Network Propagation and Air Congestion Policies
Transcript of Network Propagation and Air Congestion Policies
Network Propagation andAir Congestion Policies
Tom Lam, Clemson UniversityChristy Zhou, Clemson University
Jan. 4, 2021, ASSA/TPUG Session
Lam and Zhou Network Propagation 1 / 41
Congestion in Air Transportation
Jets line up for takeo�, March 27, 2006 at O'Hare International
Airport in Chicago, Illinois (Photo by Tim Boyle/Getty Images)
I Congestion ine�ciency wastes time: Busy airports, such as JFK
and EWR tend to be congested 10 to 20 percent of the time. At
Newark, planes average taxi times of 52 minutes during
congested periods versus 14 minutes during less busy times.
Pushback times for planes can exacerbate the situation. - Forbes
Lam and Zhou Network Propagation 1 / 41
Air Travel Time Has Been Increasing10
011
012
013
014
015
0m
inut
es
2010 2011 2012 2013 2014 2015 2016 2017year
Elapsed Time + Delayed DepartureElapsed TimeAirborne Time
A. Delayed, Elapsed, and Airbourn Time
510
1520
25m
inut
es2010 2011 2012 2013 2014 2015 2016 2017
year
Taxi TimeDelayed DepartureDelayed Arrival Time
B. Delayed and Taxi Time
Sources: Chu and Zhou (2019) using DOT On-Time Performance
Lam and Zhou Network Propagation 2 / 41
Sources of Inefficiency in Air Travel Delays
Underprovision of modern infrastructure
I A public good problem
I Positive e�ect of increasing FAA expenditure (Morris and
Winton 2017), adopting the Next Generation Air Transportation
System (NextGen) (Chu and Zhou 2019)
Congestion externality (this paper)
I Airline companies do not have incentives to schedule their �ights
in a way that internalizes the cost of delays they impose on their
competitors, causing a congestion externality
I E.g., Delta does not need to internalize the cost of delay it
imposes on United
I This e�ect can further propagate via the air tra�c network
Lam and Zhou Network Propagation 3 / 41
Congestion also matters for competition
Usual forms of competition
I price
I non-price channels such as product di�erentiation along attribute
space to substitute consumer away from competing products
Congestion - another plausible non-price channel
I Could be unintended or strategically intended
I A�ects rival demand: Consider delay as a dimension in the
product space. Delay would make rival �ights would look bad on
websites such as �ightstats.com
I A�ects rival marginal cost: delay per se is costly
I Empirical evidence: Mayer and Sinai (2003 AER); Morris and
Winston 2007 AEJ Pol)
Lam and Zhou Network Propagation 4 / 41
Hub airlines bunch flights into peaks at hubHub vs. Airport Total Flights at DFW
Sources: Mayer and Sinai (2003, AER)
Lam and Zhou Network Propagation 5 / 41
Hub airlines bunch flights into peaks at hubDeparture Density for Hub and Non-hub Carrier at DFW
Sources: Mayer and Sinai (2003, AER)
Lam and Zhou Network Propagation 6 / 41
Mostly for departure flights...Departure Density for Hub and Non-hub Carrier at DFW
Sources: Mayer and Sinai (2003, AER)
Lam and Zhou Network Propagation 7 / 41
This paperGoal: To quantify the e�ect of congestion on air delay and air travel
time - the congestion e�ect
I important for designing air tra�c policies
I important for learning the proportion of congestion that are not
internalized and how they vary across airports
Challenge - the presence of network
I Congestion e�ects are heterogeneous
I E.g., the external cost �ight A imposed on other �ights might be
higher than �ight B
I The heterogeneity arises from (1) The characteristics of the
�ights a�ected (e.g. number of passengers, destination) (2) to
what extent this e�ect propagates in the network (e.g. peak
time, busy airport)
Lam and Zhou Network Propagation 8 / 41
Suggestive evidence of heterogeneous effects
Measure the importance of a �ight using the eigenvalue centrality
using �ights from March 31 to April 10, 2018
I For a departing �ight A (as a node), its centrality depends on
the other �ights that share the runway with �ight A and
proceeding operations of those �ights. We allow all proceeding
�ights to connect to this node if they share the airport within a
90 minute window.
Lam and Zhou Network Propagation 9 / 41
Approach
Step 1. Estimate the contemporaneous e�ect of having more �ights
on departure delay and taxi-out time - the direct congestion e�ect.
