Content uploaded by Jeffrey Czajkowski
Author content
All content in this area was uploaded by Jeffrey Czajkowski on Aug 26, 2014
Content may be subject to copyright.
Is It Time to Go yet? Understanding Household Hurricane
Evacuation Decisions from a Dynamic Perspective
Jeffrey Czajkowski1
Abstract: To better understand household hurricane evacuation decisions, this paper addresses a limitation to existing hurricane evacuation
modeling aspects by developing a dynamic model of hurricane evacuation behavior. A household’s evacuation decision is framed as an
optimal stopping problem where every potential evacuation time period prior to the actual hurricane landfall, the household’s optimal choice
is to either evacuate, or to wait one more time period for a revised hurricane forecast. We build a realistic multiperiod model of evacuation that
incorporates actual forecast and evacuation cost data for our designated Gulf of Mexico region. Results from our multiperiod model are
calibrated with existing evacuation timing data from a number of hurricanes. Given the calibrated dynamic framework, a number of policy
questions that plausibly affect the timing of household evacuations are analyzed, and a deeper understanding of existing empirical outcomes
in regard to the timing of the evacuation decision is achieved. DOI: 10.1061/(ASCE)NH.1527-6996.0000037.© 2011 American Society of
Civil Engineers.
CE Database subject headings: Hurricanes; Evacuation; Decision making; Gulf of Mexico.
Author keywords: Hurricanes; evacuation; dynamic decision-making.
Introduction
As moving water causes most hurricane-related fatalities, much of
the decline in hurricane fatalities since 1950 (Kunkel et al. 1999;
Rappaport 2000;Sadowski and Sutter 2005;Blake et al. 2007)is
attributed to improvements in hurricane forecasts and warnings that
have allowed for more timely evacuations (and hence less fatalities)
from storm-surge zones (Rappaport 2000;Willoughby et al. 2007).
The decline in fatalities is even more noteworthy when considering
the fact that population in high hurricane risk coastal areas has
grown significantly over this identical timeframe. However, this
population increase must be further stressed, for despite the fact
that the overall lethality of hurricanes has declined in recent dec-
ades, the potential risk for amplified casualties and/or injuries has
actually increased because of the growing coastal populations
(Centrec 2007). The fatality numbers from Hurricane Katrina alone
in 2005 were approximately 9,000 times higher than the annual
mean fatality rate of 20, and consequently provide a poignant
example of this high-risk reality, and more prominently of the criti-
cal role of evacuation.
Despite the critical role that timely evacuation plays in lowering
hurricane and storm surge fatalities, an understanding of household
evacuation is deemed to be “extremely limited”(Dash and Gladwin
2007) from an overall perspective, and even more so when consid-
ering the incorporation of temporal aspects of the evacuation
decision-making process (Dash and Gladwin 2007;Gladwin et al.
2007). In their overview of social science research needs related to
hurricane forecasts and warnings, Gladwin et al. (2007) highlight
the need for research that leads to the “… modeling of evacuation
behavioral response in more precise and comprehensive ways,”
including capturing the dynamic nature of microscopic individual-
ized (i.e., households) decision-making.
To better understand household hurricane evacuation, the pur-
pose of this paper is to develop a dynamic model of hurricane
evacuation behavior that models evacuation behavioral response
to hurricane forecasts in a way that captures the intertemporal
aspects of the evacuation decision process. Specifically, a house-
hold’s evacuation decision is framed as an optimal stopping prob-
lem where every potential evacuation time period prior to the actual
hurricane landfall, the household’s optimal choice is either to
evacuate, or to wait one more time period for a revised hurricane
forecast. We build a realistic multiperiod model of evacuation that
is calibrated by using existing forecast and evacuation cost data for
coastal areas on the Gulf of Mexico. Then, we show how the model
can help explain actual evacuation behavior from specific hurri-
canes, and expected evacuation timing outcomes for various house-
hold types. Finally, and most significantly, this dynamic framework
is used to explore a number of relevant policy questions that plau-
sibly affect the timing of household evacuations, and sometimes
provides the rationalization for seemingly counterintuitive post-
storm assessment evacuation results. For example, why does imple-
menting contraflow actually cause some households to be less
likely to evacuate?
Whitehead (2003) estimates the probability of evacuation for
varying levels of hurricane intensity, but does so from a static per-
spective, because the timing of the probability of an evacuation for
any particular storm intensity level is not addressed. However, the
evacuation decision when faced by a hurricane threat has the three
qualities of irreversibility, uncertainty, and the ability to wait for
more information that characterize a decision process that is better
understood from a dynamic modeling approach (Dixit and Pindyck
1994). Standard empirical results from the evacuation literature
such as the traditional S-shaped evacuation response curves
(USACE 2006a) clearly indicate that certain households wait,
1Assistant Professor, Department of Economics, Austin College, 900
North Grand Avenue, Suite 61579, Sherman, TX 75090. E-mail:
jczajkowski@austincollege.edu; and Assistant Adjunct Research Professor,
Florida International Univ., Int. Hurricane Research Center.
Note. This manuscript was submitted on January 28, 2010; approved on
September 8, 2010; published online on July 29, 2010. Discussion period
open until October 1, 2011; separate discussions must be submitted for
individual papers. This paper is part of the Natural Hazards Review,
Vol. 12, No. 2, May 1, 2011. ©ASCE, ISSN 1527-6988/2011/2-0–0/
$25.00.
1
2
3
NATURAL HAZARDS REVIEW © ASCE / MAY 2011 / 1
while others evacuate, and therefore further underscore the need for
a dynamic perspective of evacuation behavior.
Moreover, modeling the evacuation decision process dynami-
cally over many time periods with households having the ability
to wait for more information, is analogous to a real-life evacuation
decision situation where the National Hurricane Center (NHC)
issues official forecast advisories every 6 h once a tropical depres-
sion, tropical storm, or hurricane has developed. While Fu and
Wilmot (2004) utilize a sequential choice model to estimate the
probability of a household evacuating or waiting in each period
of their dynamic multiperiod framework, and further use their
dynamic model results to provide clarification to the standard
evacuation timing empirical outcomes, their research differs from
ours in a number of significant ways. Importantly, we provide a
theoretical model of dynamic evacuation behavior which is neces-
sary for conducting policy analysis. Furthermore, our dynamic
model is calibrated with forecast data from a number of storms
across a number of locations which coincides directly with the
6 h NHC forecast advisory timeline. Lastly, we explicitly address
the costs of evacuation in a household’s evacuation decision.
Although Regnier and Harr (2006) have developed an analogous
dynamic evacuation decision model, they have done so from an
emergency management perspective as opposed to the household
standpoint as we have here.
This research then serves as a contrast to the existing models of
household hurricane evacuation behavior, by utilizing a theoreti-
cally driven dynamic modeling approach that provides a more real-
istic interpretation to the multiperiod evacuation decision process
through the use of forecast and evacuation cost data. As a result,
through our dynamic model we can identify a representative house-
hold’s optimal point in time to evacuate over a five-day forecast
period, given a particular storm forecast and associated costs of
evacuation. We can further flex the model to incorporate various
household types such as high-damage (e.g., mobile home) versus
low damage, or salaried versus hourly wage to investigate whether
the optimal evacuation timing becomes earlier or later according to
type, and by how much. We can also analyze how policies that
change aspects of the costs of evacuating (e.g., contraflow), or
the cost of not evacuating (e.g., improved structural mitigation)
affect the optimal timing outcome, with the results illuminating
potential unintended and/or unwanted consequences of an other-
wise well-intentioned policy. Thus, our dynamic model framework
allows us to begin to bridge the previously noted knowledge gap
between hurricane forecasts and evacuation timing behaviors in a
variety of meaningful ways.
Hurricane Evacuation Decisions as a Dynamic
Process
The evacuation decision when faced by a hurricane threat has the
three qualities of irreversibility, uncertainty, and the ability to wait
for more information that characterize a decision process that is
better understood from a dynamic modeling approach (Dixit and
Pindyck 1994). We assume that the hurricane evacuation decision
is irreversible once made, and illustrate both the uncertainty inher-
ent in a hurricane forecast and the empirical evidence demonstrat-
ing a household’s ability to wait for more hurricane information.
