Reservoir dynamics of rabies in south-east Tanzania and the roles of cross-species transmission and domestic dog vaccination

Understanding the role of different species in the transmission of multi-host pathogens, such as rabies virus, is vital for effective control strategies. Across most of sub-Saharan Africa domestic dogs Canis familiaris are considered the reservoir for rabies, but the role of wildlife has been long debated. Here we explore the multi-host transmission dynamics of rabies across south-east Tanzania. Between January 2011 and July 2019, data on probable rabies cases were collected in the regions of Lindi and Mtwara. Hospital records of animal-bite patients presenting to healthcare facilities were used as sentinels for animal contact tracing. The timing, location and species of probable rabid animals were used to reconstruct transmission trees to infer who infected whom and the relative frequencies of within- and between-species transmission. During the study, 688 probable human rabies exposures were identified, resulting in 47 deaths. Of these exposures, 389 were from domestic dogs (56.5%) and 262 from jackals (38.1%). Over the same period, 549 probable animal rabies cases were traced: 303 in domestic dogs (55.2%) and 221 in jackals (40.3%), with the remainder in domestic cats and other wildlife species. Although dog-to-dog transmission was most commonly inferred (40.5% of transmission events), a third of inferred events involved wildlife-to-wildlife transmission (32.6%), and evidence suggested some sustained transmission chains within jackal populations. A steady decline in probable rabies cases in both humans and animals coincided with the implementation of widespread domestic dog vaccination during the first 6 years of the study. Following the lapse of this program, dog rabies cases began to increase in one of the northernmost districts. Synthesis and applications. In south-east Tanzania, despite a relatively high incidence of rabies in wildlife and evidence of wildlife-to-wildlife transmission, domestic dogs remain essential to the reservoir of infection. Continued dog vaccination alongside improved surveillance would allow a fuller understanding of the role of wildlife in maintaining transmission in this area. Nonetheless, dog vaccination clearly suppressed rabies in both domestic dog and wildlife populations, reducing both public health and conservation risks and, if sustained, has potential to eliminate rabies from this region.


| INTRODUC TI ON
Rabies virus (RABV) is a true multi-host pathogen. Although typically maintained in distinct species-specific transmission cycles (Rupprecht et al., 2002), the virus is capable of infecting any mammal. Rabies is spread primarily through bites from infected animals and cross-species transmission causes disease in humans, livestock and wildlife (Hikufe et al., 2019). Every year an estimated 59,000 human rabies deaths occur , mostly in low-and middle-income countries (LMICs; WHO, 2018). The economic burden of rabies due to livestock losses is high, and rabies outbreaks within wildlife can threaten endangered species (Randall et al., 2004).
When planning control and elimination strategies for multihost pathogens, it is important to identify the populations that are essential for their persistence (Haydon et al., 2002;Roberts & Heesterbeek, 2003). Throughout this study, we use terminology defined by Haydon et al. (2002) and illustrated in Figure 1. That is, a single population capable of independently maintaining the pathogen of interest is termed a maintenance population. Where multiple interconnected host populations collectively maintain the pathogen this is termed a maintenance community. A reservoir is made up of one or more epidemiologically connected populations capable of permanently maintaining the pathogen and from which infection is transmitted to a population of concern (the target population). If a 6 years of the study. Following the lapse of this program, dog rabies cases began to increase in one of the northernmost districts.
6. Synthesis and applications. In south-east Tanzania, despite a relatively high incidence of rabies in wildlife and evidence of wildlife-to-wildlife transmission, domestic dogs remain essential to the reservoir of infection. Continued dog vaccination alongside improved surveillance would allow a fuller understanding of the role of wildlife in maintaining transmission in this area. Nonetheless, dog vaccination clearly suppressed rabies in both domestic dog and wildlife populations, reducing both public health and conservation risks and, if sustained, has potential to eliminate rabies from this region.

