Introduction Rabies has been one of the most feared diseases throughout human history and has the highest human case-fatality proportion of any infectious disease [1,2]. Every year over 7 million people receive post-exposure prophylaxis, and an estimated 55,000 people die from rabies [3] (more than yellow fever, dengue fever, or Japanese encephalitis [4]). Over 99% of these deaths occur in developing countries where rabies is endemic in domestic dog populations [5]. However, the impacts of canine rabies are often overlooked, largely because human rabies deaths are now extremely rare in Western Europe and North America, where mass vaccination successfully eliminated the disease from domestic dog populations [6]. Increasing incidence of canine rabies in Africa and Asia has prompted concerns that similar strategies may not be effective in these areas [7,8]. The critical question now is whether global elimination of domestic dog rabies is achievable. Keys to answering this question include: a quantitative understanding of the transmission dynamics of rabies in domestic dog populations, particularly the basic reproductive number, R0; a quantitative understanding of domestic dog demography; and information about the practicality and effectiveness of various vaccination strategies. While recent data support the feasibility and practicality of domestic dog vaccination strategies [9–11], there are very little quantitative data on rabies transmission dynamics [12] and the underlying demographic processes. Transmission is the most important process underlying infectious disease dynamics [13], but it is also the least understood. Rates of transmission are usually inferred from population patterns of disease incidence, but population-level analyses do not capture between-individual variation in transmission resulting from differences in behaviour, genetics, immune status, and environmental and stochastic factors, which play an important role in determining disease dynamics [14,15]. Contact tracing has been used to directly measure case-to-case transmission, and applications of the technique to emerging infections such as SARS have generated important insights into disease transmission and control in human populations [16,17], but transmission processes for diseases circulating in animal populations are much harder to study. Rabies is an acute viral encephalitis that is spread through the saliva of infected hosts [2]. Clinical manifestations vary, but the neurological phase often includes increased aggression and the tendency to bite and thereby transmit infection; rapid progression to death is inevitable [4]. These distinctive signs make transmission of rabies easier to track than that of most other diseases and provide an unusual opportunity to explore epidemiological patterns at the scale of the individual. Here, we present data on rabies transmission in two districts of rural Tanzania, Serengeti and Ngorongoro (Figure 1). We were able to monitor the spread of infection using contact-tracing methods, which were feasible due to the discrete and memorable nature of transmission events. We recorded >3,000 potential transmission events between 2002 and 2006 and reconstructed case histories of over 1,000 suspect rabid animals that illustrate heterogeneity in several aspects of transmission, including the latency, movement patterns, and biting propensity of infected individuals. Although these districts border the Serengeti ecosystem, we have argued that domestic dogs are the sole maintenance population of rabies in this community: they make up over 90% of our observations of rabid animals, and the >70 isolates that have been sequenced (from 13 host species) are all consistent with the Africa 1b canid strain [18,19]. This is one of the most extensive datasets on individual transmission events assembled in an animal population; it has potential to shed light on critical, but often elusive, details of infectious disease transmission. We also analyze data from rabies outbreaks around the world, which provide a global and historical context for the Tanzania dataset. Figure 1 The Location and Timing of Animal Rabies Cases in Serengeti and Ngorongoro Districts, Northwest Tanzania (A) Rabies cases in Serengeti (blue) and Ngorongoro (red) districts from January 2002 until December 2006. LGCA = Loliondo Game Controlled Area, NCA = Ngorongoro Conservation Area. Dark gray lines show village boundaries. Populations of humans and domestic dogs are denser in Serengeti district than Ngorongoro (Table 3). (B) Biweekly time series of rabies cases in each district. Results Epidemiological Parameters and Transmission Analyses of the contact-tracing data generated robust estimates of epidemiological parameters that have important implications for rabies control (Table 1, Figures 2 and 3, and Figure S1) and provide insight into how infectious disease transmission scales from individual behaviour to population-level dynamics. We estimated R0 for rabies in Serengeti and Ngorongoro districts directly from infectious histories, from reconstructed epidemic trees based on the spatiotemporal proximity of cases, and from the exponential rate of increase in cases at the beginning of an epidemic. Biting behaviour of rabid dogs during the course of infectious periods was highly variable (mean bites per rabid dog = 2.15, 95% confidence interval (CI) from fitting a negative binomial distribution: 1.95–2.37; variance = 5.61, CI: 4.63–6.92; shape parameter k = 1.33; CI: 1.23–1.42) (Figure 3A). The probability that an unvaccinated dog developed rabies after being bitten by an infectious animal was high (P rabies|bite = 0.49, CI: 0.45–0.52) (Table 1) if the bitten dog was not vaccinated or killed immediately after exposure. Multiplying the average number of dogs bitten per rabid dog by the probability of developing rabies following exposure gave an R0 estimate of 1.05 (CI: 0.96–1.14) (Figure 3A and Table 1). These estimates should be regarded as lower bounds, because not all transmission events were observed (this calculation excludes rabid dogs that were killed before biting other animals or that disappeared and likely corresponded to unknown or unobserved rabid dogs in other areas; see Materials and Methods). Detailed data on the timing and location of transmission events and infections allowed us to estimate the spatial infection kernel and generation interval (distances and times between source cases and their resulting infections, respectively) (Figure 2) and probabilistically reconstruct transmission networks (Videos S1 and S2). Calculating the average number of secondary cases per rabid dog during the period of exponential epidemic growth (before vaccinations were implemented) from these reconstructions gave similar R0 estimates of 1.1 in Serengeti district and 1.3 in Ngorongoro (CIs: 1.04–1.10 and 1.26–1.42, respectively) (Table 1). The more traditional approach of estimating R0, by fitting a curve to incidence data over the same interval of exponential epidemic growth, also produced similar estimates of 1.2 in Serengeti and 1.1 in Ngorongoro (CIs: 1.12–1.41 and 0.94–1.32, respectively) (Table 1 and Figure 3B). This approach is robust to underreporting (Text S1 and Figure S2) but should likewise be considered a lower bound, because some local control measures were instituted (such as tying or killing). We also estimated R0 from the intrinsic growth rate of outbreaks of domestic dog rabies elsewhere in the world (Table 2) and obtained values between 1.05 and 1.85, which are consistent with our estimates from northwest Tanzania. Table 1 Epidemiological Parameter Estimates Figure 2 Observed Frequency Distributions of Important Epidemiological Parameters (A) The incubation period, (B) the infectious period, and (C) the spatial infection kernel. The best fitting gamma distributions to the data are shown by black lines (see Materials and Methods). Figure 3 Transmission of Rabies (A) The distribution of dogs bitten per rabid dog (fitted by a negative binomial distribution with mean = 2.15 [95% CI: 1.95–2.37]; variance = 5.61 [95% CI: 4.63–6.92]; shape parameter k = 1.33 [95% CI: 1.23–1.42]; R0 ∼ 1.1). To calculate R0, we excluded dogs that were killed, tied, or those that disappeared before biting any other dogs. Variability in biting behaviour means that a small number of individuals disproportionately affect transmission and can potentially spark an epidemic, but since most individuals cause few, if any, infections, R0 is low and most introductions quickly die out (Figure 4C). (B) Exponential epidemic growth in Serengeti (blue, R0 ∼ 1.2) and Ngorongoro (red, R0 ∼ 1.1) districts. The R0 estimates from the epidemic trajectories were relatively insensitive to the period used for fitting the exponential curve. The inset shows the distribution of R0 estimates based on fitting to different regions of the time series. (C) The effective reproductive number, R, (averaged over three-month intervals) for Serengeti (blue) and Ngorongoro (red) districts measured from reconstructed epidemic trees that incorporate prior knowledge on who infected whom. Dots indicate the number of secondary cases resulting from each primary case (inferred from the composite tree of most likely links, with random jitter to avoid superposition on the y-axis). R0 estimated from these reconstructions (during the period of exponential epidemic growth) was ∼1.1 and ∼1.3 for Serengeti and Ngorongoro, respectively. Table 2 Estimates of R0 for Outbreaks of Rabies in Domestic Dog Populations around the World Table 3 Demographic Parameters and Population Attributes Estimated from Domestic Dog Populations in Northwest Tanzania Figure 4 The Impact of Vaccination on Transmission (A) The size of village-level outbreaks (defined as at least two cases not separated by more than one month, isolated cases are assumed to be non-persistent introductions) in Serengeti (blue, n = 138) and Ngorongoro (red, n = 20) districts plotted against village-specific vaccination coverage at the outbreak onset. Coverage was extrapolated from a demographic model initialized with village-specific dog population estimates and incorporating village-specific vaccination data. Gray shading and contours correspond to the probability of observing an outbreak of a particular size or less, generated from 10,000 stochastic simulations of rabies transmission for every initial vaccination coverage (contours were calculated conditional upon >1 secondary case occurring). The inset illustrates a village-level example of the susceptible reconstruction used to calculate instantaneous vaccination coverage plotted beside rabies cases in that village. (B) The