Study area and vector species
Mosquito trapping data were collected from six adjacent villages in the Ulanga and Kilombero districts of south-eastern Tanzania, namely: Kivukoni (8.2135°S, 36.6879°E), Minepa (8.2710°S, 36.6771°E), Mavimba (8.3124°S, 36.6771°E), Milola (8.3306°S, 36.6727°E), Igumbiro (8.3511°S, 36.6725°E) and Lupiro (8.385° S, 36.670°E). Data were collected over 12 months between 2015 and 2016. The valley has relatively high mosquito abundance which peaks at the end of the rainy season. The common vectors of malaria transmission are Anopheles arabiensis and Anopheles funestus [16, 24, 52]. Mosquitoes in the Culex genera are also highly abundant, with some species being potential vectors for arboviruses found in the study area [53, 54].
Data collection
Mosquito sampling was carried out in both the wet and dry seasons, using six different traps for sampling outdoor-biting mosquitoes around human dwellings. The traps were: Mosquito Magnet trap (MMX) [55], BG-Sentinel trap (BGS) [56], Suna trap (SUN) [3], Ifakara Tent Trap-C (ITT-C) [48], M-Trap (MTR) [57], M-Trap fitted with CDC Light trap (MTRC) (this study) and the Human Landing Catches (HLC) [3]. Most of these traps have been extensively described elsewhere except for the MRT fitted with a CDC light trap (MTR), which was adapted from the original exposure-free M-trap designed by Mwangungulu et al [57]. In this current study, the original MRT was divided into two compartments made of UV-resistant shade netting: one in which a human volunteer sat to attract mosquitoes and the other section in which mosquito are entered [57]. A CDC light trap was suspended inside the other section of the trap to attract more mosquitos to the light source.
The traps were located at least 100m apart. Initial trap allocation was random, but their positions were switched over successive sampling nights in a Latin square design. This way each trap was used in each position once over a seven night cycle. After completion of each cycle, the study team moved to the next village so that one round of sampling in all six villages was completed over 42 trap-nights. Six rounds of data collection were completed spanning the wettest and the driest periods of the year (252 trap nights between April 2015 and April 2016). Mosquito sampling was done overnight from 6pm to 6am. The collected mosquitoes were morphologically sorted by taxa. A subsample of An. gambiae s.l. (n = 1,405, 26% of total) were analyzed by PCR [58] to assess sibling species composition within the complex.
Model fitting
The main goal of our analyses was to create a calibration tool to evaluating outdoor mosquito traps and to validate the tool by comparing the performance of candidate trapping methods relative to HLC, being “gold standard”. In particular, we wanted to test the shape of the association between the numbers of mosquitoes collected by each trap type with those collected by the HLC. First, we pooled all the hourly collections into a single collection cup per trap per night. Then, for each of the focus mosquito groups (Culex genera, An. arabiensis and An. funestus s.l), we modeled HLC catches as a function of the catching rate of each alternative trap.
Four general linear models were developed within a Bayesian model fitting framework to allow us to test for linear and non-linear associations through increasing the levels of complexity. The Bayesian approach allowed specific constraints on the parameters based on biological plausibility; in the form of priors and uncertainty when converting the counts from alternative traps into HLC equivalent values in the form of full posteriors.
For any given trap and mosquito group, we defined the response variable (\({N}_{i})\) as the number of female mosquitoes on every \({i}^{th }\)sampling night. Preliminary investigation of data using Poisson likelihood showed over-dispersion for all the three mosquito groups. Our final models did not account for other environmental covariates at specific trap locations (e.g. temperature, humidity). We accounted for the over dispersion by using a negative binomial likelihood model formulated as a Gamma-Poisson mixture distribution [59]:
with
where the Poisson rate λi is defined by the shape of the relationship between \({N}_{i}\) and the number of mosquitoes collected with the alternative trap (\({n}_{i}, \text{T}\text{a}\text{b}\text{l}\text{e} 1\)).
Since the algebraic form of this relationship is not known, we made three mutually inclusive assumptions with specified mathematical definitions, as follows: 1) that the relationship must start at the origin (i.e. when HLC catches zero mosquitoes, the other traps will, on average also collect zero mosquitoes), 2) that the relationship is positive (i.e. no negative relationships between trap catches), and 3) that any given trap could potentially suffer from a density effect (i.e. the slope of the relationship is not constant and it can change according to the baseline abundance of mosquitoes, either just of the same mosquito group or of all mosquitoes).
To define λi we therefore formulated four possible scenarios to describe the relationship between HLCs and other trapping methods as summarized in Table 1 and Fig. 1. In Model 1, we considered a simple linear relationship between \({N}_{i}\) and \({n}_{i}\) (Table 1, Fig. 1A). In Model 2 we tested if the efficiency of the alternative trap was dependent on the density of the focal mosquito (e.g. “intra-specific” density dependence) by adding a quadratic term \({{n}_{i}}^{2}\) (Table 1, Fig. 1B). In Model 3 we tested if the captures of a given group by a given trap were dependent on the abundance of the other taxonomic groups (e.g. “inter-specific” density dependence) by adding, an interaction term between \({n}_{i}\) and the number of all the females from other mosquito groups collected with the same trap (\({m}_{i}\)) (Table 1, Fig. 1C). Model 4 was similar to Model 3, but we considered all the other \({K}_{i}\) taxonomic groups separately. Therefore it included all the pair wise interaction terms between \({n}_{i}\) and the number of females of each \({k}^{th}\) mosquito group \({(s}_{{k}_{i}})\) (Table 1, Fig. 1D). Our analysis mainly focused on three mosquito groups, but we collected a higher number of species hence \(K>3\) (S1 Additional file).
Table: 1 – Description of models used to investigate the relationships between female mosquito catches by human landing catch and the alternative traps.
The analysis was performed in the statistical environment R [60], with Bayesian model fitting to the data done using the program JAGS [61] interfaced within R via the package rjags [62]. For parameters\({ \beta }_{1}\), \({\beta }_{2}\) and \({\beta }_{k}\) we used a gamma prior (shape = 0.1, rate = 0.1). The prior for \({ \beta }_{1}\) was chosen to ensure a positive relationship between \({n}_{i}\) and\({N}_{i}\) and a positive effect of the quadratic and the interaction terms for \({\beta }_{2}\) and\({ \beta }_{k}\). To achieve convergence, the models were run for up to \(3 x {10}^{4}\) iterations. Means of posterior distributions with corresponding credible intervals were obtained for each model coefficient\(\beta\). We compared different models by their Deviance Information Criteria (DIC) and the goodness of fit of each model using pseudo \({R}^{2}\) values. Models with the lowest DIC were selected as best.
Interactive calibration tool
We designed a lookup table (Table 3) containing means of posterior predictions for different combinations of mosquito taxa, trap types and models. This allowed us to predict the expected number of a given mosquito taxa from an HLC (with credible intervals) based on the number caught in the alternative traps. We also developed an interactive online tool, in the form of an R Shiny App [63] to facilitate these evaluations. This tool provides users with an interactive graphical user interface (GUI) to select the number of captured mosquitoes for a group of interest by trap type, and to explore the predicted number of mosquitoes caught in an HLC by method.