Socio-economic correlates and management implications of livestock depredation by large carnivores in the Tarangire ecosystem, northern Tanzania

ABSTRACT Livestock depredation by large carnivores is the key source of human–carnivore conflict worldwide and entails financial losses to livestock keepers. We examined the extent and patterns of livestock depredation, the financial impacts of livestock losses and determinants of livestock depredation by large carnivores in the Tarangire ecosystem of northern Tanzania. Of 300 households surveyed, 74.7% reported losses of 1906 livestock to wild predators over 1.5 years, which represents an annual loss rate of 1.4% of their total herd. Spotted hyena (Crocuta crocuta) accounted for 70% of the total livestock loss, followed by leopard (Panthera pardus) (12%), African wild dog (Lycaon pictus) (8%), lion (Panthera leo) (7%) and cheetah (Acinonyx jubatus) (3%). This loss equated to a total financial loss of US$141,847 amounting to approximately US$633/household/year. Depredation frequency by all carnivore species increased significantly with increasing number of livestock owned, respondent’s residency time, distance from the park boundary and declined significantly with increasing education, number of herders and improved fortified boma for cattle. Livestock depredation peaked during the wet season linked to seasonal migration of wild prey. Our study suggests that improving formal and conservation awareness education, boma fortification as well as improving herding practices could help mitigate the human–carnivore conflict. EDITED BY Carsten Smith-Hall


Introduction
Livestock depredation by large carnivores is one of the most important sources of human-carnivore conflicts as well as a major challenge threatening the conservation of large carnivores around the world (Woodroffe 2000;Treves & Karanth 2003). Large carnivores range widely, in such a way that existing protected areas (hereafter PAs) are not large enough to sustain their long-term viable populations (Woodroffe & Ginsberg 1998). Hence, due to their large home ranges and high dietary protein requirements, large carnivores tend to move outside PAs and overlap with human-dominated landscapes (Treves & Karanth 2003;Patterson et al. 2004;Woodroffe et al. 2005). This close proximity to humans often results in conflict due to the damage they cause to livestock (Woodroffe & Ginsberg 1998;Patterson et al. 2004). Outside PAs, large carnivores are killed either deliberatively or accidentally, thus making the borders a 'sink' in which human-caused mortality might limit survival of predators dispersing from the PAs (Woodroffe & Ginsberg 1998;Kolowski & Holekamp 2005;Kiffner et al. 2009). The economic losses associated with livestock depredation on local communities often provoke retaliatory and preventative killing of the large predators (Ogada et al. 2003;Patterson et al. 2004), which have a substantial impact on carnivore populations and thus jeopardize conservation efforts Dickman 2008). Consequently, reducing antagonism towards large carnivores following depredation will contribute towards their conservation and promoting coexistence between humans and large carnivores.
Rates of livestock depredation by large carnivores may be influenced by environmental conditions, e.g. abundance and distribution of natural prey (Mizutani 1999), seasonal patterns (Patterson et al. 2004), socioecological factors, livestock husbandry practices and characteristics of livestock enclosures (Ogada et al. 2003). In Tanzania, five large predators (lions, Panthera leo; leopards, Panthera pardus; cheetahs, Acinonyx jubatus; spotted hyenas, Crocuta crocuta and African wild dogs, Lycaon pictus) are chiefly responsible for livestock depredation (Holmern et al. 2007;Dickman 2008;Kissui 2008). Other carnivores (striped hyenas, Hyaena hyaena and caracal, Caracal caracal) occasionally prey on livestock as well. The Tarangire ecosystem is one of the richest wildlife areas in northern Tanzania. However, habitat CONTACT Felix J. Mkonyi mkonyif@nm-aist.ac.tz The supplemental material for this article can be accessed here.
loss and fragmentation associated with increasing human population and the conversion of land for agriculture and livestock grazing (Msoffe et al. 2011) have resulted in frequent encounters of large carnivores with humans and their livestock in this landscape. Livestock depredation by large carnivores entails economic damage to livestock keepers in Tanzania. However, diseases have been reported to contribute to far more livestock losses than depredation in other Tanzanian areas (Holmern et al. 2007;Nyahongo 2007;Kissui 2008;Nyahongo & Røskaft 2012).
Previous studies in the Tarangire ecosystem have focused either on actual livestock depredation events on a small set of carnivore species (Kissui 2008;Mponzi et al. 2014) or single species conflict i.e., human-lion conflict (Lichtenfeld 2005). In addition, patterns of livestock depredation by large carnivores have been well documented in Ruaha landscape (Dickman 2008) and western Serengeti National Park (Holmern et al. 2007). In this study, we build upon these previous studies by looking more broadly at the patterns of livestock depredation as well as ecological (distance to park boundary) and socioeconomic factors (household size, education levels, number of livestock owned) influencing reported conflicts with a wide range of large carnivore species in the eastern part of the Tarangire ecosystem. There is currently limited information on these factors in our study area and across landscapes (Dickman et al. 2014;Hampson et al. 2015). Correspondingly, no empirical data are available on perceived costs of livestock depredation by large carnivores and other causes of livestock losses such as disease and theft on people's livelihoods in this ecosystem. The cost of livestock depredation may play a critical role in shaping people's attitudes and behaviour towards carnivores (Hazzah 2006;Kideghesho et al. 2007;Røskaft et al. 2007;Lyamuya, Masenga, Mbise et al. 2014, 2016Bencin et al. 2016). Furthermore, despite the existing studies on the extent of livestock depredation in Tanzania, relatively few studies have investigated the key determinants (ecological and socio-economic factors) of perceived human-carnivore conflict (Holmern et al. 2007;Dickman 2008Dickman , 2010Koziarski et al. 2016). A better understanding of the extent and patterns of livestock depredation and its drivers is important to developing the most effective conflict mitigation and conservation management strategies (Dickman 2008;Dickman et al. 2014).
Our specific objectives were to (1) determine the reported extent of conflict and patterns of livestock depredation by large carnivores in relationship to other causes of livestock losses, (2) estimate the financial livestock losses caused by large carnivores and other factors and (3) identify factors influencing livestock depredation by large carnivores. Based on our results, we suggest appropriate measures that might be taken to reduce human-carnivore conflict and contribute to improved conservation of large carnivores in the region. We hypothesized that (1) livestock depredation by large carnivores should be higher in households located closest to the PA than further away, (2) livestock depredation is influenced by socio-ecological (distance to park boundary, household size, education levels) and economic factors (number of livestock owned) and that (3) livestock husbandry practices will affect depredation rates. In this paper, we predicted that (1) livestock depredation would decline significantly with increasing distance from the park boundary (Patterson et al. 2004;Holmern et al. 2007), (2) livestock depredation would be negatively associated with social factors (household size, education levels), positively associated with economic factors (number of livestock owned) Holmern et al. 2007) and respondent's residency time (Newmark et al. 1993;Arjunan et al. 2006) and (3) livestock depredation would increase significantly during the wet season for non-resident lions, spotted hyenas, cheetahs and wild dogs in response to seasonal migration of wild prey (Kahurananga & Silkiluwasha 1997;Mponzi et al. 2014;Koziarski et al. 2016)and vary independently with season by resident species such as leopard (Kissui 2008) and (4) improved livestock husbandry practices would be negatively associated with depredation because fortified bomas and increased number of herders should result in decreased depredation (Ogada et al. 2003;Woodroffe et al. 2007;Lichtenfeld et al. 2014).