distribution of secondary cases per infectious dog as inferred from reconstructed epidemic trees in Serengeti (blue) and Ngorongoro (red) districts, plotted against vaccination coverage in the village where the primary case occurred. Random jitter was added to prevent superposition on the y-axis. (C) Probability of an outbreak being seeded by an introduced case under different levels of vaccination coverage. Due to heterogeneity in the transmission process outbreaks rarely occur when coverage is maintained above P crit. However if infections are frequently imported from outside the vaccinated region, at least 40% coverage would need to be maintained to reduce the probability of subsequent outbreaks (of at least ten cases) to 70%. Small outbreaks occurred in villages with lower coverage and the largest (and longest) outbreaks only occurred in villages with 0.5 IU/ml) in response to vaccination [32], suggesting that these factors do not impair the efficacy of dog vaccination in rural Tanzania. In addition, numerous practicalities—such as occasional failures in the cold chain, improper vaccination of animals, mistaken registrations, etc.—will all reduce the level of population immunity below the estimated vaccination coverage. Furthermore, our observations and simulations confirm that small outbreaks may occur simply by chance even when coverage exceeds P crit [33], and these are particularly likely when there is individual variation in transmission (Figure 4). Higher levels of coverage are therefore necessary to reduce the chance of outbreaks with greater certainty; especially where the risk from imported infections is highest (Figure 4C). This could be a concern if canine rabies were to be eliminated from domestic dog populations but continued to circulate in sympatric wildlife; however, canine rabies was successfully eliminated in Western Europe and North America despite the presence of wildlife hosts capable of transmission. Thousands of people die every year from this horrific and preventable disease, because the control of canine rabies has been severely neglected in developing countries [2]. Inherent inter-annual periodicity of epidemics exacerbates the situation, with rabies only intermittently perceived as problematic [6], as illustrated by the recent outbreak in China [34]. The problem of canine rabies has often been considered intractable in rural Africa, because of poor infrastructure, limited capacity, and the misperception that large populations of wild carnivores are responsible for disease persistence. Our analyses show that global control of canine rabies is entirely feasible and that successful elimination of canine rabies in many parts of the world has likely been achieved precisely because R0 is so low and institutional commitment to maintain high levels of vaccination coverage has been sustained [6]. Achieving vaccination coverage of 60% or more in dog populations in Africa is both logistically and economically feasible through annual vaccination campaigns [9–11,29]. The resultant reduction in costs of human post-exposure prophylaxis suggest that vaccination interventions targeted at domestic dog populations could translate into appreciable savings for the public health sector [3,8,29]. Furthermore, the inherently low R0 and the tractability of rabies contact-tracing indicates that once endemic rabies is controlled, elimination could be achieved through active case detection in remnant foci of infection (much like the strategy used to eradicate smallpox [35]); similar measures are proving effective in programmes to eliminate canine rabies in the Americas [36]. However, the most crucial step towards global elimination of canine rabies will be sustained commitment and coordinated efforts to maintain sufficient vaccination coverage in domestic dog populations. Materials and Methods Study areas. We collected data from two districts in northwest Tanzania: Serengeti, inhabited by multi-ethnic, agro-pastoralist communities and high-density dog populations, and Ngorongoro, a multiple-use controlled wildlife area, inhabited by low-density pastoralist communities, predominantly Maasai, and lower-density dog populations (Figure 1). Attributes of the dog populations in these districts are presented in Table 3. Wildlife populations also differ in the two districts, but domestic dogs are the focus of this study because they are the only maintenance population of rabies in the area [18]. Incidence data. Data on patients with animal-bite injuries from hospitals and dispensaries, case reports of rabid animals from livestock offices, and community-based surveillance activities were used as primary sources [18]. Visits were made to investigate incidents reported in 2002 to 2006 involving suspected rabid animals. Cases were mapped at the site of the incident (wherever possible) and villagers interviewed to evaluate the status of the biting animal, determine its case history, and identify its source of exposure and subsequent contacts (if known). The same procedure was exhaustively followed for all associated exposures/cases. Interviews were conducted with veterinary officers, local community leaders, and livestock field officers in attendance, resulting in an active reporting network. Cases were diagnosed on epidemiological and clinical criteria, adapting the “six-step” method through retrospective interviews with witnesses [37]. Rabies was suspected if an animal displayed clinical signs [37] and either (a) disappeared or died within 10 days, or (b) was killed, but had a history of a bite by another animal or was of unknown origin. Additional clinical criteria for wild carnivores (∼10% of human exposures were caused by wild animals and ∼10% of inferred transmission events involved rabid wildlife) included tameness, loss of fear of humans, diurnal activity (for nocturnal species), and unprovoked biting of objects and animals without feeding. When multiple incidents involving suspected rabid wildlife were reported on the same/consecutive days within neighbouring homesteads, we assumed a single animal was involved. Brain samples were collected and tested for confirmation wherever possible, but despite efforts to obtain diagnostic samples, most cases reported here were suspected rather than confirmed. Inadequate sample preservation such as storage at room temperature and long intervals between sample collection and testing (during which samples underwent repeated freeze-thaw cycles) probably caused specimens to deteriorate. Composite samples of each brain necessary to achieve the highest test reliability were also rarely available. Nevertheless, a high percentage of samples from suspected cases of rabies were confirmed by laboratory diagnosis (∼75%) suggesting that use of epidemiological and clinical criteria is justified and reliable [18]. Researchers are encouraged to contact the authors regarding data availability. Vaccination data. Dog vaccination campaigns in Serengeti district in 2000 resulted in low and patchy vaccination coverage (35–40% estimated from post-vaccination household surveys). Annual campaigns conducted from 2003 onwards in a 10-km zone adjacent to the western border of Serengeti National Park achieved higher coverage levels of between 40 and 80%. In 2004, the Tanzanian government conducted vaccinations in villages in Serengeti district beyond the 10-km zone reaching 55% coverage across the remainder of the district, but in subsequent years, campaigns were less systematic and conducted in fewer villages. Vaccination in Ngorongoro was restricted to small-scale localised campaigns in the district town centre until 2004, whereupon widespread annual vaccinations were implemented with overall coverage exceeding 80% [9]. Data on the number of dogs vaccinated in each village and on each campaign date were collected from 2003 onwards. Parameter estimation. The incubation period and duration of infectiousness were estimated for rabies in domestic dogs from records of when individual dogs were bitten, developed clinical signs, and were killed or died. Gamma distributions were fitted to these data using maximum likelihood with interval censoring to account for cases where the relevant data were only approximately known (Figure 2 and Table 1). To estimate the probability distribution of the generation interval, G(t), an incubation and an infectious period were drawn from their respective distributions, a “time-to-bite” deviate was drawn from a uniform distribution over the interval of the infectious period, and the two intervals were summed. There was a significant correlation between the length of the infectious and incubation periods, but significance was entirely due to a single data point; we therefore treated the distributions as independent. The spatial infection kernel K(d) was estimated by fitting a gamma distribution to the distances between known source cases and animals that they contacted. Many contacts occurred within the same, or neighbouring, homesteads. In these cases, the precise distance was not always recorded, but we assumed it was less than 100 m. We therefore replaced the probability of a contact within 100 m by the probability distribution averaged over the range 0–100 m. The basic reproductive number R0. (1) Direct estimates from infectious histories. Using maximum likelihood, we fitted a negative binomial distribution to data on biting behaviour of rabid dogs (Figure 3A). The probability of developing rabies following a bite (P rabies|bite) was estimated, excluding bitten animals that had previously been vaccinated, or that were either killed or vaccinated immediately after the bite, and binomial confidence intervals were calculated. R0 was estimated as the probability P rabies|bite multiplied by the average number of bites per rabid dog and confidence intervals were calculated using a resampling procedure. Dogs that were removed (killed or tied up) before causing secondary cases in other dogs (even if they bit people) were excluded from this calculation, as were suspect rabid dogs that either disappeared before biting other dogs or that were of unknown origin and were killed before being observed to bite other dogs (Figure 3A). We pooled data from both districts for this estimate because insufficient complete case-histories of rabid dogs (after excluding cases with interventions) were traced to accurately estimate R0 for Ngorongoro (35 versus 477 in Serengeti). We also estimated R0 directly from the distribution of secondary cases per rabid dog. Dogs that were bitten by rabid animals but did not develop rabies because of interventions (previous vaccination or being killed/vaccinated immediately