K E Y W O R D S
dog-mediated rabies, lyssavirus, One Health, spillover, surveillance, vaccination, zero by thirty, zoonoses F I G U R E 1 Potential rabies reservoir systems in south-east Tanzania. Here humans are indicated as the target population, but the target may include livestock or endangered wildlife, for example African Wild Dogs (Lycaon pictus). We investigate whether the reservoir consists of both maintenance and non-maintenance populations (a and b) transmitting infection to the nonmaintenance target (humans); or either two maintenance (c) or non-maintenance (d) populations which are capable of transmitting infection to the target if only one of the host types within a maintenance community is targeted (Roberts & Heesterbeek, 2003).
The global eradication of rinderpest is one example of the success of interventions targeted at a maintenance population. Although capable of infecting more than 40 domestic and wild artiodactyl species, cattle Bos taurus vaccination alone eradicated rinderpest virus (Youde, 2013). More recently, the importance of understanding reservoir dynamics has been highlighted by the global Guinea Worm Eradication Programme where the recent discovery of dogs as a potential maintenance host for this parasitic infection has complicated eradication efforts (Molyneux & Sankara, 2017).
Although many RABV variants exist, each variant tends to associate closely with a particular mammalian species which serves as the maintenance host for that variant. Spillover of infection to other species does occur, but sustained transmission outside the maintenance population is uncommon. Interventions targeted at the maintenance population should therefore be effective in controlling that variant. However, there are reported instances of multiple species maintaining a single RABV variant, either separately as distinct maintenance populations or together as a combined maintenance community. On some Caribbean islands, dogs and mongooses (family Herpestidae) maintain the same dog-derived RABV variant and act as a combined maintenance community (Nadin-Davis et al., 2008;Velasco-Villa et al., 2017). In such situations, interventions may need to target both populations to achieve elimination. The presence of multiple maintenance populations can also have implications for disease re-emergence. In north-eastern Mexico, a dog/coyote RABV variant was believed to have been eliminated following widespread dog vaccination. However, the variant continued to circulate in the coyote Canis latrans population and was subsequently reintroduced to dogs via dog-coyote contact. Sustained transmission was possible given the waning herd immunity from inadequate vaccination coverage (Velasco-Villa et al., 2017).
Domestic dogs are considered the maintenance hosts for RABV in Africa and Asia, and are responsible for 99% of all human rabies deaths (WHO, 2018). However, the diversity of wild carnivores across Africa has led to ongoing debate regarding a role for wildlife in maintaining rabies in this region. Across parts of South Africa and southern Namibia, independently maintained rabies cycles are reported in the bat-eared fox Otocyon megalotis (Hikufe et al., 2019;Sabeta et al., 2007;Thomson & Meredith, 1993). Jackal species frequently represent a large proportion of reported wildlife rabies cases in southern Africa (Moagabo et al., 2009;Pfukenyi et al., 2009), and in parts of Namibia, South Africa and Zimbabwe, black-backed jackals Canis mesomelas appear to play a role in maintaining transmission (Bellan et al., 2012;Bingham et al., 1999;Courtin et al., 2000;Hikufe et al., 2019;Zulu et al., 2009).
Evidence from northern Tanzania suggests that domestic dogs are the only species necessary for the maintenance of RABV in this area, although other carnivores contribute to the reservoir as nonmaintenance populations (Lembo et al., 2008). In contrast, very little is known about the transmission dynamics of RABV in south-east Tanzania. In 2010, the government of Tanzania began a 5-year rabies elimination demonstration project, which involved provision of free post-exposure prophylaxis to bite victims, mass dog vaccination and improved surveillance across 28 districts (Mpolya et al., 2017). The area was selected to include a large wildlife-protected area to enable examination of hypotheses evaluating the role of wildlife areas as buffers against infection, and/or the potential of wildlife to impede elimination efforts.
'Zero by Thirty' is an initiative backed by the United Against Rabies Coalition aiming to achieve zero human deaths world-wide from dogmediated rabies by 2030 (WHO, FAO & OIE, 2015 whether wildlife present an obstacle to elimination which will be particularly important in informing efforts to achieve the 2030 target.