While the irreversible evacuation decision assumption may not
hold in every case, e.g., severe highway congestion causing some
evacuees to return, we feel it is a reasonable assumption for most
evacuation situations. For example, mean evacuation distance trav-
eled for Hurricane Ivan was 182 miles (Morrow and Gladwin
2005); clearly not an easily reversible distance to cover.
Uncertain Hurricane Forecasts
Once a tropical depression, tropical storm, or hurricane has devel-
oped, the NHC issues an official forecast advisory every 6 h at
5:00 a.m., 11:00 a.m., 5:00 p.m., and 11:00 p.m. Two of the most
critical aspects of information contained in the NHC forecast ad-
visory are the 12, 24, 36, 48, 72, 96, and 120 h forecasts of an
approaching hurricane’s center position (track forecast), and the
maximum 1 min sustained wind speeds (intensity forecast).
Although the track and intensity forecast errors have improved over
time, there is still a high amount of variability in the accuracy of the
track and intensity forecasts themselves as the number of hours
prior to landfall increases. For example, in 2004, the average track
forecast error of 62 miles at 24 h out from landfall increased by
nearly 420% to 323 miles at 120 h out from landfall. A more rec-
ognizable illustration of the magnitude of the track forecast errors
over time is revealed in the NHC’s forecast uncertainty cone where
the diameter of the cone expands with forecast time, sometimes
affecting four or five separate U.S. states 120 h out from potential
landfall.
Given that recommended safe evacuation times for major
coastal communities are at least 30 h in advance of a hurricane’s
expected landfall (Lindell et al. 2007), the forecast errors highlight
the significant amount of uncertainty inherent in the hurricane fore-
cast that households use to make evacuation decisions during this
recommended safe evacuation timeframe. For example, the 36 h
forecast in 2005 had a track forecast error of 90 miles, and an in-
tensity error of 15 mph. Imagine a location that is within the aver-
age error forecast cone for the 36 h forecast, but it is located
80 miles east of the forecasted center of the storm, and is therefore
on the far eastern side of the cone for a hurricane moving south to
north. Also, let the 36 h forecast call for an intensity of 105 mph at
landfall, which is a category 2 (CAT 2) hurricane.
Assuming the storm actually stays within the cone 36 h later
(70% of the time it does not; Norcross 2006), it could potentially
make landfall 170 miles west of our imagined location (i.e.,
90 miles west of the forecasted storm center), effectively placing
our location out of harm’s way. On the other hand, even if the storm
does head directly toward our location it could make landfall any-
where from 90 mph to 120 mph, or from a CAT 1 up to a CAT 3
hurricane. As CAT 1 hurricanes are classified as causing minimal
damage, and CAT 3 hurricanes as causing extensive damage, the
difference in potential damage is significant. Undoubtedly, a house-
hold’s evacuation decision 36 h from landfall knowing with cer-
tainty that a CAT 3 storm will be tracking directly over it, or a
CAT 1 storm will be tracking 170 miles west of it, would be differ-
ent. And the fact that the storm is forecasted to potentially fall any-
where in between these two extremes leads to inevitable uncertainty
in regard to a household’s decision to evacuate during this 36 h time
period.
Heterogeneous Evacuation Behavior
Given the NHC track and intensity forecast uncertainty, and assum-
ing that households in the projected path of the storm are using this
information for their evacuation decision (“… almost all hurricane
forecast information the public receives is a repackaged form of
NHC data”;Regnier 2008), it is not surprising to see empirical evi-
dence suggesting heterogeneous evacuation behaviors among
households where some evacuate while others wait. For example,
Morrow and Gladwin (2005) found that for Hurricane Ivan in 2004
more than 68 h elapsed from the time the first person evacuated to
the time the last person evacuated in the Gulf region, which equates
to nearly 11 NHC forecast advisories spanning the course of three
days. Likewise, the cumulative evacuation timing curves that are
4
2/ NATURAL HAZARDS REVIEW © ASCE / MAY 2011
produced as a part of FEMA and the USACE (USACE 2006b)
poststorm assessments indicate heterogeneous evacuation behavior
of varying degrees through a variety of S-shaped evacuation
response curves. Fig. 1presents a version of the S-shaped cumu-
lative evacuation response curve from the 1995 Hurricane Opal
poststorm assessment where despite the majority of households
evacuating between 6 and 18 h before landfall, the first households
to evacuate left nearly 42 h before landfall.
There are a number of potential factors affecting the observed
variable timing of household evacuations. Fast, medium, and slow
evacuation response rates in relation to the issuance of an official
evacuation order have been observed (USACE 2006a). Thus, not
only do some households wait while others evacuate, but their
rate of waiting and evacuating vary as well, dependent upon
either household location (Lindell et al. 2005), or household type
(Fu and Wilmot 2004). For example, noncoastal locations typically
have a slower rate of evacuation compared to coastal locations, and
households with at least one household member working have a
slower rate of evacuation than those households that do not.
Dow and Cutter (pg. 15, 2002) state that the “majority of evacu-
ation trips begin during normal waking hours on the 2 days prior to
anticipated landfall”. Fu and Wilmot (2004), and Lindell et al.
(2005) also highlight heightened rates of evacuation occurring dur-
ing the daylight hours, and subsequent slowdowns during the night.
Lastly, Lindell et al. (2005) indicate that the steadier is the track of a
storm; the earlier will evacuations be induced.
However, these factors do not apply to all general evacuation
timing outcomes, nor are they able to sufficiently explain specific
evacuation timing outcomes. For example, the evacuation timing
graphs from Hurricane Ivan’s poststorm assessment illustrate in-
creased levels of evacuation beginning to occur during the night-
time hours, as opposed to slowdowns (Morrow and Gladwin 2005).
Dow and Cutter (2002) are at a loss to explain as to why for
Hurricane Floyd in 1999 48% of evacuees left between a
9:00 a.m. to 3:00 p.m. window, with so few leaving before and after
this period? Also, as no primary factor for the timing of evacuations
is identified from the empirical evidence, extrapolating which fac-
tor is predominantly driving evacuation timing for a specific out-
come is difficult when interactions between the various factors
occur, such as the issuance of evacuation orders during daylight
hours for coastal communities (Dow and Cutter 2002;Lindell et al.
2005). Hence, to better understand and explain evacuation timing
outcomes we build a realistic multiperiod dynamic model of evacu-
ation where households have the ability to evacuate, or to wait one
more time period for more information from a revised NHC
hurricane forecast advisory.
Multiperiod Model of Evacuation
Generically, the dynamic multiperiod model dictates that in each
NHC issued forecast advisory period households compare the costs
of evacuating versus the expected costs of not evacuating stemming
from the observed forecast information, and select the minimum
value of these two amounts. As a result, provided the costs of
evacuating represent the minimum value in a particular forecast ad-
visory period, households evacuate, otherwise the expected costs of
not evacuating are the minimum value and households wait one
more time period for a revised hurricane forecast. Or, in economic
terms, in each forecast advisory period households act rationally
and evacuate when the expected benefits of evacuating (i.e., the
avoided increased costs of evacuating next period and/or avoided
personal damage costs of not evacuating) are greater than the costs
of evacuating, otherwise it can be said that a positive option value to
waiting exists. Landry et al. (2007) have framed their static evacu-
ation return migration decision from a similar weighing of benefits
and costs.
More specifically, we can think of households potentially af-
fected by the storm as being placed into a discrete-time multiperiod
0
10
20
30
40
50
60
70
80
90
100
Oct 3 - 12A
(T*-5)
(T*-4)
(T*-3)
(T*-2)
Oct 4 - 12A
(T*-1)
T*
T
Date / Time
Cumulative % of Evacuees
Fig. 1. Adapted cumulative evacuation timing Hurricane Opal
5
NATURAL HAZARDS REVIEW © ASCE / MAY 2011 / 3
evacuation decision situation, where each discrete evacuation de-
cision time period is 6 h and is associated with a mutually exclusive
NHC forecast advisory. While the track and intensity of the hurri-
cane at landfall are unknown to households, we assume that the
timing of a storm’s landfall is known with certainty at time T,
and based upon the empirical evidence presented in available
evacuation response curves such as Fig. 1, that the last safe possible
time period for a household to evacuate is 6 h prior to T, denoted by
T. As the 120 h forecast is the maximum forecast time issued, let
n¼0;1;…;19 be the potential number of evacuation decision
time periods from Tover the five-day forecast period such that
we have ðT19Þ;ðT18Þ;…;ðT1Þ;Tpotential evacu-
ation decision time periods.
