Dog-mediated rabies is a serious zoonosis responsible for at least 59,000 human deaths every year, primarily in low-income countries in Asia and Africa where rabies is endemic [1
]. In these areas, over 95% of human rabies deaths result from bites by domestic dogs. Empirical data and mathematical modelling have shown that annual dog vaccination campaigns that achieve a coverage of 70% are sufficient to eliminate rabies [2
]. The global call for the elimination of dog-mediated human rabies by 2030 [4
], has prompted many countries to invest in dog vaccinations. For example, Tanzania has developed a National Rabies Control and Elimination Strategy, aiming to control dog rabies and eliminate human rabies in the country by 2030 [5
]. However, in low-income countries, dog population sizes are usually unknown, or hard to estimate, making it difficult to implement and evaluate dog vaccination campaigns [6
In most low-income countries where rabies is endemic, owned dogs are not registered by local authorities, dogs are free to roam, and dog censuses are not conducted. Insufficient knowledge of dog population sizes for planning of vaccination campaigns was reported as one of the limiting factors for ineffective rabies control in Africa [6
]. This lack of knowledge prevents countries from forecasting vaccine procurement needs, and hinders assessment of the effectiveness of mass dog vaccination campaigns. It is, therefore, important to develop practical methods for estimating dog population sizes.
Approaches for estimating dog population sizes include extrapolation from human/dog ratios derived from household surveys and from transects with counts of dogs differentiating unvaccinated and vaccinated dogs marked with collars or paint sprays [7
]. The accuracy of these methods has been questioned, as they generate very different population estimates from the same geographical areas, and surveys can be imprecise unless large numbers of households are sampled [7
]. Household surveys are restricted to assessing owned dogs, while transects capture only free-roaming (observable) dogs. In Tanzania, over 78% of dog owners reported that their dogs roam freely all the time [14
], and therefore, the majority of dogs can be observed during transects. In settings where the vast majority of dogs are owned, such as Tanzania [6
], a complete dog census, whereby each household in a community is visited, is the gold standard method to estimate the dog population, but requires considerable investment of time and resources.
Dog ownership patterns are not uniform across all communities. The distribution of dogs in different settings is linked to religious, cultural, geographical, and socioeconomic factors [9
]. For example, there tend to be fewer dogs in predominantly Muslim communities than in Christian communities [9
]. Tools to accurately estimate dog populations that take into account local cultural norms could, therefore, support the scaling up of mass dog vaccination programmes.
Our overall aim is to provide practical and effective approaches to estimate dog populations in different settings. Our first objective was to compare methods to determine which provides the most precise dog population estimates, and explain why estimates differ according to method. From this comparison, we identified that post-vaccination transects, which involve counting both vaccinated and unvaccinated dogs, provide more reliable dog population estimates than either household or school-based surveys. Our second objective was to identify factors that predict dog ownership in different settings in Tanzania, using our estimates of dog population size from transects. The aim of identifying these factors was to enable prediction of dog population sizes and densities in other parts of the country not yet subject to vaccination campaigns, which was our third objective. We assessed the performance of these factors from known populations in our study, and finally used these factors together with nationally available human census data to predict dog population sizes throughout Tanzania. Our findings should be valuable for both Tanzania and other countries as they develop, implement, and monitor their national rabies control programmes.
