departures
Aggregate
Accident Rale
(
III!
\
I
PRonABii
i955-.sq
0.0211
mrs
t%0-64
by carrii r nv iimf
1965-69
1970-74
pi.rioi:)
1975-79
1980-83
23
as
do the intra-Alaskan
carriers,
Alaska Airlines, Northern Consolidated/Wien
Consolidated /Wien Air Alaska, and Wien Airlines (prior to
Consolidated).
This illustrates the need for caution
accident rates:
higher rates
may
in
its
merger with Northern
interpreting the individual
reflect differences in the inherent risk
of areas
served by particular carriers (such as potentially riskier conditions for international
airports
and Alaskan operations), rather than differences
To
test
I
group
type of operations (domestic trunk
carriers
on the basis of four characteristics:
v. others), size
(measured
passenger-miles),
in
(measured by operating margins) and maintenance expenditures (on
For the
constant dollar per aircraft mile basis).
tests, carriers are
size, profits,
The
threshold value.
a
and maintenance
separated into "High" and "Low" groups, depending on whether
their value of the relevant variable over the time period falls
threshold
specific to each time period.
one period and the
is
The
above or below
based on natural break points
This means that a carrier
Low group
different characteristics.
that the
se.
whether groups of carriers with different characteristics exhibit
different accident probabilities.
profitability
per
in carriers' safety
the next period,
may
and may be
use of natural break points,
two groups (High and Low) may not be of equal
be
in
in
the data, and
size;
is
the High group
in different
when
a
groups for
these occur, imply
the
Low group
in
each period and for each characteristic tends to be slightly smaller than the High
group, typically comprising 30 to 40 percent of the carriers.
trunk/nontrunk
test,
domestic trunk carriers are considered the High group and
others are considered to be in the
I
random
model the number of
variable.
Low
airline accidents per
The Poisson
distribution
Abraham, and Schimmel
all
group.
problem, and has been employed widely
Barnett,
For the
in
(1979),
is
thousand departures as
a
Poisson
particularly suited to this type of
studies of accident probabilities (see
Oolbe
(1986)).
A
detailed discussion of the
)
24
distributional assumption*;
Given
and estimation technique
this distribution, the
is
contained
accident probability per 1000 departures, ^
and
that both High
I
1
,
B.
can be
From
estimated from the accident rate, N,D, for each group of carriers.
accident probabilities.
Appendix
in
these
construct likelihood ratio statistics to test the hypothesis
cnv
groups are drawn from
distribution.
The
characteristic
and each time period, table
a
common
accident probability
For each
results of these tests are reported in tabic 4.3.
4.3 reports a
"
+
"
if
the accident
probability of the High group significantly exceeds the accident probability of the
Low
group,
"-"
a
if
the accident probability of the High group
the accident probability of the
fails to reject
low
group, and
a "0" if
is
significantly
the likelihood ratio test
When
the hypothesis that the two rates are equal.
the test rejects
the equality hypothesis, the table also reports the ratio of the
Low
accident probability to the High group's accident probability.
It
that these tests reflect only bivariate correlations,
below
groups's
should be stressed
and say nothing about causal
relationships.
The
does not appear
to
statistical equality
and
last
The trunk/other
results of these tests are mixed.
distinction (column
be associated with different accident probabilities; while the
of accident rates for these two groups
is
rejected in the
periods, the difference has opposite signs in the two periods.
clear pattern of differential accident rates
is
first
Thus, no
associated with this grouping.
Similarly, size (passenger-miles,
in
column
one of the periods
is
there a significant difference between accident
rates; in only
rates for the
High and
difficult to interpret.
Low
Over
group.
4)
does not appear to affect accident
The evidence on operating margins
the 1965-69
and 1980-83 period. High
ha\c significanth' lower accident rates than do
relationship
is
I
Low
more
profit carriers
This
profit carriers.
reversed, however, in the I9S.S-59 period, and there
is
is
no significant
25
1
ABLE
4.3
TESTS OE HIE HOMOC^ENEI lY OV ACCIDENT RATES ACROSS CARRIER GROUPS
Charac teristic on Which Carriers are Grouped
Period
26
difference in accldcnl
results
may
rate*; for
suggest the \flkic
the
(if
two grnup*; during 1960-64 and 1970-79. The
further exploration of profit-accident
do not provide very strong evidence on the existence or form of that
The
results for
maintenance expenditures per mile (column
to the results for other characteristics.
