Initial IRS Impact on Mosquito Mortality

The first analysis summarises and compares the initial impact of different IRS products. Data were restricted to initial timepoints collected within 2 months of IRS application as the active ingredient decays with time, so that averaging across the whole dataset may mis-represent the initial potency of IRS as studies had different durations. Statistical models were fit to generate overall estimates of the efficacy of the chemical class. These explanatory factors included the mosquito vectors (classified at the species complex level and species level where possible, i.e. A. arabiensis, A. funestus s.l. and A. gambiae s.l.), experimental hut type (West or East African design) and hut wall substrate (cement or mud) alongside the chemical class used for the IRS (carbamate, clothianidin, organophosphate and pyrethroid). Preliminary data exploration revealed that there were too few data to perform an extensive statistical test on all covariates. To overcome this a subset of the full database was generated by removing Ifakara hut studies, wall substrates that were not mud or cement and chemistries other than pyrethroids, organophosphates, carbamates or neonicotinoids. Binomial logistic regression models were fitted to the remaining count data (N = 78) to estimate the number of mosquitoes that were dead in 24-h, had exited, blood-fed or been deterred by the IRS product. The predicted value for the proportion of mosquitoes being killed, exiting, blood-fed or deterred is calculated as: $${\pi }_i=logi{t}^{-1}ln{\pi }_i∕\left(1-{\pi }_i\right)=\frac{\exp \left({\beta }_0+{\sum }_h{\beta }_h{\mathit{X}}_{\mathit{hi}}\right)}{\left(1-\exp \left({\beta }_0+{\sum }_h{\beta }_h{\mathit{X}}_{\mathit{hi}}\right)\right)}$$ where π_i is the estimated proportion for the ith data (e.g. the proportion of mosquitoes killed), β₀ is the intercept, the subscript h denotes the covariate of interest (taking number of 1 to H) and X_h is a matrix of explanatory factors (mosquito species, hut type, substrate and chemistry sprayed) with coefficients β_h^{59 (link)}. Bayesian models were fitted using Hamiltonian Monte Carlo sampling methods^{60 (link),61} . Four chains were initialised to assess the convergence of 2000 iterations, the first 1000 of each were discarded as burn in. The posterior distributions of parameters (4000 iterations) and 90% Bayesian credible intervals were estimated, posterior checks were performed using ShinyStan (version 1.0.0)⁶² and visually confirmed to fit the data (Supplementary Fig. 2–5).