Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

DE GRUYTER Journal of Homeland Security and Emergency Management.

2019; 20180071

Abhishek Damera1 / Hemant Gehlot2 / Satish Ukkusuri2 / Pamela Murray-Tuite3 / Yue Ge4 /
Seungyoon Lee5

Estimating the Sequencing of Evacuation


Destination and Accommodation Type in
Hurricanes
1 Indian Institute of Technology Madras, Department of Civil Engineering, Chennai, India
2 Purdue University, Lyles School of Civil Engineering, West Lafayette, IN, USA, E-mail: sukkusur@purdue.edu.

https://orcid.org/0000-0001-8754-9925.
3 Clemson University, Glenn Department of Civil Engineering, Clemson, SC, USA
4 University of Central Florida, School of Public Administration, Orlando, FL, USA
5 Purdue University, Brian Lamb School of Communication, West Lafayette, IN, USA

Abstract:
Hurricanes are one of the most dangerous catastrophes faced by the USA. The associated life losses can be
reduced by proper planning and estimation of evacuation demand by emergency planners. Traditional evac-
uation demand estimation involves a sequential process of estimating various decisions such as whether to
evacuate or stay, evacuation destination, and accommodation type. The understanding of this sequence is not
complete nor restricted to strict sequential ordering. For instance, it is not clear whether the evacuation desti-
nation decision is made before the accommodation type decision, or the accommodation type decision is made
first or both are simultaneously made. In this paper, we develop a nested logit model to predict the relative
ordering of evacuation destination and accommodation type that considers both sequential and simultaneous
decision making. Household survey data from Hurricane Matthew is used for computing empirical results.
Empirical results underscore the importance of developing a nested structure among various outcomes. In ad-
dition to variables related to risk perception and household characteristics, it is found that social networks also
affect this decision-making process.
Keywords: accommodation type, hurricane evacuation destination, nested logit model, relative ordering, se-
quencing, social networks
DOI: 10.1515/jhsem-2018-0071
Automatically generated rough PDF by ProofCheck from River Valley Technologies Ltd

Introduction
Hurricanes are among the most costly and dangerous natural disasters that affect the United States. In 2016,
Hurricane Matthew was responsible for causing 585 deaths and an estimated total damage of around $10 bil-
lion. Also, more than 3 million people evacuated from their homes during this time (Stewart 2017). A house-
hold’s evacuation process is the result of multiple decisions taken to travel from home to a safe area. These
decisions include time of departure, evacuation destination, accommodation type, route choice, and others.
Much of the previous evacuation literature either considers only a single decision or considers them sequen-
tially. For instance, one of the sequences that is widely used to model the household evacuation decisions (listed
in increasing order of modelling frequency) is (1) decision to evacuate, (2) accommodation type, (3) destina-
tion choice, (4) departure time, (5) mode choice and (6) route choice (Ukkusuri et al. 2017; Gehlot et al. 2018b).
However, the sequential approach is adopted for modelling convenience rather than clearly understanding the
actual sequence in which the decisions are made. Whether decisions are made simultaneously (i.e. jointly) or
sequentially may vary across households and depend on various factors like socio-demographic factors and
social network characteristics (Tierney, Lindell, and Perry 2002; Sadri, Ukkusuri, and Gladwin 2017). This pa-
per investigates two key questions related to sequential and joint decision making for evacuations. First, if both
sequential and simultaneous decisions can be made then what is the probability of making sequential ver-
sus simultaneous decisions? In the literature, sequential and simultaneous decisions are modeled differently.
For instance, correlation parameters are used in simultaneous decision modeling for modeling relationships

Satish Ukkusuri is the corresponding author.


© 2019 Walter de Gruyter GmbH, Berlin/Boston.

Brought to you by | Göteborg University - University of Gothenburg


Authenticated 1
Download Date | 11/19/19 9:01 AM
Damera et al. DE GRUYTER

between various decisions. Different decision-making procedures (sequential decision-making and simultane-
ous decision-making) could generate different values of evacuation demand distribution across various desti-
nations, routes, modes etc. Evacuation demand refers to the number of the people from a region who travel for
the purpose of evacuation due to a disaster in a given time period. Another question that arises is whether there
is a specific relative ordering of these decisions that modelers should follow to more closely capture reality and
potential dependencies among the decisions? It is possible that there is no single sequence that applies to the
entire population and that the decision making context is complex. In that case, the estimation of evacuation
demand will involve the use of both the sequential and joint modeling techniques. Therefore, it is imperative
to understand the factors that influence this sequential decision making process.
This paper begins the exploration of joint and sequential decision making for two evacuation decisions:
accommodation type and destination. Accommodation type (denoted as Acco) indicates the type of facility in
which evacuees stay while away from home. Evacuation destination (denoted as Evac) indicates the physical
location where evacuees terminate their evacuation. These two decisions were the most commonly selected
joint decisions, as revealed by our post-Hurricane Matthew household survey that specifically inquired about
the order in which decisions are made. These two decisions also form important components of the decision
making process during hurricanes.
The objective of this paper is to determine the factors that are associated with the relative ordering of these
two decisions (Evac and Acco). Barrett, Ran, and Pillai (2000) defined evacuation destination as the location to
which the evacuee is travelling to seek safety. Barrett, Ran, and Pillai (2000) and Southworth (1991) provided as-
sumptions for modelling primarily what Lindell and Prater (2007) refer to as the proximate destination, which
is the point in the transportation network where the evacuee exits the risk area. The ultimate destination refers
to both the town and/or city and the type of accommodation where the evacuees will stay until they can re-
turn to their homes (Lindell and Prater 2007). The second part of this ultimate destination is accommodation
type (e.g. shelter, hotel, relatives’ home) that is chosen. In the literature, accommodation type has been largely
modelled before evacuation destination (Ukkusuri et al. 2017; Gehlot et al. 2018b) but this sequence needs to
be validated. It is possible that the estimated evacuation demand depends on the relative order of modelling
various decisions. So, it is important to study the factors associated with the relative ordering of these two
decisions.
Social network variables are another important influence on evacuation decision making. Social ties of an
individual have been estimated as important factors for influencing different decisions during hurricanes (Sadri,
Ukkusuri, and Gladwin 2017; Sadri et al. 2018). Hasan and Ukkusuri (2011) investigated the role that social
networks play in the propagation of warning information. They found that information spreads faster with
more intercommunity connections. First the warning information is received by the people and then they spread
the information to others (Perry, Lindell, and Greene 1981). But at the same time, the role of social ties on
evacuation destination and accommodation type decisions has not been explored, let alone the importance of
social networks in influencing the relative ordering of these decisions. In this paper, we fill this gap by analyzing
Automatically generated rough PDF by ProofCheck from River Valley Technologies Ltd

the role of social ties on the relative ordering of the two decisions.
This study is organized in the following manner. The next section reviews the past studies related to evacu-
ation destination and accommodation type and outlines the motivation for this study. After this, we discuss the
data source and sample used in empirical analysis. Then, we advance the statistical framework for the nested
model. The next section presents empirical results. The final section summarizes important findings from the
research.