In five
statistically significant
relationship
is
surprising:
Low
High and
they
relationship.
stand
in
of the six time periods, there
significant difference in the accident rates of the
While
3)
linl<s;
contrast
a
is
carrier groups.
and persistent through time, the direction of the
l.ow maintenance expenditure carriers are associated with
lower, not higher accident rates,
expenditure carriers arc half
t(^
in fact,
accident rates for
Low maintenance
two-thirds the accident rates for High maintenance
expenditure carriers. This result
may
be driven, at least
in
part,
by the typically
high maintenance expenditures of the international carriers, which vvc noted earlier
also tended to ha\'c higher accident rates.
some other
factor that tends to raise both
This association, or the presence of
maintenance expenditures per mile and
accident rates (such as aircraft age or fleet composition),
correlation between maintenance and accident rates.
may
The
create a positive
regression analysis,
estimating the effect of maintenance holding other factors constant,
on
shed
light
this result.
4.3
Regression Analys is of IndfvidiJal
In this section,
I
Ca rriers' Accident Rates
model the accident
of their traffic and financial characteristics.
correlation of
level
rates for individual carriers as a function
This permits us to estimate the
particular characteristic with accident rates, controlling for the
a
of other characteristics.
rather than as
is
may
by
a
The model should be
interpreted as a reduced form,
structural model of the accident-generating process.
quite similar to studies by
Graham and Bowes
(1979),
The
analysis
Golbe (1986), and Peter
27
Belenky (1986 unpubHshed memoranda), although the data and
statistical
techniques
differ.
The
(4.1)
basic specification of carrier
i's
accident rate
in
year
t is:
ln(TOTACCjt/DEP>VRTjt) = BO + Bl*DINTLjt + B2*TIMEit + B3*AVSTAGEit
+ B4*LNEXPERit + B5*OPMARGit., + B6*RMAINTPMji
where the variables are as defined
specification
is
in section 4.1
A
and appendix A.
log-linear
used to ensure that the estimated accident rates satisfy the non-
negativity constraint.
I
treat profitability (lagged
OPMARG)
and maintenance as exogenous variables
Some
from the standpoint of the accident generating process.
for this treatment
reject the
is
provided by Golbe (1986),
who
finds that her data fail to
hypothesis of exogeneity for profitability measures.
reduces potential simultaneity problems:
last period's profits
contaminated by costs that arc incurred due to accidents
replacement of
aircraft,
Lagged
may
profits
damage
also be appropriate since, even
The exogeneity of maintenance expenditures
if
is
and
will
profits
should not be
this period (repair or
low profit firms reduce safety
is
unlikely to be immediate.
less plausible,
to the dearth of reasonable instruments for firm-level
will explore the
Using lagged
claims, higher insurance premiums, and the like).
investments or expenditures, the impact on accidents
Future work
empirical support
but
is
maintained due
maintenance expenditures.
robustness of the results to the exogeneity assumption,
attempt to develop instrumental variables estimators for the model.
The
predictions:
variables in equation (4.1) have the following interpretations and
First, the accident rate,
(TOTACC/DEPART),
estimate of the underlying accident probability, %.
DINTL,
can be interpreted as an
a
dummy
variable for
28
international dpcrntinns.
To
operations outside the U.S.
risky,
DINTL
measured by
a
should have
any higher
the effect of
cflptiirc"N
)-l
percent change
)
on accident rates
is
the expected accident rate
if
Its effect
in
below are qualitatively similar
carrier has international operations (the results
those obtained using the proportion of international flights
variable).
rate of
TIMF: allows
kmger
the effect of
flights
year.
R2
is
(holding constant the
effect of an increase of 100 miles in the
.1*I00*B3 percent. The
log of
log specification of the relationship
(
the expected sign of B4
(EXPER).
expected sign
is
in
over time,
is
and
at the
reflects
other
all
positive; the
flight
may
rise
experience
(LNEXPER)
is
included
The
with airline experience.
is
log-
taken from the literature on learning curves
is
A
negative.
dummy
AVSTAGE
flights
I9R5) and the references cited therein).
with a B4*I00 percent change
in levels
Its
number of
the
to
average stage length on accident rates
cumulative
to capture the notion that safety levels
Joskow and Rose
place of this
expected to be negative.
right-hand side \arinhlcs) on accident rates.