During the 2014–2015 mass dog vaccination campaigns, 86,361 dogs were vaccinated in the 28 study districts, and 86,142 dogs were vaccinated in the 2015–2016 campaign. The following data collection activities were completed: (i) post-vaccination transects in ~2100 villages in 2014–2015 and in ~2600 villages in 2015–2016; (ii) household surveys in 4488 households in 2011, from 160 randomly selected villages; and (iii) school-based surveys of 8254 primary school pupils (each representing a unique household) within 115 randomly selected schools following the 2014–2015 campaign. During the 2014–2015 transects, 18,436 dogs were counted, of which 63% were observed with collars, indicating that they were vaccinated. From the school-based surveys, 2198 owned dogs were reported, corresponding to a mean of 0.7 dogs per household, and from the household surveys, 731 dogs were recorded, corresponding to a mean of 0.6 dogs per household (Table A2
Estimation of dog population sizes:
The overall dog population estimated in these study districts varied according to the survey method used (Figure 2
). From transects, we estimated a total dog population in the 27 study districts of 164,000 (95% CI 163,000–169,000 reported to three significant digits) and an overall human/dog ratio of 53.6:1 in the study districts. The estimated dog population size leads to a vaccination coverage estimate of 52% (86,000/164,000), which is lower than the direct coverage estimate from transects (63%), due to adjusting for unobserved and unvaccinated pups. By contrast, using household and school-based surveys, we estimated the dog population in the study districts to be 412,000 (CI 348,000–544,000) and 403,000 (CI 341,000–531,000), respectively. District-level estimates of dog numbers from household and school-based surveys tended to have wide 95% confidence intervals, with a mean ratio of upper to lower confidence limit of 3.5 for household, and 3.0 for school-based surveys. Transect estimates were considerably more precise, with a mean upper/lower confidence limit ratio of 1.2. Household and school-based surveys were also sometimes highly inaccurate, giving estimates that were lower than the number of vaccinated dogs (Figure 2
). Transect estimates, by contrast, cannot be lower than the number of vaccinated dogs. We therefore considered transect estimates to be more reliable, and used them to fit the model for predicting dog population sizes. There was minimal year-to-year variation in the estimated number of dogs in each district from the data that were collected in 2014–2015 versus those collected in 2015–2016 (Figure A2
Prediction of dog population sizes and densities:
using our transect estimates, we investigated the influence of district-level variables on dog population sizes. The pairwise plots between the log-scale continuous variables investigated showed that the number of households was highly correlated with the human population (Pearson’s r
= 0.96). We therefore dropped the number of households from the model. A further two variables, the number of people living in rural areas and the number of livestock-owning households, were dropped in order to reduce all VIFs to below 5 (Table A1
). These variables were also highly correlated with human population size.
All 64 possible models were fitted from the combination of the six retained variables and the models were ranked by δAICC
). The top three models were almost equivalent in predictive power (R2FPE
= 58%), and were very close in δAICC
, which ranged from 0–1.27. Our best fitting model retained three variables: the proportions of livestock keepers and of peasants, and the geographic setting (inland versus coastal and island). The proportions of livestock keepers and peasants were both positively associated with the dog/human ratio: a doubling of the proportion of livestock keepers was associated with 28% (95% CI: 14%, 44%) larger dog populations, while the equivalent effect for the proportion of peasants was 36% (95% CI: 13%, 65%), all other characteristics being equal. We also found that there were 103% (95% CI: 21%, 120%) more dogs per person in inland districts than in island and coastal districts.
Two predictor variables: the proportion of livestock keepers and the geographic setting (inland versus coastal/island) were consistently retained in the best-ranked models (Table 2
). Dropping either the setting variable or the proportion of persons employed as peasants also reduced the variance explained by 16% (R2FPE
fell from 58% to 42% in both cases, giving a partial R2FPE
of 27%). Excluding the proportion of persons employed as livestock keepers in the final model reduced R2FPE
to 30%, showing the substantial predictive power of this variable (partial R2FPE
We used permutations and bootstrapping to assess the reliability of the selected best-fitting model. Using 1000 permutations, we estimated an FDR of 28%. Bootstrapping the model selection procedure showed that two of the three variables were highly robust: proportion of livestock keepers and coastal setting were selected in 97% and 91% of bootstrapped models, respectively (Table 2
). However, despite their robustness, a model containing these two predictor variables alone performed substantially worse than the best-fitting model, as evidenced by its relatively low R2FPE
of 42% and its selection as the best model in only 6% of bootstrapped datasets. Two other variables were selected with around 50% frequency, proportion of the population employed as peasants (49%), and human population size (51%). Based on δAICC
, and the reliability analysis, we concluded that two of the three variables selected are highly robust, and that at least one other variable is required to maximise predictive power. Since the top three models are almost equivalent in terms of predictive power (R2FPE
) and δAICC
, we selected the one with only three variables selected, which also happened to be the best-fitting model. This combination of three variables (the proportions of livestock keepers and of peasants, and geographic setting) chosen for our final model was used to predict the dog population across Tanzania.
