Travel behavior changes during the COVID-19 pandemic in Japan: Analyzing the effects of risk perception and social influence on going-out self-restriction

Abstract

This article analyzes factors affecting travel behavior changes at the individual level in light of the COVID-19 pandemic in Japan, in the context of non-binding self-restriction requests. In particular, this study focuses on the effects of risk perception and social influence. A panel web-survey was conducted targeting residents of the Kanto Region, including the Tokyo Metropolis. In addition to describing the observed patterns in behavioral change, we modeled behavioral changes of four key, non-work-related activities: (i) grocery shopping, (ii) other types of shopping, (iii) eating out, and (iv) leisure. For eating-out and leisure the distinction was made between going alone or in groups. Based on the observed distributions of these activities, regression models of going-out frequency were estimated for shopping, while for eating-out and leisure, which showed considerably smaller frequencies, a discrete choice approach was used. Findings showed that as a measure of social influence, the perception of degree of self-restriction of others was associated with small reductions in shopping frequencies, and moderate yet non-negligible increases in going-out self-restriction probability for eating-out and leisure activities. Risk perception, measured as COVID-19 dread, was also associated with higher probabilities of going-out self-restriction for eating-out and leisure. These findings suggest that in the context of non-binding requests, soft measures such as campaigns to promote a reduction of non-essential travel might be more effective if they (i) properly convey the severity of the threat posed by COVID-19 as well as its coping mechanisms, and (ii) appeal to the group, rather than the individual, emphasizing the behavior (or at least the perception of behavior) of others.

Full text

Travel behavior changes during the COVID-19 pandemic in Japan: Analyzing
the effects of risk perception and social influence on going-out self-restriction
Giancarlos Parady
a,
⁎,Ayako Taniguchi
b
,Kiyoshi Takami
a
a
TheUniversityof Tokyo,Tokyo,Japan
b
Tsukuba University, Tsukuba, Japan
ABSTRACTARTICLE INFO
Article history:
Received 13 June 2020
Received in revised form 20 July 2020
Accepted 20 July 2020
Available online xxxx
This article analyzes factors affecting travel behaviorchanges at the individual level in light ofthe COVID-19 pandemic
in Japan, in the context of non-binding self-restriction requests. In particular, this study focuses on the effects of risk
perception and social influence. A panel web-survey was conducted targeting residents of the Kanto Region, including
the Tokyo Metropolis. In addition to describing the observed patterns in behavioral change, we modeled behavioral
changes of four key, non-work-related activities: (i) grocery shopping, (ii) other types of shopping, (iii) eating out,
and (iv) leisure. For eating-out and leisure the distinction was made between going alone or in groups. Based on the
observed distributions of these activities, regression models of going-out frequency were estimated for shopping,
while for eating-out and leisure, which showed considerably smaller frequencies, a discrete choice approach was
used. Findings showed that as a measure of social influence, the perception of degree of self-restriction of others
was associated with small reductions in shopping frequencies, and moderate yet non-negligible increases in going-
out self-restriction probability for eating-out and leisure activities. Risk perception, measured as COVID-19 dread,
was also associated with higher probabilities of going-out self-restriction for eating-out and leisure. These findings sug-
gest that in the context of non-binding requests, soft measures such as campaigns to promote a reduction of non-
essential travel might be more effective if they (i) properly convey the severity of the threat posed by COVID-19 as
well as its coping mechanisms, and (ii) appeal to the group, rather than the individual, emphasizing the behavior
(or at least the perception of behavior) of others.
Keywords:
COVID-19
Self-restriction
Social influence
Risk perception
Travel behavior
1. Introduction
The COVID-19 pandemic has had an unprecedented effect on people's
mobility across the globe. By the end of March 2020, more than a hundred
countries had implemented some form mobility restriction, ranging from
full or partial mandatory quarantines (usually referred to as lockdowns)
to non-binding requests for activity restrictions, such as stay-at-home re-
quests, closing certain types of businesses, canceling events, etc. (BBC,
2020).
The case of Japan falls under the latter category. Starting on February
26th with the central government request to cancel or postpone large
events and gatherings, and culminating with a declaration of state of emer-
gency on April 7th, first in seven prefectures including the Tokyo area, and
later expanded nationwide on April 16th (Table 1).
Since these measures are not mandatory or legally enforceable under
Japanese law, they revolve around self-restriction, and their success largely
dependent on compliance by the population. In this regard, using datafrom
telecommunications and IT firms, the Cabinet Office reported significant
drops in activity in large employment and commercial hubs of the Greater
Tokyo Area (defined in this article as the Tokyo Metropolis and the prefec-
tures of Chiba, Saitama and Kanagawa) both before and after the state of
emergency declaration, as summarized in Table 2.
Independently, Google used the location history data from their map ap-
plication to measure changes in visit frequency to different facilities (see
Fig. 1). A decreasing trend in visit frequencies to retail and recreation facil-
ities, workplaces and transit stations, and a consequent increase in activities
in residential areas can be observed after the issuing of the joint stay-at-
home request by the governors of Tokyo and adjacent prefectures.
Certainly, aggregate results show a change in behavioral patterns. The
question that this article then raises is, given that stay-at-home requests
are non-enforceable and non-binding, what factors affect the observed
self-restriction behavior at the individual level. The objective of this study
is thus to quantify the effect of factors that affect behavioral changes in re-
sponse to the COVID-19 pandemic, in the context of Japan. To do so, we use
data from an original panel survey targeting residents of the Kanto region,
more specifically, the Tokyo Metropolis and the prefectures of Chiba,
Saitama and Kanagawa (hereinafter the Greater Tokyo Area) and the pre-
fectures of Ibaraki, Gunma and Utsunomiya (hereinafter Northern Kanto).
Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
⁎Corresponding author.
E-mail addresses: gtroncoso@ut.t.u-tokyo.ac.jp, (G. Parady), taniguchi@risk.tsukuba.ac.jp, (A. Taniguchi), takami@ut.t.u-tokyo.ac.jp. (K. Takami).
http://dx.doi.org/10.1016/j.trip.2020.100181
2590-1982/© 2020 TheAuthors. Publishedby Elsevier Ltd. Thisis an open access article under the CC BY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.
0/).
Contents lists available at ScienceDirect
Transportation Research Interdisciplinary Perspectives
journal homepage: https://www.journals.elsevier.com/transportation-research-
interdisciplinary-perspectives