after the bite) were multiplied by P rabies|bite and added to observed secondary cases, giving an expected number of secondary cases per rabid dog in the absence of intervention and a similar estimate of R0 (1.14, CI: 1.03–1.25) (Figure S1). (2) Epidemic tree reconstruction. We used an algorithm for probabilistically constructing epidemic trees based on the location of cases in space and time [38]. For each suspected case (i), we chose a progenitor (j) at random with probability pij from all n cases preceding that case, where: G is the distribution of generation times, tij is the length of time (in days) between the occurrence of case i and its potential progenitor j (G(t) = 0 for t 1 secondary case (Figure 4A). Demographic parameters were incorporated, and 10,000 runs were completed for each starting condition. We also calculated the probability of an outbreak of a particular size or larger being seeded by one infectious case to evaluate the coverage needed to prevent outbreaks with different degrees of certainty (Figure 4C and Figure S4). If V and N denote numbers of vaccinated individuals and the total population size respectively, then vaccination coverage can be expressed as a proportion P = V/N. The number of vaccinated dogs declines following a campaign as individuals die and as vaccine-induced immunity wanes (V t = V0 e – (d+ν)t , where d is the death rate and 1/ν is the duration of vaccine-induced immunity), whereas the total population grows at the rate of population increase (N t = N 0e rt ). To prevent sustained endemic transmission, vaccination coverage must be maintained above P crit (such that R is held below 1). From our estimates of demographic parameters and R0, we calculated the proportion of the population that needs to be vaccinated, P target, to prevent vaccination coverage falling below P crit during the interval, T, between campaigns: P target = e(ν+d+r)T P crit. This formulation for estimating the coverage needed to interrupt endemic transmission given turnover in the domestic dog population assumes that immunity from vaccination lasts an average of 1/ν time units and declines exponentially. In reality, vaccine-induced immunity is likely to be closer to a fixed duration, and thus fewer dogs would be expected to lose immunity during the 1-y interval between campaigns than under the exponential model. This indicates that our estimate of P target may be slightly overestimated, although this is an important area for further investigation. Supporting Information Figure S1 R0 Estimated from Secondary Case Distributions Observed numbers of secondary cases are shown in gray. We extrapolated additional cases (using the probability of developing rabies following a bite, 0.49) that would have occurred had there been no intervention (black). The inset shows the estimated secondary case distributions (∼1.1-1.3) for dogs in Serengeti (black) and Ngorongoro (red) districts based on the reconstructed trees during the early stages of the epidemics. (3.92 MB EPS) Click here for additional data file. Figure S2 Accuracy of R0 Estimates Derived from Epidemic Trajectories (A) Estimates of R0 from fitting to trajectories of simulated epidemics plotted against the underlying R0 used in the simulations (biting behaviour was modelled using a negative binomial distribution, varying the mean number of bites per dog whilst keeping the shape parameter constant). The median R0 estimate from 1,000 realizations is shown by the solid black line and 95 percentiles are indicated by gray shading. R0 was estimated accurately across a range of underlying R0 values apart from at very low values (R0 25% from monthly time series and >800% from weekly time series, upper 95% prediction intervals of 1.71 and 2.65, respectively, versus 1.65 and 1.69, respectively). (5.58 MB EPS) Click here for additional data file. Text S1 Impacts of Under-Reporting and Incomplete Tracing on Estimation of R0 (24 KB DOC) Click here for additional data file. Text S2 Effects of Heterogeneity in Transmission Behaviour (22 KB DOC) Click here for additional data file. Video S1 Rabies Transmission in Serengeti District Inferred from Detailed Spatiotemporal Incidence Data and Estimated Epidemiological Parameters Rabies cases appear as red dots. Incubating animals appear as black dots, which turn red when clinical signs start (only animals that went on to develop rabies are shown). When a rabid animal bites another animal that will subsequently develop rabies, a black line connects the two individuals. The video is on a weekly timescale and the red arrow on the time series of rabies incidence corresponds to infectious cases during that week. (2.20 MB AVI) Click here for additional data file. Video S2 Rabies Transmission in Ngorongoro District Inferred from Detailed Spatiotemporal Incidence Data and Estimated Epidemiological Parameters Rabies cases appear as red dots. Incubating animals appear as black dots, which turn red when clinical signs start (only animals that went on to develop rabies are shown). When a rabid animal bites another animal that will subsequently develop rabies a black line connects the two individuals. The video is on a weekly timescale and the red arrow on the time series of rabies incidence corresponds to infectious cases during that week. (409 KB AVI) Click here for additional data file.