Background and Motivation


Most prior studies of evacuation-related decisions, such as the choice to evacuate/stay, departure timing, ac-
commodations, and destination, treat them as separate decisions. This approach simplifies the modeling and
data collection. A few exceptions are Fu and Wilmot (2004, 2006) who jointly examined the choice to evacu-
ate/stay and the departure time for evacuees, Yang et al. (2016) who examined the evacuate/stay decision with
the accommodation choice, Golshani et al. (2018) who jointly modeled the departure time and accommodation
choice, and Gehlot, Sadri, and Ukkusuri (2018a) who jointly modeled departure time and travel time. While
several approaches to modeling joint decision exist, few (if any) explore the factors that determine whether
decisions are made jointly or separately. In this study, the authors explore the factors that influence whether
accommodations and the location of those accommodations (destination) are made jointly or separately, and
in what sequence.
With few studies available to inform whether decisions are joint or separate, this literature review examines
the factors associated with the selection of accommodations and destinations. These factors may or may not
affect the sequence of these decisions, and are examined in the data exploration and modeling. The types of

Brought to you by | Göteborg University - University of Gothenburg


2 Authenticated
Download Date | 11/19/19 9:01 AM
DE GRUYTER Damera et al.

accommodations selected by evacuees are often considered in three primary groups, i.e. peers’ homes (fami-
ly/friends), commercial establishments (hotels/motels), and public shelters, with general preference in that or-
der (Prater, Wenger, and Grady 2000; Lindell, Lu, and Prater 2005; Lindell, Kang, and Prater 2011; Mesa-Arango
et al. 2012; Wu, Lindell, and Prater 2012; Yin et al. 2014). The relatively few studies investigating characteristics
of public shelter users found that these evacuees typically have low incomes, be ethnic minorities, and depart
later (Mileti, Sorensen, and O’Brien 1992; Baker 2000). When examining accommodation choice in general, the
significance of departure time varies across studies (Lindell et al. 2018), with non-significance in studies of
Hurricane Lili (Lindell, Kang, and Prater 2011) and Katrina and Rita (Wu, Lindell, and Prater 2012), and later
departures associated with public shelters and hotels/motels (Yin et al. 2014) for a hypothetical hurricane in
the Miami area. Those who travel greater distances are less likely to stay in public shelters (Mesa-Arango et al.
2012). Larger households are more likely to stay in public shelters or commercial facilities. Finally, higher levels
of income are associated with those who stay with peers (Wu et al. 2013) and those with lower incomes are less
likely to stay in hotels/motels (Mesa-Arango et al. 2012).
Although most prior studies of evacuation-related decisions have treated accommodation and destination
as separate decisions, Cheng, Wilmot, and Baker (2008) developed separate destination models for two types
of accommodations – peer’s homes and commercial establishments. For both models, a socio-demographic char-
acterization (Lindell et al. 2018) of the destination was included. Findings show that increased distance from the
evacuees’ home as well as higher risk at that destination decreased the likelihood of that destination zone being
selected. For the peers’ homes model, the zone being in a metropolitan area and having a greater population in-
creased the likelihood of selecting that zone. For the hotel/motel model, a greater number of these commercial
facilities and closer proximity to the interstate increased the likelihood of selecting that zone (Cheng, Wilmot,
and Baker 2008). Extensions to Cheng et al.’s work also treated destinations as zones (Cheng and Wilmot 2009;
Cheng, Wilmot, and Baker 2011). However, other studies more simply characterized destinations as within or
outside evacuees’ home counties (Prater, Wenger, and Grady 2000; Dow and Cutter 2002; Murray-Tuite et al.
2012) or in or out of state (Dow and Cutter 2002). Selected options varied by the size of the counties and states,
and presumably with the availability of accommodations and the risk levels at those destinations. In the current
study, destinations were recorded with city, state, and zip code, which could be recoded into either approach
found in the prior literature. However, the focus of this paper is on identifying the factors associated with the
sequence or joint decision of accommodations and destination. Also, as mentioned above, the influence of so-
cial ties on both the separate decision making and relative ordering of the two decisions has not been studied.
Therefore, this study makes the following contributions:

– Develops a model that predicts the relative ordering of evacuation destination and accommodation type
decisions,
– Empirically establishes the significance of considering a nested logit modelling framework to estimate the
Automatically generated rough PDF by ProofCheck from River Valley Technologies Ltd

relative ordering of the two decisions, and

– Quantifies the role of various factors including social ties in the relative ordering of the two decisions.

Data Source and Sample Selection


A self-administered mail survey was conducted with households from the metropolitan area of Jacksonville,
Florida after Hurricane Matthew. The stratified sampling procedure considered three dimensions to randomly
select households that were affected by coastal surge, high wind, and/or intensive flooding during or after the
major hurricane. Storm surge data from Hurricane Matthew advisories using the Tropical Cyclone Storm Surge
Probabilities product (then P-Surge 2.5 version) (NHC 2018) were retrieved (NWS 2018) and mapped into four
geographic areas: Shoreline, Moderate Surge Zone (20–50% probability), Weak Surge Zone (5–20% probability)
and Inland. The surge map was further overlaid with two other sampling factors – Mandatory Evacuation
Area versus Voluntary Evacuation Area (FDEM 2018), and Urban Area versus Rural Area (FactFinder 2018)
(see Figure 1). Number of households in each zip code were computed by taking all three sampling factors into
account using a random selection from the metropolitan area of over 1.3 million population.