The collected data is then used to model behavioral changes of four key,
non-work-related activities: (i) grocery shopping, (ii) other types of shop-
ping, (iii) eating out, and (iv) leisure.
The rest of this article is structured as follows. Section 2 brieflysumma-
rizes findings from the literature. Section 3 describes the survey design and
execution details.Section 4 describes the data characteristics. Section 5 pre-
sents the model results, while Section 6 discusses findingsandlimitationsof
this study.
2. Literature review
Since at the time of writing the COVID-19 pandemic is still an ongoing
crisis, the literature on the subject is still limited. As such, we present only a
brief review with a general overview of findings, in order to frame the anal-
ysis presented in this article.
Given the way of transmission of the SARS-CoV-2 virus, physical dis-
tance has emerged as a key mitigation strategy. This translates in some
form of mobility restrictions. In the case of Wuhan, China, where the first
cases were reported, control measures were drastic, but it has been shown
that these measures substantially mitigated the spread of COVID-19
(Kraemer et al., 2020;Zhang et al., 2020). However, such implementations
are not necessarily possible in many countries. In the case of Japan, where
the government cannot legally enforce lockdowns on citizens and residents,
strategies have relied largely on requests for voluntary self-restriction
which include avoiding unnecessary travel, teleworking, etc. (Shaw et al.,
2020). As such, in the context of non-enforceable self-restriction requests,
understanding what factors affect behavioral change at the individual
level is of importance.
Although not specific to pandemics, human behavior is influenced by
social norms, that is, what they perceive that others are doing or what
they think others approve of (Van Bavel et al., 2020). As such, researchers
have suggested applying the principles of behavioral change to “nudge”
people into desirable behaviors to help control the spread of COVID-19
(West et al., 2020). Stay-at-home requests, which require changing one's
travel patterns and practicing self-restriction for non-essential activities,
fall squarely into the realm of travel behavior analysis. In the field, there
is a small but growing number of articles addressing the issue of social influ-
ence on travel behavior, albeit on ordinary circumstances (Kim et al.,
2018). These studies are motivatedby the idea that the behavioral patterns
of individuals vary with the behavior of the reference group (Manski,
1993). In addition, the frequency of discretionary activities such as leisure,
have also been found to be associated with social network characteristics
(Carrasco and Miller, 2006;Parady et al., 2019;Parady et al., 2020). As
such, social influence might alsobe an important factor affecting behavioral
changes in extraordinary situations, such as the COVID-19 pandemic, an
issue we aim to evaluate in this study.
Another aspect that might be relevant to behavioral change, is the per-
ception of the risk posed by COVID-19. According to the protection motiva-
tion theory, the evaluation of the severity of a threat is one of the cognitive
processes behind the decision to engage in protective behavior (Rogers,
1983). In addition, research on risk perception has shown that there is a
gap between actual risks and the perception of risks by individuals
(Slovic, 1987). Slovic developed a simple scale of risk measurement to as-
sess the perception of risk relative to other hazards, and identified three
key factors associated with risk perception: (i) dread, (ii) unknown (unfa-
miliarity) and (iii) number of people involved.
Based on data collected through an online survey, Wise et al. (2020)
evaluated changes in risk perception in the US during the first week of
the pandemic, and found that as risk awareness grew so did the report fre-
quency of engagement in protective behavior. They also reported that prac-
ticing physical distancing was better predicted by the perceived likelihood
of personally being infected, rather than the likelihood of transmission or
severity of potential infections.
Finally, an individual trait that mightaffect behavioral response is social
anxiety. For example, Spielberger et al. (1983) proposed the State-Trait
Anxiety Inventory (STAI), where state anxiety refers to reactions related
to adverse situations in a specific moment, while trait anxiety refers to a
more stable tendency to experience anxiety in a myriad of situations. As
such, individuals with high scores in social anxiety traits might exhibit par-
ticular behavioral responses. At present, we have not found any study
attempting to evaluate the relation between social anxiety and behavioral
responses to COVID-19.
3. Survey design and execution
A panel data web-survey was conducted through Rakuten Insight, a
market research firm that maintains an internet panel of around 2.2 million
monitors in Japan. The target area was the Kanto region, specifically, the
Greater Tokyo Area and Northern Kanto.
The first wave of the survey was conducted on April 8th, one day after
the emergency state declaration, but well after the spread of the virus
(more than 4000 cases had been reported nationwide. See Table A1 in
the Appendix). The sample size was 800 persons. The sampling method
was quota sample, with an equal number of samples for each group. The
group segmentation criteria were region (Greater Tokyo Area or Northern
Kanto), gender (male or female) and age cohort (20s, 30s, 40s, 50s, and
60s and over). This resulted in 20 groups of 40 respondents each.
Table 1
Key dates regarding COVID-19 activity restriction requests in Japan.
Sources: Asahi Shimbun (2020).
Date Action taken
February 26th, 2020 The central government issues a request to cancel or postpone large-scale events
March 2nd, 2020 The central government requests the closure of all elementary, junior high, and high schools across the country
March 26th, 2020 The Tokyo Metropolis and the four adjacent prefectures in the capital region issue a joint stay-at-home request
April 7th, 2020
(late afternoon)
The central government declares a state of emergency for 7 prefectures across Japan, including the Tokyo Metropolis,
and the prefectures of Chiba, Saitama and Kanagawa
April 9th, 2020 Tokyo governor Koike defines which facilities are requested to stop operations during the state of emergency
April 11th, 2020 The central government requests a reduction of at least 70% of office commuters for the regions targeted by the emergency state declaration
April 16th, 2020 The central government expands the target of the emergency state declaration to all prefectures
Table 2
Changes in number of people in different commercial districts of the Greater Tokyo Area for April 21st relative to key dates (compiled from DOCOMO mobile spatial statistics).
Source: Cabinet Secretariat (2020).
Average changes compared to/area Shinjuku (Tokyo) Yokohama station (Kanagawa) Chiba station (Chiba) Omiya station (Saitama)
January 18 –February 14 (before the widespread of the virus) −69.7% −64.7% −57.5% −59.3%
April 6–7 (before the of emergency state declaration in the Greater Tokyo Area) −55.4% −53.1% −47.2% −45.9%
April 15–16 (before the nationwide emergency state declaration) −4.8% 0.1% −4.2% −4.9%
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
2

Fig. 1. Percentage change in visit frequency to different facilities from the 15th of February to the 17th of April. The gray vertical lines mark national holidays. The red vertical lines
mark the following key events (from left to right): 1) The Tokyo Metropolis and four adjacent prefectures issue a joint stay-at-home request, 2) The central government declares a state
of emergency for 7 prefectures across Japan, including the Greater Tokyo Area, 3) The central government expands the target of the emergency state declaration to all prefectures.
Source: Produced by authors using data from: https://www.google.com/covid19/mobility/
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
3

3.1. Survey items
The following data were collected:
1. Basic socio-demographic characteristics, including employment type
and working style.
2. Mobility tools ownership.
3. Measuresof social anxiety: following Spielberger et al. (1983),wemea-
sured two dimensions of social anxiety, namely, state anxiety and trait
anxiety. Although each of these scales involve measuring 20 indicator
variables, in orderto reduce the response burden, we used the three in-
dicators that according to research in the context of Japan (Iwamoto
et al., 1989) had the highest factor loading:
a) State anxiety:
1. “I feel anxious”.
2. “Ifeelnervous”.
3. “Iamworried”.
b) Trait anxiety:
1. “I feel that difficulties are piling up so that I cannot overcome them”.
2. “I am inclined to take things hard”.
3. “Some unimportant thought runs through my mind and bothers me”.
4. Risk perception: following Slovic (1987) we measured risk perception
of COVID-19 using three questions related to (i) levels of dread (“I think
COVID-19 is frightening”), (ii) familiarity (“I know COVID-19 well”)and
(iii) perception of controllability of the hazard (“I think COVID-19 can
be controlled”). Each item was measured using a 5-point Likert scale.
In this case too, a reduced number of indicators was used to reduce
the response burden. We also measured perception of other hazards
to evaluate the relative risk perception of COVID-19.
5. Weekly trip frequency by purpose before the spread of COVID-19 (Be-
fore February 2020). Frequency was measured on the following scale:
“don't do that activity”,“less than weekly”,“1/week”,“2/week”,“3/
week”,“4/week”,“5/week”,“6/week”,“7/week”,“few times per day”
6. Weekly trip frequency by purpose after the spread of COVID-19 and be-
fore the emergency state declaration, that is, until April 7th, 2020. This
correspond to the behavior up to the day before the first wave of the
survey was conducted.
7. Behavioral changes in response to COVID-19, including hygiene, eat-
ing etiquette, shopping, going-out self-restriction, and substitution of
face-to-face interactions with ICT interactions.
8. Perception of degree of going-out self-restriction (“gaishutsu jishuku”in
Japanese): Measured using a 7-point scale to the question “to what ex-
tent do you think you and other people are practicing going-out self-
restriction?”. This was measured for the respondents themselves, as
well as for relatives, work- and/or schoolmates, neighbours, other
friends, and the general public.
9. Social expectations regarding going-out self-restriction behavior: Mea-
sured using a 5-point Likert scale to the questions “do you think other
people are expecting you to practice going-out self-restriction?”and “do
you think other people would approve of you practicing going-out self-
restriction?”. These were measured for relatives, work- and/or school-
mates, neighbours, other friends, and the general public.
10. Subjective well-being: Measured using a 10-point scale to the question
“How happy would you say you are?”
The secondwave was conducted 2 weeks after the first wave,which was
after the declaration of the state of emergency in the Greater Tokyo Area.
Items 3 to 10 were collected again. Participation rate for the second wave
was 90.2% (722 respondents).
4. Basic data characteristics and processing
4.1. Social anxiety measures
Principal component analysis was used to reduce the dimensionality of
the data to one variable per construct. First principal components explained
78% and 80% of the variance in the data for state anxiety and trait anxiety
indicators, respectively. Note that since state anxiety levels are likely to
change in particularly adverse moments, and given that we did not observe
state anxiety levels before the spread of COVID-19, in the behavioral anal-
ysis reported in the next section, only trait anxiety, considered more stable
over time is used.
4.2. Risk perception
Following Slovic (1987),Fig. 2 plots the means of risk perception in
terms of dread, unknown and controllability for different hazards to get
an idea of the average level of perceived risk of COVID-19. As the left
panel shows, COVID-19 is second only to earthquakes in terms of dread
and is also perceived to be the least controllable of all hazards inquired.
For the behavioral analysis presented in the next section, risk perception
variables were dummy-codified so that they take value 1 if the rating was
4 or 5 and 0, otherwise. For the time period before the infection spread
(t
0
), all risk perception variables were set to zero. As an aside, the case of
elderly driving is of particular interest as it shows the effect of availability
on risk perception. With a series of cases where elderly drivers lost control
of the vehicle resulting in fatal accidents reported often in the media, the
dread perception of elder driving is high relative to the actual risk.
Fig. 2. Comparison of risk perception of different hazards (first wave results).
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
4