| Study area
The study took place between January 2011 and July 2019 within the 13 districts of the Lindi and Mtwara regions of south-east Tanzania, covering an area of 82,668 km 2 ( Figure 2). These districts were part of a World Health Organization (WHO) coordinated demonstration project funded by the Bill and Melinda Gates Foundation to pilot strategies for rabies control and elimination (Mpolya et al., 2017).
The study area contains forest reserves, plantations and the wildlifeprotected area of the Selous Game Reserve. These areas are potential habitats for wildlife species previously reported to transmit rabies to humans, such as jackals and hyenas Crocuta crocuta. The total human population in the study area for 2018 was estimated to be 2, 277,552 (United Nation Development Programme, 2018) and the dog population 49,701 extrapolated from post-vaccination transects and human:dog ratios (Sambo et al., 2018). The districts vary dramatically, ranging from <200 to >1,400 km 2 in area, with human:dog ratios spanning from <35 to >90 persons per dog and dog densities from <0.2 to >7 dogs per km 2 (Table S1).
Prior to the study, no mass dog vaccinations had been carried out in any of these districts, so dog vaccination coverage was likely negligible. However, with the WHO-coordinated demonstration project, five mass dog vaccination campaigns were implemented in each district F I G U R E 2 Study districts and locations of probable rabies cases. (a) The study area (dark grey) and protected areas where no human settlements are allowed (light grey): Selous Game Reserve in south-east Tanzania and Serengeti National Park in northern Tanzania. (b) Districts in Lindi and Mtwara regions (labelled) with estimated dog density on a 4-km 2 raster (grey shading). Urban districts within Masasi, Lindi and Mtwara are not labelled to improve readability. (c) Probable rabies cases in dogs (red) and wildlife (blue) each year in Lindi and Mtwara regions between 2010 and late 2016/early 2017 (Mpolya et al., 2017). A central point strategy was used and post-vaccination transects conducted to evaluate the campaigns (further details in Appendix S1). District-level vaccination coverages increased from a median of 23.5% in the first campaign (range across districts: 3.2%-60.1%) to 34.5% in 2014/15 (range: 22.1%-58.0%), with 33.2% coverage (range: 15.8%-54.8%) achieved during the final campaign in late 2016/early 2017 (Table S2).
Coverage within districts was highly heterogeneous ( Figure S1).

| Data collection
A mobile phone-based surveillance system was used across the area to record animal-bite victims presenting to health facilities requiring post-exposure prophylaxis (Mtema et al., 2016). Records from this system were extracted and compared with paper-based records from health facilities, and any additional animal-bite victims from health facilities that were not present in the mobile phone-based surveillance system were identified. All animal-bite victims and owners of biting animals were exhaustively traced and interviewed to obtain details of each bite incident, as described in the study by Hampson et al. (2009). Information collected during interviews included the following: the date and GPS coordinates of the person bitten; if possible, the origin of the biting animal; the species; the dog owner, if known; and whether the animal was known to have bitten other people or animals. Details regarding the animal's behaviour and the bite circumstances were used to assess whether the animal was considered likely to have been rabid. If additional biting animals or bite victims were identified during investigations, they were also traced and interviewed. Where the GPS location did not correspond to the bite location, a categorical indicator of uncertainty was assigned based on the estimated distance to the biting incident (nearby <2 km away, walking distance 2-5 km and far 5-10 km). We used the resulting data on probable rabies cases and exposures to examine trends in incidence and infer transmission within and between species.

| Parameter estimation
We define the serial interval as the interval between the onset of clinical signs in a primary case to the onset of clinical signs in a secondary case infected by the primary case. The distance kernel represents the distance between the locations of the primary and secondary cases. The probability distribution of the serial interval and distance kernel for transmission of RABV in domestic dogs were estimated using data on probable rabies cases from a long-term contact tracing study in Serengeti District, northern Tanzania. These data included the date and location of the bite incident for the primary rabid animals and the secondary cases that they infected, with information available for serial interval and distance kernel estimation in 1,139 and 958 cases respectively. The parameters from the Serengeti data were used throughout subsequent analyses, as information regarding the serial interval and distance kernel for known transmission events within the south-east Tanzania data was limited (32 and 15 cases respectively), and thus it was felt that the distributions would be better characterized using the Serengeti data. The best-fitting distributions for data available from south-east Tanzania were estimated and a likelihood ratio test was undertaken to test the null hypothesis that the parameters were the same for the Serengeti and south-east Tanzania. For both parameters, maximum likelihoodbased approaches were used for estimation, and the best-fitting parametric form was selected using Akaike's information criteria (AIC).
Details are provided in the Supporting Information (Appendix S2).

