Monitoring Public Behavior During a Pandemic Using Surveys Proof-of-Concept Via Epidemic Modelling

2025-05-02 0 0 1.81MB 41 页 10玖币

侵权投诉

Monitoring Public Behavior During a

Pandemic Using Surveys: Proof-of-Concept

Via Epidemic Modelling

Andreas Koher1, Frederik Jørgensen2, Michael Bang

Petersen2and Sune Lehmann1,3*

1DTU Compute, Technical University of Denmark.

2Department of Political Science, Aarhus University.

3Center for Social Data Science, University of Copenhagen.

*Corresponding author(s). E-mail(s): sljo@dtu.dk;

Abstract

Implementing a lockdown for disease mitigation is a balancing act:

Non-pharmaceutical interventions can reduce disease transmission signif-

icantly, but interventions also have considerable societal costs. Therefore,

decision-makers need near real-time information to calibrate the level

of restrictions. We ﬁelded daily surveys in Denmark during the sec-

ond wave of the COVID-19 pandemic to monitor public response to

the announced lockdown. A key question asked respondents to state

their number of close contacts within the past 24 hours. Here, we estab-

lish a link between survey data, mobility data, and hospitalizations

via epidemic modelling. Using Bayesian analysis, we then evaluate the

usefulness of survey responses as a tool to monitor the eﬀects of lock-

down and then compare the predictive performance to that of mobility

data. We ﬁnd that, unlike mobility, self-reported contacts decreased

signiﬁcantly in all regions before the nation-wide implementation of

non-pharmaceutical interventions and improved predicting future hos-

pitalizations compared to mobility data. A detailed analysis of contact

types indicates that contact with friends and strangers outperforms con-

tact with colleagues and family members (outside the household) on the

same prediction task. Representative surveys thus qualify as a reliable,

non-privacy invasive monitoring tool to track the implementation of non-

pharmaceutical interventions and study potential transmission paths.

Keywords: Epidemic monitoring, Mobility data, Survey data, Epidemic

modelling

arXiv:2210.01472v2 [physics.data-an] 31 Jan 2023

21 INTRODUCTION

1 Introduction

Pandemic management is a balancing act. When an outbreak of infections

ﬂares up, governments and authorities need to impose restrictions and recom-

mendations on society that are carefully calibrated to the situation. On the

one hand, during the COVID-19 pandemic, such non-pharmaceutical interven-

tions have considerable beneﬁts by changing the dominant transmission route

– close contacts between individuals – via the incentives and information they

provide [1,2]. On the other hand, these interventions have considerable costs

in the form of negative externalities relating to the economy and mental health

[3–5].

This balancing act puts authorities and governments in need of informa-

tion to continuously calibrate the level of restrictions. It is not a matter of

simply sending out a single set of instructions regarding restrictions and rec-

ommendations. Rather, authorities need to continuously receive information

about the eﬀectiveness of those restrictions and recommendations and adjust

accordingly. An obvious source of information is directly related to the epi-

demic and includes the number of infection cases, hospitalizations, and deaths.

Yet cases of infection are diﬃcult to monitor and, for example, changes in the

public’s motivation to participate in testing programs may create problems

with respect to comparisons over time [6]. Furthermore, there is a signiﬁcant

lag between the onset of interventions and hospitalizations and death counts,

which imply that it is diﬃcult to calibrate interventions on the basis of such

information. Consequently, researchers, authorities and governments world-

wide have complemented epidemiological information with information on the

direct target of the interventions: Behaviour [7,8].

In this manuscript, we assess the predictive performance of a particular

source of information about behavior during lockdowns: Population-based sur-

veys on social contacts, ﬁelded daily to representative samples of the Danish

population during the COVID-19 pandemic (see Methods for details on this

dataset). This assessment aligns with recommendations about the use of sur-

veys as epidemic monitoring tools on the basis of experiences during the SARS

epidemic in Hong Kong [9] and recommendations from the World Health Orga-

nization during the COVID-19 pandemic [10]. From a public health policy

perspective, this particular dataset is a unique test case as it was, in fact,

reported to the Danish government for this purpose on a twice-weekly basis

during the second wave of COVID-19 infections in December 2020.