4.3. Perception of degree of going-out self-restriction
Fig. 3 illustrates the perception of the degree of going-out self-
restriction of oneself and others and its changes over time. It can be seen
that with the exception of the “general public”case the mean values
(marked by the rhomboids) increased. Median values also increased, for
“family”,“friends”and “work/school friends”.
For the behavioral analysis, principal component analysis was used to
extract a single component of perception of degree of going-out self-restric-
tion of others. To guarantee that principal component scores were on the
same scale, the PCA coefficients were calculated using data from time pe-
riod t
1
(after spread - before emergency declaration) and predicted for the
time period t
2
(after spread - after emergency declaration). In addition,
for ease of understanding, these variables were normalized so that the pos-
sible range of values falls between 0 and 1. For time period t
0
(before the
infection spread) all values were set to zero.
Tables A2 and A3 in the Appendix summarize the rest of the descriptive
statistics for time invariant and time-varying covariates.
4.4. Changes in travel behavior
Regarding changes in travel behavior, Figs. 4 and 5 illustrate changes
over time of different activity frequencies. In Fig. 4, for the top plot, the
blue bars (no change from zero) show the cases where the activity was
not conducted even before the spread of the infection, for the bottom
plot, it include those that stopped an activity altogether at t
1
and kept
doing so at t
2
.Fromt
0
to t
1
, a considerable share of the sample reported de-
creases in frequency for most activities. Interestingly enough, from t
1
to t
2
a
rebound for shopping activities was observed, suggesting a sort of adjust-
ment process. It can also be seen that for activities such as eating-out and
leisure, the reductions in frequency persisted, as shown by the shares of
“no change (from zero)”and “decrease”in the bottom plot. These changes
are also a result of changes in the supply side. That is, many eating-out and
leisure establishments closed down or shortened their business hours in re-
sponse to the non-binding request of the government, in contrast to most
shopping facilities which provided more essentials services.
The overall trends across all time periods are shown in Fig. 5. In the top
figure, the blue bars show those who did not conduct a given activity at any
time period. The bottom figure is perhaps more enlightening, showing the
changes across all periods when excluding individuals with constant zero
frequencies. It can be seen that for this group, the share of decreasing trends
ranged from 34% for grocery shopping to 87% for eating out for business.
For the behavioral analysis described in the following section, the alter-
native “few times per day”was truncated at 7, and the alternative “less than
weekly”was coded to 0.5. See Table A4 in the Appendix for the descriptive
statistics of trip frequencies over time, after this transformation.
5. Modeling behavioral change
As stated in the introduction, the question that this article seeks to an-
swer is what factors affect the observed self-restriction behavior at the indi-
vidual level, given that stay-at-home requests are non-enforceable and non-
binding. Since self-restriction decisions regarding commuting and work-
related trips are not completely in the control of individuals, we will narrow
down the scope of the analysis to shopping, eating out and leisure, and ex-
clude work-and business-related trips. Furthermore, based on the observed
distributions of these activities, for shopping variables, which exhibit rela-
tively higher frequencies, regression models of going-out frequency are es-
timated. On the other hand, given that eating out and leisure frequencies
are not that high to being with, their distributions are not suitable for linear
regression analysis, as such, a discrete choice approach is used.
Based on thediscussion in the previous sections, the keyvariables in this
analysis are the principal component of perception of degree of self-restric-
tion of others and COVID-19 dread as a measure of risk perception, and to a
lesser extent the principal component of trait anxiety. Although a set of con-
trol variables such as population density at residential location, changes in
commuting frequency, mean cumulativenumber of infections by prefecture
(for the seven days up to the survey day), and basic socio-demographic
characteristics are included as control variables in the models, the discus-
sion of results will focus on these three variables.
5.1. Shopping frequency models
To model shopping frequency, instead of depending on a single model
with a single set of assumptions, we rely on three different types of models,
each with different assumptions, thus providing more robustness to the
analysis. The first two models are a pooled OLS and a random effect
model. In these models the dependent variables are the shopping trip fre-
quencies foreach time period pooled together. The random effect model es-
timates random intercepts for each respondent in order to capture
heterogeneity across individuals. These two models however require that
unobserved individual fixed effects be uncorrelated with the explanatory
variables. The third model is a first-difference regression, where both de-
pendent and independent variables are differenced over time (Δy=y
t
−
y
t−1
;Δx=x
t
−x
t−1
). Given this transformation, unobserved individual
fixed effects are allowed to be correlated with explanatory variables,
since these effects are differenced out of the equation. However, as a result,
time-constant variables cannot be included. In this case, this is not a prob-
lem since the two key variables ofthis analysis are time varying. The regres-
sion models were estimated in R 3.6.3. with the plm (Croissant and Millo,
2008)andlme4 (Bates et al., 2014) packages. Robust errors (cluster errors
at the individual level) are estimated for all models, to account for the
panel nature of the data.
Fig. 3. Perception of degree of going-out self-restriction (original scale).
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
5

Table 3 summarizes the elasticities and marginal effects of key variables
for shopping frequency models. Elasticities are estimated at mean values of
xand y. Full estimation results are reported in the Appendix (Tables A5 and
A6).
Regarding the effect of the perceptionofdegreeofself-restriction
of others on shopping frequencies, as expected, a negative association
was observed. However, the elasticity estimates are rather small,
especially for the first-difference regression estimates. Across all
models, a 1% increase in the perception of self-restriction principal
component score results in a decrease in grocery shopping frequency
of around 0.04% to 0.07%, and of around 0.03% to 0.08% for other
types of shopping.
Although we expected a negative association between COVID-19 dread
and shopping frequency, the effects of COVID-19 dread were estimated
with very high uncertainty, so no inference can be made about effect direc-
tion or magnitude. Finally, interms of trait anxiety effects, elasticities for all
models are essentially zero, suggesting no treat anxiety effect on shopping
frequency.
Goodness of fit statistics and out of sample validation statistics
suggest that for grocery shopping, the first-difference model (Adj.R
2
:
Fig. 4. Changes in travel behavior over time.
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
6

