Last updated:2022-02-10
Two sites (ravines) in Saint-Joseph municipality, southern Reunion Island
Langevin: control site
Payet: releases of sterile male mosquitoes (Ae. aegypti) coated with piriproxyfen, starting end of April 2021
Distance between Payet and Langevin site: 1,300 m
12 CO2-baited BG traps, laying on the ground: 5 in Langevin, in Payet)
Results reported from 404 trapping sessions totaling 1,785 Ae. aegypti
Study period cut into 8 equal-length periods of 27.5 days (\(\simeq\) four weeks)
Time periods (27.5 days) from 15 Dec. 20 to 21 Jul. 21
|
||||||||
---|---|---|---|---|---|---|---|---|
1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | |
Langevin | 20 | 20 | 20 | 20 | 15 | 10 | 20 | 17 |
Payet | 31 | 28 | 28 | 28 | 21 | 42 | 27 | 28 |
Goal
assess the effect of boosted SIT on the standardized apparent density (SAD) of an Ae. aegypti population
H0: constant SAD after starting boosted SIT
H1: decreasing trend in SAD, related to boosted SIT
Challenges
High variability in count data
two study sites (Payet and Langevin), with ecological differences
Boosted SIT only implemented in Payet, during a short time period (3 months)
Strong seasonal - as well as trap-to-trap, variations
Strategy
Choice of the outcome: apparent density of Ae. aegypti, standardized with respect to time period to eliminate other time trend than boosted SIT effect
Choice of a model to assess the boosted SIT effect, while providing useful information on spatial variation sources: we used a spatial (mixed-effect) Poisson model (Moraga 2019, chap. 6)
Goal
Method
In each site \(s\) (Langevin or Payet), for each period \(i\) (1 to 8), the expected apparent density \(E_{s,i}\) was the average of trap-level apparent density
For each time period \(i\), \(E_{s,i}\) was divided by the expected density in Langevin \(E_{L,i}\) to get \(E_i\), the relative expected density. The average standardized apparent density (SAD) in Langevin, as well as in Payet before starting booster SIT (5th period), was a series of 1’s
Payet data included the boosted SIT effect for periods 6 to 8, which cases, we took as the density expectations in Payet, the expected density in Langevin, multiplied by the ratio of expected densities in Payet and Langevin when starting boosted SIT
Between-trap differences, according to their location
Trap location did not change, and traps were close to each other: 58 m on average in Langevin (from 0 to 177 m), vs 92 m in Payet (0 to 177 m)
model should account for two sources of spatial variations
trap location (habitat suitability…)
trap proximity (flight range…)
To capture this correlation, site areas split into Voronoi polygons to compute neighborhood matrices used in the model
Bayesian modelling framework, with an integrated nested Laplace approximation (INLA) approach: Bayesian process approximated by analytical solutions, implemented in INLA (Rue et al. 2017)
For period \(i \in 1,..., I\\,\,\) (\(I=8\,\,\)), the response \(y_i\,\,\) is a length-\(J\,\,\) vector of standardized apparent densities (SAD), with J,,$ the number of traps in the design (\(J=12\,\,\)). The \(J\,\,\) components of \(y_i\,\,\) are \(a_{i,j} / E_i\), i.e., the apparent density divided by the expected density \(E_i\); i.e., the average density during period \(i\,\,\) for the \(J\,\,\) traps
We used a spatial Poisson model
\[y_i \sim \mathcal{P}(E_i \times m_i)\,\,\text{and}\,\, \log(m_i) = \eta_i = b \times x_i + u_j\]
\(m_i\,\,\) the expectation of \(y_i\), \(x_i\,\,\) the number of exposure months to boosted SIT:1, 2; or 3 for traps in Payet from periods 6 to 8, and 0 elsewhere (traps in Langevin at any period, and in Payet before the 6th period); \(b\,\,\) is the fixed-effect coefficient for \(x_i\)
\(u_j\,\,\) a \(J\)-length vector of random-effect coefficients (one per trap), representing the spatial variation around \(b\,\,\), with \(\sum_{j=1}^J u_j = 0\). We used a modified version of the Besag-York-Mollié model Riebler et al. (2016): \(u_j\,\,\) is split into (i) a spatially-structured component with a conditional auto-regressive (CAR) distribution, and (ii) a “white noise” component. A mixing parameter \(\Phi\) - ranging from 0 to 1, gives the CAR proportion in \(u_j\). Interpretable parameters and meaningful priors for their precision are provided by the modified version
Deviance information criterion (DIC) used to compare four assumptions about boosted SIT effect on SAD
a constant-slope model (presented above)
a constant-effect model: initial drop, then constant SAD
a three-slope model: one for each period
With Mixed-effect models, the estimation of population means must account for both fixed and random effects. Thus, for the one-slope model
\[ E(y_i) = m_i = \frac{1}{J} \left[ \sum_{j=1}^J \exp(b \times x_i + u_j) \right]\]
\[n \geq \frac{log(\alpha)}{\log(m_1)}\]
bym2
random effect and its two components to identify traps with outstanding effects
Mean and 95% credible interval
|
|||||||
---|---|---|---|---|---|---|---|
Fixed effect | Random effect | DIC | Delta DIC | SAD (1 month) | Low | High | Visualization |
Constant decay | BYM2 | 349.9 | 0.32 | 0.16 | 0.57 | ||
Constant decay | IID | 350.6 | 0.7 | 0.33 | 0.17 | 0.58 | |
Constant effect | BYM2 | 352.8 | 2.9 | 0.22 | 0.10 | 0.43 | |
Constant effect | IID | 353.5 | 3.6 | 0.23 | 0.11 | 0.44 | |
Eime-varying decay | BYM2 | 375.4 | 25.5 | 0.35 | 0.16 | 0.69 | |
Eime-varying decay | IID | 376.1 | 26.2 | 0.36 | 0.17 | 0.70 | |
Constant SAD | BYM2 | 377.9 | 28.0 | 0.41 | 0.20 | 0.71 | |
Constant SAD | IID | 378.6 | 28.7 | 0.42 | 0.22 | 0.72 |
95% credible interval
|
||||
---|---|---|---|---|
Boosted SIT exposure (months) | Fitted SAD (%) | Lower limit | Upper limit | Visualization |
0 | 1.03 | 0.88 | 1.19 | |
1 | 0.57 | 0.43 | 0.72 | |
2 | 0.32 | 0.19 | 0.49 | |
3 | 0.18 | 0.08 | 0.35 |
Besag, Julian, Jeremy York, and Annie Mollié. 1991. “Bayesian Image Restoration, with Two Applications in Spatial Statistics.” Annals of the Institute of Statistical Mathematics 43 (1): 1–20. https://doi.org/10.1007/bf00116466.
Moraga, Paula. 2019. Geospatial Health Data: Modeling and Visualization with r-INLA and Shiny. Chapman; Hall/CRC. https://doi.org/10.1201/9780429341823.
Riebler, Andrea, Sigrunn H Sørbye, Daniel Simpson, and Håvard Rue. 2016. “An Intuitive Bayesian Spatial Model for Disease Mapping That Accounts for Scaling.” Statistical Methods in Medical Research 25 (4): 1145–65. https://doi.org/10.1177/0962280216660421.
Rue, Håvard, Andrea I. Riebler, Sigrunn H. Sørbye, Janine B. Illian, Daniel P. Simpson, and Finn K. Lindgren. 2017. “Bayesian Computing with INLA: A Review.” Annual Reviews of Statistics and Its Applications 4 (March): 395–421. https://doi.org/10.1146/annurev-statistics-060116-054045.