Study area
The study was conducted in five villages (Loiborsoit, Terat, Emboret, Sukuro and Loibor Siret) of the Simanjiro Plains in Simanjiro district, northern Tanzania ( Figure 1). Simanjiro district is located between 3°52′ and 4°24′S and 36°05′ and 36°39′E and lies within the Tarangire ecosystem of northern Tanzania. The Tarangire ecosystem (20,000 km 2 ) is defined by the movements of its migratory animals and consists of Tarangire National Park (TNP) (2850 km 2 ) forming the dry season range for the migratory herds, and its wet season dispersal area and calving grounds in Monduli (including Lake Manyara National Park, Lolkisale Game Controlled Area, Manyara Ranch, Burunge and Randilen Wildlife Management Area) and Simanjiro districts (including Simanjiro plains, Mkungunero Game Reserve) (Borner 1985). The area is characterized by bimodal rainfall averaging 650 mm per annum, with short rains from November to December and the long rains from March to May. The climate is highly seasonal with the dry season (June-October) and wet season (November-May).
The Simanjiro plains are one of the most important wet season dispersal and calving ranges for wildebeests (Connochaetes taurinus) and other ungulates such as zebra (Equus burchellii), hartebeest (Alcelaphus buselaphus) and fringe eared oryx (Oryx beisacallotis). During the wet season (November-May), about 50% of the wildebeest move from the TNP to the northern plains and the other 50% to the Simanjiro plains (Morrison & Bolger 2014). The plains are also important for non-migrant herbivores such as Thomson's gazelle (Gazella thomsoni), impala (Aepyceros strepsiceros) and an important area for livestock grazing by pastoralists during the dry season (June-October) (Kahurananga & Silkiluwasha 1997