0.087, correlation
1
: 0.28) is slightly superior to both the pooled model
(Adj.R
2
: 0.076, correlation: 0.24) and the random effect model (R
2
:
0.074, correlation: 0.24). For other kinds of shopping, when looking at
goodness of fitstatisticsonly,itcannotbesaidthatthefirst-difference
model (R
2
: 0.045) outperforms the other models (R
2
: 0.048 (pooled),
0.053(random effect)). However, when looking at out-of-sample valida-
tion statistics, the first-difference model is marginally superior (correlation:
0.20, against 0.18 for the other models). It must be noted that the R-square
statistics for both shopping models are low, so the large share of the vari-
ance in the data remains unexplained.
5.2. Eating out and leisure models
For eating out and leisure, three types of discrete choice models are es-
timated. The first two models are models of going-out self-restriction. In
these models the dependent variables are the choices to “stay at home”or
“go out”for each particular purpose. Data from all time periods are pooled
Fig. 5. Trend of changes of travel behavior across all time periods.
1
Refers to average correlation between observed and predicted outcomes in out-of-sample
validation.
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
7

together. Here, “stay at home”refers to zero-valued weekly activity frequen-
cies, while “go out”implies non-zero-valued weekly activity. The first model
is a simple pooled binary logit model. The second modelis an error compo-
nent mixed binary logit model. This model captures individual-specificef-
fects that cause correlation across successive choices by the same decision
maker, by adding i.i.d. N(0,1) error components that vary across respon-
dents (not choice situations) to each alternative, and estimating an addi-
tional parameter θfor the individual errors (Hess et al., 2008). The third
model is a multinomial logit model of changes in trip frequency (decrease,
no change or increase) defined based on the time differencing of trip fre-
quencies, as explained in the previous section. For this model, all indepen-
dent variables are also first-differenced when includedin the model. Robust
errors (cluster errors at the individual level) are estimated for all models, to
account for the panel nature of the data. Since we are interested in evaluat-
ing what factors affect behavioral change, and given the large percentage of
the sample that did not engage in eating-out and leisure activities even be-
fore the spread of the infection (see Fig. 5), we subset the data to focus on
individuals who reported non-zero frequencies at t
0
. The discrete choice
models were estimated in R 3.6.3. with the Apollo package (Hess and
Palma, 2019).
Table 4 summarizes the elasticities and marginal effects of key vari-
ables. For the full estimation results see Tables A7–A9 in the Appendix.
Elasticities are calculated analytically for the logit models. Elasticities for
the mixed logit models and the marginal effects for categorical variables
are estimated via simulation.
The perception of degree of self-restriction of others is consistently asso-
ciated with increases in the probability to stay home and the probability to
reduce trip frequencies, suggesting the possible existence of a social influ-
ence mechanism for activities such as eating-out and leisure. For example,
the mixed logit elasticity estimate for eat-out alone is around 0.16, which
suggest that a 1% increase in the principal component score of self-
restriction perception will result approximately in a 0.16% (C.I. 0.03% to
0.22%) increase in the probability of staying home.
COVID-19 dread is also consistently associated with increases in the
probability to stay home and the probability to reduce trip frequencies.
Since COVID-19 dread is a binary variable, marginal effects were estimated.
These indicate the absolute change in choice probability given a change in
the explanatory variable. For example, the probability of reducing leisure
(group) frequency will increase on average by 0.12 (C.I. 0.004 to 0.228)
for a person going from not considering COVID-19 dreadful to doing so.
Regarding the trait anxiety coefficients, the mixed logit models includ-
ing these covariates resulted in non-negative-semidefinite Hessian matri-
ces, hence these variables were excluded from the final models. As such,
these effects could only be evaluated with the pooled logit models. How-
ever, as Table 4 shows, the elasticities of trait anxiety were essentially zero.
In addition to the elasticities and marginal effects presented in Table 4,
and to further illustrate the magnitude of the effects of social influence and
risk perception on behavioral changes, we conducted a simulation exercise
of changes in choice probabilities given changes in 1) the principal compo-
nent of perception of degree of self-restriction of others and 2) the dread
perception of COVID-19, holding other variables constant. Fig. 6 shows
the simulation results of the binary and mixed logits, while Fig. 7 shows
the results of the changes in trip frequency models. It can be clearly seen
that the magnitudes of these effects are not negligible.
Goodness of fit statistics suggest all models perform rather well, with
considerable improvements in log-likelihood compared to the constant
only models. Furthermore, out-of-sample mean percentage of correct pre-
dictions range from 79.6% to 84.1% for the pooled and mixed binary
Table 3
Elasticities and marginal effects of key variables of shopping frequency models.
Model Grocery shopping Other shopping
Estimate (95% C.I.) Estimate (95% C.I.)
Elasticity of perception of degree of
self-restriction of others
Pooled model −0.069 −0.109 to −0.029 −0.084 −0.116 to −0.052
Random effect model −0.036 −0.087 to 0.014 −0.046 −0.100 to 0.007
First difference model −0.035 −0.067 to −0.004 −0.028 −0.041 to −0.015
Elasticity of trait anxiety Pooled model 0.0001 0.0001 to −0.0001 −0.0001 0.0000 to −0.0001
Random effect model 0 0.0002 to −0.0001 −0.0001 0.0000 to −0.0002
First difference model –– – –
Marginal effect of COVID-19 dread Pooled model 0.100 −0.095 to 0.295 −0.110 −0.288 to 0.067
Random effect model −0.014 −0.240 to 0.212 −0.089 −0.289 to 0.111
First difference model −0.017 −0.319 to 0.285 −0.083 −0.331 to 0.165
Elasticities estimated at mean values of x and y.
Bold font indicates coefficients with p-value <0.1.
Table 4
Elasticities and marginal effects of key variables of discrete choice models.
Eat out (alone) Eat out (Private) Leisure (alone) Leisure (group)
Estimate 95% C.I. Estimate 95% C.I. Estimate 95% C.I. Estimate 95% C.I.
Pooled binary logit of self-restriction behavior
Elasticity of perception of degree of self-restriction of others:
stay home 0.169 0.032 to 0.307 0.107 0.015 to 0.200 0.101 −0.020 to 0.222 0.120 0.013 to 0.227
Elasticity of trait anxiety: stay home 0.000 −0.001 to 0.000 0.000 0.000 to 0.000 0.001 −0.001 to 0.004 0.000 0.000 to 0.000
Marginal effect of COVID-19 dread: stay home 0.107 0.025 to 0.179 0.092 0.028 to 0.145 0.051 −0.019 to 0.110 0.076 0.004 to 0.134
Error component mixed binary logit of self-restriction behavior
Elasticity of perception of degree of self-restriction of others:
stay home 0.155 0.030 to 0.219 0.096 0.009 to 0.139 0.082 −0.050 to 0.148 0.114 0.004 to 0.159
Elasticity of trait anxiety: stay home –– –– –– ––
Marginal effect of COVID-19 dread: stay home 0.099 0.019 to 0.169 0.095 0.029 to 0.148 0.062 −0.009 to 0.122 0.048 −0.023 to 0.107
MNL of changes in frequency
Elasticity of perception of degree of self-restriction of others: increase −0.101 −0.271 to 0.070 −0.120 −0.374 to 0.1 34 −0.026 −0.219 to 0.167 0.075 −0.234 to 0.384
Elasticity of perception of degree of self-restriction of others: decrease 0.060 −0.023 to 0.142 0.083 0.001 to 0.165 0.080 −0.005 to 0.165 0.131 0.051 to 0.212
Marginal effect of COVID-19 dread: increase −0.043 −0.072 to 0.023 −0.009 −0. 035 to 0.054 −0.039 −0.067 to 0.025 −0.012 −0.039 to 0.075
Marginal effect of COVID-19 dread: decrease 0.158 0.027 to 0.277 0.142 0.019 to 0.254 0.157 0.027 to 0.274 0.120 0.004 to 0.228
Bold font indicates coefficients with p-value <0.1.
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
8