In each (Tn) evacuation decision time period, households
face the binary choice of either to evacuate and incur a known cost
of evacuation, or to wait one more time period for a revised hurri-
cane forecast. We further assume that if at any (Tn) period the
decision has been made to evacuate; this decision is not reversible
as evacuation is assumed to be immediate and evacuation costs are
sunk. Moreover, given the unknown track and intensity of the hur-
ricane at landfall this binary choice is predicated upon the track and
intensity forecast information contained in the (Tn) forecast
advisory. We construct a hurricane forecast risk index, denoted
θðTnÞ, that captures both the intensity and track forecast informa-
tion into a single value that households use for their binary evacu-
ation decision. In this way, our risk index variable is similar to the
2006 wind speed probability products issued by the NHC in that it
provides a single and less complicated source of information on the
probability of winds of a certain strength affecting a given location.
Of course, as presented in the section on hurricane evacuation de-
cisions as a dynamic process hurricane track and intensity forecasts
contain a significant amount of uncertainty, with the degree of un-
certainty decreasing as (Tn) approaches T. Accordingly, the
constructed risk index is a random variable which we assume fol-
lows a Markov process such that in the current period the proba-
bility that a particular realization of any of the possible jrisk index
values occurs, θj
ðTnÞ, depends only on the risk index values from
the previous period.
Explicitly then for forecast advisory periods n¼1;…;19, the
household evacuation decision in each (Tn) period given risk
index θðTnÞis either to evacuate immediately incurring cost of
evacuation cEVðTnÞ, or to wait one period for more informa-
tion from the expected updated forecast of the risk index condi-
tional upon the current period forecasted risk index, ½EðTnÞ
ðθðTnþ1ÞÞjθðTnÞ, and the possibility of evacuating during period
(Tnþ1) with associated costs of evacuation cEVðTnþ1Þ. For
n¼0, the household evacuation decision in period Tgiven the
risk index θTis either to evacuate immediately incurring cost
of evacuation cEVðTÞ, or to wait and simply ride out the storm at
Tincurring the expected costs of not evacuating cNEVðTÞwhich
are a function of the risk index at landfall and conditional upon
θT,ET½cðθTÞNEV jθT. For a more technical version of this dynamic
model as a formal optimal stopping problem, interested readers are
referred to Czajkowski (2007).
Model Inputs
To solve our multiperiod dynamic model of evacuation decision-
making, three main data inputs are needed: (1) for n¼0;
1;…;19, the possible jhurricane risk indexes, θðTnÞ, and their as-
sociated probability distributions; (2) for n¼0;1;…;19, the costs
of evacuation, cEVðTnÞ; and (3) for T, the expected costs of not evacu-
ating, cNEVðTÞ. The construction of these inputs is detailed below.
Hurricane Risk Index
We construct our hurricane risk indexes from actual historical storm
forecast advisory and realized landfall data stemming from 19
storms affecting 15 coastal locations in the Gulf of Mexico during
1992–2005. Specifically, we use Hurrevac to stipulate a 900 naut-
ical mile (NM) by 180 NM Gulf of Mexico region that includes the
15 coastal locations listed in Table 1. We select the 19 historical
storm tracks from 1992–2005 passing through this region listed
in Table 2. For the years 2004 and 2005, historical storm tracks
for storms achieving either tropical storm or hurricane strength
are utilized, while for 1992–2003 only those storms making land-
fall as a hurricane are utilized. Although we are only utilizing data
from 19 storms, a healthy mixture of storm intensity levels and
storm tracks are included.
To construct the hurricane forecast risk indexes, we need to
combine intensity and track forecast information provided in the
Table 1. 15 Coastal Gulf of Mexico Locations by County/Parish (Nearest
Major City)
# State Locations
1 TX Calhoun County (Port Lavaca/Port O Connor)
2 Brazoria County (Freeport)
3 Galveston County (Galveston)
4 Jefferson County (Port Arthur)
5 LA Iberia Parish (New Iberia)
6 St. Charles Parish (New Orleans)
7 Plaquemines Parish (Buras)
8 MS Harrison County (Gulfport)
9 AL Mobile County (Mobile)
10 FL Escambia County (Pensacola)
11 Bay County (Panama City)
12 Franklin County (Apalachicola)
13 Wakulla County (St. Marks)
14 Levy County (Cedar Key)
15 Hillsborough County (Tampa)
Table 2. 19 Identified Gulf of Mexico Storms
# Year Storm Landfall CAT Max CAT
1 2005 Arlene 0 0
2 Cindy 0 0
3 Dennis 3 4
4 Katrina 4 5
5 Rita 3 5
6 2004 Bonnie 0 0
7 Charley 4 4
8 Frances 0 4
9Ivan35
10 Matthew 0 0
11 2003 Claudette 1 1
12 2002 Lili 1 4
13 1998 Earl 1 2
14 Georges 2 4
15 1997 Danny 1 1
16 1995 Allison 1 1
17 Erin 1 1
18 Opal 3 4
19 1992 Andrew 3 5
6
4/ NATURAL HAZARDS REVIEW © ASCE / MAY 2011
forecast advisories for these storms into a single value. We assume
that households focus on the forecasted Saffir-Simpson Hurricane
Scale (SSHS) category level of the hurricane as opposed to the
storm’s specific wind speed when thinking about the hurricane’s
intensity, and therefore given the specific sustained wind speed val-
ues of the center of the storm provided in the forecast advisory, we
utilize the associated SSHS category (CAT) 1–5 values for the risk
index intensity forecast component. In addition, for storms with
forecasted wind speeds not at hurricane strength, we have an asso-
ciated 0 CAT value. Strike probabilities are the percentage chance
that the center of the storm will cross within 75 miles of a specified
location and are issued in conjunction with the forecast advisories
(although as of 2006 the strike probabilities are no longer issued,
replaced by the wind speed probability products). Similarly, we
assume that households focus on their specific location strike prob-
ability as opposed to the exact center of the storm latitude and
longitude coordinates provided in the forecast advisory when think-
ing about the hurricane’s track, and therefore utilize strike proba-
bilities for the risk index track forecast component. The 15 coastal
locations are given explicit strike probabilities in Hurrevac and the
NHC forecast archives. By using the forecasted intensity CAT lev-
els in conjunction with the location strike probabilities, for each of
the 19 storms we construct hurricane forecast risk indexes for each
of the 15 coastal locations along the Gulf of Mexico for each
(Tn) period with n¼0;1;…;19, and for the realized landfall
period T.
At landfall, our constructed risk index variable is straight-
forward to illustrate as θis no longer a forecast, but rather a reali-
zed value for any number of identified locations. For example, in
2005 Hurricane Dennis made landfall in Pensacola, Florida as a
CAT 3 hurricane with corresponding period Trealized CAT
intensity level ¼3 for all 15 locations, and strike probability ¼
100%for Pensacola, and ¼0%for all other 14 locations. There-
fore, θT¼3 for Pensacola—100% ðstrike probabilityÞ×3ðCATÞ.
However, θT¼0 for all other locations along the Gulf Coast that at
some point had the possibility of being struck by Hurricane Dennis,
such as Port Arthur, Texas-0% ðstrike probabilityÞ×3ðCATÞ.
Thus, at landfall θTnot only includes constructed values 1 through
5 corresponding to the five SSHS CAT levels, but also a constructed
value 0 that can either correspond to actual CAT 1–5 hurricanes that
do not make landfall at a particular location, or storms that make
landfall at a particular location but fall below the CAT 1 hurricane
designation.