Brought to you by | Göteborg University - University of Gothenburg


Authenticated 3
Download Date | 11/19/19 9:01 AM
Damera et al. DE GRUYTER

Figure 1: Household Survey Sampling Map of Jacksonville (FL) Metropolitan Area.

During the summer and fall in 2017, four waves of mailings (i.e. three complete survey packets and a post-
card reminder) were implemented using the standard procedure recommended by Dillman, Smyth, and Chris-
tian (2016). In each wave, the survey questionnaire was assembled into four different versions where the four
Automatically generated rough PDF by ProofCheck from River Valley Technologies Ltd

major blocks of questions (i.e. uncertainty factors and evacuation behaviors, social network information, intra-
family decision making factors, and information sources) were placed in different sequences with household
demographic questions at the end. This design choice was made in consideration of survey fatigue, in which
there may potentially be low response rates for questions at the latter part of the survey. In other words, for
each wave, four batches (each has 1250 packets) of the survey were mailed to the four quarters of the sampled
households. The same survey questionnaire was also made available for respondents to complete online at
SurveyMonkey.com if one preferred web entries with a given survey ID for each questionnaire in the mail.
The obtained data set had variables relating to hurricane risk perception, evacuation, socio-economic and
social network characteristics of the respondents. After removing incomplete responses from both mail and web
responses, a total of 589 valid responses were present in the final dataset. Out of these responses, only 248 people
responded to the question corresponding to the sequencing of decisions. These 248 responses were cleaned of
missing values by list-wise removing the observations where people did not respond, refused to answer or did
not know about questions. After conducting this procedure, the cleaned data had 100 observations.
The survey also gathered social network information by asking the respondents to provide information
about the people with whom they closely interacted during hurricane Matthew. Personal research design ap-
proach (PNRD) approach (Halgin and Borgatti 2012) was used to obtain the ego-centric social network data
of the individuals. In the first step of PNRD, a list of alters (social network members) of the ego (respondent)
was created. Respondents were asked to think of up to five alters. In the next step, the attributes (e.g. age, gen-
der, relationship duration of ego with alters, how frequently ego communicates with alters) of each alter were
collected. Using the responses of alters, various measures of an individual’s social network were computed.
For instance, gender heterogeneity measure was calculated based on Blau’s heterogeneity index embedded as
a part of the E-Net program, a network analysis software designed for egocentric networks (Borgatti 2009).
In order to collect information about the order in which decisions are taken, the following question was
asked: “In what sequence did you make the following travel related decisions? Please indicate the sequence

Brought to you by | Göteborg University - University of Gothenburg


4 Authenticated
Download Date | 11/19/19 9:01 AM
DE GRUYTER Damera et al.

next to the decision starting with 1 for the first decision. If you made some decisions at the same time, assign
them the same number.” Six types of decisions were given: (a) Evacuation, (b) Departure time, (c) Accommo-
dations (public shelter, friend/relative’s home, hotel/motel, etc.), (d) Evacuation destination (city/town), (e)
Evacuation travel mode (private vehicle, public transit, ride with someone else), and (f) Evacuation route (roads
used).
Table 1 presents the descriptive statistics of explanatory variables (with abbreviations in italics) used in the
final model specification with corresponding mean, standard deviation, minimum, and maximum.

Table 1: Descriptive Statistics of Explanatory Variables.


Variable and corresponding description Mean Standard Minimum Maximum
deviation
Household characteristics
 Abo65: Equal to 1 if there is at least one individual in house above 65 0.27 0.44 0 1
years of age, 0 otherwise
 Adults: Equal to 1 if there are more than two individuals in house 0.17 0.38 0 1
above 18 years of age (i.e. if there are at least three adults in the
house), 0 otherwise
Risk perception
 ConWind: Extent of concern about dam-age to home from wind 0.81 0.39 0 1
(equal to 1 if concerned more than moderate extent, 0 otherwise)
 ConExp: Extent of concern about out of pocket expenses for gas, 0.11 0.31 0 1
food, and lodging while away from home (equal to 1 if concerned
more than moderate extent, 0 otherwise)
 ConUti: Extent of concern about disruption to utility services (equal 0.57 0.50 0 1
to 1 if concerned more than moderate extent, 0 otherwise)
Social network characteristics
 FreSD: Standard deviation of the frequency of communication with 0.52 0.56 0 2
different alters
 TotFreq: Sum of frequency (1 indicates very seldom, 5 indicates very 11.64 5.60 1 25
often) of communication with all (up to 5) alters
AvgAge: Average age of alters 52.39 10.39 23 75
 MaxYear: Maximum years of relationship with alters 34.32 18.72 1 75
 GendHet: Gender heterogeneity of alters (higher value denotes 0.24 0.23 0 0.5
higher gender heterogeneity among alters)
Automatically generated rough PDF by ProofCheck from River Valley Technologies Ltd

Methodology
The methodology used here to model the relative ordering of Acco and Evac comes under the framework of dis-
crete choice modelling. We model three outcomes: first deciding evacuation destination (denoted as EvacFirst),
first deciding accommodation type (denoted as AccoFirst) and simultaneously deciding both (Simult).1 These
three outcomes are modeled as nominal variables. Figure 2 presents the distribution of the three outcomes
(with the values above the bins representing the percentage of each outcome in the final set of observations).
Note that discrete choice modeling is able to capture the quintessential features of the independent variables
(e.g. correlation between them, marginal effects, and statistical significance of the variables), and predict the
probability of choosing different outcomes depending on the independent variables.

Figure 2: Distribution of Relative Ordering Categories for Evacuation Destination and Accommodation Type Decisions.