logit models, and 66.6% to 72.1% for the frequency change models. It is
also worth pointing out that while the mixed logit model outperformed
the pooled binary logit model in terms of goodness of fit, the binary logit
was superior, albeit very slightly, in terms of out-of-sample validation.
6. Discussion and conclusion
In this article we analyzed factors affecting travel behavior changes at
the individual level during the COVID-19 pandemic in Japan using data
from a panel web-survey conducted in the Kanto Region. In line with
aggregate results reported by IT firms, a significant drop in activity levels
was observed. We showed that as a measure of social influence, the percep-
tion of degree of self-restriction of others was consistently associated with
reductions in activity levels across all activity types. Since the dependent
variables of the estimated models are different, only a rough comparison
of effect magnitudes is possible, but for shopping frequency, effects were
rather small, while in terms of probability of self-restriction (staying
home) and probability of trip frequency reduction for eating-out and leisure
activities, effects were moderate, yet non-negligible. This can be explained
by the fact that grocery shopping, and to some extent other types of
Fig. 6. Simulation of the effects of perception of degree of self-restriction of others
and COVID-19 dread on going-out self-restriction (“stay home”)choiceprobability
for eating-out and leisure activities. Other covariates are fixed as follows: time
period = t
1
. All continuous variable set to mean values. All categorical variables
set to reference categories.
Fig. 7. Simulation of the effects of perception of degree of self-restriction of others
and COVID-19 dread on changes in activity frequency choice probability for eating-
out and leisure activities. Other covariates are fixed as follows: time period: 1st
period (t
1
-t
0
). All continuous variable set to mean values. All categorical variables
set to reference categories.
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
9

shopping (excluding of course, leisure and luxury shopping) fall under the
category of essential activities, and consequently they are less susceptible
to self-restriction social effects. On the other hand, activities like eating-
out or leisure are not, at least inthe short term, absolutely essential forsub-
sistence and thus likely to be more susceptible to self-restriction social
effects.
A similar argument can be made for risk perception, which, measured as
COVID-19 dread, was also found to be associated with non-negligible in-
creases in the probability to stay home and the probability to reduce trip
frequencies. Since shopping activities are necessary for subsistence, they
must be conducted with some regularity in spite of the perceived risk. As
such, it is plausible that individuals would opt to take protective measures
such as wearing masks, physical distancing when in stores, etc. instead of
reducing shopping frequency.
Regarding social anxiety measures, effects were essentially zero, for all
types of activities. Although this might be a result of the use of a limited
measure of trait anxiety (as opposed to using the full STAI scale), further re-
search is necessary to evaluate this.
These findings have important policy implications for Japan, where
there are no legal mechanisms to enforce a “city lockdown.”Our findings
suggest that in the context of non-binding requests, soft measures such as
campaigns to promote a reduction of non-essential travel might be more ef-
fective if they(i) properly conveythe severity of the threatposed by COVID-
19 as well as its coping mechanisms, and (ii) appeal to the group, rather
than the individual, emphasizing the behavior (or at least the perception
of behavior) of others.
Fear arousal has long been a subject of study in social psychology as a
mean to promote attitudinal or behavioral change. The protection motiva-
tion theory by Rogers (1983), for example, posits two cognitive processes
behind the decision to engage in protective behavior against a threat. The
first one is the evaluation of the severity of the threat (threat appraisal),
where fear arousal plays a role by affecting the cognitiveappraisal of the se-
verity of the threat. The second one is the evaluation of the ability to cope
with such threat (coping appraisal), which takes into consideration not
only the efficacy of the response but also its cost.
As a policy issue, the agency doing the persuasion has the responsibility
to guarantee that the fears claimed are legitimate and the coping measures
effective, by relying on sound science and effective science communication.
Too strong a fear appeal, however, might result in discrimination and
prejudice, as well as undesirable behavior, such as what has come to be
known in Japan as the “self-restriction police”where private individuals ag-
gressively call-out or “force”individuals and businesses that do not respond
to voluntary requests of self-restriction to comply. This underscores the
need to achieve a difficult balance where the severity of the threat is effec-
tively communicated, without encouraging discrimination and prejudice
against individuals who don't comply with the government requests.
Sufficient policy support is also necessary to minimize the cost of coping
behavior by compensating individuals and businesses for loss of income
that might result from self-restricting behavior.
Finally, this study has two important limitations. The first one is that
given the non-probabilistic sampling method, the generalization poten-
tial of findings is limited. The main reason to do so was that we wanted
to capture behavior as close to real time as possible. Based on past expe-
rience, using the basic resident register (the most comprehensive sam-
pling frame available in Japan) for a probabilistic sample would have
takenmonths,andcapturingbehavior retrospectively might have re-
sulted in a different bias due to difficulties associated with accurately
remembering past behavior, and in particular, behavioral changes. As
such, additional studies are necessary to validate the results presented
here. Additional studies are also needed to evaluate to what extent
the findings reported here replicate in different cities and/or socio-
cultural contexts.
The second limitation is that due to the nature of the survey instrument,
behaviorwas measured as trip frequency and trip chaining effects could not
be captured. Regardless, the findings presented in this article, while pend-
ing further validation, provide some insight on the effects of social influ-
ence and risk perception on behavioral changes during the COVID-19
pandemic in Japan.
Declaration of competing interest
The authors report no conflict of interest.
Acknowledgements
This study was funded by kakenhi Grants-in-Aid for Scientific Research
20H02266 and 17K18947 from the Japan Society for the Promotion of
Science.
Appendix A
Table A1
Cumulative number of COVID-19 infections by prefecture.
Region Prefecture Wave 1 Wave 2
Apr.1st Apr.8th Apr.16th Apr.23rd
All Japan –2107 4168 8442 11,172
Northern Kanto Ibaraki 24 77 119 151
Tochigi 14 15 39 52
Gunma 19 29 107 131
Greater Tokyo Area Saitama 98 209 479 714
Chiba 164 284 525 725
Tokyo 527 1203 2457 3452
Kanagawa 120 279 614 841
As reported by the Asahi Shimbun based on documents from the Ministry of Health, Labour and Welfare.
Table A2
Descriptive statistics of time fixed covariates.
Variable n Mean SD Med. Min. Max.
Male 722 0.508
Worker 722 0.723
Age
20–39 722 0.377
40–59 722 0.422
Over 60 722 0.201
Household size
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
10