Although similar to the θTdiscretization, the discretization of
θðTnÞfor n¼0;1;…;19 is more complex. First, for each specific
forecasted wind speed in each 12, 24, 36, 48, 72, 96, and 120 h
intensity forecast of each storm forecast advisory, an average range
of probable wind speed values is determined through the use of the
associated average intensity errors from 1996–2005. For example, a
126 mph (CAT 3) wind speed was forecasted in the 36 h intensity
forecast of the July 9, 2005, 5:00 am Hurricane Dennis advisory
(advisory #19 where mph converted from knots). Given the average
36 h intensity forecast errors, this specific forecasted CAT 3 wind
speed value from a probabilistic perspective is a CAT 3 65% of the
time, and a CAT 4 35% of the time. Next, these determined average
range of probable wind speeds for each 12, 24, 36, 48, 72, 96, and
120 h forecast are combined with the associated strike probability
for each location (total strike probability provided was utilized re-
gardless of the specific forecast hour). For example, the July 9,
2005, 5:00 a.m. advisory gave Pensacola, Florida a 32% total strike
probability. Combining this strike probability with the 36 h average
range of probable wind speeds results in a 68% chance that
Hurricane Dennis will miss Pensacola (the center of the storm will
not come within 75 mi), but if it strikes Pensacola, there is a 21%
chance it will strike as a CAT 3 hurricane, or an 11% chance it will
strike as a CAT 4. Lastly, the appropriate 12, 24, 36, 48, 72, 96, and
120 h forecast from each advisory is associated with the known
landfall timing of the storm, and is assigned the related multiperiod
model (Tn) period. For example, given landfall at July 10, 2005
at 3:00 p.m., the appropriate (T5) forecast from the July 9,
2005, 5:00 a.m. advisory is the 36 h forecast.
Combining all of this information, risk indexes per storm, loca-
tion, and period are calculated by multiplying the probability of a
strike by CAT level for the selected landfall forecast periods by the
0, 1, 2, 3, 4, and 5 SSHS CAT levels. Table 3provides an example
of the (Tn) period, n¼0;1;…;11, generated risk indexes for
Pensacola, for Hurricane Dennis. For example, the (T5) risk
index is estimated by ð0×:68Þþð1×:0Þþð2×:0Þþð3×:21Þ
þð4×:11Þþð5×:0Þ¼1:07. In this way we are weighing the
higher CAT storms more heavily, and we assume this is appropriate
given the exponential increase of damages along the SSHS. Table 3
illustrates that although our generated risk indexes are based upon
the SSHS, the uncertainty in the intensity and track information
does not allow for a direct comparison. For instance, the actual
(T5) forecast is predicting a CAT 3 hurricane (126 mph)
making landfall somewhere in the Gulf region. However, the con-
structed risk index by location which incorporates the uncertainty
of the track and the average intensity errors for a 36 h forecast,
equates to a 1.07 value for Pensacola at the (T5) evacuation
decision period. For Pensacola, the risk index evolves from 0.54
Table 3. Pensacola, Florida Risk Indexes for Hurricane Dennis
(Tn) Advisory Forecast (hr) Forecasted landfall date Wind (mph) 0 1 2 3 4 5 Risk index
(T11) 72 7/10 2 p.m. 121 82% 0% 5% 8% 5% 0% 0.54
(T10) 72 7/10 8 p.m. 115 81% 1% 6% 9% 3% 0% 0.51
(T9) 48 7/10 2 a.m. 132 82% 0% 0% 8% 10% 0% 0.64
(T8) 48 7/10 8 a.m. 132 80% 0% 0% 9% 11% 0% 0.71
(T7) 48 7/10 2 p.m. 126 78% 0% 1% 12% 8% 0% 0.73
(T6) 48 7/10 8 p.m. 115 75% 0% 9% 14% 2% 0% 0.67
(T5) 36 7/10 2 p.m. 126 68% 0% 0% 21% 11% 0% 1.07
(T4) 24 7/10 8 a.m. 109 64% 0% 20% 16% 0% 0% 0.88
(T3) 24 7/10 2 p.m. 126 65% 0% 0% 24% 11% 0% 1.16
(T2) 12 7/10 8 a.m. 138 62% 0% 0% 1% 37% 0% 1.51
(T1) 12 7/10 2 p.m. 144 50% 0% 0% 0% 50% 0% 2.00
T7/10 11 a.m. 138 6% 0% 0% 0% 94% 0% 3.76
Landfall (T) 7/10 3 p.m. 120
NATURAL HAZARDS REVIEW © ASCE / MAY 2011 / 5
at (T11) to 3.76 at (T), with landfall being 3.00 at T. For com-
parison purposes, the risk index for Iberia Parish in Louisiana
evolves from 0.18 at (T11), to 0.12 at (T1), and finally
0.0 at (T).
Because these constructed risk index values are random varia-
bles that we assume follow a Markov process, related Markov tran-
sition probability matrices are also constructed for each (Tn)
period. For example, given a Trisk index value of θTwithin
the range of [2–2.5], this value has the probability of transitioning
into a θTvalue at landfall of 0 ¼67%,1¼5%,2¼20%,3¼8%,
4¼0%,5¼0%. To construct these per period matrices, the 15
specific site hurricane forecast risk indexes are aggregated by year,
and the yearly probability matrices from 1992–2005 are then
aggregated into the final probability transition matrices per each
period. Because data limitations and maximum strike probabilities
are constrained by the NHC to be 60–80%, 35–50%, 20–25%,
13–18%, and 10% for the 12, 24, 36, 48, and 72 h forecasts respec-
tively, as nbecomes larger, more risk indexes with unavailable and/
or undefined values become more persistent in our probability tran-
sition matrices. For example, in period (T4) the maximum risk
index value is (1.0–1.5), while for period (T11) the maxi-
mum risk index value is (0.5–0.75). A complete listing and more
specific details on the construction of the hurricane risk indexes and
their associated probability transition matrices is provided in
Czajkowski (2007).
Costs of Evacuation
The costs of evacuation include evacuation travel and time costs,
direct costs incurred while away (food, lodging, entertainment),
and lost wages (Whitehead 2003). Although some of these imme-
diate evacuation costs, such as lost wages and portions of direct
costs, may actually decrease with waiting as tapproaches T,we
assume that certain costs, e.g., the crowdedness of the roads, dis-
tance needed to travel for adequate lodging, gasoline scarcity, etc.,
will increase rapidly enough so as to offset these declining costs.
Therefore, we assume that the longer a household waits to evacuate,
the more difficult, and hence more expensive the overall evacuation
will become such that the costs of evacuation are increasing as
(Tn) approaches T. Indeed, if cEVðTnÞare not increasing over
time, there would be an incentive for households to simply wait
until the last possible minute to evacuate. As we have already seen,
the empirical evidence from the S-shaped evacuation response
curves does not point to such a last minute evacuation result for
all households.
We use the CAT 3 Hurricane Bonnie evacuation costs produced
by Whitehead (2003) and the evacuation costs/data from Hurri-
canes Ivan (CAT 4), Charley (CAT 3), Frances (CAT 2), and Jeanne
(CAT 3) produced by USACE (2006b) to derive a household’s
average costs of evacuation given a CAT 3 storm. To incorporate
how these evacuation costs change over time we make the follow-
ing assumptions: (1) travel and travel time costs from (T19) to
Tand direct costs from only (T3) to Tto reflect the scarcity
of items such as food, lodging, and gasoline when waiting until
24 h prior to landfall, increase following the increases occurring
in the average cumulative timing of evacuations across a number
of existing studies, i.e., based upon how many other people are on
the road; (2) direct costs prior to (T3) are steady per day, but
decrease between days; and (3) lost income costs for (T19) to
Tare steady per day, but decrease between days. The specific
numeric results of the costs increases and decreases by cost cat-
egory following from our assumptions for periods (T11) to
Tare presented in Table 4. From Table 4, overall costs of evacu-
ation for a CAT 3 hurricane increase from $454 for evacuation at
period (T11) to $526 for evacuation at period T, but are not
increasing linearly. In fact, we actually see a decline in overall costs
between periods (T8) and (T7), and between periods
(T4) and (T3).