The identi�cation arises from:
I (1) Variation in number of unscheduled �ights
I (2) A selection-over-unobservable approach (SOO). Exhaustive
sets of �xed e�ects using the rich high-frequency air travel data:
OAG data from 2014 to 2017
Step 2. Quantify how the direct e�ect propagate via the network -
the indirect congestion e�ect
I Simulating a shock of �ight delay to the �ght network
The summation of them represent the overall marginal congestion
cost of an unexpected �ghts on delay and taxi-out time
Lam and Zhou Network Propagation 10 / 41
Approach
Step 1. Estimate the contemporaneous e�ect of having more �ights
on departure delay and taxi-out time - the direct congestion e�ect.
The identi�cation arises from:
I (1) Variation in number of unscheduled �ights
I (2) A selection-over-unobservable approach (SOO). Exhaustive
sets of �xed e�ects using the rich high-frequency air travel data:
OAG data from 2014 to 2017
Step 2. Quantify how the direct e�ect propagate via the network -
the indirect congestion e�ect
I Simulating a shock of �ight delay to the �ght network
The summation of them represent the overall marginal congestion
cost of an unexpected �ghts on delay and taxi-out time
Lam and Zhou Network Propagation 10 / 41
Our estimates would allow us toAssess how second-best congestion price can approximate the �rst
best
I Methods as in Sallee (2019)
I Current policies: airport-by-weight-speci�c landing fee at
speci�c airports
I Are centrality good enough to approximate the social cost?
I Other policy options regarding unscheduled �ights
Improve the delay multipliers in the FAA's Cost-Bene�t Guideline
I Current delay multipliers by FAA: Airport-speci�c parameters
for top 20 airports using simple counting
I Should construct from estimated marginal e�ect of congestion
I Should be more heterogeneous; should account for the network
e�ects
Lam and Zhou Network Propagation 11 / 41
Empirical Model
Step 1: Estimate the direct congestion e�ect
For a �ight i departing from the origin airport o in a given year y and
month m on a given outgate time t (down to a particular minute):
yiot = βnactot + φob + φym + εifot (1)
I yiot - departure delay (min), taxi-out (min)
I nactot - the number of unscheduled �ights taxiing out from the
airport o at the actual outgate time tI Data: all domestic and international �ights from and to top 40
US airports from 2014 to 2017 using OAG historical �ight
dataset
Lam and Zhou Network Propagation 12 / 41
Unscheduled flightsI Charter airlines and unscheduled airlines do not set months or years
ahead of the time as scheduled carriers
I Commercial charter �ights - can buy a seat or reserve for a groupI Private charter �ights - can rent a trip or own a private jetI Our data: 5.3% are unscheduled, within which 32.9% are private
Lam and Zhou Network Propagation 13 / 41
Summary Statistics: Main X Variable
Lam and Zhou Network Propagation 14 / 41
Summary Statistics: Other X Variables
Lam and Zhou Network Propagation 15 / 41
Empirical ModelStep 1: Estimate the direct congestion e�ect
For a �ight i departing from the origin airport o in a given year y and
month m on a given outgate time t (down to a particular minute):
yiot = βnactot + φob + φym + εifot (1)
I φob - the main airport-timeblock �xed e�ects:
origin airport o × time block b (which is year × quarter-in-a-year
× day-of-a-week × hour-of-a-day × 15-minute-slot-in-an-hour)
- e.g., compare
�ights at ATL Friday Nov. 01 12:00-12:15pm
�ights at ATL Friday Nov. 22 12:00-12:15pm
I φym - year by month �xed e�ects
Lam and Zhou Network Propagation 16 / 41
Identifying assumptions
No additional unobservables correlated with unexpected number of
unscheduled �ights and air tra�c delay
I Airports and terminal towers do not schedule a dynamic shift of
sta� members to respond to unexpected increases of operations
every day
I Peak time blocks (15-min) vary across days
I Airlines have to schedule their �ights to the computer reservation
system (CRS) every quarter ahead of time (Forbes 2008 IJIO)
Lam and Zhou Network Propagation 17 / 41
The Direct Congestion EffectDeparture Delay
Lam and Zhou Network Propagation 18 / 41
The Direct Congestion EffectTaxi-out
Lam and Zhou Network Propagation 19 / 41
RobustnessAlternative and Additional Fixed E�ects
I Add route-by-carrier �xed e�ects Go
I Add route-by-carrier-by-year-quarter �xed e�ects Go