(see
in
for a logarithmic decline in accident rates
B2*100 percent per
more
the extent that international operations are
a positive coefricient.
IOn*(exp(Rl
risk associated with
1
percent change
the accident rate.
the coefficient will imply a
Given
experience
in
When
experience
B4*100 percent change
this rationale,
associated
is
measured
is
in
the
accident rate for an additional million miles of cumulative airline experience.
OPMARG.
argument
negative.
the operating margin,
is
a
measure of carrier
that lower profits induce lower levels of safety
A
a .l*B.*i*IOO
change
in
the operating margin from
percent change
in
expenditures per aircraft mile
of variation
in
the accident rate.
in
1982 dollars,
is
profitability.
the
If
correct, R5 should be
is
to 10 percent (.10) will predict
Finally,
RMAINTPM,
maintenance
included to control for the effects
maintenance expenditures on accident
rates.
The standard
financial
health argument ('financifllly stressed carriers shade on maintenance, reducing
29
safety") suggests a negative sign for this variable, although the results in section
4.2 from simple correlations between maintenance expenditures
A
found the opposite.
one dollar increase
in
be associated with a Bfi*19f) percent change
EXPER;
in
will
accident rates.
measuring experience as the
replacing the profit measure
and financial health indicators from the
rates
maintenance expenditures per mile
Variations on this basic equation include:
experience,
and accident
OPMARG
COMPUSTAT
level
of
with other profitability
data
set,
such as return on
equity (ROE), the current ratio, the equity-to-total capital ratio, and the market-to-
book
ratio; replacing
for each
TIME
two or three year
with a series of time fixed effects (separate intercepts
period),
and measuring maintenance by
total real
expenditures.
Table 4.4 reports
results for five variations of
equation (4.1) for the period
1955 through 19R2, for which complete data were available. These equations are
estimated over 26 carriers, excluding the former intrastate carriers, for
whom
cumulative experience measures could not be calculated. The variations reported
table 4.4 are representative of a
much broader range
in
of specifications that have
been estimated.
Column
except
1
is
the basic specification in (4.1).
RMAINTPM
have the expected
estimates are not precisely identified.
profitability,
measured
in
signs,
expenditures, on the accident rate.
although a number of the point
The focus of
these regressions
The
In this equation, all variables
the analysis
is
the infiuence of
by the operating margin and maintenance
coefficent on
OPMARG
is
negative,
implying that higher operating margins (profits) are associated with lower accident
rates.
A
change
in
the operating margin from
3.16 (2.95) percent reduction
on the coefficient makes
it
in
accident rates.
difficult to
to 10 percent
However, the
is
associated with a
large standard error
determine the precise effect of
OPMARG,
30
I
RI
\n] V 4 4
r.RISSFON ANAI YSIS
IQ.S.S
Variable
OI"
-
1
C
ARRIT R ACrinPNI RATI
982
S
31
and the hypothesis that the true
conventional levels of
coefficient
the carrier groupings
in
RMAINTPM
statistical significance.
but of the wrong sign. The result
observed
zero could not be rejected at
is
is
statistically significant,
consistent, however, with the patterns
table 4.3.
in
is
This suggests that we need better
explanators for maintenance expenses, or better controls for omitted variables that
may
drive both accidents and maintenance.
and the potential endogeneity of the
Given these problems with maintenance
variable, a
number of
the variations on
equation 4.1 omit maintenance expenditures.
TIME
has a very small negative effect on accident rates, corresponding to a
half percent decline per year, and
AVSTAGE
is
from zero.
statistically indistinguishable
has a positive effect on accidents:
an increase
in
the average stage
length from 500 to 1000 miles would raise the expected accident rate by 3.3 (.07)
percent.
For the 1980-83 sample aggregate accident rate of .003
would correspond
an increase
LNEXPER
on
coefficient
to
the accident rate from .003 to .0031.
in
suggests a quite pronounced learning effect:
airline experience reduces the accident rate
in part reflect
in table 4.2, this
by 24.3
(2.6) percent.
The
doubling
While
this
may
nonlinear time effects (as experience trends strongly through time),
the persistence of this finding across a wide variety of specifications, including
inclusion of individual year intercepts in place of
effects are not purely time effects.