From our final model, we predicted considerable variability in dog population sizes across all 168 districts in Tanzania (Figure 3
). Dog population estimates ranged from just 630 dogs in Kusini district on the island of Zanzibar, to 45,000 dogs in the inland district of Nzega. Overall, the final model predicted a dog population of 2,316,000 (95% CI 1,573,000–3,122,000) in Tanzania in 2014–2015. In 2014, a human population of 47,831,000 was projected from the Tanzania population census [16
]. Taken together with our predicted dog population size (2,316,000 dogs), we obtain an overall human/dog ratio of 20.7:1 in Tanzania. Generally, the highest dog ownership was predicted from inland districts, especially those dominated by rural livestock keepers. Human/dog ratios for each study district are presented in Table A2
Dog density from the predicted dog population sizes for each of the 168 districts also varied greatly, ranging from 0.14 to 113 dogs per square kilometre (Figure 4
). Liwale district (an inland study district which has an area of ~15,000 km2
) had the lowest dog density of 0.14 per km2
, while Bukoba urban (an inland district with an area of 30 km2
) had the highest dog density of 113 dogs per km2
. Based on Tanzania’s total land area and predicted dog population, we determined the mean density of dogs to be 8.81 per square kilometre (interquartile range: 2.24–9.23 per km2
). Predicted densities were highest in northern and northeastern Tanzania, and lowest in the southern and central-west parts of the country (Figure 4
Knowledge of the size of dog populations is critical for the planning and implementation of effective canine rabies control strategies. Before implementation of vaccination campaigns, this information is useful for determining the personnel required, and for the procurement of vaccines and other supplies. After implementation, this information is required to evaluate the intervention in terms of the vaccination coverage achieved. Our study uses data from well-studied dog populations in Tanzania, where vaccination campaigns have been conducted to predict the size of dog populations in new areas, and where vaccination campaigns will hopefully be scaled up in the future.
Although several methods have been used to calculate the size of dog populations [34
], these methods have not been comprehensively compared to assess which generates the most precise dog population estimates. From our comparison, we found that post-vaccination transects generated more precise and reliable estimates than either household or school-based surveys in Tanzania. Transects are, however, only reliable if all dogs (owned and unowned, free roaming or restricted) have an equal chance of being counted. In many sub-Saharan countries, most owned dogs are free roaming, so this assumption would hold. In a previous survey in Tanzania, most dog owners reported that they did not tie or cage their dogs. Those who reported restricting their dogs were from urban areas, while in rural areas, the vast majority of dog owners reported that they do not restrict their dogs at any time [14
]. Hence, transects are appropriate for estimating dog population sizes in Tanzania, but in countries where a large proportion of dogs are kept indoors, this method would not be appropriate.
The imprecision of the dog population estimates from household and school-based surveys was due to their low sampling effort compared with post-vaccination transects [13
]. Although sample size at the household level was not low (30 households per village and ~100 families per school), only six out of an average of 95 villages per district were sampled, and the precision of estimates from hierarchical sampling designs is expected to be dominated by sample size at the highest level [35
]. The overall proportions of the population surveyed were also small (4488/441,000 households from household surveys versus 8254/441,000 households from school-based surveys), despite the large numbers of surveys conducted, whereas for the post-vaccination transects, ~2500 villages were sampled. The practical consequence of imprecision in the estimates from household and school-based surveys for the implementation of mass dog vaccination is that the upper bound of the range of plausible dog population sizes is, on average, 3 times the lower bound, compared with 1.18 for transects.