Furthermore, these data are unique in another respect: They constitute an

open and ‘citizen science’ [11] alternative to the most used source of informa-

tion on pandemic behavior: Mobility data. As we detail below, mobility data

as a source of information may be problematic from both a methodological and

policy perspective. Mobility data provides a proxy for close contacts between

people and has been heavily utilized by researchers and public health institu-

tions [8,12–14]. Mobility data quantiﬁes the population’s movement patterns

and is unobtrusively obtained in a number of ways, for example, via people’s

2020-12-01 2020-12-15 2021-01-01 2021-01-15 2021-02-01 2021-02-15

0.6

0.9

1.2

1.5

1.8

Reproduction Number Rt

lockdown announced

partial lockdown

full lockdown

95% CI

median

2020-12-01 2020-12-15 2021-01-01 2021-01-15 2021-02-01 2021-02-15

−60

−40

−20

relative change [%]

Self-Reported Survey Data

40th %tile (3 contacts)

50th %tile (5 contacts)

60th %tile (7 contacts)

70th %tile (10 contacts)

80th %tile (15 contacts)

90th %tile (25 contacts)

2020-12-01 2020-12-15 2021-01-01 2021-01-15 2021-02-01 2021-02-15

−60

−40

−20

relative change [%]

Mobility

Self-Reported Survey Data

Google Mobility Data

Apple Mobility Data

Telco Mobility Data

Fig. 1 Panel A: inferred reproduction number from national hospitalizations. Panel B:

Comparison between thresholds that deﬁne risk-taking behaviour: The percentile gives a

number of contacts nthat deﬁnes risk-taking behaviour. The time-series present the daily

fraction of individuals P(#total contacts ≥n) that report at least ncontacts. Panel C:

Comparison between risk-taking behaviour with a threshold at the 70th percentile (self-

reported survey data), Google mobility, Apple mobility, and telecommunication data (Telco).

smart phones and provided to researchers and governments via private compa-

nies such as Google [15]. This reliance, however, can and has raised concerns.

First, in many cases, it implies that pandemic management and research relies

on the willingness of private companies to share information during a critical

crisis. Second, citizens themselves may be concerned about real or perceived

privacy issues related to the sharing of data with authorities [16,17]. Given

the importance of public trust for successful pandemic management [18], such

concerns – if widespread – can complicate pandemic control. Third, data from

companies such as Google, Facebook and local phone companies may not be

representative of the population of interest: The entire population of the coun-

try. Rather than being invited on the basis of traditional sampling methods,

people opt-in to the services of diﬀerent companies and, hence, the data from

any single company is likely a biased sample. Fourth, the movements of peo-

ple in society as captured by mobility data is only a proxy of the quantity of

interest: Actual close encounters between individuals that drive the pandemic.

For these reasons, it is key to assess alternative sources of information

about public behavior such as nationally representative surveys of the adult

population. In principle, surveys could alleviate the problems relating to the

collection and validity of mobility data. Survey research is a centuries old

41 INTRODUCTION

low-cost methodology that can be utilized by public actors and that relies on

well-established procedures for obtaining representative information on private

behaviours in voluntary and anonymous ways [19].

At the same time, data from surveys come with their own methodological

complications. As documented by decades of research, people may not accu-

rately report on their own behaviour [20]. Survey answers during the pandemic

may be biased by, for example, self-presentational concerns and inaccurate

memory. While research on survey reports of behaviour during the pandemic

suggests that self-presentational concerns may not aﬀect survey estimates [21],

memory biases may (although such biases are likely small for salient social

behavior) [22]. Even with such biases, however, surveys may be fully capable

to serve as an informative monitoring tool. The key quantity to monitor is

change in aggregate behaviour over time. If reporting biases are randomly dis-

tributed within the population, aggregation will provide an unbiased estimate.

Even if this is not the case, changes in the survey data will still accurately

reﬂect changes in population behaviour as long as reporting biases are stable

within the relevant time period.

