We used a data-model assimilation technique based on the Bayesian Melding (BM) algorithm to calibrate and identify locality-specific LF transmission models based on the baseline mf prevalence and vector biting intensity data observed in each of our study sites9 (link). This is done by “melding” the observed baseline data from each site (Supplementary Table 2 ) with model-generated outputs in order to learn or parameterize models for describing the localized parasite transmission dynamics. The fitted models from each site were then used to quantify the various quantities of interest to this study, viz. estimations of mf thresholds and threshold biting rates, predictions of the impact of various MDA interventions with and without vector control on mf prevalence, and calculations of the probabilities of transmission interruption and recrudescence from using the WHO-set TAS thresholds versus model-derived threshold values once mf prevalences are forecast to cross below these thresholds in each study site.
The BM procedure begins by first specifying a range of plausible parameter values to generate distributions of parameter priors. We then randomly sample from those prior distributions to generate 200,000 parameter vectors, which are then used with the observed ABR in a site to generate predictions of baseline age-specific prevalences. The Sampling Importance Resampling (SIR) algorithm is then used to select N (typically N = 500) parameter vectors, θ, or models applicable to a site based on their likelihoods for describing the observed local baseline prevalence data. This BM fitting procedure normally relies on observed baseline age profiles of mf prevalence9 (link), but, in the present analysis, these data were available only for DoakanTofa and Piapung, while overall community-level mf prevalences were available for the other sites (Mossasso, Kirare, Peneng, Dozanso). In this scenario, the observed overall prevalences from these sites were transformed into theoretical age-infection profiles using: (1) the national demographic profile applicable to the site in question, and (2) by conversion of the community-level mf prevalence to reflect either a plateau, concave or linear age-infection profile known typically to occur in LF endemic regions15 (link). The derived age-prevalence infection data were then used in the model fitting procedures described above, which also effectively allowed the integration of partially observed data into the present LF model.
The BM procedure begins by first specifying a range of plausible parameter values to generate distributions of parameter priors. We then randomly sample from those prior distributions to generate 200,000 parameter vectors, which are then used with the observed ABR in a site to generate predictions of baseline age-specific prevalences. The Sampling Importance Resampling (SIR) algorithm is then used to select N (typically N = 500) parameter vectors, θ, or models applicable to a site based on their likelihoods for describing the observed local baseline prevalence data. This BM fitting procedure normally relies on observed baseline age profiles of mf prevalence9 (link), but, in the present analysis, these data were available only for DoakanTofa and Piapung, while overall community-level mf prevalences were available for the other sites (Mossasso, Kirare, Peneng, Dozanso). In this scenario, the observed overall prevalences from these sites were transformed into theoretical age-infection profiles using: (1) the national demographic profile applicable to the site in question, and (2) by conversion of the community-level mf prevalence to reflect either a plateau, concave or linear age-infection profile known typically to occur in LF endemic regions15 (link). The derived age-prevalence infection data were then used in the model fitting procedures described above, which also effectively allowed the integration of partially observed data into the present LF model.
Free full text: Click here