I Add aircraft model (e.g., B777) �xed e�ects Go
I Use scheduled outgate time to de�ne �xed e�ects Go
More Conservative Clustered Standard Error
I Origin airport-by-quarter Go
I Origin airport-by-year Go
I Origin airport Go
To be added
I Add route-by-terminal �xed e�ects
I Control of number of runways
Lam and Zhou Network Propagation 20 / 41
Additional resultsAlternative Speci�cations: Include
I nact (unscheduled �ight at actual outgate time) and
nactreg (regular �ight at actual outgate time) Go
I nact (unscheduled �ight at actual outgate time) and
ninair (unscheduled �ight at actual in-air time) Go
Additional heterogeneous e�ects
I By domestic vs international �ight Go
I By domestic top 40 airports vs others Go
I By if the �ight is the �rst operation of an aircraft in a day Go
I By hub vs non-hub carrier at hub airports
To be added
I Consider capacity constraint: Allow for hockey-stick style of marginal
e�ect
Lam and Zhou Network Propagation 21 / 41
The indirect effectGoal: Simulate how delays and taxi-out time propagates in
subsequent minutes within the airport
Lam and Zhou Network Propagation 22 / 41
The indirect effectIn the future: Add how delays and taxi-out time propagate across
locations
Lam and Zhou Network Propagation 23 / 41
Simulate the indirect effect from τ to {τ + 1, τ + 2...}
Arrange all the scheduled �ights in an airport into 1-minute slots. Shock
the airport by adding extra �ights at a given timeslot τ by ∆�ights
I 1. At τ , set ∆�ights = 1I 2. Compute the expected outgate time and in-air time of this
additional ∆�ight using �xed e�ects and residuals
I 3. Use step 2 to compute in which time slots we will have ∆nact = 1I 4. For each a�ected time slots, simulate direct congestion e�ect of
�ights in those time slots. Using β̂ and ∆n to compute the new
outgate time and in-air time for these �ights
I 5. Re-calculate ∆n caused by step 4 all time slots
I 6. Repeat steps 4 � 5 until no additional �ight is a�ected
I 7. Sum up all repeated steps 4 � 6 (except for the �rst-round direct
e�ect) as the indirect e�ect
Lam and Zhou Network Propagation 24 / 41
Simulate the indirect effect from τ to {τ + 1, τ + 2...}
Arrange all the scheduled �ights in an airport into 1-minute slots. Shock
the airport by adding extra �ights at a given timeslot τ by ∆�ights
I 1. At τ , set ∆�ights = 1I 2. Compute the expected outgate time and in-air time of this
additional ∆�ight using �xed e�ects and residuals
I 3. Use step 2 to compute in which time slots we will have ∆nact = 1I 4. For each a�ected time slots, simulate direct congestion e�ect of
�ights in those time slots. Using β̂ and ∆n to compute the new
outgate time and in-air time for these �ights
I 5. Re-calculate ∆n caused by step 4 all time slots
I 6. Repeat steps 4 � 5 until no additional �ight is a�ected
I 7. Sum up all repeated steps 4 � 6 (except for the �rst-round direct
e�ect) as the indirect e�ect
Lam and Zhou Network Propagation 25 / 41
Simulate the indirect effect from τ to {τ + 1, τ + 2...}
Arrange all the scheduled �ights in an airport into 1-minute slots. Shock
the airport by adding extra �ights at a given timeslot τ by ∆�ights
I 1. At τ , set ∆�ights = 1I 2. Compute the expected outgate time and in-air time of this
additional ∆�ight using �xed e�ects and residuals
I 3. Use step 2 to compute in which time slots we will have ∆nact = 1I 4. For each a�ected time slots, simulate direct congestion e�ect of
�ights in those time slots. Using β̂ and ∆n to compute the new
outgate time and in-air time for these �ights
I 5. Re-calculate ∆n caused by step 4 all time slots
I 6. Repeat steps 4 � 5 until no additional �ight is a�ected
I 7. Sum up all repeated steps 4 � 6 (except for the �rst-round direct
e�ect) as the indirect e�ect
Lam and Zhou Network Propagation 26 / 41
Simulate the indirect effect from τ to {τ + 1, τ + 2...}
Arrange all the scheduled �ights in an airport into 1-minute slots. Shock
the airport by adding extra �ights at a given timeslot τ by ∆�ights
I 1. At τ , set ∆�ights = 1I 2. Compute the expected outgate time and in-air time of this
additional ∆�ight using �xed e�ects and residuals
I 3. Use step 2 to compute in which time slots we will have ∆nact = 1I 4. For each a�ected time slots, simulate direct congestion e�ect of
�ights in those time slots. Using β̂ and ∆n to compute the new
outgate time and in-air time for these �ights.