Brought to you by | Göteborg University - University of Gothenburg


Authenticated 5
Download Date | 11/19/19 9:01 AM
Damera et al. DE GRUYTER

There are different modelling techniques to model nominal variables, including the multinomial logit,
nested logit, and random parameter logit models. The simplest modelling technique is to use a multinomial
logit model (MNL) where all the outcomes/decisions are treated at the same level. We suspected that Evac-
First and AccoFirst may share some unobserved effects, i.e., the disturbance terms may not be independently
and identically distributed (IID) and independent from irrelevant alternatives (IIA). This indicates that the rel-
ative probability of choosing an alternative changes if an outcome option is removed from the outcome set
and the results may be erroneous. To overcome the IIA limitations, McFadden (1981) developed generalized
extreme value models.2 The outcomes that are suspected to share unobserved effects are grouped into nests
in the model. In our case, EvacFirst and AccoFirst are grouped into a nest. Since the outcome probabilities are
determined by difference in the utility functions, the outcomes which share unobserved effects under a nest
cancel out. The Nested Logit Model, has the following structure (Washington, Karlaftis, and Mannering 2003):

exp(𝛽𝑖 𝑋𝑖𝑛 + ∅𝑖 𝐿𝑖𝑛 )


𝑃𝑛 (𝑖) = (1)
∑∀𝐼 exp(𝛽𝐼 𝑋𝐼𝑛 + ∅𝐼 𝐿𝐼𝑛 )

exp(𝛽𝑗/𝑖 𝑋𝑛 )
𝑃𝑛 (𝑗/𝑖) = (2)
∑∀𝐽 exp(𝛽𝐽/𝑖 𝑋𝐽𝑛 )

⎛ ⎞
𝐿𝑆𝑖𝑛 = 𝐿𝑁 ⎜
⎜∑ exp(𝛽𝐽/𝑖 𝑋𝐽𝑛 )⎟
⎟ (3)
⎝ ∀𝐽 ⎠
where X are vectors of measurable independent variables that determine the probability of discrete outcomes,
β are vectors of estimable parameters, Pn (i) is the unconditional probability of observation n having discrete
outcome i, Pn (j/i) is the probability of observation n having discrete outcome j conditioned on the outcome
being in the outcome category i (e.g., for the nested structure in Figure 3, the outcome category i is either
sequential or simultaneous and Pn (j/i) is the binary logit model of choice between AccoFirst and EvacFirst),
J is the conditional set of outcomes (conditioned on i) and I is the unconditional set of outcomes (the upper
two branches of Figure 3), LSin is the inclusive value (logsum), LN is natural logarithm, and ∅i is an estimable
parameter corresponding to the inclusive value. Therefore, the unconditional probability of having outcome j
is,
Automatically generated rough PDF by ProofCheck from River Valley Technologies Ltd

Figure 3: Nested Logit Structure for Relative Ordering of the Two Decisions.

𝑃𝑛 (𝑗) = 𝑃𝑛 (𝑗/𝑖)𝑃𝑛 (𝑖) (4)


The estimation of the nested model is done by simultaneously estimating all the nests using full-information
maximum likelihood (Washington, Karlaftis, and Mannering 2003). The likelihood function was optimized in
NLOGIT (Greene 2002).

Empirical Analysis
In this section, the model specification of the Nested logit model and the estimation results are presented.
Before testing the significance of various variables in the model it was ensured that endogenous variables are

Brought to you by | Göteborg University - University of Gothenburg


6 Authenticated
Download Date | 11/19/19 9:01 AM
DE GRUYTER Damera et al.

discarded. For example, variables such as route choice, mode choice, accommodation type, departure time that
can be affected by a different relative ordering of Evac and Acco were not tested as independent variables. The
choice of variables for potential inclusion in the model is also guided by previous theoretical and empirical
works on evacuation destination and accommodation type, and intuitive arguments regarding the effect of
exogenous variables on these variables. As mentioned in the background and motivation section, the size of a
household determines the type of accommodation a household prefers. The preference for accommodation type
could also influence the destination and vice versa. Therefore, we decided to include household characteristics
variables such as Abo65 and Adults in the model. Cheng, Wilmot, and Baker (2008) observed that perceived risk
affects evacuation destination. They found that hurricane threatened areas are less attractive as destinations
than those outside the at-risk areas. In this study, we extend the importance of risk perception variables (such
as ConWind, ConExp and ConUti) on the relative ordering of Evac and Acco. As mentioned before, the influence
of social network variables on Evac and Acco has not been studied, let alone the effect of sequencing on the
two decisions. But, the effect of social network variables on evacuation decision has been quantified (Sadri,
Ukkusuri, and Gladwin 2017). Thus, we expect that there could possibly be significant effects of social network
members on relative ordering of Evac and Acco. We arrived at the final specification model by systematically
eliminating variables found to be insignificant at the 90% confidence level. In Table 2, correlation between the
various significant explanatory variables in the final specification are presented. It is found that there is no pair
of variables with high correlation and therefore, multicollinearity is not an issue in the estimated model. The
details and implications of the estimated results are subsequently presented.

Table 2: Correlation Matrix.


FreSD Abo65 ConWind TotFreq ConExp AvgAge MaxYear Adults ConUti GendHet
FreSD 1.00 −0.10 0.04 0.28 −0.07 −0.15 0.02 −0.12 −0.13 0.33
Abo65 1.00 −0.16 −0.14 −0.15 0.27 0.10 −0.22 −0.11 −0.07
ConWind 1.00 −0.02 0.17 0.12 0.06 0.02 0.15 −0.01
TotFreq 1.00 0.07 −0.09 0.30 −0.01 0.04 0.40
ConExp 1.00 −0.08 −0.05 0.27 0.25 −0.12
AvgAge 1.00 0.01 0.04 0.04 −0.02
MaxYear 1.00 −0.08 0.02 0.06
Adults 1.00 0.13 −0.01
ConUti 1.00 0.05
GendHet 1.00

Table 3 presents the estimated coefficients of the variables that were found to be significant and the asso-
ciated p-values for both the lower and upper nests of Figure 3 (e.g., constants were not found to be significant
Automatically generated rough PDF by ProofCheck from River Valley Technologies Ltd

and are therefore omitted). The variables in the lower nest corresponds to AccoFirst. Note that the variables
that we considered do not vary across alternate outcomes, so they are estimated at most from n − 1 outcomes
(Washington, Karlaftis, and Mannering 2003) with n being 2 in our case. Also, the variables in the upper nest
correspond to the simultaneous decision Simult, for the same reason as stated before.

Table 3: Coefficient Estimates.