In addition to being imprecise, estimates from household and school-based surveys were often inconsistent (Figure 2
), either with each other (for example, Kisarawe and Wete districts), or with the number of dogs vaccinated (for example, Mkuranga and Chakechake districts). The dog population may, therefore, be either over- or underestimated by projecting from the mean number of dogs per household, given limited sampling and considerable house-to-house variation in dog ownership. For example, Kinondoni district in Dar es Salaam has over 400,000 households with 0.32 dogs per household estimated from the household survey, which would suggest a dog population of 141,197, a large overestimate (Table A2
The number of vaccinated dogs (numerators), together with the size of dog populations derived from household and school-based surveys (denominators), were used to calculate the vaccination coverage, and compared against coverage estimated directly from these surveys. We found large discrepancies depending upon the method used. For example, in 2014–2015, around 86,000 dogs were vaccinated, but school-based surveys estimated a dog population of 399,000, which would have led to just 22% coverage (Table A2
). By contrast, direct estimates of dog vaccination coverage from school-based and household surveys were both around 56% (Table A2
). In some districts, the discrepancy was clearly implausible, for example, in Chake Chake, Pemba, where more than double the number of dogs estimated from school-based surveys were vaccinated (Table A2
). These discrepancies have practical impacts on rabies control, and demonstrate the need for careful post-vaccination evaluations, without reliance on dog population survey estimates only.
Our multivariable regression analysis identified the proportion of livestock keepers and geographical setting as robust variables for predicting dog population sizes, consistent with previous studies on factors influencing dog ownership in different parts of the world [34
]. In most African countries, dogs are reported to play a role in protecting livestock, explaining why livestock keeping was such an important variable [9
]. Other factors reported to influence dog ownership in Africa include socioeconomic status, livelihood (which could include livestock keeping), culture, and religious beliefs linked to different settings. The reason for fewer dogs in Tanzanian coastal and island areas could be linked to the predominantly Muslim communities in these areas, as Muslims are reported to own fewer dogs [9
]. Our study demonstrates the need to consider such factors in planning vaccination campaigns, and in understanding dog rabies incidence, control, and prevention, more generally.
We found that despite the robustness of these two variables (livestock keeping and geographical setting), by themselves, they accounted for only about two-thirds of the predictive power of the best-fitting model. A third variable was also needed to improve predictive power. Our model validation showed that the proportion of peasants or the human population were almost equivalent in the final model (R2FPE
values of 58% and 55%, respectively). If applying this model to new settings, the decision as to which variable to use will depend on the availability of data on either of these variables. We also do not propose evaluating campaign success by working backwards from the total dog population estimates derived from this predictive model, because the uncertainty in this estimate is inflated from the large variation in district population sizes. To evaluate coverage across districts, we would suggest estimating mean coverage directly from the village-level transect measures. District-level estimates of dog populations (Figure 2
), which are based on these coverage estimates, are quite precise, as are directly derived district-level estimates of coverage ([13
], 58–65% not adjusted for puppies).
Our study was consistent with previous findings in Tanzania, which reported more dogs in mainland compared to island and coastal areas [9
]. However, our overall human/dog ratio estimate of 20.7:1 was higher than previous studies in Africa, which ranged from 3:1 to 15:1 [6
]. This suggests that human/dog ratios extrapolated from household or school-based surveys could be unreliable when extrapolated to district or national level. Initially, the study area was estimated to have about 400,000 dogs [5
], based on reported human/dog ratios [9
], which was much higher than the number of dogs subsequently estimated with post-vaccination transects (164,000 (95% CI 163,000–169,000). The lower number of dogs in our study suggests that dog vaccine requirements in Africa might be less than previous estimates. A study in Uganda also found lower numbers of dogs than previously estimated [43
]. If this pattern holds across more countries, the lower number of dogs provides further incentives for African governments to undertake vaccination programmes, as the target of 70% could be more easily achieved [43
]. However, our data were largely collected from southeast Tanzania, where there are fewer pastoralists who tend to own more dogs [44
], and from coastal or island districts (~50% of study districts) which tend to have fewer dogs. This suggests that additional data (dog vaccination and transect surveys) from other populations (inside and outside of Tanzania) would be valuable to further refine and validate this predictive approach.