I 5. Re-calculate ∆n caused by step 4 all time slots
I 6. Repeat steps 4 � 5 until no additional �ight is a�ected
I 7. Sum up all repeated steps 4 � 6 (except for the �rst-round direct
e�ect) as the indirect e�ect
Lam and Zhou Network Propagation 27 / 41
Simulate the indirect effect from τ to {τ + 1, τ + 2...}
Arrange all the scheduled �ights in an airport into 1-minute slots. Shock
the airport by adding extra �ights at a given timeslot τ by ∆�ights
I 1. At τ , set ∆�ights = 1I 2. Compute the expected outgate time and in-air time of this
additional ∆�ight using �xed e�ects and residuals
I 3. Use step 2 to compute in which time slots we will have ∆nact = 1I 4. For each a�ected time slots, simulate direct congestion e�ect of
�ights in those time slots. Using β̂ and ∆n to compute the new
outgate time and in-air time for these �ights.
I 5. Re-calculate ∆n caused by step 4 all time slots
I 6. Repeat steps 4 � 5 until no additional �ight is a�ected
I 7. Sum up all repeated steps 4 � 6 (except for the �rst-round direct
e�ect) as the indirect e�ect
Lam and Zhou Network Propagation 28 / 41
Numbers of ∆flights to Shock
I We will set ∆�ights = 2, 4, or 10 to illustrate
I 2 ≈ the s.d. of nact
I 10 << the s.d. of nactreg
Lam and Zhou Network Propagation 29 / 41
Direct and indirect congestion effectsATL Saturday March 1, 2014
I Outcome: number of �ights a�ected (delay to the next minute,
or have to wait more on the runway)
I ∆�ights = 2
050
100
150
200
250
Num
ber o
f flig
hts
affe
cted
(del
ay +
taxi
out)
5pm 6am 7am 8am 9am 10am11am12pm 1pm 2pm 3pm 4pm 5pm 6pm 7pm 8pm 9pm 10pm11pm12amOutgate time in a day
Direct + propagated effectDirect effect
Lam and Zhou Network Propagation 30 / 41
Direct and indirect congestion effectsATL Saturday March 1, 2014
I Outcome: minutes of �ights a�ected (minutes of additional
departure delay or additional taxi-out time)
I ∆�ights = 2
010
020
030
040
0M
inut
es o
f flig
hts
affe
cted
(del
ay +
taxi
out)
5pm 6am 7am 8am 9am 10am11am12pm 1pm 2pm 3pm 4pm 5pm 6pm 7pm 8pm 9pm 10pm11pm12amOutgate time in a day
Direct + propagated effectDirect effect
Lam and Zhou Network Propagation 31 / 41
Direct and indirect congestion effectsATL Saturday March 1, 2014
I Outcome: minutes of �ights a�ected (minutes of additional
departure delay or additional taxi-out time)
I ∆�ights = 4
010
0020
0030
0040
0050
00M
inut
es o
f flig
hts
affe
cted
(del
ay +
taxi
out)
5pm 6am 7am 8am 9am 10am11am12pm 1pm 2pm 3pm 4pm 5pm 6pm 7pm 8pm 9pm 10pm11pm12amOutgate time in a day
Direct + propagated effectDirect effect
Lam and Zhou Network Propagation 32 / 41
Direct and indirect congestion effectsATL Saturday March 1, 2014
I Outcome: minutes of �ights a�ected (minutes of additional
departure delay or additional taxi-out time)
I ∆�ights = 10
050
0010
000
1500
0M
inut
es o
f flig
hts
affe
cted
(del
ay +
taxi
out)
5pm 6am 7am 8am 9am 10am11am12pm 1pm 2pm 3pm 4pm 5pm 6pm 7pm 8pm 9pm 10pm11pm12amOutgate time in a day
Direct + propagated effectDirect effect
Lam and Zhou Network Propagation 33 / 41
Direct and indirect congestion effectsATL Monday March 3, 2014
I Outcome: minutes of �ights a�ected (minutes of additional
departure delay or additional taxi-out time)
I ∆�ights = 4
050
0010
000
1500
0M
inut
es o
f flig
hts
affe
cted
(del
ay +
taxi
out)
5pm 6am 7am 8am 9am 10am11am12pm 1pm 2pm 3pm 4pm 5pm 6pm 7pm 8pm 9pm 10pm11pm12amOutgate time in a day
Direct + propagated effectDirect effect
Lam and Zhou Network Propagation 34 / 41
Direct and indirect congestion effectsORD Saturday March 1, 2014
I Outcome: minutes of �ights a�ected (minutes of additional
departure delay or additional taxi-out time)
I ∆�ights = 4
010
0020
0030
0040
00M
inut
es o
f flig
hts
affe
cted
(del
ay +
taxi
out)
5pm 6am 7am 8am 9am 10am11am12pm 1pm 2pm 3pm 4pm 5pm 6pm 7pm 8pm 9pm 10pm11pm12amOutgate time in a day
Direct + propagated effectDirect effect
Lam and Zhou Network Propagation 35 / 41
Direct and indirect congestion effectsSEA Saturday March 1, 2014
I Outcome: minutes of �ights a�ected (minutes of additional
departure delay or additional taxi-out time)
I ∆�ights = 4
010
020
030
040
0M
inut
es o
f flig
hts
affe
cted
(del
ay +
taxi
out)
5pm 6am 7am 8am 9am 10am11am12pm 1pm 2pm 3pm 4pm 5pm 6pm 7pm 8pm 9pm 10pm11pm12amOutgate time in a day
Direct + propagated effectDirect effect
Lam and Zhou Network Propagation 36 / 41
Direct and indirect congestion effectsAUS Saturday March 1, 2014
I Outcome: minutes of �ights a�ected (minutes of additional
departure delay or additional taxi-out time)
I ∆�ights = 4
020
4060
8010
0M
inut
es o
f flig
hts
affe
cted
(del
ay +
taxi
out)
5pm 6am 7am 8am 9am 10am11am12pm 1pm 2pm 3pm 4pm 5pm 6pm 7pm 8pm 9pm 10pm11pm12amOutgate time in a day
Direct + propagated effectDirect effect
Lam and Zhou Network Propagation 37 / 41
Direct and indirect congestion effectsMDW Saturday March 1, 2014
I Outcome: minutes of �ights a�ected (minutes of additional
departure delay or additional taxi-out time)
I ∆�ights = 4
050
100
150
Min
utes
of f
light
s af
fect
ed (d
elay
+ ta
xiou
t)
5pm 6am 7am 8am 9am 10am11am12pm 1pm 2pm 3pm 4pm 5pm 6pm 7pm 8pm 9pm 10pm11pm12amOutgate time in a day
Direct + propagated effectDirect effect
Lam and Zhou Network Propagation 38 / 41
Summary + next (for simulation)
Summary:
I Sizable direct congestion e�ect
I Sizable and (very) heterogeneous propagated e�ect
Construct �ner network structure:
I Include propagation across airports, i.e., allow �ight A delayed
at its departure airport in t to delay at its destination airport
I Include propagation across operations for an aircraft, i.e., allow
operation n to a�ect its next operation n + 1I Include propagation across �ights sharing the same departure
gate (using the newly acquired data)
I Account for heteroskedasticity
Lam and Zhou Network Propagation 39 / 41
Next for policy implications
Next for policy implications
I Fraction of delay internalized vs externalized
I Second-best congestion prices
I Delay multipliers
Thank you!
Lam and Zhou Network Propagation 40 / 41
Alternative Specification I
Back
Lam and Zhou Network Propagation 1 / 12
Alternative Specification II
Back
Lam and Zhou Network Propagation 2 / 12
Additional FE I: Carrier-by-Route FE
Back
Lam and Zhou Network Propagation 3 / 12
Additional FE II:Carrier-by-Route-by-Year-Quarter FE
Back
Lam and Zhou Network Propagation 4 / 12
Additional FE III: Aircraft Model e.g.,Boeing 777
Back
Lam and Zhou Network Propagation 5 / 12
Alternative FE Timing Definition: Use theScheduled Outgate Time for the Main FEs
Back
Lam and Zhou Network Propagation 6 / 12
Alternative Cluster SE I: Airport-by-Quarter
Back
Lam and Zhou Network Propagation 7 / 12
Alternative Cluster SE II: Airport-by-Year
Back
Lam and Zhou Network Propagation 8 / 12
Alternative Cluster SE III: Airport
Back
Lam and Zhou Network Propagation 9 / 12
Heterogeneous Effect: Domestic vs Int’l
Back
Lam and Zhou Network Propagation 10 / 12
Heterogeneous Effect: US Top 40 vs Others
Back
Lam and Zhou Network Propagation 11 / 12
Heterogeneous Effect: First Flight in a Day
Back
Lam and Zhou Network Propagation 12 / 12