Thus far, the overall increasing costs of evacuation that we have
determined are for a CAT 3 hurricane. But we also further assume
that these costs would be less for CAT 1 and 2 and more for CAT 4
and 5 hurricanes as higher CAT levels induce more evacuees. We
use data from Lindell et al. (2002) on the predicted increases in the
number of cars and associated number of hours to evacuate along
the Texas Gulf Coast for CAT 1 to CAT 5 hurricanes to estimate the
varying levels of average CAT 1 to CAT 5 evacuation costs from
our derived average CAT 3 evacuation cost base. Fig. 2illustrates
the combined results of our cost of evacuation methodology. The
difference in evacuation costs are most significant between moving
from minor hurricanes (CAT 1 and 2) to a major hurricane, and
again the declines in overall costs for periods (T7) and
(T3) are clearly illustrated. Lastly, the derived costs of evacu-
ation for all periods need to be modified to coincide with the
defined risk index levels. A further detailed account of the deriva-
tion of the costs of evacuation is provided in Czajkowski
(2007).
Expected Costs of Not Evacuating (Personal Hurricane
Damage)
If a household chooses not to evacuate at T, and given that the
hurricane ultimately makes landfall at their location, they will
be forced to ride out the storm which has an associated probability
of being injured, or even killed. We use existing data from the
Multihazard Mitigation Council’s (MMC) study to assess future
savings from implementing mitigation activities related to natural
hazards (MMC 2005) to assign these probabilities for CAT 1 to
CAT 5 hurricanes, and estimate the expected costs of not evacuat-
ing from a hurricane (i.e., the value of avoided injury/death). As
part of the study, cost of injury data and rates of injury statistics
due to hurricanes were collected. We use both of these pieces of
information to generate our costs of not evacuating. The cost of
injuries used in their study (converted to 2004 dollars) are:
minor ¼$6;303, moderate ¼$51;471, serious ¼$189;076,
severe ¼$619;478, and critical ¼$2;521;008. Actual hurricane
injury rates for three hurricanes were provided in the report:
Andrew (CAT 3 in Louisiana) 0:2%; Opal (CAT 3) 0:0%;
and Isabel (CAT 2) 0:9%.
From the existing rates of injury, we take a conservative prob-
ability of injury for CAT 3 storms to be 0.45%. We further use the
Table 4. Increasing/Decreasing Evacuation Costs for CAT 3 Hurricane
CAT 3 (T11) (T10) (T9) (T8) (T7) (T6) (T5) (T4) (T3) (T2) (T1) T
Direct $193 $193 $193 $193 $128 $128 $128 $128 $153 $184 $215 $256
Travel costs $12 $16 $21 $25 $28 $33 $40 $50 $59 $71 $83 $99
Travel time $11 $15 $20 $23 $26 $31 $37 $46 $55 $67 $78 $92
Lost wages $238 $238 $238 $238 $159 $159 $159 $159 $79 $79 $79 $79
Total $454 $461 $472 $479 $341 $351 $364 $383 $347 $401 $456 $526
6/ NATURAL HAZARDS REVIEW © ASCE / MAY 2011
fact that damages along the SSHS are generally thought to follow
an exponential form to ascertain our probabilities of injury for
CAT 1 ¼0:050%,CAT2¼0:200%,CAT3¼0:45%,CAT
4¼0:85%, and CAT 5 ¼0:95%hurricanes. The generated prob-
abilities of injury from hurricanes are then multiplied by each of the
cost of injury values to obtain an expected cost of not evacuating by
CAT level. The expected costs of not evacuating are presented in
Table 5, with costs ranging from $1,694 for a CAT 1 to $32,182 for
aCAT5.
Concerning our derived expected costs of not evacuating: first,
we make no distinction between perceived costs and actual
expected costs of not evacuating, assuming that perceived costs
are unbiased estimates of actual costs across all evacuees. Secondly,
we have assumed that the probability of injury by each CAT level is
the same for all types of injuries. For example, in a CAT 3 storm we
have assumed the probability of incurring a minor injury is
0.45%, and that the probability of incurring a critical injury is also
0.45%. This is a limitation stemming from our available data.
Furthermore, we have not made any distinction for the probability
of injury depending upon location. Both of these issues should
be addressed in future research. Thirdly, we have assumed that
the probability of injury from a tropical storm is 0.0%, and there-
fore the expected cost is $0 despite the fact that tropical storms have
produced injuries and deaths. Lastly, we have abstracted away
from explicitly accounting for fatality probabilities and fatality
avoided damages for two reasons: (1) cost of injuries from the
critical severity level ¼$2:5 million and therefore easily fall
within traditional value of statistical life estimates of $1–10 million;
and (2) probability of death from a hurricane is low compared to the
injury probabilities above. For example, Hurricane Andrew (CAT
5) killed 14 people out of 1.9 million, or a fatality rate of.
000007368 (MMC 2005). As hurricanes average 20 deaths per year
in the U.S., this would be a relatively high rate.
$-
$100
$200
$300
$400
$500
$600
(T*-11)
(T*-10)
(T*-9)
(T*-8)
(T*-7)
(T*-6)
(T*-5)
(T*-4)
(T*-3)
(T*-2)
(T*-1)
T*
Evac Costs
CAT 1
CAT 2
CAT 3
CAT 4
CAT 5
Fig. 2. Derived CAT 1 to CAT 5 costs of evacuation
Table 5. Expected Costs of Not Evacuating
CAT Prob of injury Expected Cost ($)
0 0.000% $0
1 0.050% $1,694
2 0.200% $6,775
3 0.450% $15,244
4 0.850% $28,795
5 0.950% $32,182
NATURAL HAZARDS REVIEW © ASCE / MAY 2011 / 7
Solution and Results
General Solution and Results
By using the previous model, input of the multiperiod dynamic
model of evacuation is solved through backward recursion from
period T, the last safe possible time to evacuate, for a general
household in the Gulf of Mexico region (see Czajkowski 2007
for a more technical description of this solution). The intuition be-
hind the solution to the multiperiod dynamic model is that for cer-
tain forms of cEVðTnÞa unique cutoff for households exists where
waiting is optimal on one side of the forecast, and evacuating on the
other. Fig. 3illustrates the solved dynamic model evacuation cutoff
result for period Tonly with all the Trisk index values repre-
sented along the x-axis and the dollar value at time Tof having
a forecasted risk index of θðTÞrepresented along the y-axis. By
moving along the x-axis of Fig. 3we see that in period Tit is
optimal for an average household in our Gulf of Mexico region
to evacuate for storms that have a forecast risk index ≥1:0 because
to the right of this risk index value the expected costs of not evacu-
ating are greater than the costs of evacuating in period T. In other
words, an evacuation cutoff point exists in period Tfor forecast
risk index values ≥1:0, where waiting is optimal on one side of the
forecast, and evacuating on the other.The maximum risk index for
period Tis 4.0.
Fig. 4presents the specific θðTnÞevacuation cutoff results for
all (Tn) periods, n¼0;1;…;11, with time to landfall repre-
sented along the x-axis and the forecast risk index values repre-
sented along the y-axis. Note, no specific dollar value is
represented in this figure as was the case in Fig. 3. In addition
to the specific θðTnÞevacuation cutoff result for each (Tn)
periods, n¼0;1;…;11, the maximum risk index determined
for each of these periods is also presented. Two things should
be readily apparent from this figure: 1) each discrete-time period
does not have an associated evacuation region; and 2) the maximum
risk index value is increasing over time as the storm is making its
way closer to landfall moving from (T11) to (T) on the x-axis.
The specified evacuation region in Fig. 4is the area above
the evacuation cutoff line but below the maximum risk index line.