Finally,
TIME,
DINTL
is
suggests that experience
of the predicted sign,
suggesting a 4.7 (6.5) percent higher accident rate for international carriers, but
is
statistically insiginificant.
The
variations in columns 2 through 5
qualitative conclusions.
Although
level
its
0PMARG
is
do not substantially
affect the
negative in virtually every specification.
point estimate remains fairly imprecise, for specifications using the
of experience (such as regression (3)) and those omitting experience (such as
32
regression
(5)")
hypothesis that the coefficient
\vc reject the
is
zero at the 95
percent confidence level or better (using a two-tailed hypothesis
regressions (3)
and
(5).
comes from
the rejection of zero
However, preliminary
not a smaller standard error.
a larger point estimate,
results for
Hausman,
estimation of accident probfibilities following
maximum
LNEXPER
specifications, including those with
all
range the results
evidence of
in (3)
and
lc\cl
of maintenance expenditures.
RMAINTPM.
although
variations (only one of which
constant but imprecise
magnitude
results all
in all
is
(with point estimates
appear
is
it
remains
RMAINT,
to suggest at least
for
in
the
weak
the
LNEXPER
EXPER
The
and
is
most
TIME
coefficient on
specifications,
is
slightly larger in
or omitting experience.
The
effect of
roughly constant throughout the variations, with the exception of the
equations that omit experience, as represented by
in specification.
insignificant
performs somewhat better
statistically insignificant in
reported here).
the equations using
in
AVSTAGE
variations
OPMARG
a profitability-accident relationship.
The
than does
These
(5)).
likelihood
Hall and Griliches (1984)
suggest statistically significant negative estimates of the coefficient on
almost
For both
test).
and
DINTL
(5).
LNEXPER
varies between positive
significant in the variations;
it
typically
is
robust to most
and negative,
adds
little
to the
explanatory power of the regressions.
To
explore the profit-accident relationship further.
various subsamplcs of firms and years.
the relationship through time.
19.').*;-
The
first test
The sample was
1964, I96.S-1974. and I97.S-I982.
The
I
estimated equation
4.1
explored the robustness of
di\-idcd into three time periods:
results for the first
and
third period are
quite similar to the aggregate results for variations of the basic model; those for
the middle time period dlffbr primarily
columns
I
through
3.
in
the
for
OPMARG
coefficient.
Table
reports these results for the basic specification (4.1).
4..'^,
33
IABLF4.5
si'Nsri'ivn Y
TFsrs of basic rfgression results
34
OPMARG
ha'^ a large
periods; this effect
essentially
no
is
effect
negative cfTect on accident rates during the early and late
during the 1965-74 period, although
large standard error.
The
second
local service,
is
estimated with
a quite
stability
through time.
explored whether the results were sensitive to pooling trunk,
test
and
it
has
coefficients of the remaining variables (except
maintenance) exhibit considerable
A
OPMARG
the 1953-1964 period.
stflti^ticallv signiTicant in
Alaskan /Hawaiian
territorial
carriers.
Variations of equation (4.1)
were estimated for each of these groups separately, over the entire 1955-1982
period.
The
results
were quite similar
to the
whole sample
results for
specifications, with the exception of the local service regressions.
through 6
in
local .service,
most
Columns
4
tabic 4.5 report the results for the basic specification (4.1) for trunk,
and territorial/Alaskan/Hawaiian
carriers.
OPMARG
has a large
negative effect for trunk and territorial carriers (statistically significant for the
latter group).
i^.
It
local service carriers.
however, estimated with
The remaining
a large positive coefficient for the
coefficients (except
maintenance) are
relatively constant across groups.
Finally,
firms with
I
explored
a richer set
COMPUSTAT
post- 1967 period.
For
this
data.
of financial measures for the subsamplc of
This restricted the sample to sixteen firms
subsamplc. the basic variations on
(4.1)
in the
were estimated,
as well as \ariations that included return on equity, current ratios, market-to-book
ratios,
and equity leverage
ratios to
measure
profitability
and financial health.
these regressions, none of the financial measures-- including
statistically significant,
and many were estimated with the wrong
other variables were broadly consistent with the
These
OPMARG-
results prox'irie
full
sample
sign.
For
was
Most of the
results.
mixed cxidence on the robustness of the specification of
the accident-generating process estimated here.
The
relationship seems
in
large
35
part stable across nutnernus variations on the basic specification and across a
variety of subsamples of the dataset.
positive profits-safety relationship.
