How does the paragraph expand on the central idea that malaria?

A global decline of malaria in recent decades has led to a push for complete national elimination of human malaria parasites in 21 countries by 2020 (WHO, 2018). In elimination and pre-elimination settings, malaria transmission is often highly heterogeneous geographically (Bousema et al., 2012; Carter et al., 2000; Sturrock et al., 2016), and policy-makers must focus on reducing transmission in remaining endemic foci, while protecting areas where malaria has been effectively controlled from imported infections that threaten to reintroduce parasites and reignite transmission (Cotter et al., 2013; Churcher et al., 2014). These tasks require different interventions, so understanding how patterns of regular travel to and from malaria endemic regions of a country contribute to the spread of malaria is a critical component of elimination planning. Measuring the spatial spread of malaria remains technically challenging, however, due to frequent asymptomatic and undetectable infections (Lindblade et al., 2013), as well as the inherent challenges associated with epidemiological surveillance in rural areas among highly mobile populations.

Decisions about how to allocate resources for malaria control and elimination are generally based on reports from hospitals, clinics, community health workers (CHWs) and non-governmental organizations (NGOs) around the country, which provide a measure of the incidence of symptomatic cases (WHO, 2007; Moonen et al., 2010). In areas with high mobility, however, the reporting health facility may not accurately capture local transmission, instead detecting imported infections that have been acquired elsewhere. Traditionally, patient travel surveys have been the principal method for identifying imported cases, but surveys can be unreliable and limited in scope, and importation cannot always be confirmed even when a travel history is accurate (Wesolowski et al., 2014). We and others have used spatially explicit epidemiological models– for example, parameterized using mobile phone data– to estimate ‘sources and sinks’ of parasites and the impact of travel on the spread of malaria (Wesolowski et al., 2012). These methods show promise, but do not provide highly spatially resolved estimates in regions with low cell tower density, such as in forested or sparsely populated regions, and may suffer from ownership biases.

There is also a growing interest in the routine use of sequencing to genotype parasites, which may provide insights into the spatial spread of malaria, particularly in low transmission settings where imported infections that sustain remaining hotspots must be addressed and in places where drug resistance poses a threat (Amato et al., 2018; Busby et al., 2016; Imwong et al., 2017; Karunaweera et al., 2014; MalariaGEN Plasmodium falciparum Community Project et al., 2016; Miotto et al., 2015; Obaldia et al., 2015; Taylor et al., 2017; Tun et al., 2015). However, in spite of decreasing costs, the use of genetic data is currently constrained by the lack of appropriate methods for analysis. Plasmodium falciparum infections are often composed of multiple different parasite clones, which undergo sexual recombination in the mosquito (Mzilahowa et al., 2007). As a result, phylogenetic and population genetic methods normally applied to the analysis of non-recombining organisms cannot be used. In addition to applying human population genetic tools to malaria parasites (Chang et al., 2012; Lawson et al., 2012; Miotto et al., 2013; Patterson et al., 2006; Pritchard et al., 2000), an emerging field of promising new approaches, such as THE REAL McCOIL, hmmIBD, or isoRelate, are being developed to explicitly address these characteristics of malaria infection (Taylor et al., 2017; Chang et al., 2017; Henden et al., 2018; Schaffner et al., 2018).

Conceptually, genetic and epidemiological approaches should complement each other – genetic data encodes information about the relatedness of parasites and mixing patterns between different parasite populations, while epidemiological data provides insights into clinical cases and the patterns of transmission. For viral pathogens like the influenza virus, where recombination does not obscure relationships between lineages, phylodynamic frameworks have been developed for inferring migration patterns and identifying the factors contributing to the migration rate (Lemey et al., 2014; Nelson et al., 2015; De Maio et al., 2015; Vaughan et al., 2014). Although it is possible to measure population differentiation on continental scales for malaria (Miotto et al., 2015), elimination programs often require inferences on smaller spatial scales, which are currently more challenging because of frequent recombination and the complex epidemiological dynamics of the parasite. In particular, on local spatial scales genetic signals may not be easily resolved because parasites are likely to be highly related, and comparisons between many pairs of populations become computationally prohibitive (De Maio et al., 2015; Vaughan et al., 2014; Rasmussen et al., 2014). Developing methods that leverage insights from genomic data, epidemiological information, and mobility data sources such as travel survey or mobile phone data, is therefore an important goal for malaria, particularly if the collection of clinical samples for genetic analysis and travel data can be incorporated into routine surveillance workflows to support decision making for control programs.

The National Malaria Elimination Programme (NMEP) in Bangladesh currently estimates the burden of malaria using symptomatic case counts aggregated at administrative level 3 (upazila) or 4 (union). Typical of malaria in Southeast Asian countries, much of the transmission is thought to occur in forests and forest fringes where the principal vectors abound (Bhatia et al., 2013). The difficulties of surveillance in hard-to-reach forested areas complicate decisions about where to target transmission-reducing interventions (often focused on vector control) and how to find and treat imported cases as Bangladesh moves towards elimination. Moreover, Bangladesh – one of the countries that plan to eliminate malaria nationally – borders the Greater Mekong Subregion (GMS) (WHO, 2017b; Haque et al., 2014), where evidence of resistance to first line treatment for malaria (artemisinin combination therapy, or ACT) has been found in all countries (WHO, 2017a). Although resistance to artemisinin and to ACT partner drugs has not been found to date locally, Bangladesh is a potential point of passage in the spread of drug resistance out of the GMS, so monitoring drug resistance markers is crucial.

Epidemiological models of malaria designed to understand spatial transmission are often based on clinical incidence and seroprevalence estimates; data and methods required for integrating insights about importation from other data sources, including genetic data, are currently lacking (Daniels et al., 2015). Mobility patterns derived from mobile phone data are difficult to access, and rarely coincide temporally with epidemiological and genetic data, but could in theory provide real-time estimates of population movements. We collected and analyzed paired genetic, travel survey, and epidemiological malaria case data across the Chittagong Hill Tracts (CHT) region of Bangladesh, in combination with estimated population mobility patterns from mobile phone calling data. We developed a simple metric to quantify genetic mixing from parasite genetic barcode data. Geographic patterns of genetic mixing were then compared to estimates of malaria importation from a mathematical model, based on clinical epidemiological data and on mobility data from travel surveys and mobile phone records. We show that, taken together, these data sources provide evidence of heterogeneous transmission outside the high-incidence forested areas, and of substantial importation of parasites throughout the CHT. We propose that this type of data integration and analytical pipeline could help national control programs effectively target resources for malaria elimination.

The map of clinical incidence from the NMEP for 2015 and 2016 shown in Figure 1A in the CHT supports the hypothesis that transmission is highest in the eastern, forested region, similar to other countries in the GMS with forested border areas (Figure 1B). The location of the reporting health clinic may not reflect local acquisition of infection, however, particularly in regions like Bangladesh where laborers, forest-workers, and military personnel may travel frequently in and out of high transmission areas, but report to clinics in low transmission areas. We aimed to map remaining transmission foci in the CHT; in particular to identify areas outside forested regions where transmission is being sustained, rather than imported from the forest, since these areas should be targeted by transmission reduction interventions. To this end, we enrolled 2090 patients with confirmed malaria, attending 57 hospitals and clinics across the malaria-endemic region (see Figure 1—figure supplement 1 for a map of sample distribution). All patients answered a travel questionnaire and their dried blood spots were collected on filter paper. In total, parasite genotypes of 1412 patients were obtained successfully. These data were combined with reported incidence for the region, and data on aggregated daily movement patterns from 2.4 million mobile phone subscribers over 6 months across the CHT (see Materials and methods).

(A) The average monthly incidence per 1000 population of P. falciparium malaria from 2015 to 2016 as reported by the NMEP. Incidence was highest in the eastern portion of the CHT (shown in relation to the country borders) and decreased westward. (B) The forest coverage (%). (C) Unions sharing at least one parasite with identical genetic barcodes. (D) Top 50% of most traveled routes reported between pairs of locations from the travel survey data. (E) Top 1% of routes traveled between pairs of locations from the mobile phone data. Unions were colored grey where data was collected on genetic (C), travel survey (D) or mobile phone data (E).

We first analyzed each type of data separately, focusing on quantifying signals of spatial connectivity and on measuring transmission intensity (Figure 1C,D,E). Given the importance of distinguishing between the forested border regions – where most malaria is thought to occur – and less forested areas elsewhere, for clarity we defined forested areas as those with >50% forest coverage. In the travel survey, 31% (N = 654) of malaria positive individuals reported living outside forested areas. Of these individuals, the majority did not report any travel to forested areas (66%, N = 434), suggesting that either individuals were acquiring infections outside these areas, or that the travel survey did not capture all forest travel. The survey indicated a general separation between the north and south portions of the endemic region, with fewer people traveling between these two regions. Within the regions, however, substantial travel was reported between the high incidence forested areas and the lower incidence ones (Figure 1D). Similarly, by simply connecting those pairs of unions where we identified parasites with identical genetic barcodes, a broad north-south division of the CHT region clearly emerged, as did connections between forested and non-forested areas (Figure 1C). Temporally, parasite pairs with identical genetic barcodes were significantly closer in time than those that were not identical (medians = 16.5 vs. 145 days, Mann-Whitney test p-value<2×10−16). Conversely, we found that other genetic diversity metrics, except for drug resistance-related markers, showed limited spatial resolution (see Figure 1—figure supplements 2 and 3). Roughly consistent with the north-south divide, mobility estimates derived from mobile phone calling data suggested a high amount of travel in the coastal areas around the cities of Chittagong and Cox’s Bazar, with little travel between these areas (Figure 1E). This separation into two main transmission foci was consistent with the larger geographic distance between these areas of the CHT, but provided little insight into the extent of transmission heterogeneity on more local spatial scales.

To assess whether patterns of genetic relatedness among parasites showed signals consistent with survey- or mobile phone-based estimates of mobility, we analyzed all parasite pairs, along a continuum of genetic similarity, with respect to mobility outcomes. As expected, genetically similar parasites (fewer differences in genetic barcode comparisons) were more likely to come from the same residence unions (Figure 2A). For parasites from different unions, more related parasites were more likely to come from unions where direct travel had been reported in the survey (Figure 2B; see Materials and methods). To ensure this signal was not obscured by local effects, since both travel and genetic similarity is expected to be high between neighboring unions, we separated the data into unions less than 20 km apart and unions more than 20 km apart (Figure 2—figure supplement 1). The association shown in Figure 2B was mainly driven by unions that were ≥20 km apart, indicating that travel is responsible for closely related parasites in more geographically separated unions. Remarkably, closely related parasites that were neither from the same location nor from locations with direct travel reported were more likely to come from unions with indirect travel reported to a shared third location (Figure 2C). Using the travel survey data at the individual patient level instead of population level, we found that individuals infected with closely related parasites were significantly more likely to report living, working, or traveling to the same locations, confirming the correlation between the genetic and travel survey data (Figure 2—figure supplement 2). Comparing the genetic data to travel measured using mobile phone data, we found that genetically similar parasites that were not from the same locations were more likely to come from locations with greater inferred connectivity between them (Figure 2D), even though the majority of travel measured was in less forested areas. Therefore, although the travel survey and mobile phone data represent different geographic and population coverage, and disparate sample sizes (Figure 2—figure supplement 3), spatial signatures of connectivity were generally consistent with the genetic data.

We compared genetic data with population-level travel survey data and mobile phone data under four scenarios: being in the same union (A), coming from unions with direct travel reported in travel survey (B) or indirect travel (C), and coming from unions with high direct travel inferred from mobile phone data (D) (See Materials and methods for details). We calculated the proportion of pairs under these four scenarios when SNP differences were lower than given thresholds (red) and compared them with 100 permutation results (grey). (A) Sample pairs with smaller SNP differences were more likely to be from the same union than random permutations; (B) if they did not live in the same union, they were more likely to be from unions with direct travel; (C) if neither of these conditions held, then they were more likely to be from unions with indirect travel. (D) If sample pairs were not from the same union, those with a smaller SNP difference were more likely to be from unions with higher direct travel (>0.1%) estimated from mobile phone data.

https://doi.org/10.7554/eLife.43481.006

To map transmission heterogeneities on smaller spatial scales, these rich data sources can be analyzed using more sophisticated approaches. We developed a metric designed to quantify the amount of parasite genetic mixing – indicative of frequent parasite importation and recombination between different parasites from different locations – in different regions. This index was calculated for each location with respect to all other locations as a risk ratio; specifically, the ratio of risks that parasites were genetically similar given that they were far away and nearby. To identify this index, we used the empirical relationship between genetic similarity and geographic distance for this region, so the thresholds defined would be setting-specific if rolled out routinely as part of surveillance programs (see Materials and methods). In brief, we quantified the decline of parasite genetic similarity with geographic distance, uncovering a robust signal; pairs of parasites sampled from unions that are geographically closer were more likely to be genetically similar (Figure 3A). In addition to the enrichment of genetically similar parasites in spatial proximity, there was a temporal signal: parasite pairs with SNP differences smaller than 10% were significantly closer in time than those with SNP differences greater than 10% (medians = 14 vs. 145 days, Mann-Whitney test p-value<2×10−16). Using the empirical relationship between geographic and genetic distances, we assumed a uniform prior on geographic distance and calculated the probability of geographic distances between patients given different levels of genetic similarity (Figure 3—figure supplement 1; see Materials and methods). This allowed us to define a CHT-specific ‘genetic mixing index’ to identify higher sharing of parasite barcodes between samples from relatively distant locations, indicative of high gene flow or importation (Figure 3B and Figure 3—figure supplement 3; see Materials and methods). Figure 3B maps this index across the CHT, showing heterogeneous patterns of mixing in the southwestern region, with high mixing regions indicating frequently imported infections, and low genetic mixing suggesting predominantly local transmission. In contrast, we did not observe elevated genetic mixing in the north (see robustness analysis in Figure 3—figure supplement 2 and Supplementary file 1 for 95% confidence intervals of the index).

(A) The association between genetic data and geographic distance was only obvious for small SNP differences. Pairs of parasites sampled from unions that are geographically closer were more likely to be genetically similar. The proportion of intermediate or high SNP differences did not vary much with geographic distance. (B) The genetic mixing index for each location. Unions were colored white if they did not include genetic data and grey if they included genetic data but their genetic mixing index was not identifiable due to lack of samples that were both nearby and genetically similar. High genetic mixing index suggests high parasite flow or importation.

https://doi.org/10.7554/eLife.43481.010

Genetic measures based on SNP barcodes were able to identify unions with a high likelihood of importation events, but not the origin of imported infections since sharing genetic similarity did not provide information about the direction of movement. Epidemiological models that describe directional travel patterns may be used to combine reported incidence with data from travel surveys or mobile phone data to reconstruct routes of importation. We used mobility estimates from the survey and from mobile phone data to infer the amount of travel between locations in a simple epidemiological model that predicts the rate of infections moving between unions and mapped the resulting malaria source and sink locations (see Materials and methods, Figure 4A,B, and Figure 4—figure supplements 1 and 2). We compared the relative contribution of imported infections to local transmission in each union and predicted that locations in the southwestern area of the CHT had the highest proportion of imported cases, using both the travel survey and mobile phone data (Figure 4A,B). The mobility patterns derived from mobile phone data suggested more infections were being imported overall, however qualitatively both mobile phone data and travel surveys identified overlapping areas of the country where the top locations of importations were occurring. The genetic mixing index therefore provides an independent confirmation that extensive parasite flow is occurring in the highly populated southwestern region of the CHT, while the epidemiological model – parameterized using travel survey data and mobile phone data – provides insights into the sources of imported infections.

(A–B) The estimated proportions of non-locally acquired parasites. We estimated the percentage of infections in each union that were acquired in other destination unions (red) versus locally acquired (blue). For each union where travel data was available from the travel survey (A) or mobile phone data (B), the percentage was shown. Unions were colored according to their forest coverage (light green for low, dark green for high). (C) We classified genetic mixing index >0.1 as high and ≤0.1 as neutral. The estimated proportion of imported infections from the travel survey data (top panel) or mobile phone data (bottom panel) was higher for unions classified as a high genetic mixing index, while not statistically significant (p-value>0.05).

https://doi.org/10.7554/eLife.43481.014

Combining unions with a high genetic mixing index together with unions with a high proportion of imported cases estimated from the epidemiological models, we created an updated risk map for malaria transmission in the CHT (Figure 5D). Both the genetic mixing index and the epidemiological models suggest frequent mixing in the southwest of the CHT, in Cox’s Bazar district (Figures 3B and 4A,B). The epidemiological model predictions were roughly consistent with the unions exhibiting a high genetic mixing index (Figure 4C). Additionally, we identified frequent importation within the southwest region from epidemiological models parameterized using both travel survey and mobile phone data. In the model parameterized using travel survey data, parasites were predicted to be frequently imported from the forested areas of the CHT to the more populated, lower transmission areas, particularly unions within Cox’s Bazar district (Figure 5A, Figure 5—figure supplements 1 and 2). The model parameterized by mobile phone data also identified the forested areas as a source, and in addition, suggested that parasites were imported from more populated, lower transmission areas in the west (Figure 5B). Since travel surveys and mobile data had quite different geographic coverage and different biases (Figure 1C,D), some discrepancies are expected. Nevertheless, these analyses support the ideas that a) forested areas are an important source of imported infections, consistent with current dogma, but b) human movement around densely populated regions with relatively low transmission appears to contribute substantially to heterogeneous transmission across the southwest of the CHT (Figure 5C). This means that resources for elimination should be targeted in these areas, in addition to the forested regions. The updated map (Figure 5D) highlights the regions which were not identified in the incidence map (Figure 1A), for example. Our results also suggest that importation routes occur separately in the north and south of the Chittagong region (Figure 5C and Figure 5—figure supplement 1), implying that a staged elimination strategy could be effective.

(A–B) We estimated parasite flows between unions using epidemiological models parameterized by the travel survey (A) or mobile phone data (B). The top routes of parasite flows (the top 25% for travel survey and the top 1% for mobile phone) and their origins and destinations were shown by lines and dots, respectively. Unions were colored grey if they included travel survey (A) or mobile phone data (B). The top routes of parasite flows accounted for 86.4% (travel survey) and 87.8% (mobile phone) of the total importation. (C) The combined map showing top parasite importation routes from the travel survey (nodes colored blue), mobile phone data (nodes colored purple), or both (nodes colored black with a yellow outline). Unions were shown in red if they had a high genetic mixing index. (D) An updated risk map for malaria transmission in the CHT: unions with high genetic mixing index (>0.1) or a high proportion of imported cases estimated from the epidemiological models (blue: top 25% from the travel survey, top 1% from the mobile phone data), high incidence areas (red: the average monthly incidence per 1000 population >4), and unions that have both high incidence and a high importation level (purple).

https://doi.org/10.7554/eLife.43481.017

Routine surveillance for malaria among control programs is primarily based on prevalence estimates and/or clinical incidence reported by hospitals, clinics, community health workers and non-governmental organizations. The geographic distribution of clinical cases may not reflect patterns of transmission, particularly in areas hosting highly mobile populations, where importation of parasites is common. Identifying the true foci of transmission is particularly important in elimination settings, such as Bangladesh. A number of approaches have been suggested to augment risk maps based on incidence data, including molecular surveillance and epidemiological models parameterized by new data streams, such as mobile phone data. Here, our goal was to assess whether a combination of these approaches could provide actionable insights for control programs. As a complement to a simple map of clinical incidence, or model-based geospatial maps based on incidence and prevalence estimates (https://map.ox.ac.uk), this combined approach illustrated that, while the eastern forested regions are indeed contributing imported infections to the lower transmission areas in the southwest, substantial local importation and genetic mixing is also occurring in these highly populated areas along the southwestern part of the CHT. This suggests that, while targeting interventions to forested areas is a key strategy for elimination, it might not be successful unless the mixing between low transmission settings in the southwest region is also addressed. We propose that, with appropriate sampling strategies, parasite genetic analysis could provide actionable insights for national control programs and help evaluate the success of elimination programs.

Our samples came from symptomatic patients who presented at a health facility, which would be the norm for routine reporting to the NMEP. We therefore did not include asymptomatic or subclinical infections, and it is unknown to what extent the parasite in symptomatic patients are representative of the entire parasite population in this area (Searle et al., 2017). Although we were able to enroll patients at 57 health facilities that covered a wide geographic area, these study sites only represent 59% of all CHT health facilities which report data to the NMEP. In several health facility catchment areas, non-government organizations provide malaria treatment, testing, and education outside health facilities as part of a malaria outreach initiative. In these areas, we may be under-estimating true incidence, since fewer people seek malaria treatment at the government health clinics. Although these biases may impact epidemiological modeling results, the genetic mixing index – which may lack power when sample size is small, but is not biased by underestimating incidence – can provide complementary insights.

In this transmission setting and geography, genetic differentiation based on genetic barcodes between parasite populations cannot be easily distinguished by commonly used methods, such as average pairwise difference or FST (Figure 1—figure supplement 3). Consistent with our previous studies showing that the proportion of nearly identical genetic sequences is more sensitive to recent migration of pathogens (Taylor et al., 2017; Chang et al., 2016), in our data the proportion of parasites with nearly identical genetic barcodes was highly associated with geographic distance. This allowed us to develop a genetic mixing index, which may provide context-specific information about gene flow. However, the thresholds used for calculating this index would need to be determined in each setting independently, because the association between geographic distance and genetic similarity can vary between settings. This index could also be used at the individual level, to identify infections that are likely to be imported. We found three cases with a genetic mixing index greater than 1, suggesting that they were likely to be imported and, consistently, there was reported travel between their residence unions and the unions they shared similar barcodes with. The reliance of the genetic mixing index and other spatial approaches on nearly-identical parasites – which are rare, but hold the most meaningful genetic signals – means that small sample sizes and large uncertainty are inevitable challenges to their routine use. Although increasing sample size would improve our estimates, adding more SNPs to the analysis is unlikely to reduce uncertainty of the index, since the patterns of relatedness are an intrinsic feature of malaria epidemiology in this region. Nevertheless, in combination with other modeling approaches, genetic data can yield insights that are not attainable any other way – particularly in the context of elimination planning. As molecular surveillance becomes more widespread, clearly establishing the strengths and limitations of genetic data will be critical.

Mobile phone data and travel surveys are different from each other in terms of scale, geographic coverage, and granularity of data about human mobility in this context (Wesolowski et al., 2018). Our travel survey provides detailed, individual-level information about infected individuals, although its quality is dependent on the respondent’s ability to accurately recall their travel over the past two months. Although we included specific questions to help with recall bias, such as asking explicitly about travel to the forested areas, these biases likely still exist and may have limited our ability to infer parasite movement. Overall, travel survey data were sparse at the union level, and many individuals reported no travel (287 out of 2,090), or only travel within a single union (1240 out of 2,090). In contrast, mobile phones provide large volumes of data about the connectivity of the general population, but no individual-level information, and only in locations where cell towers exist. The contrasting geographic coverage and scale of the two sources of data on human mobility – mobile phone and travel survey data – lead to different estimates of parasite importation. Mobile phone data provides insights into large-scale population flows, but is limited to regions with cell towers, while travel survey data provide detailed information about a small number of malaria-infected individuals. Both data sources are biased, but together can provide complementary insights into the population dynamics underlying the spread of parasites in the region. Given the difficulties associated with measuring population mobility and the importance of understanding importation in countries aiming for elimination, multiple (albeit biased) estimates of human travel provide the most complete picture of parasite flows. Given the biases in the incidence data (discussed above), in the epidemiological model, we chose a simple approach without having to estimate the underlying transmission or directly incorporating vectoral capacity. Remarkably, travel patterns from both data sources were consistent with results from parasite genetic analyses, confirming their complementarity.

Every epidemiological and genetic data set has inherent limitations and biases, but multiple data layers can complement each other effectively to provide insights into different components of malaria transmission. This study represents a first example of the utility of combining these sources of information in the context of identifying imported infections, and we believe it shows that cost-effective sequencing approaches can be combined with simple epidemiological models to provide a more complete picture of malaria transmission on a subnational level.

All genetic data are included in Supplementary file 4 and all travel matrices are included in Supplementary file 5.

Thank you for submitting your article "The geography of malaria elimination in Bangladesh: combining data layers to estimate the spatial spread of parasites" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by Neil Ferguson as the Senior Editor and Reviewing Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Edward A. Wenger (Reviewer #1); Oliver Brady (Reviewer #3).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

The authors provide a logical framework for combining multiple data sources relating to travel history and parasite genetics, resulting in a single updated risk map for the Chittagong Hill Tracts region of Bangladesh. This updated map is informative of source-sink dynamics beyond simple prevalence/incidence measures produced by traditional epidemiological approaches and is therefore likely to be useful to NMEPs. Combining data layers in this way is a non-trivial problem due to the idiosyncrasies of different types of travel data and the complex signal in Plasmodium genetic data, and so this represents a significant step forward in terms of demonstrating the applied utility of Plasmodium genetic data over large spatial scales.

The manuscript is clear and well-written. Nevertheless, we have some comments that should be addressed:

First, this paper would benefit from more context in the presentation and interpretation of genetic signatures:

a) How do complex infections or inferred within-sample diversity vary among samples and between geographic regions? Does this tell a complementary story to the parasite pairwise relatedness?

b) Does presenting the relatedness information as a graph structure rather than bulk properties of pairwise relatedness provide any additional insights?

c) Can the authors give more context for choices of relatedness thresholds – 10% "nearly identical" (subsection “The association between genetic data, travel survey, and mobile phone data” last paragraph) and 17.5% similarity (subsection “Genetic analysis”, last paragraph). How do these relate to e.g. 1 or 2 outcrossing events with random parents drawn from population-level minor allele frequencies?

The big difference in coverage of travel survey and mobile phone data makes the assessment of their comparability / complementariness quite difficult to assess as a reader. While we appreciate there is not much the authors can do to change this, there are changes to the way the data are presented that could improve this, e.g. Figure 4 would be easier to interpret differences between the two data sources if only areas for which both travel survey and mobile phone data were available. The discussion could also elaborate more on the implications of this disparity in coverage.

Given the huge amount of work that has gone into this analysis, the final risk map (Figure 5D) is disappointingly non-specific. To what extent is this east-west flow of parasites a new result, or is it already known to some degree? If resources were limited, what would be the suggestion about how best to use them? Can top importation routes be predicted, and how much of total importation would they account for?

We have some concerns about the new genetic mixing index. This is presented as an odds ratio on samples being in near vs. far spatial unions given their level of genetic similarity, but does this not have the "exposure" and the "outcome" the wrong way around, i.e. what we observe is genetic similarity as a function of distance, not the converse? The symmetry property of odds ratios means that numerically swapping these round (i.e. Prob (genetically similar | far)) would result in the same final value, but would seem to make more intuitive sense.

There are many ways of constructing a statistic to measure the relationship between geographic and genetic distance, and so some statistical justification is needed for the use of an odds ratio. I'm sure the authors are aware that odds ratios are typically avoided in general due to frequent misinterpretation as risk ratios and their extreme behaviour in some parts of the parameter space (a small difference in probabilities could lead to a very small or very large difference in odds). For example, if the "rare disease" assumption is being used to justify this statistic then this should be stated. If no justification can be given then perhaps a risk-ratio would be more appropriate.

Some representation of uncertainty in the genetic mixing index should be given, either analytically or by simulation. Extreme uncertainty is another known weakness of odds ratios, hence it needs to be addressed.

There is a tendency to swap between talking about genetic similarity as a function of distance (e.g. subsection “Genetic and model-based evidence of transmission in low incidence areas”) and distance as a function of genetic similarity (e.g. Figure 3A) without appreciating that this involves a transformation. For example, when talking about the probability of geographic distance between parasite pairs (Figure 3) there must have been an application of Bayes' rule with a particular prior on distance, but this is not formally stated. Note that using the raw counts without a prior is equivalent to assuming a uniform prior on distance, which is almost certainly not appropriate here (we would not expect points to be uniformly separated a priori).

How sensitive are values of the genetic mixing index to non-uniform sampling of clinical infections or genetic sequences from different transmission regions or subpopulations?

Note – we understand and appreciate the intent of the genetic mixing index and feel it could be a valuable addition to the analytical toolbox, but it does need to pass certain basic statistical checks to avoid issues later on.

Reviewer #1:

This an important body of work, combining different sources of mobility data – travel surveys, mobile-phone records, parasite genetics – to quantify epidemiological and programmatic endpoints.

1) This paper would benefit from significantly more context in the presentation and interpretation of genetic signatures.

a) How do complex infections or inferred within-sample diversity vary among samples and between geographic regions? Does this tell a complementary story to the parasite pairwise relatedness?

b) Does presenting the relatedness information as a graph structure rather than bulk properties of pairwise relatedness provide any additional insights?

c) Can the authors give more context for choices of relatedness thresholds – 10% "nearly identical" (subsection “The association between genetic data, travel survey, and mobile phone data” last paragraph) and 17.5% similarity (subsection “Genetic analysis”, last paragraph). How do these relate to e.g. 1 or 2 outcrossing events with random parents drawn from population-level minor allele frequencies?

2) How sensitive are mixing-metric values to non-uniform sampling of clinical infections or genetic sequences from different transmission regions or subpopulations?

Reviewer #2:

The authors provide a logical framework for combining multiple data sources relating to travel history and parasite genetics, resulting in a single updated risk map for the Chittagong Hill Tracts region of Bangladesh. This updated map is informative of source-sink dynamics beyond simple prevalence/incidence measures produced by traditional epidemiological approaches, and is therefore likely to be useful to NMEPs. Combining data layers in this way is a non-trivial problem due to the idiosyncrasies of different types of travel data, and the complex signal in Plasmodium genetic data which is just starting to be leveraged effectively. This manuscript therefore represents a significant step forward in terms of demonstrating the applied utility of Plasmodium genetic data over large spatial scales.

One potential point of weakness in the analysis is the presentation of the new genetic mixing index. This statistic is presented as an odds ratio on samples being in near vs. far spatial unions given their level of genetic similarity. First, does this not have the "exposure" and the "outcome" the wrong way round, i.e. what we observe is genetic similarity as a function of distance, not the converse? The symmetry property of odds ratios means that numerically swapping these round (i.e. Prob (genetically similar | far)) would result in the same calculation, but would seem to make more intuitive sense.

Second, there are many possible ways of constructing a statistic to measure the relationship between geographic and genetic distance, and so some statistical justification is needed for the use of an odds ratio. I'm sure the authors are aware that odds ratios are typically avoided in general due to frequent misinterpretation as risk ratios and their extreme behaviour in some parts of the parameter space (a small difference in probabilities could lead to a very small or very large difference in odds). For example, if the "rare disease" assumption is being used to justify this statistic then this should be stated. If no justification can be given then perhaps a risk-ratio would be more appropriate. Some representation of uncertainty should also be given, as again this is a weakness of odds ratios. Keep in mind that this statistic may well be taken up by the community and applied/interpreted in other scenarios, and therefore it is important to ensure statistical robustness at this stage.

On a similar note, there is a tendency to swap between talking about genetic similarity as a function of distance (e.g. subsection “Genetic and model-based evidence of transmission in low incidence areas”) and distance as a function of genetic similarity (e.g. Figure 3A) without appreciating that this involves a transformation. For example, when talking about the probability of geographic distance between parasite pairs (Figure 3) there must have been an application of Bayes' rule with a particular prior on distance to obtain this. Note that using the raw counts without a prior is equivalent to assuming a uniform prior on distance, which is almost certainly not appropriate here.

I could not find a reference or SI detailing the 101 SNP barcode used. Are all these SNPs in linkage equilibrium? Given that other investigators are likely to apply the genetic mixing index on larger data sets, some comment on the effect of LD on this statistic would be welcome.

Other than these statistical issues I have no major concerns with the manuscript, which is well-written and makes a significant contribution to the field.

Reviewer #3:

This manuscript by Chang et al. details an interesting assessment of the combination of travel and genetic data to assess malaria elimination strategy in Bangladesh. It is highly novel, timely and interesting as well as being well described and reproducible. I have only a few comments on what I found an interesting, if a little difficult to follow, analysis.

In the first part of the results it is shown that human movement is important for defining the network of parasite similarity. However in the second part, the authors use a genetic similarity index that only uses simple geographic distance. I understand that the authors want to propose a easily calculatable index, but it would be nice to see the best risk map possible for Bangladesh given the data available.

The big difference in coverage of travel survey and mobile phone data makes the assessment of their comparability / complementariness quite difficult to assess as a reader. While I appreciate there is not much the authors can do to change this, there are changes to the way the data are presented that could improve this, e.g. Figure 4 – would be easier to interpret differences between the two data sources if only areas for which both travel survey and mobile phone data were available. The discussion could also elaborate more on the implications of this disparity in coverage.

Given the huge amount of through work that has gone into this analysis, the final risk map (Figure 5D) is disappointingly non-specific. If resources were limited, what would be the suggestion about how best to use them? Can top importation routes be predicted, and how much of total importation would they account for?

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Mapping imported malaria in Bangladesh using parasite genetic and human mobility data" for further consideration at eLife. Your revised article has been favorably evaluated by Neil Ferguson as the Senior and Reviewing Editor, and three reviewers.

The manuscript has been improved but two reviewers require some remaining issues to be addressed before acceptance, as outlined below.

I would also concur with reviewer 2 that the SNP positions need to be included in this paper or a citation given to a published source. A preprint on bioRxiv would suffice. However, if a preprint is not yet published, please include the positions here.

Reviewer #1:

Thanks to the authors for the welcome additions and responses to reviewer feedback.

A reiteration of the substantial concern that publishing SNP positions with genetic sequences is an integral part of the work, especially in a journal dedicated to "improv[ing] research communication through open science and open technology innovation".

Reviewer #2:

I am happy that the new version does a good job of addressing the initial concerns, and I appreciate the work done by the authors.

My only remaining concern is with the application of Bayes rule to obtain Prob (distance | similarity). First, I agree with the choice to swap Figures 3 and Figure 3—figure supplement 1. But the main text reads that there was "no prior" on geographic distance – there was a prior, it was uniform as stated in the Materials and methods section. It is impossible to apply Bayes rule without the use of a prior, and a uniform prior does not represent no information, it actually represents quite specific information. If the intention is to let the genetic data "speak" and not worry about the prior probability of edges being a certain distance then perhaps it would be better to report this as a raw likelihood, i.e. Prob (similarity | distance). Alternatively if the authors really want to stick with the inverted probability then this could be smoothed over with a statement to the effect of "a uniform prior allows the genetic data alone to speak, and more realistic priors can be incorporated by weighting based on the known sampling distances". At the moment it is unfortunately misleading, because the probability of a randomly chosen pair of samples being a given distance apart based on their genetic similarity almost certainly is not given by Figure 3—figure supplement 1.

Reviewer #3:

The reviewers have sufficiently addressed all my comments in this revision

[…] The manuscript is clear and well-written. Nevertheless, we have some comments that should be addressed:

First, this paper would benefit from more context in the presentation and interpretation of genetic signatures:

a) How do complex infections or inferred within-sample diversity vary among samples and between geographic regions? Does this tell a complementary story to the parasite pairwise relatedness?

With the number of possible genetic analyses, we could have conducted and the space limit, we decided to constrain the number of genetic results we presented. However, we agree that we should present the COI results, since these were indicative of different transmission settings in other contexts (for example, Uganda [Chang et al., 2017]). We have now expanded the Results to show additional genetic results.

In particular, we now include a supplementary figure (Figure 1—figure supplement 3C) that shows the variation in the complexity of infection between geographic regions. In general, the complexity of infection was low (we did not observe an average COI greater than 2 in any union) and the differences between locations were small, and therefore did not provide additional insights into malaria transmission.

We mention that genetic measures shown in Figure 1—figure supplement 3 provided no or very limited spatial resolution in the Results as follows:

“Conversely, we found that other genetic diversity metrics, except for drug resistance-related markers, showed limited spatial resolution (see Figure 1—figure supplement 2 and 3).”

b) Does presenting the relatedness information as a graph structure rather than bulk properties of pairwise relatedness provide any additional insights?

In Figure 1—figure supplement 3E, we showed that average pairwise difference is not clearly associated with geographic distance. To compare with the graph structure shown in Figure 1C, we added one additional supplementary figure (Figure 1—figure supplement 3B) to show the average pairwise difference in the map, which did not show any obvious geographic pattern. These patterns of relatedness across these populations mean that the most informative signals are found among highly related parasites rather than in average pairwise relatedness in our study, and this likely also applies to proximate populations in other low transmission settings as well.

c) Can the authors give more context for choices of relatedness thresholds – 10% "nearly identical" (subsection “The association between genetic data, travel survey, and mobile phone data” last paragraph) and 17.5% similarity (subsection “Genetic analysis”, last paragraph). How do these relate to e.g. 1 or 2 outcrossing events with random parents drawn from population-level minor allele frequencies?

The expected similarity based on population-level allele frequencies is 55.8%, and therefore one outcrossing event corresponds to 77.9% parent-offspring similarity (0.779= 0.5+(1-0.5)*0.558), which is equivalent to 22.1% difference. This means the SNP difference thresholds of 10% and 17.5% used in our study were within one outcrossing events and represent high similarity and recent shared ancestors. We added this information to give more context for these thresholds in the Methods and Materials as follows:

“Based on the estimated population allele frequencies, expected pairwise similarity between two random parents is 0.558 and therefore the expected parent-offspring similarity after one outcrossing event is 0.779 (= 0.5+0.5*0.558). This indicates that the SNP difference thresholds of 10% and 17.5% used here are conservative, falling within one outcrossing event (i.e. lower than 1–0.779=22.1%) and represent high similarity and recent ancestor sharing.”

The big difference in coverage of travel survey and mobile phone data makes the assessment of their comparability / complementariness quite difficult to assess as a reader. While we appreciate there is not much the authors can do to change this, there are changes to the way the data are presented that could improve this, e.g. Figure 4 would be easier to interpret differences between the two data sources if only areas for which both travel survey and mobile phone data were available. The discussion could also elaborate more on the implications of this disparity in coverage.

We have added a new figure where only areas with both travel survey and mobile phone data were plotted (Figure 4—figure supplement 1). We also now discuss the implications of the disparity in coverage in the Discussion as follows:

“The contrasting geographic coverage and scale of the two sources of data on human mobility – mobile phone and travel survey data – lead to different estimates of parasite importation. […] Given the difficulties associated with measuring population mobility and the importance of understanding importation in countries aiming for elimination, multiple (albeit biased) estimates of human travel provide the most complete picture of parasite flows.”

Given the huge amount of work that has gone into this analysis, the final risk map (Figure 5D) is disappointingly non-specific. To what extent is this east-west flow of parasites a new result, or is it already known to some degree? If resources were limited, what would be the suggestion about how best to use them? Can top importation routes be predicted, and how much of total importation would they account for?

The dogma in Bangladesh is that nearly all malaria originates in the forest. While it is true that incidence is highest in the forest regions, this analysis highlights the importance of spread to the west, and the high volume of human travel around these regions. Our analysis suggests that the sheer volume of travel around the other lower transmission regions in the western part of Chittagong is likely to be sustaining transmission, and that these areas may be easier to target in any case. In terms of resource allocation, our study suggests that surveillance and control in lower transmission regions in southwestern Chittagong may be worthwhile, but that the importation routes that are most important are separate from north to south – therefore, it should be possible to conduct a staged elimination program focusing on each region separately. We have added the following section to the Results:

“This means that resources for elimination should be targeted in these areas, in addition to the forested regions. […] Our results also suggest that importation routes occur separately in the north and south of the Chittagong region, implying that a staged elimination strategy could be effective.”

With regards to the top importation routes: these are plotted in Figure 5 ABC and Figure 5—figure supplement 1. The total importation top routes account for is 86.4% and 87.8% for the epidemiological models parameterized by the travel survey and mobile phone data respectively. We have added the amount of total importation that the top importation routes account for in the legends of Figure 5 and Figure 5—figure supplement 1 as follows:

Figure 5:

“The top routes of parasite flows (the top 25% for travel survey and the top 1% for mobile phone) and their origins and destinations were shown by lines and dots, respectively. The top routes of parasite flows account for 86.4% (travel survey) and 87.8% (mobile phone) of the total importation.”

Figure 5—figure supplement 1:

“The top 10 routes of parasite flows using epidemiological models parameterized by the (A) travel survey (blue) or (B) mobile phone data (purple). These routes accounted for 39.1% (travel survey) and 40.2% (mobile phone) of the total importation.”

We have some concerns about the new genetic mixing index. This is presented as an odds ratio on samples being in near vs. far spatial unions given their level of genetic similarity, but does this not have the "exposure" and the "outcome" the wrong way around, i.e. what we observe is genetic similarity as a function of distance, not the converse? The symmetry property of odds ratios means that numerically swapping these round (i.e. Prob (genetically similar | far)) would result in the same final value, but would seem to make more intuitive sense.

After reading reviewers comments, we agree, and have decided to define genetic mixing index that is similar to a risk ratio rather than an odds ratio. We have revised the formula, and it is now Prob (genetically similar | far distance). Please see the responses to the comments below for more details.

There are many ways of constructing a statistic to measure the relationship between geographic and genetic distance, and so some statistical justification is needed for the use of an odds ratio. I'm sure the authors are aware that odds ratios are typically avoided in general due to frequent misinterpretation as risk ratios and their extreme behaviour in some parts of the parameter space (a small difference in probabilities could lead to a very small or very large difference in odds). For example, if the "rare disease" assumption is being used to justify this statistic then this should be stated. If no justification can be given then perhaps a risk-ratio would be more appropriate.

We did make the “rare disease” assumption because the chance of observing genetically similar pairs is very low. We compared our original genetic mixing index with risk ratios and the values are nearly identical (the largest difference is 0.00133 and the average difference is 0.00019), but we agree with the reviewer that risk-ratio would be more appropriate and therefore modify our definition of genetic mixing index in the Materials and methods as follows:

“

Genetic mixing indexi=Probgenetically similar | farProbgenetically similar | close=nf,snf,s+nf,nsnc,snc,s+nc,ns.

[…] The numbers were then aggregated across individuals within each union i: nf,s, nf,ns, nc,s, and nc,ns represent the number of samples that are far (f) or close (c) and genetically similar (s) or not genetically similar (ns).”

Some representation of uncertainty in the genetic mixing index should be given, either analytically or by simulation. Extreme uncertainty is another known weakness of odds ratios, hence it needs to be addressed.

We agree that uncertainty is an important factor. We have now reported 95% confidence intervals of genetic mixing index in Supplementary file 1 and mentioned this in the Results as follows:

“Figure 3B maps this index across the CHT, showing heterogeneous patterns of mixing in the southwestern region, with high mixing regions indicating frequently imported infections, and low genetic mixing suggesting predominantly local transmission. In contrast, we did not observe elevated genetic mixing in the north (see robustness analysis in Figure 3—figure supplement 2 and Supplementary file 1 for 95% confidence intervals of the index).”

The formula used to obtain confidence intervals was described in the Materials and methods as follows:

“The 95% confidence intervals were calculated as follows (Katz et al., 1978):

Explog(geneticmixingindex) ± 1.961nf,s-1nf,s+nf,ns+1nc,s-1nc,s+nc,ns.”

We emphasized that the uncertainty of genetic mixing index needs to be noted in the Discussion as follows:

“The reliance of the genetic mixing index and other spatial approaches on nearly-identical parasites – which are rare, but hold the most meaningful genetic signals – means that small sample sizes and large uncertainty are inevitable challenges to their routine use. […] As molecular surveillance becomes more widespread, clearly establishing the strengths and limitations of genetic data will be critical.”

There is a tendency to swap between talking about genetic similarity as a function of distance (e.g. subsection “Genetic and model-based evidence of transmission in low incidence areas”) and distance as a function of genetic similarity (e.g. Figure 3A) without appreciating that this involves a transformation. For example, when talking about the probability of geographic distance between parasite pairs (Figure 3) there must have been an application of Bayes' rule with a particular prior on distance, but this is not formally stated. Note that using the raw counts without a prior is equivalent to assuming a uniform prior on distance, which is almost certainly not appropriate here (we would not expect points to be uniformly separated a priori).

We agree with the reviewer that for most readers it is more intuitive to think about the probabilities of genetic similarity given geographic distance, and decided to replace Figure 3A with the original Figure 3—figure supplement 1 and move Figure 3A to a supplementary figure. We further include the following methodological updates below.

We chose a uniform prior since we wanted to show what genetic distance alone tells us without knowing where samples are collected, and to explicitly compare this with the “true” geographic distance. In particular, we developed the method to provide information about future samples as they are collected by the control program, who will want to have some idea whether the transmission event was local or imported. When the SNP difference between samples is small, these isolates are likely to be from nearby locations (Figure 3—figure supplement 1) or, if the true geographic distance is large, it is likely a non-local transmission event. On the other hand, because the probability of a large SNP difference does not vary much with geographic distance (Figure 3A), when the observed SNP difference is large (in our case >30%), we cannot infer anything about importation (all the colors in the sixth column of the figure have similar proportions). We have included an explanation in the Materials and methods as follows:

“To establish what genetic distance alone tells us about the spatial relationships between isolates, we assume a uniform prior with geographic distance. The inferred geographic distances (based on genetic data) can then be compared to the true geographic distance between samples collected by the control program, to assess whether isolates are imported or not.”

We have modified the text to make it clear that we assumed uniform prior as follows:

“Using the empirical relationship between geographic and genetic distances, we assumed no prior on geographic distance and calculated the probability of geographic distances between patients given different levels of genetic similarity (Figure 3—figure supplement 1; see Materials and methods).”

The details on how we calculated the probability of geographic distance using Bayes’ rule was in the Materials and methods as follows:

“Quantifying the probability of a geographic distance given a SNP difference

To investigate the relationship between geographic and SNP distance, we considered seven geographic distance windows D= (0, 0–10, 10–15, 15–20, 20–40, 40–100, >100 km) and six SNP difference windows S= (0, 0–10, 10–17.5, 17.5–25, 25–30, 30–35%). […] We then calculated the probability of a geographic distance d given a SNP difference s,Probds, by applying Bayes' rule and assuming uniform prior as follows:

Probds=Prob(d,s)Prob(s)=ProbsdProb(d)∑alld∈DProbsdProb(d)⇒uniform priorProbsd∑alld∈DProbsd.”

How sensitive are values of the genetic mixing index to non-uniform sampling of clinical infections or genetic sequences from different transmission regions or subpopulations?

The genetic mixing index is a “relative probability” based on the proportion of genetically similar parasite pairs between places that are geographically close or far. Because the expected value of proportion does not vary with sample size, genetic mixing index is not expected to be sensitive to non-uniform sampling from different transmission settings. However, for subpopulations with smaller sample sizes, the confidence interval of the genetic mixing index is expected to be higher or the index may not be identifiable. We performed subsampling and showed in Figure 3—figure supplement 2that the genetic mixing index remained qualitatively similar.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Mapping imported malaria in Bangladesh using parasite genetic and human mobility data" for further consideration at eLife. Your revised article has been favorably evaluated by Neil Ferguson as the Senior and Reviewing Editor, and three reviewers.

The manuscript has been improved but two reviewers require some remaining issues to be addressed before acceptance, as outlined below.

I would also concur with reviewer 2 that the SNP positions need to be included in this paper or a citation given to a published source. A preprint on bioRxiv would suffice. However, if a preprint is not yet published, please include the positions here.

We have now included the positions in Supplementary file 6 and cited the file in the Materials and methods as follows:

“Genotypes were produced using the mass-spectrometry based platform from Agena (Supplementary file 4 and 6).”

Reviewer #1:

Thanks to the authors for the welcome additions and responses to reviewer feedback.

A reiteration of the substantial concern that publishing SNP positions with genetic sequences is an integral part of the work, especially in a journal dedicated to "improv[ing] research communication through open science and open technology innovation".

We have now included the positions in Supplementary file 6.

Reviewer #2:

I am happy that the new version does a good job of addressing the initial concerns, and I appreciate the work done by the authors.

My only remaining concern is with the application of Bayes rule to obtain Prob (distance | similarity). First, I agree with the choice to swap Figures 3 and Figure 3—figure supplement 1. But the main text reads that there was "no prior" on geographic distance – there was a prior, it was uniform as stated in the Materials and methods section. It is impossible to apply Bayes rule without the use of a prior, and a uniform prior does not represent no information, it actually represents quite specific information. If the intention is to let the genetic data "speak" and not worry about the prior probability of edges being a certain distance then perhaps it would be better to report this as a raw likelihood, i.e. Prob (similarity | distance). Alternatively if the authors really want to stick with the inverted probability then this could be smoothed over with a statement to the effect of "a uniform prior allows the genetic data alone to speak, and more realistic priors can be incorporated by weighting based on the known sampling distances". At the moment it is unfortunately misleading, because the probability of a randomly chosen pair of samples being a given distance apart based on their genetic similarity almost certainly is not given by Figure 3—figure supplement 1.

We have changed “no prior” to “uniform prior” in the main text and added the sentence the reviewer suggested in the Materials and methods to make our statement clear as follows:

“Here, a uniform prior allows the genetic data alone to speak; depending on the question of interest, more realistic priors can be incorporated by weighting based on the known sampling distances.”

Article citation count generated by polling the highest count across the following sources: Crossref, Scopus, PubMed Central.

(2019)

Mapping imported malaria in Bangladesh using parasite genetic and human mobility data