In forecast advisory periods (T2) to Tthe evacuation re-
gion for an average household at a representative Gulf of
Mexico location corresponds to: 1:0≤forecast risk index
<maximum risk Index value. For example, if the (T1)
forecast risk index ¼1:5 for one of the 15 specified Gulf of
Mexico locations, it is rational for an average household in this
location to be evacuating during this (T1) time period as
1:5>1:0. Likewise, if the (T1) forecast risk index ¼0:5 for
one of the 15 specified Gulf of Mexico locations, it is rational
for an average household in this location to wait one more time
period for the revised Thurricane forecast as 0:5<1:0. In
period (T3) the evacuation region corresponds to: 0:75 ≤
forecast risk index <maximum risk index ¼1:5. In forecast ad-
visory time periods (T11) to (T4) an evacuation cutoff line
does not exist, and therefore it is always optimal for an average
household to wait one more time period for a revised hurricane
forecast during these time periods, i.e., prior to 24 h out from
landfall. Again, given that recommended safe evacuation times
for major coastal communities are at least 30 h in advance of a
hurricane’s expected landfall, these optimal private household
evacuation results do not coincide well with the desired socially
optimal evacuation timing outcome. One potential way of thinking
about this issue is that early household evacuation provides external
benefits to other households in terms of reduced evacuation costs
later. However, these external benefits are likely ignored in the pri-
vate household evacuation decision. Similar to the undersupply of
other private goods that provide external benefits such as the level
of education, incentives need to be provided to households to in-
duce early evacuation. A preliminary assessment of potential incen-
tives is provided in the policy analysis section of this paper.
The maximum risk index value increases over time because the
track uncertainty is conveyed in the forecast advisories via the
strike probabilities. For example, given low strike probabilities
three days out from landfall, a CAT 5 storm that is just making
its way into the Gulf of Mexico at this time is not able to achieve
a risk index value >0:75. However, as the storm makes its way
0
250
500
750
1000
1250
1500
1750
2000
0
0 - 0.5 0.5 - 1 1 - 1.5 1.5 - 2 2 - 2.5 2.5 - 3 3 - 3.5 3.5 - 4
Risk Index
$
Evacuation Costs Expected Costs of not Evacuating
Fig. 3. Value at Tof risk index θT
8/ NATURAL HAZARDS REVIEW © ASCE / MAY 2011
closer to landfall and the strike probabilities become higher, the risk
index value increases over time. These rising risk index values help
provide the intuition behind the nonexistent evacuation region prior
to period (T3). One can simply compare the evacuation cost
results of Fig. 2versus the maximum risk index values in Fig. 4
to clearly see that evacuation costs are sufficiently high 72 h out
from landfall, whereas the risk values and consequently the ex-
pected costs of not evacuating are constrained during these earlier
time periods. Accordingly, households have an incentive to wait for
a revised hurricane forecast during earlier time periods as the model
solution formally shows.
Empirical Robustness
Although our results thus far have been general, i.e., for an average
household at an representative location in our defined Gulf of
Mexico region, we can use available evacuation timing graphs
and our per period, per location constructed risk indexes to evaluate
how well our model does in explaining actual evacuation timing
outcomes for particular storms.
Hurricane Opal: In October, 1995 Hurricane Opal made landfall
as a strong CAT 3 hurricane over Pensacola, Florida. Fig. 1illus-
trates the aggregated evacuation timing for evacuees from Alabama
(Baldwin and Mobile counties) and FL (Bay, Escambia, Okaloosa,
Santa Rosa, and Walton counties) from the 1995 postassessment
report (USACE 2006b). Fig. 1shows the 50th percentile evacuee
leaving during period (T1), and the slope of the curve steep-
ening considerably over this time period. Forecasts for Opal from
periods (T11) to (T3) called for a minor hurricane at land-
fall, but forecasts from (T2) to Tcalled for a major hurricane
at landfall.
Table 6presents the per period risk indexes by location, ranked
in descending order by T, for Hurricane Opal. Following from the
general results of our multiperiod model, evacuation is rational be-
ginning in period (T1) for average households in our defined
locations of Pensacola (Escambia County) and Panama City (Bay
County), Florida, and for Mobile (Mobile County), Alabama, as
risk indexes for these locations are all ≥1:0 and are highlighted
in Table 6. These results coincide well with the actual evacuation
timing as illustrated in Fig. 1, where the steepest slope of the evacu-
ation timing curve and the 50th percentile are occurring in period
(T1), and overall evacuation is really beginning in earnest dur-
ing this timeframe. In this case, the results of our multiperiod model
offer an explanation for the relatively late (12 h prior to landfall for
a major hurricane) evacuation response.
Similar exercises are performed for Hurricanes Ivan, Charley
and Lili in Czajkowski (2007) with comparable positive results
for Ivan and Charley and mixed results for Lili. In all 4 cases,
evacuation is predicted only in those locations where actual evacu-
ation occurred according to the poststorm assessment survey data.
For example, Hurricane Charley made landfall in Southwest
Florida, and our model predicts evacuations for locations close
to the eventual landfall such as Tampa, while not predicting evac-
uations for locations not in close proximity such as Mobile,
Alabama. In 3 of the 4 cases, our model correctly predicts evacu-
ation for an average household, which correspond to the 50th per-
centile on the evacuation timing graphs. However, the 10–50% of
cumulative evacuations occurring between periods (T8) and
(T4) for Hurricanes Lili and Ivan are not predicted from our
model, at least not for an average household.
Range of Solutions by Various Household Types
In addition to evaluating our model’s robustness against actual
evacuation timing outcomes for an average household, we can also
evaluate it through expected evacuation outcomes by various
household types. For example, let us assume two household types:
high-damage (e.g., coastal location or mobile home) versus low
damage (e.g., inland location or nonmobile home), where high
(low) damage households have a greater (lower) probability of
being injured than the probabilities from the model inputs section
provided in Table 5. Intuitively, compared to an average damage
household type, high (low) damage household types should evacu-
ate earlier (later) in general and also be more (less) willing to evacu-
ate for lower (higher) risk index storms.
0.00
1.00
2.00
3.00
4.00
5.00
(T*-11) (T*-10) (T*-9) (T*-8) (T*-7) (T*-6) (T*-5) (T*-4) (T*-3) (T*-2) (T*-1) T*
Time to Landfall
Forecast Risk Index
Evac Cutoff Max Risk Index
Evacuation
Region
Waiting Region
3 days out 2 days out 1 day out
Fig. 4. Average household optimal evacuation results
NATURAL HAZARDS REVIEW © ASCE / MAY 2011 / 9
Table 7summarizes the model results for high and low damage
household types along with model result summaries for various
other household types such as high cost versus low cost households
and tourists. For example, we flex the dynamic model to incorpo-
rate high damage household types by increasing the probability of
injury by five times that of the average household. Consequently,
the costs of not evacuating increase for high damage household
types and the evacuation region generally expands for lower risk
indexes in periods (T3) to T. Furthermore, if the probability
of injury has increased significantly, one would also expect the
number of evacuees to increase leading to higher rates of evacu-
ation costs increases for all periods compared to those used for
an average household. When the higher probability of injury is
combined with evacuation cost increases two times that of the gen-
eral model, not only does the evacuation region expand for periods
(T3) to T, but earlier evacuation is induced for periods
(T7) to (T5). From Table 7we see that overall, results from
the multiperiod model do a good job of predicting expected evacu-
ation timing outcomes for various household types. What’s more,
these predicted results have the potential to explain the 10–50% of
cumulative evacuations occurring between periods (T8) and
(T4) for Hurricanes Lili and Ivan that our average household
results from the general multiperiod model could not. However, we
do note the need for caution in by using predictive validity as a
criterion for assessing a model’s adequacy.
Policy Implications
The dynamic modeling framework is most relevant in beginning to
understand the implications of policies that plausibly affect the tim-
ing of household evacuations. In this section we provide a prelimi-
nary assessment of a number of potential hurricane policies meant
to affect the timing of evacuation.