Several household surveys have been conducted in Tanzania, generating lower human/dog ratios than we found from transect-based estimates. For example, in Iringa urban, the human/dog ratios were estimated to be 14 [8
], versus our transect estimate for Iringa urban of 34. Meanwhile, in Kilombero and Ulanga districts, human/dog ratios were estimated to be 12 and 29, respectively, from households surveys [15
], in contrast to our transect estimates of 21 and 18, respectively. Variation was also reported by geographical setting, with human/dog ratios estimated to on average be 7.6:1 in rural-inland areas, 10.8:1 in rural-coastal areas, 27.1:1 in urban-coastal areas, and 14.4:1 in urban-inland areas [9
]. These estimates derived from household surveys are likely to be affected by limited sampling. Mark-recapture studies are useful for estimating numbers of dogs [7
]. Our study suggests that transects, when done in association with dog vaccinations at scale, can capture population variability. However, there is still a need for predictive methods for working in areas where dog vaccinations have yet to be conducted, such as the model that we developed.
Our predictive model could be used to make preliminary predictions of dog numbers in other countries that are similar to Tanzania, with respect to dog-owning practices i.e., where most dogs are free roaming and there are very few unowned dogs. This model can give a starting point for settings with no dog population size estimates even prior to any vaccination campaigns (and transects), given available data on the proportions of livestock keepers and of peasants or on human population sizes. Such preliminary dog population estimates could provide a baseline for planning mass dog vaccinations. The wide confidence intervals of these model estimates may initially mean procurement of excess or insufficient numbers of vaccine vials. However, overprocurement should not be problematic, as the vaccines can be stored for long periods (normally three years) for use in future campaigns. Moreover, subsequently, the vaccination and transect data generated during vaccination campaigns should be used to refine dog population size estimates. Conducting post-vaccination transects in every village is, however, labour-intensive and costly [13
], so should not necessarily be undertaken every year. We recommend conducting transects at least in the first years of the undertaking campaigns, to refine dog population estimates. Awareness and participation of dog owners typically increase in the first few years of a rabies control programme [5
], so transects may help to refine estimates in the second or third campaigns. However, in subsequent years, established denominators can be used to evaluate the performance of vaccination campaigns as substantial changes in dog population sizes are not expected (Figure A2
). We do also recommend repeating transects after several years given dog population growth, and conducting transects in areas where control programmes have been less successful than expected, so that any coverage gaps that may be limiting progress can be identified [13
Our overall estimates of dog density were higher those reported from elsewhere in Africa [37
]. This was probably because we excluded water bodies and protected areas when calculating densities. Our dog density map highlights districts with high dog densities that should be prioritized in the scaling up of dog vaccinations (only two districts, Moshi urban and Zanzibar urban, had densities exceeding 120 dogs/km2
, Table A2
), and districts where dog densities might be too low to support rabies transmission without importation from other districts [47
Our study had several limitations. The main limitation was that we could not externally validate our predictive model, due to a lack of reliable data on dog numbers outside the study area, with the exception of Serengeti district. In addition, our study areas did not cover many inland livestock keeping districts. Our surveys were also completed at different times, with household surveys conducted in 2011, transects immediately following dog vaccination campaigns in 2014/15, and school-based surveys within 2 months of these dog vaccinations. These differences may have affected our estimates. Transects do also have several limitations. They may result in the recounting of dogs [51
], but we tried to avoid this in the design of transect paths. Larger villages also require more time to complete and those with more subvillages were less well sampled. Only observable dogs are counted from transects, which results in systematic biases, such as poor observation of pups [7
], but we tried to adjust for this using pup/adult ratios. The pup/adult ratio was calculated from a dog census completed in Serengeti district data between 2008 and 2015. Although this dog census was conducted over multiple years, we do not expect that the dog population structure changed very much during this period. Notwithstanding these limitations, we found that transects were fast and relatively low cost to complete at scale, sampling populations more representatively than other approaches that were limited in spatial scope. We recommend that marking of vaccinated dogs (visible markers/collars) should be included as part of mass dog vaccination campaigns, and that transects should be completed immediately after vaccination campaigns, aiming to cover the centre and the periphery of villages, as coverage has been reported to decrease with the distance to the vaccination point [17
]. Estimates should also be adjusted to account for not observing pups.