Table A2 (continued)
Variable n Mean SD Med. Min. Max.
1 person 722 0.215
2 persons 722 0.309
3 persons or larger 722 0.476
Marital status
Single 722 0.274
Married 722 0.623
Divorced 722 0.086
Widower 722 0.015
Other 722 0.001
Prefecture
Northern
Kanto
Gunma 722 0.154
Ibaraki 722 0.195
Tochigi 722 0.150
Greater Tokyo
Area
Chiba 722 0.086
Kanagawa 722 0.127
Saitama 722 0.068
Tokyo 722 0.220
Car available in household 722 0.737
Population density at home location 1000/sq.km 722 7.505 7.448 4.576 0.000 34.092
Principal component of state anxiety 722 0.014 1.520 0.412 −3.664 2.487
Principal component of trait anxiety 722 −0.004 1.515 0.045 −2.984 3.073
Table A3
Descriptive statistics of time-varying covariates.
Variable n Mean SD Med. Min. Max.
t
0
(before infection spread)
Risk perception
Unknown (1 if value is 4 to 5, 0 otherwise) 722 0
Dread (1 if value is 4 to 5, 0 otherwise) 722 0
Controllability (1 if value is 4 to 5, 0 otherwise) 722 0
Social influence
PC of self-restriction perception by others 722 00000
PC of self-restriction expectation by others 722 00000
PC of self-restriction approval by others 722 00000
Wt
1
(after spread - before emergency declaration)
Risk perception
Unknown (1 if value is 4 to 5, 0 otherwise) 722 0.229
Dread (1 if value is 4 to 5, 0 otherwise) 722 0.871
Controllability (1 if value is 4 to 5, 0 otherwise) 722 0.217
Social influence
PC of self-restriction perception by others 722 0.527 0.219 0.531 0 1
PC of self-restriction expectation by others 722 0.602 0.223 0.589 0 1
PC of self-restriction approval by others 722 0.678 0.226 0.750 0 1
t
2
(after spread - after emergency declaration)
Risk perception
Unknown (1 if value is 4 to 5, 0 otherwise) 722 0.251
Dread (1 if value is 4 to 5, 0 otherwise) 722 0.893
Controllability (1 if value is 4 to 5, 0 otherwise) 722 0.215
Social influence
PC of self-restriction perception by others 722 0.563 0.212 0.570 0 1
PC of self-restriction expectation by others 722 0.641 0.228 0.645 0 1
PC of self-restriction approval by others 722 0.705 0.234 0.750 0 1
Table A4
Descriptive statistics of trip frequencies over time.
Variable n Mean SD Med. Min. Max.
t
0
(before infection spread)
Trip frequencies
Commute 559 4.403 1.873 5 0 7
Business/errands 722 1.332 2.141 0 0 7
Grocery shopping 722 2.409 1.774 2 0 7
Other shopping 722 1.140 1.223 1 0 7
Eat-out (alone) 722 0.713 1.237 0 0 7
Eat-out (work-related) 722 0.346 0.709 0 0 7
Eat-out (private groups) 722 0.541 0.843 0.5 0 7
Leisure (alone) 722 0.543 1.032 0 0 7
Leisure (groups) 722 0.562 0.832 0.5 0 7
t
1
(after spread - before emergency declaration)
Trip frequencies
(continued on next page)
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
11

Table A4 (continued)
Variable n Mean SD Med. Min. Max.
Commute 559 3.581 2.257 5 0 7
Business/errands 722 0.945 1.859 0 0 7
Grocery shopping 722 1.869 1.603 2 0 7
Other shopping 722 0.787 1.083 0.5 0 7
Eat-out (alone) 722 0.327 0.905 0 0 7
Eat-out (work-related) 722 0.130 0.543 0 0 7
Eat-out (private groups) 722 0.227 0.657 0 0 7
Leisure (alone) 722 0.253 0.804 0 0 7
Leisure (groups) 722 0.212 0.597 0 0 7
t
2
(after spread - after emergency declaration)
Trip frequencies
Commute 710 2.102 2.358 0.75 0 7
Business/errands 722 0.896 1.812 0 0 7
Grocery shopping 722 2.038 1.541 2 0 7
Other shopping 722 0.773 1.118 0.5 0 7
Eat-out (alone) 722 0.295 0.962 0 0 7
Eat-out (work-related) 722 0.083 0.549 0 0 7
Eat-out (private groups) 722 0.122 0.568 0 0 7
Leisure (alone) 722 0.184 0.771 0 0 7
Leisure (groups) 722 0.122 0.530 0 0 7
Model estimation results
Table A5
Grocery shopping frequency models.
Variable Grocery shopping frequency model
Pooled model Random effect model First-difference model
Est. t value Est. t value Est. t value
(Intercept) 2.392 12.052 2.376 9.625 −0.370 −2.131
Period 2 –Period 1 (first-differenced) ––––0.605 3.568
Period 1 (after spread - after emergency declaration) −0.380 −4.289 −0.372 −3.043
Time-invariant
Period 2 (after spread - before emergency declaration) −0.157 −1.466 −0.137 −0.946
Male −0.458 −2.547 −0.460 −4.326
Age: 20–39 −0.488 −3.074 −0.490 −2.922
Age: 40–59 −0.171 −1.232 −0.171 −0.991
Household size: 1 person 0.129 0.976 0.133 0.810
Household size: 2 persons 0.025 0.258 0.028 0.225
Car in household 0.016 0.112 0.021 0.135
Population density 1/sq.km. 0.034 4.865 0.035 3.712
Commuting frequency 0.059 3.639 0.060 2.544 0.067 3.277
Mean cumulative number of infections/100 *
1
*
1
−0.003 −0.532 −0.002 −4.379
PC of perception of degree of self-restriction of others −0.398 −3.376 −0.210 −1.401 −0.204 −2.179
COVID-19 dread 0.100 1.007 −0.014 −0.120 −0.017 −0.109
PC of trait anxiety −0.022 −0.787 −0.020 −0.582 Time-invariant
Sample details
Number of observations 2166 2166 1444
Number of individuals 722 722 722
Time periods 3 3 2
Goodness of fit statistics
R-square 0.0761 Marginal: 0.074 0.087
Conditional: 0.669
Adjusted R-square 0.0705 –0.083
F-statistic (df., p-value) 13.64
(d.f.:13, 2152, p< 0.01)
–27.21
(df.:5, 1438, p < 0.01)
Intra-class correlation (individual) –0.64 –
Out-of-sample validation statistic
(Repeated learning test, r= 500, calibration sample = 90%)
Average correlation 0.240 0.235 0.283
RMSE
2
1.468 1.477 1.032
Robust standard errors (cluster errors at the individual level) are used to account for the panel nature of the data.
Bold font indicates coefficients with p-value <0.1.
1
Robust standard error could not be computed, hence excluded from model.
2
RMSE not directly comparable for first-difference model and the remaining two as the dependent variable is different.
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
12