Variable Coefficient p-Value
Lower nest
 Abo65 −2.60 0.008
 ConWind 1.79 0.008
 ConExp −1.54 0.098
 FreSD 1.85 0.007
 TotFreq −0.16 0.008
Upper nest
 Adults −1.89 0.033
 ConUti −1.10 0.040
 MaxYear −0.05 0.003
 AvgAge 0.05 0.003
 GendHet −2.15 0.094

First, we discuss the estimated coefficients for the lower nest. These coefficients govern the probability of
EvacFirst versus AccoFirst given that a sequential decision is chosen. The negative coefficient of Abo65 indicates
that if there is at least one individual in the household above the age of 65 years then the probability to de-
cide evacuation destination before the accommodation type increases. This can be attributed to older people

Brought to you by | Göteborg University - University of Gothenburg


Authenticated 7
Download Date | 11/19/19 9:01 AM
Damera et al. DE GRUYTER

evacuating as early as possible to first find an evacuation destination and then they later decide the type of ac-
commodation (Bytheway 2007), so the household might first decide the evacuation location and then the type
of accommodation. It is also possible that these people have a location with which they are familiar (e.g., before
retiring to Florida) and are comfortable there and pick the accommodations from possible friends/family or
hotel later. Evacuation behavior also is significantly affected by perceived risk in addition to socio-demographic
factors (Dixit, Wilmot, and Wolshon 2012). The positive coefficient for ConWind implies that if a person is con-
cerned about damage of the home by wind then it is more likely that they prioritize Acco over Evac as compared
to other people who do not have such concerns. Such a person might want to evacuate to an accommodation
type that is wind resistant, so that it ensures safety for them and their family. It might also be because people
who are concerned about wind damage are likely to be those in mobile homes and other vulnerable housing,
and since these people tend to be poorer than the average, they are more likely to choose the accommodation
of shelter first. The negative coefficient for ConExp suggests that a person who is worried about out of pocket
expenses for gas, food, lodging etc., while being away from home has a greater likelihood of first selecting Evac
over Acco as compared to people who are not concerned about this issue. That might be because cost sensitive
individuals prefer a city/location where the expenses are minimized (Mesa-Arango et al. 2012). Potentially,
these individuals may select closer locations to minimize fuel expenses. Therefore, location of the destination
would have priority over the type of accommodation.
Social network variables are also found to be significant in affecting the probabilities of sequential decision-
making alternatives. Overall, these findings support the previous literature that suggests that existing social
networks shape individuals’ evacuation and general disaster experiences (Lee et al., in press; Li et al. 2010; Litt
2008; Sadri et al. 2018). The positive coefficient for FreSD implies that when there is higher variation in the fre-
quency of communication with the alters of a person, then the probability of first deciding accommodation type
increases. This might be because individuals may decide to evacuate to one of their alters’ place if there is high
level of communication with some alters in comparison to others. However, they might later decide on which
one of their alters’ (with whom they communicate more frequently) homes to evacuate to, leading to a delay in
the decision making of evacuation location. That is, more variation in the frequency of communication might
represent narrowing down with whom to stay. Therefore, accommodation type might be decided before evac-
uation destination. TotFreq has a negative coefficient suggesting that it is more likely that Evac gets priority over
Acco with an increase in the overall frequency a person communicates with his alters. This could be attributed
to the fact that individuals who frequently communicate with their alters may have received information about
various evacuation destinations, some of which their alters might have evacuated to, therefore they may more
quickly decide on their destination. This might lead to first deciding Evac over Acco. It is also possible that some
of the alters with whom the person communicates more frequently live close by, so Evac is decided before Acco.
Now we discuss the coefficients in the upper nest, i.e., the coefficients governing the probability of sequen-
tial versus simultaneous decision making. The negative coefficient corresponding to Adults indicates that the
probability of taking a sequential decision increases (as compared to the simultaneous decision making) if there
Automatically generated rough PDF by ProofCheck from River Valley Technologies Ltd

are more than two adults in the household. That might be because if there are large number of people of this
age group in the household, it may become tough to simultaneously make decisions on both Evac and Acco,
considering the fact that with more adults, there are likely more people in the household whose opinions are
considered. The complexity of reaching consensus is reduced if they take each component (Evac or Acco) in
turn. The negative coefficient for ConUti implies that if someone is concerned about the disruption of utility
services to his house then his probability of making sequential decisions increases. Potentially, those who are
concerned about the loss of utilities might have the same concern of facing hardships related to utilities for the
place they want to evacuate to. Therefore, they would have greater likelihood of taking the two decisions one
by one rather than making a simultaneous decision because it might be difficult to satisfy the preferences of
both Evac and Acco while taking into account the concern related to utilities. Thus, choosing one decision after
another will allow proper decision-making.
Now we discuss the interpretation of social network variables on deciding sequential versus simultaneous
decisions. The negative coefficient of MaxYear indicates that as the maximum duration for which an alter is
known increases, the probability of choosing sequential decision increases. A person might give higher weight
to the advice given by an alter with whom he/she shared the longest relationship. This alter might have a pref-
erence of first choosing either Evac or Acco, therefore influencing the person/ego to take a sequential decision.
We observe a positive coefficient for the variable AvgAge, which implies that as average age of alters increases,
the ego has higher likelihood of choosing Simult. That might be because higher AvgAge is a proxy for experience
(with hurricanes possibly) in a person’s social network. Experienced people might understand the benefits of
both Evac and Acco and may influence towards taking a simultaneous decision. Potentially, experienced people
might already have a plan in place and thus they can jointly decide the two decisions. The negative coefficient
for GendHet suggests that as the gender heterogeneity of alters increases, the probability of choosing sequential
decisions increases. Decision making for men and women might be different, so increasing the number of al-

Brought to you by | Göteborg University - University of Gothenburg


8 Authenticated
Download Date | 11/19/19 9:01 AM
DE GRUYTER Damera et al.

ters of one gender might generate preference for choosing one decision before the other, leading to sequential
decision making. That is, mixing of genders might mean a greater diversity of perspectives. It has been found
that women have higher risk perception than men (Fu and Wilmot 2006). Perhaps accounting for the diversity
of perspectives requires a simplified decision making approach - one by one.
Table 4 presents the overall fitness measures of the estimated model. The inclusive value parameter is 0.51,
with a p-value of 0.09. The inclusive parameter should lie in the interval (0, 1) and should be significantly
different from zero and one. The significance of this parameter in our case establishes the importance of using
a nested structure to model the relative ordering of Evac and Acco (Washington, Karlaftis, and Mannering 2003).
The likelihood ratio test statistic is,

Table 4: Summary of Goodness of Fit Measures.