On this basis, the purpose of the present manuscript is, ﬁrst, to examine

the degree to which survey data provide useful diagnostic information about

the trajectory of behavior during a lockdown and, second, to compare its use-

fulness to information arising from mobility data. To this end, we focus on

a narrow period around Denmark’s lockdown during the second wave of the

COVID-19 epidemic in the Fall of 2020, i.e., prior to vaccine roll-out when it

was crucial for authorities to closely monitor public behavior. We demonstrate

the usefulness of survey data on a narrow window of time because the chang-

ing nature of factors such as seasonal eﬀects, new variants, vaccines, changing

masking eﬀorts, etc., make it diﬃcult to model COVID-19 transmission across

long periods without making a large number of assumptions [6]. See also Sec. 3

for a discussion on the limitations of our survey data. In spite of the limited

scope, we believe that the study remains relevant for policy makers because

it allows to monitor public behaviour at a crucial moment, when policy mak-

ers should not be forced to rely on proximity or mobility data from private

companies in the absence of timely incidence data.

Speciﬁcally, we ask whether a) daily representative surveys regarding the

number of close social contacts and b) mobility data allow us to track changes

in the observed number of hospitalizations in response to the lockdown.

In addition, to further probe the usefulness of survey data, we provide a

ﬁne-grained analysis of how diﬀerent types of social contacts relate to hospital-

izations. Our results shed new light on the usefulness of survey data. Previous

studies during the COVID-19 pandemic have documented high degrees of over-

lap between self-reported survey data on social behavior and mobility data,

but have not assessed whether these data sources contain useful information

for predicting transmission dynamics [23,24]. One study did compare the pre-

dictive power of mobility data to survey data on the psychosocial antecedents

of behavior [25] and found that mobility data was more predictive than the

survey data of COVID-19 transmission dynamics. Here, we provide a more

balanced test by comparing the predictive value of mobility data and survey

data when directly focused on self-reported behavior rather than simply its

psychosocial antecedents.

2 Results

We establish the link between survey data, mobility data, and hospitalizations

via state-of-the-art epidemic modeling, which uses the behavioural survey and

mobility data as an input to capture underlying infectious activity [26,27].

Speciﬁcally we extend the semi-mechanistic Bayesian model from Flaxman et

al. [27,28] to jointly model the epidemic spreading within the ﬁve regions of

Denmark. Where possible, we use partial pooling of parameters to share infor-

mation across regions and thus reduce region speciﬁc biases. We parametrize

the regional reproduction number Rtwith a single predictor Xtfrom our survey

or the mobility data, respectively, for each realization of a model:

log(Rt) = log(R0) + eXt(1)

The regional reproduction number at time tderives from the initial value R0

and the scaled predictor eXtwith a logarithmic link-function (see Methods for

full details on the model).

We compare the predictive performance of each data stream using leave-

one-out cross-validation (LOO). LOO works by ﬁtting the model to the

observed hospitalizations excluding a single observation and comparing the

prediction of the unseen observation against the observed real-world data.

Repeating this process over all observations, allows one to estimate the model

performance on out-of-sample data with a theoretically principled method that

accounts for uncertainties [29]. In practice, this would result in an immense

computational eﬀort and therefore, we use an eﬃcient estimation of LOO

based on pareto-smoothed importance sampling [30]. In order to compare the

predictive performance of, say self-reported survey against mobility, we calcu-

late the LOO score for each model parametrization and consider the diﬀerence

signiﬁcant if it exceeds the 95% CI.

Because we are interested in the use of behavioural data as a guide

for decision-making, our inference focuses on the key period of the second

wave from 1-December-2020, i.e., about one week before Denmark’s lockdown

announcement, to 20-February-2021 when vaccinations accelerated across the

country. The period captures a sharp increase and eventual decline in hos-

pitalizations during the second wave of Denmark’s Covid-19 pandemic (see

Supplementary Fig. S1).

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

MonitoringPublicBehaviorDuringaPandemicUsingSurveys:Proof-of-ConceptViaEpidemicModellingAndreasKoher1,FrederikJrgensen2,MichaelBangPetersen2andSuneLehmann1,3*1DTUCompute,TechnicalUniversityofDenmark.2DepartmentofPoliticalScience,AarhusUniversity.3CenterforSocialDataScience,UniversityofCopenhagen.*C...

展开>> 收起<<

Monitoring Public Behavior During a Pandemic Using Surveys Proof-of-Concept Via Epidemic Modelling.pdf

共41页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Monitoring Public Behavior During a Pandemic Using Surveys Proof-of-Concept Via Epidemic Modelling

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: