1
Individual and contextual risk factors for chikungunya virus infection: the
2
SEROCHIK cross-sectional population-based study
3
A. Fred1,2, A. Fianu1, M. Béral3,4, V. Guernier5,6, D. Sissoko7,8, M. Méchain7,8, A. Michault9†,
4
V. Boisson10, B.-A. Gaüzère9,11, F. Favier1‡, D. Malvy7,8‡, P. Gérardin1,12*; SEROCHIK group#
5
1
INSERM CIC1410, CHU Réunion, Saint Pierre, Reunion, France
6
2
Department of Social and Preventive Medicine, School of Public Health, Montreal University,
7
Montreal, Canada
8
3
9
Sainte Clotilde, Reunion, France
UMR 1309 CMAEE “Contrôle des maladies animales, exotiques et émergentes”, CIRAD,
10
4
Ministère de l'Agriculture, Direction Régionale de l’Agriculture et de la Forêt, Dijon, France
11
5
Université de La Réunion, CRVOI “Centre de Recherche et de Veille de l’océan Indien”,
12
CYROI, Sainte Clotilde, Reunion, France
13
6
14
Queensland, Australia
15
7
16
Bordeaux, France
17
8
18
Centre (INSERM U1219, Université de Bordeaux, ISPED), Bordeaux, France
19
9
20
10
Polyvalent Intensive Care Unit, CHU Réunion, Saint Pierre, Reunion, France
21
11
Polyvalent Intensive Care Unit, CHU Réunion, Saint Denis, Reunion, France
22
12
Université de La Réunion, CNRS 9192, INSERM U1187, IRD 249, CHU Reunion, UM 134
23
PIMIT “Processus Infectieux en Milieu Insulaire Tropical”, CYROI, Sainte Clotilde, Reunion,
24
France
Australian Institute of Tropical Health and Medicine, James Cook University, Townsville,
Department of Tropical Medicine and Clinical International Health, CHU Bordeaux,
Infectious Diseases in Low Income Countries (IDLIC), Bordeaux Population Health Research
Bacteriology, Virology and Parasitology lab, CHU Réunion, Saint Pierre, Reunion, France
25 1
26
†
Deceased. ‡ These authors contributed equally.
27 28
Running head title: Individual and contextual risk factors for chikungunya
29 30
* Correspondence and reprints: Dr Patrick Gérardin, INSERM CIC1410, CHU Réunion,
31
GHSR, BP 350, 97448 Saint Pierre, Reunion, France. email:
[email protected]
32 33
Manuscript 32 pages: Summary: 200 words. Body text: 17 pages, 4,487 words, 49 references
34
Figures: 1 (+1 additional). Tables: 3 (+7 additional).
2
35
Summary
36
The purpose of the study was to weigh the community burden of chikungunya determinants on
37
Reunion island. Risk factors were investigated within a subset of 2,101 adult persons from a
38
population-based cross-sectional serosurvey, using Poisson regression models for dichotomous
39
outcomes. Design-based risk ratios and population attributable fractions (PAF) were generated
40
distinguishing individual and contextual (i.e., that affect individuals collectively) determinants.
41
The disease burden attributable to contextual determinants was twice that of individual
42
determinants (overall PAF value 89.5% versus 44.1%). In a model regrouping both categories
43
of determinants, the independent risk factors were by decreasing PAF values: an interaction
44
term between the reporting of a chikungunya history in the neighbourhood and individual house
45
(PAF 45.9%), a maximal temperature of the month preceding the infection higher than 28.5°C
46
(PAF 25.7%), a socio-economically disadvantaged neighbourhood (PAF 19.0%), altitude of
47
dwelling (PAF 13.1%), cumulated rainfalls of the month preceding the infection higher than 65
48
mm (PAF 12.6%), occupational inactivity (PAF 11.6%), poor knowledge on chikungunya
49
transmission (PAF 7.3%) and obesity/overweight (PAF 5.2%). Taken together, these covariates
50
and their underlying causative factors uncovered 80.8% of chikungunya at population level.
51
Our findings lend support to a major role of contextual risk factors in chikungunya virus
52
outbreaks.
53 54
Key words: chikungunya, risk factor, disease burden, population attributable fraction, cross-
55
sectional study.
3
56
Introduction
57
Chikungunya virus (CHIKV) is responsible for chikungunya fever (CHIKF), a dengue-like
58
illness characterized by persistent incapacitating polyarthralgia [1,2]. Between its first
59
description in Tanzania in the early fifties [3], and its re-emergence in Lamu island (Kenya) in
60
2004 and subsequent spread in the Indian ocean in 2005-2006 [4], CHIKF was regarded as a
61
rather harmless neglected tropical disease limited to developing countries. Since the occurrence
62
of atypical and severe forms of CHIKF [5-8], and its more recent spread to Europe [9] and to
63
the Americas [10], CHIKV has increasingly been recognized as a leading threat for public
64
health, through the global expansion of its mosquito vectors, Aedes (Ae) albopictus and Ae
65
aegypti.
66
Between March 2005 and August 2006, Reunion island experienced a major outbreak of
67
CHIKF [11] due to the circulation of a new genotype of the African variant of the virus,
68
renamed Indian ocean lineage, together with the permissive adaption of the local competent
69
vector, Ae albopictus [12]. At the end of the epidemic, a population-based serosurvey,
70
highlighted a huge penetration of the virus in the community with a burden of approximately
71
266,000 clinical cases (attack rate:35%) and 300,000 infected persons (prevalence rate:
72
38.2%) [4,11]. So far, the reasons why transmission was halted at a prevalence of 40% remain
73
to be elicited, as such a figure contradicts prediction of herd immunity that estimated the
74
protective level at above 70% [13].
75
Indeed, the understanding of the community burden of CHIKF is a puzzling problem and
76
constitutes a challenge for epidemiologists and public health stakeholders aiming to implement
77
control measures in the event of future outbreaks.
78
So far, a few population-based cross-sectional serosurveys have proposed the investigation
79
of CHIKF determinants in distinguishing individual and contextual risk factors [14-17]. None
80
of these has been able to guide decision-making based on appropriate public health impact
4
81
measures such as population attributable fractions (PAF). With exception of one study [18], all
82
previous studies have based identification of CHIKF determinants on logistic regression
83
approach taking prevalence odds ratio, best known as the odds ratio (OR), as the “effect
84
measure” of interest [19]. These however represent only proxies of the log-binomial model and
85
prevalence proportion ratio (PPR), the gold standard method and estimator in cross-sectional
86
studies, because the OR often tends to overstate the PPR when the outcome is common [19],
87
and hence the OR is more difficult to interpret (i.e., the former can be interpreted only in terms
88
of strength of association while the latter is multiplicative and can be interpreted in terms of
89
multiplying/dividing the prevalence estimates) [20]. In this context, given the propensity of the
90
OR to overestimate the PPR [21], we chose the Poisson regression model and the incidence rate
91
ratio (IRR), as two proxies of the log-binomial model and PPR, respectively. We also used
92
PAF, free of hypotheses on causation, with the aim of uncovering risk patterns hosting the
93
determinants of increased CHIKV susceptibility to be investigated in further study of causal
94
pathways.
95
The objective of the present study was to weigh the community burden of CHIKF risk factors
96
using IRR and related PAF. Determinants were identified to be from primarily individual or
97
contextual origin, and investigated separately before being pooled. This purpose was chosen to
98
guide further prioritization of the most effective public health interventions for mitigating future
99
outbreaks, based on amendable determinants.
5
100
Methods
101
The SEROCHIK cross-sectional serosurvey was conducted between August 17th and October
102
20th, 2006, shortly after the end of the chikungunya epidemic on Reunion island.
103
A brief description of the study setting is displayed in Appendix 1.
104
Survey design and procedures
105
The number of subjects needed to obtain a representative random sample of the general
106
population was estimated to be 2,640, assuming 35%±2% expected prevalence [4], an alpha
107
risk of 5%, 20% proportion of refusal or absenteeism and a cluster effect.
108
The sample was built using a 2-level probability sampling procedure. At level-1, a random
109
draw for households (primary sampling unit) was carried out. Selection was conducted in six
110
putative strata defined by the crossed combinations of the dwelling type (individual/collective
111
2-20 units/collective >20 units) and size of municipality (≤10,000/>10,000 inhabitants).
112
Overall, five strata were obtained after removing an empty stratum. At level-2, following a
113
predefined Kish method [22], the field investigator randomly assigned one “index-person” (the
114
Kish individual) per household to be interviewed within the selected dwellings. In accordance,
115
a sampling fraction ranging from 0.071 to 1 was generated.
116
To ensure that the sample was representative of the underlying population, we corrected the
117
sample for age, gender, dwelling, and residence area based on socio-demographic information
118
provided by the 2006 census data. Thus, a set of weights was used to increase or decrease the
119
influence of under-/over- represented groups selected from the source population.
120
Data collection
121
Structured questionnaires, administered by 25 field investigators, aimed at collecting data on
122
demographics, health, knowledge on transmission and prevention of CHIKF, dwelling and
123
close environment of residence. These questionnaires are listed in Appendix 2.
6
124
Individual characteristics included gender, age, occupation, chronic disease, body mass
125
index (BMI), four items dedicated to the knowledge of CHIKV transmission (scored 0-4), and
126
eleven preventive behaviours against Ae albopictus (scored 0-11).
127
Contextual household variables included the type and altitude of dwelling, household size,
128
area of residence, recent history of CHIKF in the neighbourhood and a social deprivation proxy-
129
index characterizing the socio-economic status of the municipality at ecological level.
130
The construction of this deprivation index is detailed in Appendix 3.
131
Outcome measure
132
Each participant consented to a fingertip prick and the outcome measured was the result
133
(positive or negative) of CHIKV-specific IgG ELISA serology detected on filter paper-
134
absorbed blood (Whatman 903® Protein Saver™ Card, Schleicher & Schuell) [11]. The full
135
procedure of blood testing is detailed in Appendix 4.
136
Additional climate variables
137
After the geo-referencing of each participant's address at IRIS ("Ilots Regroupés pour
138
informations statistisques") level, the smallest geographical unit for acquiring aggregated socio-
139
environmental information in France, monthly meteorological data gathered from the 21 closest
140
weather stations were obtained to complete the questionnaire data, using Margouill@ website
141
(http://www.margouilla.net).
142
The climatic parameters considered were “cumulated rainfall”, average, maximal and
143
minimal temperatures, and average solar radiation. The rationale for choosing these variables
144
is explained in Appendix 5.
145
For CHIKV-infected (CHIK+) symptomatic subjects, we inputted average values of the
146
closest station from the data recorded in the two months preceding the first CHIKF case
147
included in the IRIS. For CHIKV-naïve (CHIK-) individuals, we used the average values of the
148
closest station between the first and the last cases reported in the same IRIS.
7
149
Statistical analysis
150
Given the subjective nature and need for intelligibility of some individual variables, comparison
151
of outcome was restricted to people aged 15 years or more to ensure reliable data input based
152
on personal information. First, we examined associations between CHIKF (design-based
153
seroprevalence weighted by sampling fraction) and both individual and contextual
154
characteristics. Design-based crude IRR and 95% confidence intervals (CI) were assessed using
155
2 likelihood ratio tests weighted by sampling fraction.
156
Second, we investigated independent risk factors for CHIKF in a 4-step multivariate Poisson
157
regression for dichotomous outcome, modelling the fixed effects (no random intercept to allow
158
contextual effects). In the first step, we fitted two distinct full models with respect to the
159
intrinsic nature of the variables, individual or contextual (grouping within households) with all
160
the relevant covariates found in bivariate analysis. In the “individual-variable” model, we
161
forced gender and age, potentially associated with un-adjustable determinants to minimize
162
residual confounding. In the second step, we fitted one unique parsimonious, “explicative”
163
model with all the putative risk factors previously identified using a manual stepwise backward
164
elimination procedure (P value < 0.05 to be retained in the model). All interaction terms
165
between the variables included in this model were tested using the Mantel-Haënszel method. In
166
the third step, we studied the contribution of climate variables. Finally, in the fourth step, we
167
added the significant climate variables to the precedent “explicative” model. All the covariates
168
within this final “decision-making” model were set as binary. At each step, we determined the
169
pseudo-likelihood of the Poisson regression model using a survey-adjusted Wald test.
170
For all these analyses, our purpose was to weigh the “explicative” part of each model using
171
the combined adjusted population attributable fractions (PAF) for cross-sectional data, which
172
were produced from each specific PAF value (for overall PAF calculations, see Appendix 6).
173
PAF indicates the percentage of cases that would not occur in a population if the factor was
8
174
eliminated. Subsequently, we gauged each key determinant of the final “decision-making”
175
model according to its amendable potential, and classified the risk factors into amendable and
176
non-amendable, if they appeared to be causal.
177
All the analyses were performed using Stata14® (StataCorp. 2015, College Station, USA)
178
excluding observations with missing data.
179
Ethics and funding
180
The SEROCHIK serosurvey was approved by the ethical committee for studies with human
181
subjects of Bordeaux (No 2006/47) and the National Commission for Informatics and Liberty,
182
the French Data Protection Authority. All participants provided their informed consent to
183
answer the questionnaire and for blood collection. Parent or guardian of all child participants
184
provided informed consent on their behalf.
185
The study was funded by the National Institute of Health and Medical Research. The funding
186
source had no role in study design, data collection, analysis and interpretation. The study
187
respected the STROBE statement (Checklist in Appendix 7).
9
188
Results
189
Study population
190
The selection of the study population is presented in the figure 1. Of the 3,032 subjects sampled
191
at level-1, 2,442 Kish individuals were surveyed by field investigators and 2,101 participants
192
aged 15 years or more were analysed. This sample was representative of the habitat in terms of
193
areas of residence and altitude of dwelling place. Beyond these contextual features, it was also
194
representative of individual characteristics such as obesity, hypertension, diabetes, and asthma.
195
Of note, our population was a shifted towards the overrepresentation of females (59.8% versus
196
49.3%, P