Measure Value
Log-likelihood at convergence (LL(c)) −79.82
Restricted log-likelihood (LL(r)) −116.45
Number of parameters for adjustment (K) 10
Number of observations 100
ρ2 0.31
Adjusted ρ2 0.26
Inclusive value parameter (p-value) 0.51 (0.09)

𝜒 2 = −2[𝐿𝐿(𝑟) − 𝐿𝐿(𝑐)] (5)

where, LL(c) is Log-likelihood at convergence, and LL(r) is restricted Log-likelihood. In the restricted Log-
likelihood, exogenous variable parameters are set to zero. It is found that χ 2 is equal to 73.3 and is greater than
the 90% confidence level statistic, 15.99. This result clearly rejects the null hypothesis that exogenous variable
parameters are zero. Other measures for overall fit are ρ2 and adjusted ρ2 statistics. The values of these statistics
obtained in our study are reasonable given the complexity of nested logit model (Washington, Karlaftis, and
Mannering 2003).

Summary and Conclusions


Automatically generated rough PDF by ProofCheck from River Valley Technologies Ltd

This paper presents a statistical model for estimating relative ordering of evacuation destination and accom-
modation type decisions in hurricanes. In the literature, the focus has been to separately model these two
decisions but understanding of the relative ordering of these two decisions is missing. Our unique dataset
indicates that all permutations of the relative ordering exist: destination first, accommodation first, and simul-
taneous. From the number of observations in each category, there is no strongly dominant pattern, suggesting
that the decision-making process varies across the population. To investigate the factors that might influence
the decision-making process, we estimated a nested logit model with the upper nest containing two branches: 1)
sequential and 2) simultaneous decision making. The sequential decision making branch had two sub-branches:
1) first deciding evacuation destination and then deciding accommodation type and 2) first deciding accommo-
dation type and then deciding evacuation destination. Also, most of the literature sequentially models the two
decisions (Gehlot et al. 2018b; Ukkusuri et al. 2017), but our model provides a non-zero probability of simul-
taneously choosing the two decisions. In addition, we found social network variables to be significant in the
estimation of this relative ordering. This makes a contribution to the hurricane evacuation literature as there is
limited research focusing on the effects of social network characteristics on evacuation behavior (Huang, Lin-
dell, and Prater 2016; Perry, Lindell, and Greene 1981; Sadri, Ukkusuri, and Gladwin 2017). Some of the key
insights from the study include:

– With an increase in the variation of communication frequency with alters, the probability of first choosing
accommodation type increases, given that a sequential decision is made. But, with an increase in the total
communication frequency with alters, the probability of first choosing evacuation destination increases,
provided sequential decision making happens.
– If there is at least one person in the household above the age of 65 years then the probability of first choosing
evacuation destination increases, provided sequential decision making happens.

Brought to you by | Göteborg University - University of Gothenburg


Authenticated 9
Download Date | 11/19/19 9:01 AM
Damera et al. DE GRUYTER

– If a person is concerned about the out of pocket expenses for food, gas, etc. while staying away from home then
the likelihood of first choosing evacuation destination increases, given that sequential decisions are made.
But, if a person is concerned about damage to the home by wind while evacuating then the probability of
first choosing accommodation type increases, given sequential decision making is made.

– The probability of taking sequential decisions increases as compared to simultaneous decision if there are
more than two adults in the household.
– If a person is concerned about the loss of utilities at home, then the likelihood of choosing sequential decisions
increases.
– An increase in the maximum duration for knowing an alter leads to an increase in the probability of choosing
sequential decisions. But, an increase in the average age of alters leads to an increase in the probability of
making simultaneous decisions.
– If gender heterogeneity of alters increases, then the probability of making sequential decisions increases.

The above insights would help hurricane evacuation emergency planners to develop better evacuation policies.
For instance, if it is known that many people in an area prefer choosing evacuation destination before accommo-
dation type and their alters also live in the same region then proper management of traffic network efficiency
can minimize the possibility of traffic jams (Urbina and Wolshon 2003; Dixit and Wolshon 2014). It could also
lead to better estimation of evacuation demand. Since estimation of evacuation demand conventionally requires
sequential estimation of multiple choices, the estimation of probabilities of simultaneous versus sequential de-
cisions using our model could lead to more accurate demand estimation but this requires further exploration
in the future. We also suggest an outreach program to encourage people to talk to their peers at the beginning
of each hurricane season to arrange accommodation in those peers’ homes. This could affect both the type of
accommodations and influence the likelihood of simultaneously selecting accommodations and destinations.
One limitation of this study is that the sample size is small, which is also the limitation of many studies
that rely on surveys collecting social network data (Sadri, Lee, and Ukkusuri 2015; Gehlot, Sadri, and Ukkusuri
2018a). While the sample might not adequately represent the broader population, the low response rate does
not mean every parameter of the studied variables is biased. Reports by Curtin, Presser, and Singer (2000)
and Keeter et al. (2000) indicate low response rates do not appear to bias central tendency estimates such as
means and proportions, although perhaps a larger response rate might have found more significant correlation
and coefficients. Thus, in this case, the central tendency statistics from this study might be comparable to other
hurricane evacuation studies, however further investigation into this issue is required in the future. In addition,
it is not plausible to expect respondents to remember all the details of the decision making process that went
into their evacuation process. However, this is a limitation of most evacuation related surveys, as we need to
Automatically generated rough PDF by ProofCheck from River Valley Technologies Ltd

wait for people to return and conditions to become normal before the surveys can be mailed. Literature suggests
that extreme events or unusual experiences tend to be stored in memory for a longer time (Tourangeau 1999).
Thus, it is likely that recalling evacuation during hurricanes would have been relatively easier than recalling
routine events which happened a long time ago. Also, the insights obtained from this study may not generalize
to all hurricane evacuations but the study offers meaningful observations for the regions that share similar
characteristics as our survey area. Finally, there can be many future extensions of this study. A similar modeling
approach can be extended to other decisions like time of departure, route choice, mode choice etc. Also, a
modeling framework that applies the insights from this model to estimate evacuation demand would be a
good extension. In addition, future evacuation and census surveys should collect social networks information
so that the effects of social ties on evacuation behavior can be further investigated.