Table A6
Other shopping frequency models.
Variable Other shopping frequency model
Pooled model Random effect model First-difference model
Est. t value Est. t value Est. t value
(Intercept) 1.087 9.183 1.048 6.455 −0.138 −1.471
Period 2 –Period 1 (first-differenced) ––––0.194 1.854
Period 1 (after spread - after emergency declaration) 0.051 0.543 −0.071 −0.504
Time-invariant
Period 2 (after spread - before emergency declaration) 0.136 1.198 0.014 0.084
Male 0.013 0.125 −0.001 −0.016
Age: 20–39 −0.126 −1.279 −0.143 −1.253
Age: 40–59 −0.133 −1.460 −0.161 −1.427
Household size: 1 person −0.184 −2.057 −0.186 −2.026
Household size: 2 persons −0.093 −1.320 −0.089 −1.057
Car in household −0.001 −0.009 −0.010 −0.110
Population density 1/sq.km. 0.005 1.295 0.005 0.857
Commuting frequency 0.052 3.864 0.074 4.243 0.089 4.599
Mean cumulative number of infections/100 −0.007 −4.673 −0.006 −1.444 *
1
*
1
PC of perception of degree of self-restriction of others −0.487 −5.210 −0.269 −1.708 −0.162 −4.101
COVID-19 dread −0.110 −1.221 −0.089 −0.875 −0.083 −0.658
PC of trait anxiety 0.043 2.867 0.043 2.007 Time-invariant
Sample details
Number of observations 2166 2166 1444
Number of individuals 722 722 722
Time periods 3 3 2
Goodness of fit statistics
R-square 0.048 Marginal: 0.053 0.045
Conditional: 0.477
Adjusted R-square 0.042 –0.042
F-statistic (df., p-value) 7.70
(d.f.:14, 2151, p < 0.01)
–16.94
(d.f.: 4, 1439, p< 0.01)
Intra-class correlation (individual) –0.45 –
Out-of-sample validation statistics
(Repeated learning test, r = 500, calibration sample = 90%)
Average correlation 0.179 0.181 0.198
RMSE
2
1.139 1.144 0.979
Robust standard errors (cluster errors at the individual level) are used to account for the panel nature of the data.
Bold font indicates coefficients with p-value <0.1.
1
Robust standard error could not be computed, hence excluded from model.
2
RMSE not directly comparable for first-difference model and the remaining to as the dependent variable is different.
Table A7
Pooled binary logit models of going-out self-restriction behavior for eating-out and leisure activities.
Variable Eat-out (alone) Eat-out (groups) Leisure (alone) Leisure (groups)
Est. t value Est. t value Est. t value Est. t value
ASC go out (ref) 0 0 0 0
ASC stay home −12.887 −84.138 −12.405 −82.977 −12.660 −70.099 −12.522 −74.488
Period 1: stay home 12.928 22.435 12.407 23.056 13.176 23.236 12.951 20.649
Period 2: stay home 13.592 22.862 13.583 24.372 14.218 24.111 14.058 21.504
Male: stay home*
1
−0.994 −4.722 −0.240 −1.217 −0.899 −4.310 −0.615 −3.131
Age 20–39: stay home*
1
−0.273 −0.848 −0.628 −2.098 −0.313 −1.082 −0.678 −2.174
Age 40–59: stay home*
1
0.085 0.275 0.064 0.207 0.144 0.503 −0.302 −0.973
Household size =2: stay home*
1
−0.133 −0.558 −0.395 −1.719 −0.253 −1.090 −0.427 −1.911
Household size =1: stay home*
1
−0.773 −2.970 −0.015 −0.052 −0.565 −2.149 −0.242 −0.864
Car in household: stay home*
1
0.085 0.321 −0.169 −0.568 −0.359 −1.372 −0.259 −0.957
Commuting frequency: stay home −0.049 −1.179 −0.027 −0.675 −0.043 −1.017 0.009 0.230
Population density (1/sq.km): stay home*
1
−0.014 −0.753 −0.025 −1.374 0.028 1.650 0.005 0.281
Mean cumulative number of infections/100: stay home*
1
0.019 1.101 0.066 2.961 0.020 1.170 0.026 1.426
PC of perception of degree of self-restriction of others: stay home 0.997 2.410 0.921 2.266 0.722 1.629 0.939 2.195
COVID-19 dread: stay home 0.766 2.521 0.779 2.717 0.405 1.456 0.615 2.071
PC of trait anxiety: stay home*
1
−0.046 −0.656 0.088 1.256 0.079 1.248 0.014 0.221
Sample details
Number of observations 1080 1239 1038 1287
Number of individuals 360 413 346 429
Number of choice situations 3 3 3 3
Observed choice ratio
Go out 0.42 0.50 0.46 0.49
Stay home 0.58 0.50 0.54 0.51
Goodness of fit statistics
(continued on next page)
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
13

Table A7 (continued)
Variable Eat-out (alone) Eat-out (groups) Leisure (alone) Leisure (groups)
Est. t value Est. t value Est. t value Est. t value
LL(0) −748.60 −858.81 −719.49 −892.08
LL(C) −736.10 −858.80 −716.40 −891.80
LL(β)−418.02 −401.29 −371.29 −442.43
rho-square (β) 0.44 0.53 0.48 0.50
Adjusted rho-square (β) 0.42 0.52 0.46 0.49
Out-of-sample validation statistics
(Repeated learning test, r= 100, calibration sample = 90%)
Mean percentage correct 80.2% 84.8% 81.2% 83.5%
Mean RMSE 0.0014 0.0009 0.0012 0.0010
Robust standard errors (cluster errors at the individual level) are used to account for the panel nature of the data.
Bold font indicates coefficients with p-value <0.1.
1
All time-invariant coefficients are estimated for the post-infection spread periods only (t
1
,t
2
).
Table A8
Error component mixed binary logit models of going-out self-restriction behavior for eating-out and leisure activities.
Variable Eat-out (alone) Eat-out (groups) Leisure (alone) Leisure (groups)
Est. t value Est. t value Est. t value Est. t value
ASC go out (ref) 0 0 0 0
ASC stay home −12.669 −20.153 −20.000 −58.258 −19.275 −32.999 −17.879 −20.658
Period 1: stay home 12.912 10.854 19.954 37.463 19.904 22.401 19.022 13.167
Period 2: stay home 14.062 11.037 21.725 32.522 21.325 20.067 20.935 13.076
Male: stay home*
1
−1.592 −4.506 −0.341 −1.186 −1.224 −4.195 −1.035 −2.975
Age 20–39: stay home*
1
−0.571 −1.136 −0.742 −1.743 −0.321 −0.844 −1.214 −2.273
Age 40–59: stay home*
1
0.016 0.032 0.163 0.367 0.244 0.649 −0.585 −1.102
Household size =2: stay home*
1
−1.212 −2.828 −0.069 −0.168 −0.758 −2.123 −0.375 −0.785
Household size =1: stay home*
1
−0.232 −0.611 −0.574 −1.720 −0.360 −1.160 −0.752 −1.908
Car in household: stay home*
1
0.173 0.404 −0.229 −0.535 −0.468 −1.334 −0.345 −0.748
Commuting frequency: stay home −0.075 −1.154 −0.042 −0.706 −0.048 −0.855 −0.003 −0.051
Population density (1/sq.km): stay home*
1
−0.016 −0.540 −0.027 −1.018 0.039 1.677 0.011 0.378
Mean cumulative number of infections/100: stay home*
1
0.019 0.797 0.071 2.480 0.022 1.018 0.033 1.272
PC of perception of degree of self-restriction of others: stay home 1.490 2.317 1.207 2.114 0.778 1.335 1.547 2.017
COVID-19 dread: stay home 1.136 2.403 1.152 2.763 0.648 1.726 0.682 1.365
PC of trait anxiety: stay home*
1
*
2
*******
θ2.025 6.562 1.728 5.418 1.429 5.150 2.293 6.361
Sample details
Number of observations 1080 1239 1038 1287
Number of individuals 360 413 346 429
Number of choice situations 3 3 3 3
Observed choice ratio
Go out 0.42 0.50 0.46 0.49
Stay home 0.58 0.50 0.54 0.51
Goodness of fit statistics
LL(0) −748.60 −858.81 −719.49 −892.08
LL(C) −736.10 −858.80 −716.40 −891.80
LL(β)−397.13 −391.30 −364.08 −418.51
rho-square (β) 0.47 0.54 0.49 0.53
Adjusted rho-square (β) 0.45 0.53 0.47 0.51
Out-of-sample validation statistics
(Repeated learning test, r= 100, calibration sample = 90%)
Mean percentage correct 79.6% 84.1% 80.9% 83.4%
Mean RMSE 0.0014 0.0009 0.0012 0.0010
Robust standard errors (cluster errors at the individual level) are used to account for the panel nature of the data.
Bold font indicates coefficients with p-value <0.1.
1
All time-invariant coefficients are estimated for the post-infection spread periods only (t
1
,t
2
).
2
Parameters resulted in resulted in non-negative-semidefinite Hessian matrices, hence excluded from final model.
Table A9
Discrete choice models of changes in frequency of eating-out and leisure activities.
Variable Eat-out (alone) Eat-out (groups) Leisure (alone) Leisure (groups)
Est. t value Est. t value Est. t value Est. t value
ASC increase −1.414 −2.160 −1.623 −2.767 −1.425 −2.552 −2.237 −2.461
ASC no change (ref.) 0.000 0.000 0.000 0.000
ASC decrease 0.148 0.439 0.146 0.428 0.160 0.440 0.000 0.000
Period 2: increase −0.303 −0.427 −1.012 −1.559 −0.209 −0.361 −0.017 −0.018
Period 2: decrease −1.022 −2.910 −1.096 −3.065 −0.792 −2.042 −1.084 −3.164
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
14

