Air Safety:
End of the Golden Age?
Arnold Barnett
How should we measure
aviation safety?
“NTSB studies show that, from 1993
through 1996, scheduled US carriers
averaged only 0.2 fatal accidents per
100,000 flight hours, less than half
the fatal accidents rate for the fouryear period a decade earlier.
--Wall Street Journal, 8/11/97
Two problems with the
statistic fatal accidents per
100,000 flight hours:
The numerator and
the denominator!
• The generic term “fatal accident”
blurs the distinction between a crash
that kills one passenger out of 300
and another that kills 300 out of 300.
• Measuring activity by “flying hours”
misses the point that most accidents
occur on landing or takeoff.
Airline Report Cards
(Air Travelers Association)
Score = 100 - 10,000 (Z/N)
where Z = number of fatal crashes over 1987-96
N = number of flights over 1987-96 (1000’s)
Letter Grades: A: 90 -100
C: 70-79.9
B: 80-89.9
D: 60-69.9 F: below 60
This grading system:
• Again uses the generic term “fatal
accident”
• Is quite arbitrary, and sometimes assigns
vastly different grades because of
statistically meaningless differences
Example:
• Airline #1
200,000 flights over 1987-96
No fatal accidents
Letter Grade: A
• Airline #2
200,000 flights over 1987-96
One fatal accident, which kills one passenger out of 100
Letter Grade: F
What about hull losses
per 100,000 departures?
(This is a popular one.)
Consider two hull losses this year:
• Southwest Airlines, Boeing 737, Burbank, CA
Passengers on board: 137
Passengers killed: 0
• Alaska Airlines, MD-80, off Los Angeles
Passengers on Board: 83
Passengers Killed: 83
No difference?
Measure of Safety Performance
Over a Past Period:
Death Risk Per
Randomly Chosen
Flight
Question:
If a person chooses a flight at
random from among those of
interest (e.g. UK domestic jet
flights over the period 1990-95),
what is the probability that he
will not survive it?
This death risk per flight statistic
has conceptual advantages
compared to the other statistics
just discussed.
What Conceptual Advantages?
• Ignores length and duration of flight,
which are virtually unrelated to
mortality risk
• Weights each crash by the
percentage of passengers killed
• Easy to calculate and understand
First-World Domestic Jet Services
Death Risk per Flight, 1990-99:
1 in 13 million
At a mortality risk of 1 in 13
million per flight, a passenger
who took one flight per day
would on average travel for
36,000 years before dying in a
plane crash.
Passenger Mortality Risk for Various
World-wide Jet Services, 1990-99
Type of Service
Death Risk per Flight
First-World Domestic
1 in 13 million
International within
First World
1 in 6 million
International Between First
And Developing Worlds
1 in 1 million
Within Developing World
1 in 500,000
Passenger Mortality Risk Arising from Criminal/Terrorist
Acts, Scheduled First-World Jet Services Over 1990-99
Type of Service
US:
Domestic
International
First World Outside US:
Domestic
International
Death Risk per Flight
0
0
0
1 in 2 billion
This record is all the more
remarkable because of several
successful acts of sabotage in
the late 1980’s.
Two Possible Reasons for the Quiet Decade:
• The desire to do harm to First-World
air travelers genuinely diminished.
• Improved security measures deterred
some potential attacks and foiled others.
Unfortunately, neither
of these explanations is
especially convincing.
October 2000:
“Fear of terrorist attacks after
the explosion of violence in the
Middle East hammered global
airline shares Friday.”
--Reuters
Passenger Mortality Risk Arising from Runway Collisions,
Scheduled First-World Jet Services Over 1990-99
Type of Service
Death Risk per Flight
US:
Domestic
International
1 in 100 million
0
First World Outside US:
Domestic
International
0
0
We were asked to investigate
the following question:
How great a threat do US runway
accidents pose to domestic airport
operations in the next two decades?
If there are N aircraft
operations at a given
airport in a given year,
then:
To a first approximation, one
might expect that the risk of a
runway accident would vary
2
with N .
Why?
1) The number of flights that could
(theoretically) collide is (N2-N)/2.
2) The Quadratic Model is conceptually attractive.
2
3)N is widely used in airspace collision-risk models.
But, to the extent possible, it is
desirable to go beyond merely
stating conjectures, and to test
hypotheses and “approximations”
against empirical evidence.
A most interesting data set
The 40 US runway incursions in 1997 that:
(1) were judged by experts to have
“extremely high” accident potential
and
(2) took place under known conditions of
reduced visibility (night, sunrise/sunset).
2
N -hypothesis
The
passed a
Chi-squared statistical test
with flying colors.
(The test was based on the spread of the 40
dangerous events across US airports.)
Intriguingly, the hypotheses
that dangerous events varied
across airports with either N or
N3 did not pass Chi-squared tests.
Overall, runway collisions
over the next two decades
could take 600 lives among
US jet passengers, and cause
200 serious injuries.
(Mid-range projection)
Estimated Runway Collision Death Risk
per Flight, US Domestic Jets 2003-22:
1 in 25 million
(Four times the actual risk in the 1990’s.)
Passenger Mortality Risk Arising from Mid-Air Collisions,
Scheduled First-World Jet Services Over 1990-99
Type of Service
US:
Domestic
International
First World Outside US:
Domestic
International
Death Risk per Flight
0
0
0
0
(Based on 100 million flights)
Is it safe to adopt
free-flight?
(Operations Research, Nov 2000)
B
A
F
E
C
Present Routings: A-E-F-B and C-E-F-D
Free Flight Routings: A-B and C-D
D
Under certain assumptions,
free-flight would:
• Reduce the likelihood of path
intersections
• Tend to reduce the crossing angles of
paths that intersect
Why is the latter point important?
Because:
At present, emergency warnings go
off in air-traffic control towers when
two planes come within five miles of
one another, regardless of the angle
at which they are converging.
Consider two planes on a
collision course that have just
come within five miles of one
another.
For example:
5
3.5
3.5
C
“Resolution Time” As a Function of
Angle of Convergence
Convergence Angle
180º
90°
30º
Time to Resolve Emergency
18 seconds
25
10
70 5
30°
(Assume both planes going 500 mph)
Because of fewer path
crossings and longer times to
resolve emergencies, the
geometric consequences of freeflight might act to reduce midair collision risk.
However:
All these apparent benefits of freeflight could be more than outweighed
by a decline in “situational awareness”
on the part of air traffic controllers.
How Does It All
Add Up?
Aviation Safety:
Time To Stop Worrying?
Arnold Barnett