| Transmission trees
The estimates of the serial interval G and distance kernel K, described above were used within a previously developed algorithm (Hampson et al., 2009) to generate putative epidemic trees. We define a 'progenitor' as a case that was inferred to be the source of infection for another case. For each probable case i, a progenitor j was chosen at random with probability P ij from all cases within south-east Tanzania with a date of onset of clinical signs prior to the date of onset of the case (n), where: t ij is the days between the onset of clinical signs in case i and its potential progenitor j; and d ij is the distance between the locations of cases i and j. For probable cases in wildlife or for dogs where the owner was not known, the convolution of two distance kernels was used to better incorporate the greater uncertainty in reported locations for these cases.
Due to uncertainty around the dates and locations of some cases, 50,000 bootstrapped datasets of plausible progenitors were generated. In each iteration, for cases with uncertainty around the date of onset of clinical signs and/or their location, dates and/or locations were selected randomly from a uniform distribution within the period or radius of uncertainty respectively. For each case, the case selected most frequently as the progenitor within the 50,000 bootstrapped datasets was considered the most likely progenitor.
Cases from all species were included in the analysis. As data to estimate the serial interval and distance kernel for wildlife were lacking, we assumed these distributions for wildlife were the same as those for domestic dogs.

| Assessing within-and between-species transmission
The algorithm assigning progenitors does not account for unobserved cases. Attempts to adjust for unobserved cases were made , by analysing a subset of inferred transmissions considered most likely to represent direct transmission. Only inferred transmissions below the 99th percentile value of the serial interval and the convolution of two distance kernel distributions were analysed to assess within-and between-species transmission, corresponding to cut-off values of 156 days and 9,803 m respectively. Transmissions with serial intervals and/or distances above these cut-off values were considered less likely to represent direct transmission.
Within this subset of inferred transmissions, relative frequencies of within-and between-species transmission were estimated.
Weighted random sampling was used to select a single progenitor for each case from the set of bootstrapped progenitors for that case (selected with replacement from all cases) and the species recorded and used to construct a contingency table of inferred transmissions.
Fisher's exact test statistic was calculated to test whether the inferred levels of inter-species transmission would be expected under random mixing. This procedure was repeated 1,000 times, and median levels of inferred transmission and p-values were calculated.
To assess the robustness of the transmission tree results, we conducted sensitivity analyses for different scenarios. These included The mean cluster size (including clusters of one) per 6-month period from the first case recorded was calculated and a weighted linear spline regression performed to test for a temporal trend. A 6-month period was selected to allow full use of the data while allowing a long enough time window for clusters to be observed. Sensitivity analyses were performed using periods of 3 months and 1 year, with the 2019 data excluded from the 1-year analysis, as data for the full year were not available.

| Regression analysis of monthly incidence by species
Negative binomial regression models were fitted to the monthly probable rabies cases observed among all species, among domestic animals only and among wildlife only. Linear splines were used within the regression analyses where visual inspection of the data suggested a change in trend. The correlation between the monthly time series of cases in domestic dogs and in jackals was also examined, evaluating lags of 0 to 11 months for both time series.
2.3.6 | Logistic regression of cases in relation to population composition We further examined whether the proportion of wildlife cases within a district was related to their relative availability within the susceptible population, focusing on only dogs and jackals (95% of all cases) and including data from other districts where cases had been traced using the same methods, specifically from Serengeti district The susceptible population (jackals and dogs) was estimated as follows. Jackals were assigned to grid cells at a density of 0.3 per km 2 for all districts following a literature review (Durant et al., 2011;Maddox, 2003;Yarnell et al., 2013), except to cells with human population density below 2.5 per km² or over 500 per km² (using gridded population data from: https://www.world pop.org/). Areas with lower human densities were excluded assuming negligible case detection from these largely uninhabited settings; the upper density limit was applied assuming unsuitable habitat. Alternative lower lim- Using logistic regression, the proportion of probable cases (those in both dogs and jackals) that occurred in jackals was regressed against the proportion of the susceptible population consisting of jackals to assess evidence for a relationship.
All analyses were undertaken using the R statistical computing language (R Core Team, 2018).

| RE SULTS
Over the 9-year study period, 688 human exposures to probable rabid animals were recorded within the Lindi and Mtwara regions.
Of these exposures, 47 (6.8%) resulted in death due to probable rabies (none were laboratory confirmed). The number of probable animal rabies cases recorded over the same period was 549, com-   (Table S3).
Over time, probable rabies exposures from all species, and probable animal rabies cases in both domestic animals and wildlife de-

| Parameter estimates
The best-fitting distribution to the serial intervals recorded from Serengeti District, northern Tanzania was a log-normal with mu and sigma parameters of 2.80 and 0.97, respectively, correspond-

| Transmission trees
Of the 549 inferred transmissions, 304 had values within the 99th percentile of the distributions for the serial interval and the convolved distance kernel and were included in subsequent analysis. Very little effect was seen on the percentage of wildlife-to-wildlife transmission (Table S4). Subsampling dog cases appeared to reduce the percentage of transmission inferred to occur from dog-to-dog TA B L E 1 Probable animal rabies cases, human exposures and human rabies deaths by infecting species detected from January 2011 to July 2019 in Lindi and Mtwara regions. In addition, four people died from bite injuries from probable rabid hyenas (3) and a probable rabid jackal (1)

| Chains of transmission and cluster size
Chains of transmission were constructed from the most likely  Figure 5). Sensitivity analyses using 3month and 1-year periods were consistent, with statistically significant decreases in cluster size over the first 6-and-a-half and 7 years of the study respectively. After the initial period, cluster size increases significantly using 3-month periods, but this increase is not significant using 1-year periods likely due to the omission of the incomplete 2019 data.
Of the 32 cases where the progenitor was known, the correct biting animal was not always assigned with the highest bootstrap probability. In 16 of 32 cases, the biting animal was correctly identified with less than 5% bootstrap probability. All of these cases involved dogs that were part of clusters within households. The algorithm identified a different dog but always one within the same household and cluster, meaning the assigned species-to-species transmission was correct.  Figure S7).

| Regression analysis of monthly incidence
When assessing correlations between monthly domestic dog and jackal cases with lags ranging from 0 to 11 months, all scenarios  Table S5.

| Logistic regression of cases in relation to population composition
Logistic regression suggested a statistically significant positive relationship (p < 0.001, Figure 6) between jackals as a proportion of the susceptible population and the proportion of probable cases that were in jackals (when jackals were distributed across areas with >2.5 and <500 humans/km 2 ). Results obtained using minimum and maximum district-level vaccination coverages, different cut-offs for human densities used for estimating jackal populations and different jackal density estimates were all statistically significant (Appendix S5 and Figure S8).

| D ISCUSS I ON
Less than a decade remains to achieve the 'Zero by Thirty' global target of zero deaths from dog-mediated rabies by 2030. Here we present data from a previously unstudied area of south-east Tanzania following the introduction of large-scale dog vaccination.
We examine whether wildlife could present an obstacle to rabies only species in which rabies appears capable of persisting (Lembo et al., 2008), here we find much greater involvement of jackals, but still conclude that targeting dogs through mass vaccination should eliminate rabies in this area.
One challenge faced during this study was limited information on jackal populations. We extrapolated jackal numbers using density estimates from studies elsewhere in Africa, but this approach does not incorporate geographical population differences. More accurate jackal numbers would underpin a more confident assessment F I G U R E 5 Trend in mean cluster size per 6-month period from January 2011 until July 2019. Mean cluster sizes for each 6-month period (orange) with the fitted linear spline regression (green) and associated 95% CI. A statistically significant downward trend in mean cluster size was observed over the initial 6-and-a-half years of the study (p = 0.001, reduction in mean cluster size of 0.12 (95% CI: 0.06-0.17) per 6-month period), followed by a statistically significant increase from July to December 2017 (p = 0.028, increase in mean cluster size of 0.52 (95% CI: 0.10-0.93) per 6-month period) This, in turn, could provide further insight into which species drive transmission and whether assortative mixing underpins transmission pathways or if transmission depends more on the availability of susceptible animals regardless of species. Additional data will be needed to conclude whether jackals can maintain RABV independently over the longer term. Our conclusion that where more jackals are present, rabies incidence in these populations is correspondingly higher is robust to the range of estimates of jackal density considered.
A further limitation related to case detection and confirmation.
A low proportion of probable rabies cases were confirmed through laboratory diagnosis. Of the 549 clinically diagnosed animals, samples were collected in only two cases, both of which tested positive.
The low rate of sample submission was primarily due to delays in reporting across this large area such that on follow-up the animal had been lost or the carcass decomposed. Low rates of sample submission also meant that genomic data were not available. We considered that assigned progenitors in 304 of the 549 probable cases were likely to represent direct transmission. This suggests that despite intensive effort, over one third of cases were not observed (i.e. no progenitor found for the remaining 245 cases). We also note that case locations were associated with uncertainty (31.5% within an estimated 0-2 km radius of the precise location, 22.0% within 5 km and 5.1% within 5-10 km). Despite unobserved transmission, our results were nonetheless robust under sensitivity analyses.
Genetic sequencing should prove useful in resolving transmission chains by determining whether RABV lineages include cases in both dogs and wildlife and for identifying introductions of RABV via human-mediated dog movement. Although several RABV lineages have been detected across Tanzania, there is currently no evidence of species-specific lineage associations (Brunker et al., 2015;Lembo et al., 2008). Translocations of dogs has been shown to be important in the spread of RABV (Denduangboripant et al., 2005) and genomic approaches have revealed substantial human-mediated RABV movement in Tanzania (Brunker et al., 2015), which may explain how some apparently unconnected clusters and cases arose.
Overall, our study provides insights into the epidemiology of rabies in multi-host communities and highlights the potential importance of wildlife as sources of rabies exposure. Our data highlight the frequent transmission of rabies from sparsely distributed domestic dog populations, to and from sympatric wildlife, specifically jackals ( Figure 2). Yet even in areas with relatively high proportions of wildlife cases, domestic dog vaccination still reduced the risk to humans. Maintaining dog vaccination campaigns in LMICs is challenging, and we show that if vaccination campaigns are not maintained, resurgence of rabies can rapidly occur. Herd immunity wanes quickly with high demographic turnover in the dog population, and infection circulating in nearby populations can seed introductions (Hampson et al., 2009;Zinsstag et al., 2017). Continued dog vaccination is needed to eliminate rabies from south-east Tanzania and should shed more light on the involvement of wildlife in rabies maintenance. A useful future extension would be to estimate the type reproduction number for dogs and wildlife in these regions. The type reproduction number can be used to estimate the control effort required to eliminate an infectious disease within a maintenance community when control is targeted at a subset of hosts (Roberts & Heesterbeek, 2003). Its application to rabies control was recently illustrated in relation to vaccination strategies targeting owned, freeroaming and stray dogs (Leung & Davis, 2017  Organization in partnership with the Government of Tanzania.

CO N FLI C T O F I NTE R E S T
The authors have declared no conflict of interest. wrote the first draft, and all authors provided feedback and guidance.

DATA AVA I L A B I L I T Y S TAT E M E N T
The code for our analyses are available on our public Github public repository (https://github.com/Lusha siK/Stz_rabies_reser voir), as well as anonymized data which allow replication of most results.
The anonymized data are available via the Dryad Digital Repository https://doi.org/10.5061/dryad.bg79c npbg . Our full dataset contains georeferences that are available on request from the authors following ethical approval for secondary reuse.