Table A9 (continued)
Variable Eat-out (alone) Eat-out (groups) Leisure (alone) Leisure (groups)
Est. t value Est. t value Est. t value Est. t value
Mean cumulative number of infections/100: increase 0.009 0.393 −0.020 −0.737 −0.027 −1.150 −0.045 −1.515
Mean cumulative number of infections/100: decrease 0.011 0.933 0.004 0.372 −0.010 −0.917 −0.005 −0.487
Commuting frequency increase: increase 1.086 2.531 1.551 2.947 0.813 1.770 1.099 2.285
Commuting frequency decrease: decrease 0.325 1.878 0.382 2.337 0.169 1.028 0.377 2.379
PC of perception of degree of self-restriction of others: increase −0.828 −1.160 −0.534 −0.927 −0.182 −0.262 0.441 0.474
PC of perception of degree of self-restriction of others: decrease 0.589 1.418 0.880 1.985 0.803 1.838 1.389 3.209
COVID-19 dread: increase −0.361 −0.736 0.150 0.281 −0.300 −0.619 0.005 0.008
COVID-19 dread: decrease 0.667 2.189 0.676 2.315 0.663 2.220 0.590 2.037
Sample details
Number of observations 720 826 692 858
Number of individuals 360 413 346 429
Number of choice situations 2 2 2 2
Observed choice ratio
Increase 0.08 0.04 0.07 0.04
No change 0.40 0.42 0.40 0.43
Decrease 0.52 0.53 0.53 0.52
Goodness of fit statistics
LL(0) −791.00 −907.45 −760.24 −942.61
LL(C) −651.18 −690.21 −617.02 −720.45
LL(β)−569.02 −586.72 −545.66 −593.79
rho-square (β) 0.28 0.35 0.28 0.37
Adjusted rho-square (β) 0.27 0.34 0.27 0.36
Out-of-sample validation statistics
(Repeated learning test, r = 100, calibration sample = 90%)
Mean percentage correct 66.6% 69.7% 64.9% 72.1%
Mean percentage clearly right (t= 0.5) 66.6% 69.7% 64.9% 72.1%
Mean percentage clearly wrong (t = 0.5) 33.4% 30.3% 35.1% 27.9%
Robust standard errors (cluster errors at the individual level) are used to account for the panel nature of the data.
Bold font indicates coefficients with p-value <0.1.
References
Asahi Shimbun, 2020. (Shincor ona) obaashuuto okirunoka (Will an oversho ot occur?).
https://digital.asahi.com/articles/photo/AS20200405000166.html. (Accessed 10 April
2020). .
Bates, D., Mächler, M., Bolker, B. and Walker, S. (2014) “Fitting linear mixed-effects models
using lme4,”arXiv preprint arXiv:1406.5823.
BBC, 2020. Coronavirus: The World in Lockdown in Maps and Charts. https://www.bbc.com/
news/wor ld-52103 747. (Accessed 31 May 2020). .
Cabinet Secretariat, 2020. CabinetSecretariat: COVID-19 Information and Resources. https://
corona.go.jp/. (Accessed 25 April 2020). .
Carrasco,J. A. and Miller, E. J. (2006) “Exploring the propensity to performsocial activities: a
social network approach,”Transportati on. Springe r, 33(5), pp . 463–480.
Croissant, Y., Millo, G. and others (2008) “Panel data econometrics in R: The plm package,”J.
Stat. Softw., 27(2), pp. 1–43.
Hess, S., Palma, D., 2019. Apollo: A flexible, powerful and customisable freeware package for
choice model estimation and application. J. Choice Model. 32, 100170 Elsevier.
Hess, S., Rose, J.M., Hensher, D.A., 2008. Asymmetric preference formation in willingness to
pay estimates in discrete choice models. Transportation Research Part E: Logistics and
Transportation Review 44 (5), 847–863.
Iwamoto, M., Dodo, H., Yoneda, J., Ishii, F., Goto, H., Ueda, Y., Morie, T.,1989. A study of the
State-Trait Anxiety Inventory (STAI) and its application to noise stress. Nippon Eiseigaku
Zasshi (Japane se Journal of Hygiene) 43 (6), 1116 –1123 The Japanese Society for
Hygiene.
Kim, J., Rasouli, S., Timmermans, H.J., 2018. Social networks, social influence and activity-
travel behaviour: a review of models and em pirical evidence. Transp. Rev. 38 (4),
499–523 Taylor &Francis.
Kraemer, M. U., Yang, C.-H., Gutierrez, B., Wu, C.-H., Klein, B., Pigott, D. M., Du Plessis, L.,
Faria, N. R., Li, R., Hanage, W. P. and others (2020) “The effect of human mobility and
control measures on the COVID-19 epidemic in China,”Science. American Association
for the Advancement of Science, 368(6490), pp. 493–497.
Manski, C. F. (1993) “Identification of endogenous social effects:The reflection problem,”The
review of economic studies. Wiley-Blackwell, 60(3), pp. 531–542.
Parady, G., Takami, K., Harata, N., 2020. Egocentric social networksand social interactions in
the Greater Tokyo Area. Transportat ion, 1–26 https://doi.org/10.1007/s11116-020-
10079-y Springer, in press.
Parady, G. T., Katayama, G., Yamazaki, H., Yamanami, T., Takami, K. and Harata, N. (2019)
“Analysis of social networks, social interactions, and out-of-home leisure activity genera-
tion: Evidence from Japan,”Transportation. Springer, 4 6(3), pp. 53 7–562.
Rogers, R.W., 1983. Cognitive and psycho logical processes in fear appeals and at titude
change: A revised theory of protection motivation. Social psychophysiology: A source-
book. Guilford Press, pp. 153–176.
Shaw, R., Kim, Y., Hua, J., 2020. Governance, technology and citize n behavior in pan-
demic: Lessons from COVID-19 in East Asia. Progress in disaster science. Elsevier,
p. 100090.
Slovic, P., 1987. Perception of risk. Science 236 (4799), 280–285 American Association for
the Advancement of Science.
Spielberger, C. D., Gorsuch, R. L.,Lushene, R., Vagg,P. and Jacobs, G. (1983)“Consulting Psy-
chologists Press,”Palo Alto, CA, 19831983.
Van Bavel, J. J., Baicker, K., Boggio, P. S., Capraro, V., Cichocka, A., Cikara, M., Crockett, M.
J., Crum, A. J., Douglas, K. M., Druckman, J. N. and others (2020) “Using social and be-
havioural science to support COVID-19 pandemic response,”Nat. Hum. Behav.. Nature
Publishing Group, pp. 1–12.
West, R., Michie, S., Rubin, G.J., Amlôt, R., 2020. Applying principl es of behaviour
change to reduce SARS-CoV-2 transmission. Nat. Hum. Behav. 1–9 Nature Publishing
Group.
Wise, T., Zbozinek, T. D., Michelini, G., Hagan, C. C. and others (2020) “Changes in Risk Per-
ception and Protective Behavior During the First Week of the COVID-19 Pandemic in the
United States.”PsyArXiv.
Zhang, J., Litvinova, M., Liang, Y., Wang, Y., Wang, W., Zhao, S., Wu, Q., Merler, S., Viboud,
C., Vespignani, A. and others (2020) “Changes in contact patterns shape the dynamics of
the COVID-19 outbreak in China,”Science. American Association for the Advancement of
Science.
G. Parady et al. Transportation Research Interdisciplinary Perspectives 7 (2020) 100181
15