2.2. Experimental setup

We employed an experimental design in which each LF model was prepared, calibrated and operated by the respective modelling group, following which the relevant simulation outputs from each single model were provided for use in constructing the multi-model LF ensemble. This experimental setup comprised the following steps. First, the three groups were provided with LF baseline infection and post-intervention data from three community sites chosen to represent the vector-mediated transmission dynamics specific to each of the three major LF endemic regions of Africa (primarily Anopheles-mediated transmission), Papua New Guinea – PNG – (Anopheles) and India (Culex) (Singh et al., 2013, Njenga et al., 2008, Rajagopalan et al., 1989, Subramanian et al., 1989, Das et al., 1992, Rajagopalan et al., 1988) (Table 2A and B). The groups were asked to calibrate their models to the baseline microfilariae (mf) age-prevalence data (=“training” data) from these sites, and to provide an ensemble of simulations for the construction and analysis of the multi-model ensemble. Each model aimed to generate 500 fits, or model members, but the number of initial simulations drawn by each group varied from 10,000 (LYMFASIM) to 200,000 (EPIFIL) as a result of differences in the fitting procedures followed and computational intricacies of the three models (see part A of the SI). We deem these as single-model ensembles, which are calibrated and parameterized either using Monte Carlo or stochastic perturbation methods (see parts A and B of the SI). The groups were then asked to use their respective single-model members that fell within the bounds of the weighted multi-model ensemble to carry out simulations of the observed MDA in Malindi, Africa, and Nanaha, PNG, and the effects of the integrated vector management (IVM), which was carried out in Pondicherry, India (Table 2B) (Singh et al., 2013, Njenga et al., 2008, Rajagopalan et al., 1989, Subramanian et al., 1989, Das et al., 1992, Rajagopalan et al., 1988). These simulations were provided for validation against the community mf prevalence data obtained over the durations of the interventions performed in each site (=“validation” data). An overview of the methodology used in this study for calibrating and validating the LF multi-model ensembles constructed for each study site is shown in Fig. 1.

Overview of the methodology used for developing the lymphatic filariasis (LF) multi-model ensemble. All K (=3) single LF models were 1) compared prior to inclusion and 2) trained with baseline data on LF infection (Table 2A) to produce a collection of N_k simulations for each of the three LF model types. Each constituent model was 3) assigned a weight reflecting its relative performance in reproducing the baseline age-mf prevalence data in each site. The weights were used to construct the LF multi-model ensemble. To validate the multi-model ensemble and forecast the effects of LF interventions, 4) simulations were generated by the multi-model ensemble and compared with mf age-prevalence data obtained during the intervention period in each site (Table 2B). The four processes outlined are represented by orange boxes. The different types of models (SM and MM) are represented as blue and green diamonds, respectively, and the corresponding simulation outcomes are represented as rectangles of the same color.

A. Baseline survey data for the three study sites. B. Follow-up survey data for three study sites.

In a second phase of the study, each group was invited to recalibrate their models to a set of overall mf prevalence values thought to represent the currently expected mean infection levels in the regions of Africa, PNG and India, and to provide both their fitted single-model ensembles at baseline and resulting predictions of the effects of MDA in the form of timelines to cross below the WHO-set LF elimination target of 1% mf for each of these scenarios. These single-model ensembles and predictions were then combined using the ensemble construction methodology developed in this study to present an analysis of a multi-model-based generation of these timelines in each region compared to those predicted by each single model.