Table 6. Hurricane Opal Risk Indexes by Location
Evacuate for risk index: Wait Wait Wait Wait Wait Wait Wait Wait >:75 >1>1>1
State Locations T11 T10 T9T8T7T6T5T4T3T2T1TLandfall
FL Pensacola 0.04 0.04 0.03 0.07 0.22 0.30 0.32 0.40 0.54 0.67 1.40 2.80 3.00
FL Panama City 0.03 0.04 0.02 0.05 0.22 0.26 0.28 0.36 0.56 0.70 1.08 1.84 0.00
AL Mobile 0.04 0.04 0.04 0.07 0.24 0.30 0.32 0.38 0.46 0.58 1.15 1.76 0.00
FL Apalachicola 0.03 0.04 0.02 0.05 0.20 0.26 0.26 0.34 0.54 0.67 0.83 0.92 0.00
MS Gulfport 0.05 0.05 0.04 0.09 0.24 0.30 0.34 0.38 0.39 0.49 0.86 0.60 0.00
FL St. Marks 0.02 0.03 0.00 0.04 0.18 0.24 0.24 0.30 0.49 0.61 0.65 0.60 0.00
LA Buras 0.06 0.07 0.06 0.12 0.28 0.34 0.38 0.42 0.41 0.52 0.76 0.28 0.00
LA New Orleans 0.06 0.06 0.06 0.12 0.26 0.30 0.34 0.34 0.29 0.36 0.29 0.00 0.00
FL Cedar Key 0.02 0.03 0.00 0.04 0.16 0.20 0.20 0.24 0.37 0.46 0.22 0.00 0.00
FL Tampa 0.02 0.03 0.00 0.04 0.16 0.18 0.16 0.18 0.24 0.30 0.07 0.00 0.00
LA New Iberia 0.06 0.06 0.07 0.12 0.24 0.24 0.30 0.22 0.15 0.18 0.00 0.00 0.00
TX Port Arthur 0.06 0.05 0.07 0.12 0.18 0.12 0.16 0.08 0.05 0.06 0.00 0.00 0.00
TX Galveston 0.07 0.06 0.09 0.14 0.16 0.08 0.12 0.04 0.00 0.00 0.00 0.00 0.00
TX Freeport 0.07 0.06 0.10 0.14 0.14 0.06 0.10 0.00 0.00 0.00 0.00 0.00 0.00
TX Port Lavaca 0.07 0.06 0.10 0.14 0.10 0.00 0.04 0.00 0.00 0.00 0.00 0.00 0.00
Table 7. Model Results by Various Household Types
Household type Expected result Model input change Model result
High damage Earlier evacuation and more willing to
evacuate for lower risk indexes
Probability of injury increases by 5 times Evacuation region generally expands for lower
risk indexes in periods (T3) to T
Probability of injury increase by 5 times
and 2 times rate of evacuation cost
increase
Evacuation region expand for periods (T3)
to T, and earlier evacuation is induced for
periods (T7) to (T5)
Low damage Later evacuation and less willing to
evacuate for higher risk indexes
Probability of injury decreases by half Evacuation region generally contracts for
higher risk indexes in periods (T3) to T
Probability of injury decrease by half and
half the rate of evacuation cost increase
Not much change
High evacuation
cost
Later evacuation and less willing to
evacuate for higher risk indexes
Overall cost of evacuation increase
by 2 times
Evacuation region generally contracts for
higher risk indexes in periods (T3) to T
Overall cost of evacuation increase
by 2 times and half the rate of evacuation
cost increases
Evacuation region contracts even further for
higher risk indexes in periods (T3) to T
Low evacuation
cost
Earlier evacuation and more willing to
evacuate for lower risk indexes
Overall cost of evacuation decrease by 1=2 Evacuation region expands only slightly for
lower risk indexes in period T
Overall cost of evacuation decrease by 1=2
and two times the rate of evacuation cost
increases
Evacuation region expand for periods (T3)
to T, and earlier evacuation is induced for
periods (T7) to (T5)
7
10 / NATURAL HAZARDS REVIEW © ASCE / MAY 2011
Overall Evacuation Cost Reduction
For policy makers and emergency managers interested in having
households evacuate as early as possible, the costs of evacuation
are a key constraint. Given that the costs of evacuation consist
of the various components of direct, travel-related, and lost income
costs, a variety of policy initiatives may be available to reduce
costs. We use the dynamic model to test the effect of reducing
the overall costs of evacuation equally across all components by
25, 50, and 80%, while holding all other variables constant. The
results indicate that large cost reductions are needed, as much as
80% of the original, to induce evacuation for lower risk indexes,
and even these significant cost reductions do not induce earlier
evacuation for periods prior to (T3). This suggests that a policy
aimed at simply reducing the overall costs of evacuation does not
induce early evacuation. Potentially then, a more targeted evacu-
ation cost reduction, or a nonevacuation cost-based policy such
as an improved forecast, may be a more appropriate strategy to
achieve earlier evacuations by the average household.
Targeted Evacuation Cost Reduction
The costs of lost income are one component of evacuation costs that
potentially can be targeted by policy makers as not only are these
costs the largest component of our specified evacuation costs, but
they also delineate two separate household types with someone in
the household having to work in hourly versus salaried worker
household types. We assume that salaried workers have more flex-
ibility in their decision to evacuate with any missed days of work
not equating to lost income, which we assume not be true for hourly
workers. Fig. 5illustrates the results from our multiperiod model
with the costs of lost income eliminated, demonstrating a divergent
salaried versus hourly worker outcome. As the evacuation region
expands significantly from one to two days out from landfall, we
see that when the costs of lost income are eliminated from the
evacuation decision it easier to evacuate earlier.
What other possible targeted policies might make evacuation
decisions more equitable such as a focus on the reduction of direct
costs? For example, assume a policy that focuses on reducing much
of the direct costs of evacuation through the use of improved shel-
ters that provide meals, showers, etc. Households that use the
improved shelters (which we assume to typically be hourly worker
household types) have the possibility of having much of their direct
costs eliminated. However, when direct costs are completely elim-
inated from the multiperiod model, little earlier evacuation is
induced. In this case, a policy that gives hourly workers more
evacuation options once they have evacuated is not effective in min-
imizing the divergent salaried versus hourly worker outcome. In
order for the divergence to be addressed, policies need to be
directed at making it easier for hourly workers to leave, such as
a policy that provides incentives for employers to pay hourly work-
ers for lost work time due to hurricane evacuations.
Cost Profile
Other more targeted policies intended to induce earlier evacuation
could focus on reducing the rates at which direct and travel-related
costs increase over time such as the use of contraflow or the in-
creased availability of shelters. Fig. 6illustrates the affect on evacu-
ation timing if these types of policies are implemented and our
assumed rates of per period cost increases from (T11) to T
are decreased by half. Decreasing the rate at which the costs of
evacuation increase over time leads to a contraction of the evacu-
ation region, and also to no earlier evacuation being induced prior
to period (T3). This outcome helps to explain the empirical re-
sult that Morrow and Gladwin (2005) found for Hurricane Ivan
where when contraflow was implemented, a quarter of respondents
indicated this made them less likely to evacuate. Importantly, these
results also show that when the ability to wait is a part of house-
hold’s decision to evacuate, timing results may run opposite of the
intended policy goals.
Conversely, when the rates of travel and direct costs increase
over time, earlier evacuation is induced. Fig. 6also illustrates this
result assuming the rates have increased by two times our original
assumptions, with earlier evacuation shown for periods (T7) to
0.00
1.00
2.00
3.00
4.00
5.00
(T*-11) (T*-10) (T*-9) (T*-8) (T*-7) (T*-6) (T*-5) (T*-4) (T*-3) (T*-2) (T*-1) T*
Time to Landfall
Forecast Risk Index
Costs with Income Costs without Income Max Risk Index
Expanded
Evacuation
Region
Fig. 5. Optimal evacuation region excluding lost income costs
NATURAL HAZARDS REVIEW © ASCE / MAY 2011 / 11
(T5). Paradoxically, this result also coincides with another
Hurricane Ivan finding discussed by Morrow and Gladwin
(2005) where the implementation of the contraflow actually caused
additional problems in traffic flow (which can be construed as a rate
increase), and 60% of those evacuees that used the contraflow route
indicated that they would leave earlier next time.
Value of an Improved Forecast
Recall that the evacuation response for Hurricane Opal was rela-
tively late with the vast majority of evacuations occurring within
12 h of the actual hurricane landfall. From Table 6, there are
two ways that an earlier evacuation say in period (T3) could
have occurred for the eventual landfall locations of Pensacola or
Panama City: (1) lowering of the general (T3) evacuation cutoff
point from risk index values ≥0:75 to risk index values ≥0:50, the
current Pensacola/Panama City risk index values; or (2) increasing
the location specific (T3) risk index values achieved for
Pensacola and Panama City to ≥0:75, the current evacuation cutoff
point.
