The major result of this study of immediate import to LF elimination is our finding of the occurrence of significant differences in the population dynamics and the resulting transmission breakpoint estimates between the nine endemic villages investigated. Although differences in the transmission dynamics of this parasitic disease have been investigated before, they have primarily focused on uncovering the impact of a priori
proposed drivers of such differences, such as community vector biting rates and acquired immunity [20
]. However, this study is the first to use empirical data to disclose the key transmission parameters that underlie observed site-specific differences in filarial transmission dynamics and the resulting endpoints for terminating parasite transmission. The results highlight two important conclusions on this point. First, they support previous results, from sensitivity analyses of our model, that transmission breakpoints in each of the two major filarial infection systems are likely to be highly sensitive to variations in site-specific ecological factors underlying infection dynamics. Second, they confirm that such factors are primarily related to the degree of infection aggregation, as well as the magnitude of acquired immunity occurring within endemic communities [8
], although differences in the parasite establishment rate- a more intrinsic biological parameter, the values of which are likely to depend on the strength of immunity operating in a community [20
] - may also, to a lesser degree, govern between-community differences in the values of such breakpoints.
The Monte Carlo-based sampling technique used for model fitting has also allowed the first estimates of the extant uncertainty in breakpoint values for eliminating LF. Thus, we found, for example, that worm breakpoints, when aggregated over the models for each vector, resulted in a higher spread of values for the anopheline than the culicine models (the median mf prevalence breakpoint value was 0.76% for anopheline models as opposed to 0.23% for culicine). Although this vector-specific discrepancy in the median value for this breakpoint has previously been suggested [8
] (and is due to the presence of two facilitation density dependences (mosquito uptake and worm mating probability) for the anopheline against only one (mating probability) for the culicine-mediated filarial infection dynamics), the characterization of variability in the breakpoint values, even within a local setting, is a new outcome of this study. While the uncertainty estimates for both breakpoints and for R0
obtained in this study primarily reflect epistemic uncertainty regarding parameter values and their distributional patterns (which we expect to refine by model updating with data), the possible existence of a range or cloud of extinction breakpoints within a setting, nonetheless, supports the notion that stochastic variability in infection parameters will, in reality, give rise to a distribution rather than a simple point estimate for these variables in natural communities [10
]. Nevertheless, the data-driven finding here substantiates the theoretical conjecture [7
] that it may be easier, if all other factors are held constant, to eliminate anopheline rather than culicine filariasis in the field. Future work should include further data sets for both culicine and anopheline-mediated filariasis in order to increase the statistical validity of these findings.
Our simulations of the impact of the WHO-recommended drug administration strategy (six annual mass treatments with either DEC/ALB or invermectin/ALB and a population coverage of 80%) demonstrate the likely failure of a fixed global strategy that ignores local extinction dynamics. Not only may such a strategy produce a great deal of site-to-site variability in the prospects of achieving filariasis elimination but a consideration of the drivers of transmission, such as community ABR and R0
values, may also have limited usefulness in predicting the likely success of time-bound intervention strategies for accomplishing parasite elimination, especially in those areas where the values of both these variables are high. The importance of this result for filariasis elimination programmes is clear: because the complex dynamics governing parasite transmission may cause filarial transmission breakpoints to vary between communities, any effort which aims to achieve the elimination of this disease must be based on estimates of local thresholds [8
The results of our parameter estimation have demonstrated that, although a greater knowledge of the natural variability occurring in key filariasis transmission parameters can clearly be gained using the BM model fitting approach, successful parameter updating is critically related to the quality of the available data. Thus, the most informative data are those that show low levels of variability, suggesting that in order to be useful, parameter estimation requires that either good quality data are collected and subjected to analysis or else a hierarchical multi-level framework should be developed that allows the combining of data from different communities with as similar transmission characteristics as possible. The Bayesian approach employed here will also allow for the future inclusion of further information or data, such as treatment follow-up data from local sites, which may be used sequentially to refine the model-fitting process and, hence, update parameter estimates [27
]. Such updating of the present models with more site-specific and follow-up data may also eventually enable us to determine which components, or even model structure, is necessary to obtain the most parsimonious description of the host-parasite system in different endemic localities [28
There are further limitations to our modelling approach which need to be borne in mind when interpreting the present results. The most important of these is that, although our deterministic modelling framework has yielded important insights into the extinction dynamics of LF as a result of mass drug interventions, stochastic analogues of our models would clearly enable the investigation of a greater number of sources of extinction, including the role of pure demographic effects and the impact of external drivers of population dynamics such as varying environmental or climate variables. In addition, future work must not only provide a better understanding of the forms and parameter values of the density-dependent processes that need to be included in the model to explain data in different communities, but must also show how these functions may, in turn, interact with different interventions in order to govern the specificity of the parasite population response to control. The use of longitudinal follow-up data in conjunction with model updating procedures, such as the Bayesian estimation procedure described here, will allow an analytical framework to achieve this objective.
Despite these caveats, the present findings point to important implications for the design of filariasis elimination programmes. First, the difficulty of measuring the key local parameters (for example, infection aggregation, acquired immunity), critical to differences in estimated transmission breakpoint values, implies that the core difficulty in eliminating complex dynamical diseases, such as filariasis, is fundamentally related to the problem of how best to develop elimination strategies in the face of endpoint uncertainty in different sites. While adaptive management strategies, whereby data from each site or from endemically homogeneous regions could be used to develop and apply local strategies, would provide the optimal solution [29
], this is unlikely to be practically possible in most endemic settings. This implies a need to consider strategies developed and used in other fields (for example, engineering) for managing complex dynamical systems [9
]. The first of these might be to rely on achieving local elimination on the assumption that good local elimination everywhere implies good ultimate elimination overall, as long as the local interventions and elimination targets are well chosen. This approach could start by splitting the overall problem into a hierarchy of levels, with objectives for local, short-term elimination initially set at a higher level - for example, achieving disease control first [31
] - and then expanded on a longer time scale to accomplish parasite transmission interruption [27
]. The second tactic may be to avoid focusing solely on meeting the objective of uncertain elimination and exploit the ability of even a relatively poor model to give fairly good guidance to promote good parasite system transitions (for example, parasite control or even elimination) and prevent bad transitions (for example, infection re-emergence following control) [9
]. Previously, we have shown that including vector control with MDA can, by increasing the worm breakpoint threshold value, reduce the resilience of the endemic state and, by raising the re-emergence infection threshold, promote the resilience of the parasite-free state, and hence, can play this resilience-enhancing role in sustaining LF elimination [8