Funding

Funder Name: Division of Civil, Mechanical and Manufacturing Innovation, Funder Id:
http://dx.doi.org/10.13039/100000147, Grant Number: 1520338.

Notes
1 Note that EvacFirst, AccoFirst and Simult are the possible sequencing outcomes of the variables Evac and Acco. Thus, the sequencing
outcomes are different from the variables Evac and Acco.
2 Both random parameter logit and nested Logit models address the IIA limitation. However, when we tested the random parameter
model, it did not turn out to significant in comparison to the fixed parameter model, and therefore we chose a fixed parameter model.

Brought to you by | Göteborg University - University of Gothenburg


10 Authenticated
Download Date | 11/19/19 9:01 AM
DE GRUYTER Damera et al.

References
Baker, E. J. 2000. “Hurricane evacuation in the United States.” Storms 1:306–319.
Barrett, B., Ran, B., and Pillai, R. 2000. “Developing a Dynamic Traffic Management Modeling Framework for Hurricane Evacuation.” Trans-
portation Research Record: Journal of the Transportation Research Board 1733:115–121.
Borgatti, S. 2009. E-NET 0.023 [Computer Program]. Harvard, MA: Analytic Technologies.
Bytheway, B. 2007. “The Evacuation of Older People: The Case of Hurricane Katrina.” In Annual Conference of the Royal Geographical Society and
Institute of British Geographers, London, Vol. 7, 2012. http://understandingkatrina.ssrc.org/Bytheway, accessed April.
Cheng, G., and Wilmot, C. G. 2009. “Time-Dependent Travel Cost Impact on Hurricane Evacuation Destination Choice Models.” Technical
Report.
Cheng, G., Wilmot, C. G., and Baker, E. J. 2008. “A Destination Choice Model for Hurricane Evacuation.” In 87th Annual Meeting of the Trans-
portation Research Board, 13–17. Washington, DC.
Cheng, G., Wilmot, C., and Baker, E. 2011. “Dynamic Gravity Model for Hurricane Evacuation Planning.” Transportation Research Record: Journal
of the Transportation Research Board 2234:125–134.
Curtin, R., Presser, S., and Singer, E. 2000. “The Effects of Response Rate Changes on the Index of Consumer Sentiment.” Public Opinion Quar-
terly 64:413–428.
Dillman, D. A., Smyth, J. D., and Christian, L. M. 2016. “Internet, Phone, Mail and Mixed-Mode Surveys: The Tailored Design Method.” Reis
154:161–176.
Dixit, V., and Wolshon, B. 2014. “Evacuation Traffic Dynamics.” Transportation Research Part C: Emerging Technologies 49:114–125.
Dixit, V., Wilmot, C., and Wolshon, B. 2012. “Modeling Risk Attitudes in Evacuation Departure Choices.” Transportation Research Record: Journal
of the Transportation Research Board 2312:159–163.
Dow, K., and Cutter, S. L. 2002. “Emerging Hurricane Evacuation Issues: Hurricane Floyd and South Carolina.” Natural Hazards Review 3(1):12–
18.
FactFinder 2018. “American Factfinder – Search.” Accessed on 07/19/2018. https://factfinder.census.gov/faces/nav/jsf/pages/ searchre-
sults.xhtml?refresh=t.
FDEM 2018. “Evacuation Orders – Florida Disaster.” Accessed on 07/19/2018. https://www.floridadisaster.org/evacuation-orders.
Fu, H., and Wilmot, C. 2004. “Sequential Logit Dynamic Travel Demand Model for Hurricane Evacuation.” Transportation Research Record:
Journal of the Transportation Research Board 1882:19–26.
Fu, H., and Wilmot, C. 2006. “Modeling Hurricane Evacuation Traffic: Development of a Time-Dependent Hurricane Evacuation Demand
Model.” Technical Report, Louisiana Transportation Research Center.
Gehlot, H., Sadri, A. M., and Ukkusuri, S. V. 2018a. “Joint Modeling of Evacuation Departure and Travel Times in Hurricanes.” Transportation,
1–22.
Gehlot, H., Zhan, X., Qian, X., Thompson, C., Kulkarni, M., and Ukkusuri, S. V. 2018b. “A-RESCUE 2.0: A High-Fidelity, Parallel, Agent-Based
Evacuation Simulator.” Journal of Computing in Civil Engineering 33(2), 04018059.
Golshani, N., Shabanpour, R., Auld, J., Mohammadian, A. K., and Ley, H. 2018. “Modeling Evacuation Destination and Departure Time
Choices for No-Notice Emergency Events.” Technical Report.
Greene, W. H. 2002. NLOGIT: Version 3.0: Reference Guide. Econometric Software.
Halgin, D. S., and Borgatti, S. P. 2012. “An Introduction to Personal Network Analysis and Tie Churn Statistics Using E-NET.” Connections
Automatically generated rough PDF by ProofCheck from River Valley Technologies Ltd