This provides suggestive evidence of a
However, there are enough aberrations
dictate further research into the robustness of the results.
attempt to
test the stability
the results to variations
This effort
is
in
Future work
to
will
of equation (4.1) through time and the sensitivity of
the functional form of the profitability specification.
likely to involve collection of additional
measures of financial condition for the
full
sample of
instruments of maintenance expenditures, and
data to provide alternative
carriers, better indicators
a richer set
and
of controls for accident
rates.
5.
CONCLUSION
The
analysis provided
in this
safety in a deregulated environment.
study provides mixed implications for airline
On
the one hand, the aggregate safety
performance of the industry, as measured by both
fatal
and nonfatal accidents,
superb, and shows no sign of deterioration since deregulation.
any change
reduction
in
in
If
there has been
accident performance relative to trends under regulation,
accident levels, not an increase.
to deregulation.
and soundness of the
On
air safety
it is
a
For three of the eight years under
deregulation, for example, there have been no passenger fatalities-- as
no such years prior
is
compared
to
This provides strong evidence on the stability
system.
the other hand, this study finds evidence that financial condition
correlated with accident rates at the level of individual carriers.
may
be
In the presence
of controls for cumulative airline fiight experience and other operating
characteristics, higher operating
rates.
This finding,
if it
margins appear to be correlated with lower accident
proves to be robust, suggests grounds for possible
36
concern.
In pnrticiilar. tn the extent that the
less insulation
deregulated airline market involves
of carriers from the effects of market forces,
more
intense scrutiny
of the safety practices and performance of fmancialiy marginal carriers
desirable.
Note that the
results
do
not,
undesirable, even on safety grounds.
(sec
may
however, imply that deregulation
The enormous
is
social benefits of deregulation
Morrison and Winston (1986)) and the maintenance of an outstanding
record both appear achievable.
be
air safety
APPENDIX A
DATA DESCRIPTION AND SOURCES
1.
Accident data
Board (NTSB)
An
:
accident
is
defined by the National Transportation Safety
as "an occurrence associated with the operation of an aircraft
which
takes place between the time any person boards the aircraft with the intention of
flight until
such time as
persons have disembarked,
all
death or serious injury as
a result
of being
in
or
upon the
contact with the aircraft or anything attached thereto, or
receives substantial
which any person suffers
in
aircraft or
in
by
which the
direct
aircraft
damage" (U.S. Department of Transportation, 1980,
p. 18).
Individual air carrier accident data are from the U.S. Civil Aeronautics Board
(CAB), Resume of Accidents, U.S. Air Carriers, Rotorcraft and Large General
Aviation Aircraft (annual. 19.53-1959); U.S.
CAB,
Statistical
U.S. Air Carrier Accidents (annual, 1960-1965); U.S.
Review and
NTSB, Annual Review
Aircraft Accident Data, U.S. Air Carrier Operations (succeeds the
publications; various years. 1966-1982); the U.S.
Briefs of
NTSB,
CAB
of
accident
Preliminary Analysis of
Aircraft Accident Data, U.S. Civil Aviation (various years, 1979-82), and the
NTSB
Accident Briefs (unpublished computer printout, for 1983-1984). These data were
available from 1954 through 1984.
2.
Traffic data
and
:
Annual
aircraft miles
airline
(MILES)
in
system revenue departures
millions are from the U.S.
Statistics (various issues. 1954-1983).
as
MILES/DEPART,
and
is
CAB,
Average stage length
measured
in
(DEPART)
in
Air Carrier Traffic
(AVSTAGE)
thousands of miles.
thousands
is
computed
38
Financial data
3.
Annual
:
CAB.
arc frcim the U.S.
airline sy«;tem operating
Air Carrier Fina ncial Statistics (various issues, 1954-1982).
All dollar-dcnominntcd variables arc
ratio
(OPRAT)
revenues and operating expenses
in
The system operating
millions of dollars.
calculated as (operating expcnses)/(operating revenues).
is
OPMARG.
operating margin,
l-OPRAT.
is
Real net income
(RINC)
transformed
to
I9R2 constant dollars using the implicit
Capital structure xariablcs are calculated from
GNP
is
deflator.
COMPUSTAT
data, and reflect
CURRENT
financial data for the entire corporation, not just air carrier operations.
is
the current ratio, calculated as current assets/current
the ratio of market xaluc of
(ROF)
return on cquit\'
is
common
computed
is
equity)/(market \aluc of
common
book value of long-term
debt).
through 1985 for
4.
Maintenance
a
:
MKTBOOK
common
equity.
preferred equity.
common
COMPUSTAT
+
data are available from 1967
airline carriers.
Maintenance expenditures are from the U.S. CAB, Air Carrier
Because of changes
in
the
financial statistics reports, these data are available only from 1956 on.
maintenance expenditures
in
millions of 1982 dollars,
expenditures are escalated using the implicit
GNP
dcfiated maintenance cost series.
RMAINTl
RMAINTPM
divided by the
is
defiator.
real
CAB
s
RMAINTl
is
where nominal
RMAINT2
nominal maintenance expenditures by an average mechanics wage
mile, calculated as
The
equity + redemption value of preferred equity
The
sub-sample of
common and
as (market value of
Financial Statistics (various issues. 1956-1982).
real
is
calculated as net income before extraordinary items and
(EQRATIO)
ratio
liabilities.
equity to book value of
discontinued operations divided by the book value of
The equity
calculated as
is
the difference between operating revenues and operating expenses, and
The
divides
series to obtain a
maintenance expenditures per
number of
aircraft miles.
39
The mechanics wage
mechanics wage
is
series
is
constructed from two sources.
the mid-pnint of the hourly
Lockheed-California, obtained from:
Statistics,
wage range
for
Aircraft Corp.] and Machinist s Union,
is
the
Grade
5
mechanics at
U.S. Department of Labor, Bureau of Labor
Wage Chronolo fty: L ockheed-California Company
For 1961-1983, the wage
For 1955-60, the
[Division of Lockheed
March 1937-October 1977
.
Bulletin 1904, 1976.
motor vehicle mechanics average wage from the
U.S. Department of Labor. Bureau of Labor Statistics,
Handbook
of Labor Statistics
(1975- Reference Edition. Bulletin 1865, Table 109, p. 285 for 1961-74; Bulletin 2217,
June 1985, Table 96.
5.
Experience
cumulative
:
p.
277 for 1975-83).
Airline experience
MILES
(EXPER)
in
year
flown, from 1954 through year
t.
t
is
calculated as the
APPENDIX
MODEL FOR ACCIDENTS
A PROHABII.ITY
A
B
number of
natural stochastic spccificatinn for the
air carrier accidents
is
based on the Poissnn probability distribution. The Poisson distribution recognizes
the infrequent
as a
and
discrete natures of accidents,
model of accident probabilities
carrier accidents (Barnctt.
purposes,
is is
and has been applied extensively
wide variety of contexts, including
in a
Abraham, and Schimmcl
most plausible
to specify the
number
Golbc
(1979),
of accidents
is
in
(in
in a
year as a
an accident, and that the number of
the binomial distribution
s
some
flights
is
experiences
\j, its
large,
convergence to the Poisson
number of departures
accidents
nj
in a
year as
rij,
which takes advantage of
in
Pr(n - np -
The
thousands)
we can express
the limit.)
i's
expected
in a
number
year as Dj, and
the probability that firm
of
its
i
accidents during the year as:
(B.l)
The maximum
(in
(This
probability, p, of being involved
Using the Poisson distribution, and denoting firm
number of
number of
thousands), rather than to specify an accident rate per year.
equivalent to assuming that each flight has
accidents as
For our
(1986)).
function of an (unknown) accident rate per thousand departures and the
departures
air
likelihood estimator
analysis
in
[exp(- XiDj)](X|Di)" /n!
(MLE)
for the accident rate Xj
is
Xj
=
nj/Dj.
section 4.2 relies on this simple parameterization of carrier's
accident probabilities to investigate whether different types of air carriers have
different underlying accident probabilities.
To
illustrate this technique, consider a
-
41
test
of the equality of accident rates for domestic trunlc carriers and
carriers.
Let
N| be
number of
the total
departures for group
(domestic trunk carriers), and
1
corresponding totals for group 2
=
= X2
Al
=
X2
^°f ^^^
= ^0
MLE
The
Tlie
•
group
tielonglng to
i
for X
MLE
"•"
|
and
for
all
i
test for
and X2 are
for xi
XO ^ ^^I
'^
2),
X
x
1
D] be
the total
N2 and D2
let
We
other carriers).
