Table 1.
Annual mf prevalence survey and MDA data for three LF endemic sites.
Fig 1.
Schematic diagram showing the sequential fitting procedure for updating models and predictions by incorporating longitudinal data.
In all scenarios, the initial EPIFIL models were initialized with parameter priors and a chi-square fitting criterion was applied to select those models which represent the baseline mf prevalence data sufficiently well (α = 0.05). The accepted models were then used to simulate the impact of interventions on mf prevalence. The chi-square fitting criterion was sequentially applied to refine the selection of models according to the post-MDA mf prevalence data included in the fitting scenario. The fitted parameters from selection of acceptable models at each data point were used to predict timelines to achieve 1% mf prevalence. The scenarios noted in the blue boxes indicate the final relevant updating step before using the fitted parameters to predict timelines to achieve 1% mf in that data fitting scenario.
Table 2.
Model predictions of timelines to achieve 1% mf prevalence and corresponding information metrics.
Fig 2.
Comparison of the distributions of predicted timelines to LF elimination from the three models for Kirare, Tanzania.
This visual comparison shows that the predictions coming from the model-only simulations (scenario 0) have the widest spread in their distributions for all three models compared to model predictions obtained via constraining using subsequent data scenarios. Pairwise Kolmogorov-Smirnov tests for equal distributions were performed on the results from each model to evaluate whether updating the models with sequential data changed the distribution of predictions. Significance was determined using the Benjamini-Hochberg procedure for controlling the false discovery rate (q = 0.05). Apart from scenarios 2 and 3 for EPIFIL and scenarios 3 and 4 for LYMFASIM, all distributions were significantly different from one another (see S2 Supplementary Information for results from the villages of Alagramam and Peneng).
Fig 3.
Comparison of model-predicted timelines from model-only simulations and the lowest entropy simulations in each site.
The boxplots show that by calibrating the models to data streams, more precise predictions are able to be made regarding timelines to achieve 1% mf prevalence across all models and sites. The results of pairwise F-tests for variance, performed to compare the weighted variance in timelines to achieve 1% mf prevalence between model-only simulations (scenario 0) and the lowest entropy simulations (best scenario) (see Table 2), show that the predictions for the best scenarios are significantly different from the predictions for the model-only simulations. Significance was determined using the Benjamini-Hochberg procedure for controlling the false discovery rate (q = 0.05). For EPIFIL, LYMFASIM and TRANSFIL, the best scenarios are scenarios 4, 3, and 4 for Kirare, scenarios 4, 4, and 3 for Alagramam, and scenarios 3, 3, and 3 for Peneng, respectively.
Fig 4.
Pooled predictions of the timelines to reach 1% mf from three LF models.
The shaded regions show the weighted 95% percentile interval from the composite predictions of all three models of the timelines required to cross the WHO 1% elimination threshold for all five scenarios. The black dots indicate upper and lower bounds (weighted 2.5th and 97.5th percentiles) of the composite predictions from all three models for each scenario. The range of predictions is tightest when the models were constrained with data from scenarios 3 and 4.
Fig 5.
Parameter constraint achieved through the coupling of EPIFIL with data.
Overall parameter constraint was measured as the ratio of the mean standard deviation of the fitted parameter distributions to that of the prior parameter distributions. Values < 1 indicate that the fitted parameter space was constrained compared to the prior parameter space. The results show that the fitted parameter space for Kirare and Peneng was more constrained by calibrating the model to data compared to the model-only scenario, but this was not the case for Alagramam.
Table 3.
Spearman parameter correlations for scenarios 1 (lower left triangle) and 3 (upper right triangle) for Alagramam, India.
Fig 6.
Weighted variance and entropy values of EPIFIL predictions of LF elimination timelines using optimal MDA coverage in each study site.
For all sites, either scenario 3 or 4 had the lowest entropies, and scenario 4 was not significantly different from scenario 3 for Kirare and Alagramam. These results were not statistically different from the results given 65% coverage (see Table 2), suggesting that the data stream associated with the lowest entropy is robust to changes in the interventions simulated. Scenarios where the weighted variance or entropy were not significantly different from the lowest entropy scenario are noted with the abbreviation NS. Significance was determined using the Benjamini-Hochberg procedure for controlling the false discovery rate (q = 0.05).
Table 4.
Predictions of timelines to achieve 1% mf in Villupuram district, India, considering extended post-MDA data.
Table 5.
Annual mf prevalence survey and MDA data for Dokan Tofa, Nigeria.
Table 6.
EPIFIL predictions of timelines to achieve 1% mf prevalence in Dokan Tofa, Nigeria.