The previous nontargeted cost of evacuation analysis suggests
that to lower the general (T3) cutoff point to risk index values
≥0:50, the overall evacuation costs need to be reduced by 80%, or
a cost of $307 per household. With approximately 50,000 house-
holds in these two locations, total cost reductions necessary to
induce an earlier evacuation 24 h out from landfall therefore equate
to approximately $15 million. The NHC strike probabilities for
Panama City and Pensacola in this period were 22% and 23%
respectively. If these strike probabilities had been increased to
31%, risk index values would have been high enough for it to
be rational for household to evacuate during period (T3),
i.e., ≥0:75. Therefore, in the case of Hurricane Opal, the difference
between the cost necessary to improve the strike probabilities from
22% to 31% 24 h before landfall, and the $15 million cost of evacu-
ation reduction is the value of an improved forecast that induces an
earlier evacuation 24 h out from landfall.
Conclusions
This paper addresses a limitation to existing household hurricane
evacuation modeling aspects, and hence an understanding of the
household hurricane evacuation, by developing a dynamic model
of hurricane evacuation behavior. In every potential evacuation
time period prior to the actual hurricane landfall within the dynamic
model, a household’s optimal choice is to either evacuate, or to wait
one more time period for a revised NHC hurricane forecast. The
dynamic framework reflects a realistic multiperiod setup incorpo-
rating existing forecast and evacuation cost data to explain actual
evacuation behavior for our designated Gulf of Mexico region.
Despite a number of assumptions made in developing the model
along with data limitations, the evacuation timing results from
our general model do a relatively good job of understanding and
explaining actual evacuation timing outcomes by location from
specific hurricanes, and expected evacuation timing outcomes
for various household types. Consequently and most significantly,
the dynamic framework is used to explore a number of relevant
policy questions that plausibly affect the timing of household evac-
uations, sometimes providing the rationalization for seemingly
counterintuitive poststorm assessment evacuation results. For ex-
ample, would building more and better shelters induce earlier
evacuation? Or, why does implementing contraflow actually cause
some households to be less likely to evacuate? Thus, this analysis
has begun to address the need for modeling hurricane evacuation
behavioral responses in more precise and comprehensive ways, lay-
ing a foundation for continued development in this regard.
Acknowledgments
I thank Peter Thompson, Stephen Leatherman, Daniel Sutter, and
Richard Woodward for their comments and suggestions. I am also
grateful for helpful comments from numerous conference partici-
pants at Texas A&M Agricultural Economics Department, South-
ern Economic Association 2007 Annual Meeting, Hazards and
0.00
1.00
2.00
3.00
4.00
5.00
(T*-11) (T*-10) (T*-9) (T*-8) (T*-7) (T*-6) (T*-5) (T*-4) (T*-3) (T*-2) (T*-1) T*
Time to Landfall
Forecast Risk Index
Original Costs Cost Rates (50%) Max Risk Index Cost Rates (200%)
Waiting Region
Evacuation
Region
Fig. 6. Optimal evacuation region for different rates of cost increases
12 / NATURAL HAZARDS REVIEW © ASCE / MAY 2011
Disaster 2007 Researchers Meeting, the 2007 Florida Governor’s
Hurricane Conference, and the 2006 and 2007 Florida Hurricane
Alliance Workshop. Lastly, I want to acknowledge financial sup-
port provided by Florida International University and NOAA.
References
Blake, E., Rappaport, E., and Landsea, C. (2007). “The deadliest, costliest,
and most intense United States tropical cyclones from 1851 to 2006
(and other frequently requested hurricane facts).”NOAA Tech-
nical Memorandum NWS TPC-5, National Weather Service National
Hurricane Center, Miami, Florida.
Centrec Consulting Group, LLC. (2007). “An investigation of the economic
and social value of selected NOAA data and products for geostationary
operational environmental satellites (goes).”Technical Rep. Prepared
for NOAA’s National Climatic Data Center, Centrec Consulting Group,
Savoy, IL.
Czajkowski, J. (2007). “Is it time to go yet? Dynamically modeling hurri-
cane evacuation decisions.”Florida International University Inter-
national Hurricane Research Center Tech. Rep.,〈http://www.ihc.fiu
.edu/technical_reports/index.htm〉.
Dash, N., and Gladwin, H. (2007). “Evacuation decision making and
behavioral responses:individual and household.”Nat. Hazards Rev.,
8(3), 69–77.
Dixit, A., and Pindyck, R. (1994). Investment under uncertainty, Princeton
University Press, Princeton, NJ.
Dow, K., and Cutter, S. (2002). “Emerging hurricane evacuation issues:
Hurricane Floyd and South Carolina.”Nat. Hazards Rev., 3(1), 12–18.
Fu, H., and Wilmot, C. (2004). “A sequential logit dynamic travel demand
model for hurricane evacuation.”Paper Prepared for the TRB 2004
Annual Meeting, January 2004, Louisiana Transportation Research
Center, Baton Rouge, LA.
Gladwin, H., Lazo, J., Morrow, B., Peacock, W., and Willoughby, H.
(2007). “Social science research needs for the hurricane forecast and
warning system.”Nat. Hazards Rev., 8(3), 87–95.
Hurrevac [computer software]. Sea Island Software, Richmond, VA.〈http://
www.hurrevac.com/〉.
Kunkel, K., Pielke, R., Changnon, S. (1999). “Temporal fluctuations in
weather and climate extremes that cause economic and human health
impacts: a review.”Bull. Am. Meteorol. Soc., 80(6), 1077–1098.
Landry, C., Bin, O., Hindsley, P., Whitehead, J., and Wilson, K. (2007).
“Going home: Evacuation-migration decisions of Hurricane Katrina
survivors.”South. Econ. J., 74(2), 326–343.
Lindell, M., Lu, J., and Prater, C. (2005). “Household decision making and
evacuation in response to Hurricane Lili.”Nat. Hazards Rev., 6(4),
171–179.
Lindell, M., Prater, C., and Peacock, W. (2007). “Organizational commu-
nication and decision making in hurricane emergencies.”Nat. Hazards
Rev.,8,50.
Lindell, M., Prater, C., and Wu, J. (2002). Hurricane evacuation time
estimates for the Texas Gulf Coast, Texas A&M Univ., College Station,
TX.
Multihazard Mitigation Council (MMC). (2005). “Natural hazard
mitigation saves: an independent study to assess the future savings
from mitigation activities.”National Institute of Building Sciences,
Washington D.C.
Morrow, B., and Gladwin, H. (2005). “Hurricane Ivan behavioral analysis.”
Federal Emergency Management Agency and U.S. Army Corps of
Engineers.
National Hurricane Center (NHC). http://www.nhc.noaa.gov/index.shtml.
Norcross, B. (2006). Hurricane almanac 2006, St. Martin’s Griffin, New
York.
Rappaport, E. (2000). “Loss of life in the united states associated with
recent atlantic tropical cyclones.”Bull. Am. Meteorol. Soc., 81(9),
2065–2073.
Regnier, E. (2008). “Public evacuation decisions and hurricane track
uncertainty.”Manage. Sci., 54(1), 16–28.
Regnier, E., and Harr, P. (2006). “A dynamic decision model applied to
hurricane landfall.”Weather and Forecasting, 21, 764–780.
Sadowski, N., and Sutter, D. (2005). “Hurricane fatalities and hurricane
damages: Are safer hurricanes more damaging?.”South. Econ. J.,72
(2), 422–432.
USACE. (2006a). http://chps.sam.usace.army.mil/USHESdata/Louisiana/
SWLA/cd/html/chapters/chapter04/chapter_04_results.html.
USACE. (2006b). http://chps.sam.usace.army.mil/ushesdata/
Post_Storm_Assessment_page.htm.
Whitehead, J. (2003). “One million dollars per mile? The opportunity
cost of hurricane evacuation.”Ocean and coastal management, 46,
1069–1083.
Willoughby, H., Rappaport, E., and Marks, F. (2007). “Hurricane fore-
casting: The state of the art.”Nat. Hazards Rev., 8(3), 45–49.
8
9
10
11
12
13
14
15
16
NATURAL HAZARDS REVIEW © ASCE / MAY 2011 / 13