32(1):37–48.
Hasan, S., and Ukkusuri, S. V. 2011. “A Threshold Model of Social Contagion Process for Evacuation Decision Making.” Transportation Research
Part B: Methodological 45(10):1590–1605.
Huang, S.-K., Lindell, M. K., and Prater, C. S. 2016. “Who Leaves and Who Stays? A Review and Statistical Meta-Analysis of Hurricane Evacua-
tion Studies.” Environment and Behavior 48(8):991–1029.
Keeter, S., Miller, C., Kohut, A., Groves, R. M., and Presser, S. 2000. “Consequences of Reducing Nonresponse in a National Telephone Survey.”
Public Opinion Quarterly 64: 125–148.
Lee, S., Sadri, A., Ukkusuri, S., Clawson, R. A, and Seipel, J. (in press). Why Different Forms of Social Support? The Network Structure of Impro-
vised Social Support Ties Surrounding Households During Post-Disaster Recovery. Natural Hazards Review.
Li, W., Airriess, C. A., Chen, A. C.-C., Leong, K. J., and Keith, V. 2010. “Katrina and Migration: Evacuation and Return by African Americans and
Vietnamese Americans in An Eastern New Orleans Suburb.” The Professional Geographer 62(1):103–118.
Lindell, M. K., and Prater, C. S. 2007. “Critical Behavioral Assumptions in Evacuation Time Estimate Analysis for Private Vehicles: Examples
from Hurricane Research and Planning.” Journal of Urban Planning and Development 133(1):18–29.
Lindell, M. K., Lu, J.-C., and Prater, C. S. 2005. “Household Decision Making and Evacuation in Response to Hurricane Lili.” Natural Hazards
Review 6(4):171–179.
Lindell, M. K., Kang, J. E., and Prater, C. S. 2011. “The Logistics of Household Hurricane Evacuation.” Natural Hazards 58(3):1093–1109.
Lindell, M., Murray-Tuite, P., Wolshon, B., and Baker, E. 2018. Large-Scale Evacuation: The Analysis, Modeling, and Management of Emergency Relo-
cation from Hazardous Areas. Boca Raton: Production at CRC Press . DOI: https://doi.org/10.4324/9781315119045 .
Litt, J. 2008. “Getting Out or Staying Put: An African American Women’s Network in Evacuation from Katrina.” NWSA Journal 20(3):32–48.
McFadden, D. 1981. “Econometric Models of Probabilistic Choice.” In Structural Analysis of Discrete Data with Econometric Applications, 198–272.
Mesa-Arango, R., Hasan, S., Ukkusuri, S. V., and Murray-Tuite, P. 2012. “Householdlevel Model for Hurricane Evacuation Destination Type
Choice Using Hurricane Ivan Data.” Natural Hazards Review 14(1):11–20.
Mileti, D. S., Sorensen, J. H., and O’Brien, P. W. 1992. “Toward An Explanation of Mass Care Shelter Use in Evacuations.” International Journal of
Mass Emergencies and Disasters 10(1):25–42.
Murray-Tuite, P., Yin, W., Ukkusuri, S., and Gladwin, H. 2012. “Changes in Evacuation Decisions Between Hurricanes Ivan And Katrina.”
Transportation Research Record: Journal of the Transportation Research Board 2312:98–107.

Brought to you by | Göteborg University - University of Gothenburg


Authenticated 11
Download Date | 11/19/19 9:01 AM
Damera et al. DE GRUYTER

NHC. 2018. “Tropical Cyclone Storm Surge Probabilities.” Accessed on 07/19/2018. https://www.nhc.noaa.gov/surge/psurge.php.
NWS. 2018. “P-Surge 2.6.” Accessed on 07/19/2018. http://slosh.nws.noaa.gov/psurge2.0/index.php?S=Matthew2016&Adv=29&Ty=e10&Z
=m1&D=agl&Ti=cum&Msg=17&Help=about.
Perry, R. W., Lindell, M. K., and Greene, M. R. 1981. Evacuation Planning in Emergency Management.
Prater, C., Wenger, D., and Grady, K. 2000. Hurricane Bret Post Storm Assessment: A Review of the Utilization of Hurricane Evacuation Studies and
Information Dissemination. College Station, Texas: Texas A&M Univ. Hazard Reduction & Recovery Center.
Sadri, A. M., Lee, S., and Ukkusuri, S. V. 2015. “Modeling Social Network Influence on Joint Trip Frequency for Regular Activity Travel Deci-
sions.” Transportation Research Record: Journal of the Transportation Research Board 2495:83–93.
Sadri, A. M., Ukkusuri, S. V., and Gladwin, H. 2017. “The Role of Social Networks and Information Sources on Hurricane Evacuation Decision
Making.” Natural Hazards Review 18(3):04017005.
Sadri, A. M., Ukkusuri, S. V., Lee, S., Clawson, R., Aldrich, D., Nelson, M. S., Seipel, J., and Kelly, D. 2018. “The Role of Social Capital, Personal
Networks, and Emergency Responders in Post-Disaster Recovery and Resilience: A Study of Rural Communities in Indiana.” Natural
Hazards 90(3):1377–1406.
Southworth, F. 1991. Regional Evacuation Modeling: A State-of-the-Art Review.
Stewart, S. R. 2017. National Hurricane Center Tropical Cyclone Report: Hurricane Matthew. National Hurricane Center.
Tierney, K. J., Lindell, M. K., and Perry, R. W. 2002. “Facing the Unexpected: Disaster Preparedness and Response in the United States.” Disas-
ter Prevention and Management: An International Journal 11(3):222–222.
Tourangeau, R. 1999. “Remembering What Happened: Memory Errors and Survey Reports.” In The Science of Self-Report (pp. 41–60). Psychol-
ogy Press.
Ukkusuri, S. V., Hasan, S., Luong, B., Doan, K., Zhan, X., Murray-Tuite, P., and Yin, W. 2017. “A-RESCUE: An Agent Based Regional Evacuation
Simulator Coupled with User Enriched Behavior.” Networks and Spatial Economics 17(1):197–223.
Urbina, E., and Wolshon, B. 2003. National Review of Hurricane Evacuation Plans and Policies: A Comparison and Contrast of STATE prac-
tices. Transportation Research Part A: Policy and Practice 37(3):257–275.
Washington, S., Karlaftis, M. G., and Mannering, F. L. 2003. Statistical and Econometric Methods for Transportation Data Analysis. Boca Raton, FL:
Chapman & Hall/CRC.
Wu, H.-C., Lindell, M. K., and Prater, C. S. 2012. “Logistics of Hurricane Evacuation in Hurricanes Katrina and Rita.” Transportation Research Part
F: Traffic Psychology and Behaviour 15(4):445–461.
Wu, H., Lindell, M., Prater, C., and Huang, S. 2013. “Logistics of Hurricane Evacuation in Hurricane Ike.” Logistics: Perspectives, Approaches and
Challenges. Nova Science, Hauppauge, 127–140.
Yang, H., Morgul, E. F., Ozbay, K., and Xie, K. 2016. “Modeling Evacuation Behavior Under Hurricane Conditions.” Transportation Research
Record: Journal of the Transportation Research Board 2599:63–69.
Yin, W., Murray-Tuite, P., Ukkusuri, S. V., and Gladwin, H. 2014. “An Agent-Based Modeling System for Travel Demand Simulation for Hurri-
cane Evacuation.” Transportation Research Part C: Emerging Technologies 42:44–59.
Automatically generated rough PDF by ProofCheck from River Valley Technologies Ltd

Brought to you by | Göteborg University - University of Gothenburg


12 Authenticated
Download Date | 11/19/19 9:01 AM

You might also like