(all
of homogeneity within eaoh group (Xj
Xi
accidents and
will
other
all
number of
be the
impose the assumption
belonging to group
1;
homogeneity across groups:
= Nj/Dj and X2 = N2/D2
N2)/(Dj + D2). To
test for
homogeneity, we
construct the ratio of the likelihood under the null hypothesis that groups
are
drawn from
the
same probability
=
alternative hypothesis thai X]
LR =
(B.2)
X
the log of
in the
LR
maximum
X2.
This likelihood ratio
and 2
is:
o^^'^^2)exp(-Xo(Di + D2))
exp(-xiDi-x2D2)
b and^
likelihood estimates of xo> X
2
and taking
yields a log-llkelihood of:
LogLR = (N| + N2)ln((Nj + N2)/(Di +
(B.3)
1
distribution to the likelihood under the
X^l X^2
Substituting
•
D2))
N|ln(N/Di)- N2ln(N2/D2)
The
test statistic
-2LogLR
In section 4.3
group have
distributed as a chi-square(I)
relax the
assumption that
identical accident rates.
carrier's operating
vector Xj.
we
is
We
and financial
We
all
variable.
carriers within a particular
parameterize x
characteristics.
random
1
as a function of a
Denote these characteristics as the
=
parameterize the accident rate per thousand flights asx
i
exp(XiB). This parameterization ensures that the estimated accident rates satisfy
42
the n(>n-ncgati\ity restriction on x
greater than or equal tn zero).
the expected
(i.e..
number
of accidents must he
This parameterization of the accident rate
is
estimated by ordinary least squares fOLS) regression of the form:
InfNj/D)) = XjR
(B.4)
This
is
a
thousand
simple linear equation
in
The elements of
flights.
the log of the observed accident rate per
the estimated coefficient vector
BqIj^
interpretation that a one unit change in the corresponding variable in
a B(,|^*IOO percent
change
in
OLS
In particular, the
mean of
implies hctciosccdasticity
the
OLS
One
assumptions
a
in
will
will
no longer be an
Poisson distribution (Aj)
equation (B.4).
X
will lead to
Note, however, that
the accident probability.
accidents arc distributed as Poisson.
have the
is
efficient estimator.
equal to
The standard
if
its
variance, which
errors calculated under
be inconsistent estimates of the true standard errors.
solution to this heteroscedasticity
to use the
is
square root of the number of
accidents as the dependent variable, which will have a constant variance under the
Poisson assumption (sec Onlbc fl98fS)V
of the non-negativity restriction on x
specification of (B.4).
The problem
later version of this paper.
•
This
will not,
however, ensure satisfaction
therefore maintain the log-linear
•
of heteroscedasticity will be addressed
In addition,
maximum
likelihood estimates of the
parameters of the accident generating process, along the
Hausman,
Hall and Griliches (19R4) and
being explored.
Poisson process.
These techniques
in a
Cameron and
lines
developed
in
Trivedi (1986), arc currently
explicitly treat the heteroscedasticity of the
REFERENCES
Inc. (1986) An Evaluation of the Relationship Between
Aircarrier Financial Condition and Safety Posture for U. S. Department of
Transportation, Federal Aviation Administration, Washington, D.C.
Advanced Technology,
,
"The Market for Lemons: Qualitative Uncertainty and the
Market Mechanisms," Quarterly Journal of Economics vol. 84 (3) pp. 488-500.
Akerlof, G. (1970)
Allen, F. (1984) "Reputation and Product Quality,"
(3) pp. 311-327.
Rand
Journal of Economics vol. 15
Barnett, A., Abraham, M. and Schimmel, V. (1979) "Airline Safety:
Findings," Man ag ement Science vol. 25 (II) 1045-1056.
Some Empirical
Barnett, A. and Lofaso, A.J. (1983) "After the Crash: The Passenger Response to
the DC- 10 Disaster," Management Science vol 29 (11) pp. 1225-1236.
Bulow,
J. and Shoven. J. (1978) "The Bankruptcy Decision,"
Economics vol. 9 (2) pp. 437-456.
Journal of
Bell
Cameron, A.C. and Trivcdi, P.K. (1986) "Econometric Models Based on Count Data:
Comparisons and Applications of Some Estimators and Tests," Journal of
Applied Econometric s
vol.
I
(1) pp. 29-53.
Chalk, A.J. (1985) "A New Proposal for the Reform of Commercial Air Crash
Litigation," Journal of Air Law and Commerce vol. 50, pp. 219-252.
(1986) "Market Forces and Aircraft Safety: The Case of the DC-10,"
Economic Inquiry vol. 24 (1) pp. 43-60.
Feaver, D.B. (1986)
(March
"How
Safe are the Skies?," Washington Post
,
Weekly Magazine,
31).
Fromm, G.
(1968) "Aviation Safety,"
590-618.
Law and Contemporary Problems
"What Deregulation Has Meant
Gattuso, J.L. (1986)
for Airline Safety,"
Backgrounder The Heritage Foundation (November
,
vol.
33
(3) pp.
The
12).
Golbe, D.L. (1981) "The Effects of Imminent Bankruptcy on Stockholder Risk
Preferences and Behavior," Bell Journal of Economics vol. 12 (1) pp. 321-328.
(1986) "Safety and Profits in the Airline Industry," Journal of Industrial
Economics vol 34 (3) pp. 305-318.
44
nnci Bowes, M. (I'^?^) Do Finances Influence Airline Safety.
Maintenance, an d Services? The Public Research Institute of the Center
Naval An.'ihsis. Alexandri.-i. Va.
Graham. D.R
.
for
Hausman.
J.. Hall. B.H., and Grilichcs. Z. (1984) "Econometric Models for Count Data
with an Application to the Pnicnts-R(^D Relationship." Economc trica vol. 52 (4)
pp. 909-93S.
I. and
Rose. N.I.. {I98S) The Effects of Technological Change,
Experience, and En\ ironrnenlal Regulation on the Construction Cost of Coalburning Generating I Inits/' Rand Jounia]j)f_Econ(^^
Ki(l). pp. 1-27.
.loskow, P.
Kahn, A.E. and Kaspcr. P.M. I9S6) Airline Safety Without Re-Rcgulation.'
York Times (April 14) p I.IQ.
(
KB
The Role of Market Forces
and Lcffler,
(1 9X1)
Contractual Pcifoimancc." .lourn al of Political Economy
Klein, B.
in
Assuring
89 (4) pp.
vol.
New
(y\5-
641.
Magnuson.
E. (1987) "Be C-ucful
Out There." Time
\ol.
129 (No.
2,
Jan. 12) pp. 24-
32.
Main.
.1.
(I98.S)
"Making
It
Safer to Fly." Fortune pp. 20-26.
,
aiul Shugart. W.F. Ill (1986) "Has Deregulation of Air Travel
Affected Air Safety?. Working Paper 101, Center for the Study of American
Business, Washington 'nixersity. St. Louis.
McKcnzic. R.B.
I
Morrison, S. and Winston, C. (1986) The Econom ic Effec ts of Airline Dereg ul ation
Brookings Institution, Washington, D,C.
Nance.
.I..I.
(1986) Blind Trust. William
Morrow and
Co..
New
.
York.
Nolan, D. (1986) "Airline Safclv: The Shocking Truth," Discover vol
7 (10) pp. 30-
53.
.Ir. and Zorn. ( .K. (1984)
Deregulation and Commuter Airline Safety,"
Journal o f Air Law and Commerce \'ol 49 (2) pp. 3 1.'^-33.5.
Oster, C.V.
Pakcs, A. and Grilichcs. Z. (1980)
Look," Economic Letters
vol.
R&D
Patents and
.S, pp. 377-381.
at
the Firm Level:
A
First
Shapiro, C. (1982) "Product Quality and Seller Reputation," Bell Journal of Economics
vol. 13 (1) pp.
20-3^
(1983) "Premiums
foi
Higher Quality Products as Returns
\ol. 98 (4) pp. 6.*i9-680.
to Reputations,"
Quar terly Journal of Econprnics
Thaver, F.C. (1986) "The Emerging Dangers of Deregulation,"
23). p. III.
3.
New York Times
.
(Feb.
45
U.S. Department of Transportation, Federal Aviation Administration (various years)
Annual Report on the Effect of the Airline Deregulation Act on the Level of
Air Safety Annual reports, 1980-1984, Washington, D. C.
,
(
Date Due
"^9
DEC 2 8
IT
Lib-26-67
Mil LIBRARIES
3
TDflD
DOS 131 DS4