Interview methods and questionnaire design
We used semi-structured questionnaire design to obtain data on perceived human-carnivore conflict. We consulted the village leaders and generated a numbered list of all eligible bomas included in the survey. Then, we selected 60 bomas from each village at random. The questionnaire survey was conducted between June and July in 2014. In designing the questionnaire, we followed the similar format used by Maddox (2003) in northern Tanzania and by Dickman (2008) in southern Tanzania (see Appendix S1, Supporting information). Pre-testing of the questionnaire was conducted on a sample of 15 respondents and revisions were made on the questionnaire to ensure clarity of the questions before the actual data collection started. The questionnaire contained both closed-ended as well as open-ended questions in order to gain more information on participant's attitudes and reasoning. We preserved the confidentiality of the respondent during the interviews. A questionnaire was administered in person by the principal investigator with the help of a local assistant and translator to 300 respondents. Within each boma, we counted the total number of households and utilized a random number generator to select a single household.
Where possible, respondents were selected from any of the three subjects (i.e. the head of the family [usually a man], the head's wife or elder son according to seniority). Women deferred to men in seniority, so respondents were predominantly male, but interviews were conducted with women where they were comfortable to do so. During the interviews, we tested the respondents' knowledge of focal carnivores using the cards of coloured photographs.
The final questionnaire contained six main sections; however, only two are applicable in this part of the study (Appendix S1). The first section focused on information relating to respondents' sociodemographics such as respondent's gender, age, ethnicity, religious beliefs, household size, education level, occupation, residency time, livestock holding, income sources and details on livestock number and type lost to wild predators compared with other causes in the month preceding the survey. The average market values of livestock species by age category (preferably adults) were obtained from livestock traders and the prices were translated to US$ at the exchange rate of the time of the survey (1US$ = 1659 TZS, June 2014). We estimated the direct economic or financial losses of livestock to wild predators and other causes of livestock loss per household based on the prevailing market price of livestock at the time of conducting this survey. The average market value for cattle, calf, small stock and donkey was US$372, US$120, US$48 and US$90, respectively ( Table 2). The second section focused on questions about the characteristics of livestock depredation. Respondents were asked to estimate the number and type of livestock they had lost in the previous one and a half years (2013 to July 2014) to wild predators, including place of attack, time and season of attack.
The focus of this study was the reported livestock loss to the wild predators; therefore, we assume that the losses attributed to wild predators were often exaggerated, either deliberately or due to the unintentional attribution of livestock deaths to wild predators. We assumed a 1.5-year period conservative enough for respondents to recall the depredation incidents, and the financial costs are estimates based on these incidents. All respondents were adults (≥18 years of age) who could freely express themselves. The household was chosen as the sampling unit, adapting Maddox (2003) and Dickman (2008), and interviews were restricted to one respondent per household. The questionnaire was conducted in a local language (i.e. Swahili languagewith the aid of a translator speaking Maasai where needed) and took approximately 1 h to complete. The research was cleared by the Tanzanian authorities. The Tanzania Commission for Science and Technology reviewed and approved the research protocol (ref. no. 2014-370-NA-97-20). Verbal Informed Consent was obtained from all the subjects prior to participation and data were kept anonymously.

Data analysis
Continuous variables were analysed using standard descriptive statistics (means, standard deviations [SD], ranges, percentages and frequencies of counts, tables and charts). Categorical variables including gender, occupation and education level were converted into a set of dichotomous, dummy-coded variables. The intensity of livestock depredation expressed as the total number of livestock reportedly killed by all predators and by each predator species separately at bomas and in the grazing area was used as a response or dependent variable within generalized linear mixed models (GLMMs) with a Poisson error distribution and a log-link function.
We included the number of owned livestock expressed in Tropical Livestock Units (TLU), respondent gender (male vs. female), respondent age (years), education level (none vs. primary, secondary and tertiary pooled), residency time (number of years since the respondent had arrived in the area), household size expressed in adult equivalent units (AEU), number of herders, distance (km) from the park boundary (measured as the nearest distance between household and the park boundary using ArcGIS v.10.1 [ESRI, Redlands, USA]) and boma type (fortified or unfortified) as predictor variables in our models. Since households from one village were not statistically independent of each other, we included the village ID as a random effect. Therefore, we used GLMMs to determine the nature of the potential relationship between response variables and all the potential predictor variables. Further descriptive statistics of explanatory variables used in the GLMMs are presented in Table S1, Supporting information. To control for variation in household size, the household size was measured in terms of AEU. The adult-equivalent conversion factors for the number of people in the household by Latham 1965, cited by Collier et al. (1990), were used to determine the AEU as presented in Table S2, Supporting information. First, in order to calculate AEU, the sex and age of surveyed household members were compiled (Cavendish 2002). Second, the AEU by age and sex were summed up for all people in the household to compute the total AEU for the particular household.
For better comparison of herd sizes across households and to account for differences in size and value of different livestock species, we converted number of reported livestock to standard units (i.e. TLUs; see also Table 1). The following conversion factors were used for each species of livestock: one head of cattle = 0.7, one goat or sheep = 0.1 and one donkey = 0.5 (Jahnke 1982;LEAD/FAO 1999). Total TLU = Livestock Nr × TLU factor. The overall TLU per respondent was then adjusted to 1 TLU being equivalent to an animal with a body weight of 250 kg (Jahnke 1982;LEAD/FAO 1999). Because the factors influencing depredation in the grazing area were distinct from those that influence depredation at bomas, we performed analyses separately for the two distinct contexts. Therefore, we analysed explanatory variables separately for each predator species and for all predators combined and eventually running six separate model sets with all possible variable combinations within GLMMs (Tables S4-S6, Supporting information).
We checked for multicollinearity of the predictor variables using Spearman's correlation coefficients (r s ) for all possible variable pairs. We chose a cutoff of r s ≥ 0.6 to indicate high collinearity between predictor variables (Zuur et al. 2010). Using this approach resulted in the exclusion of one of the highly correlated variables from the analysis (see Table S1, Supporting information).
Age was correlated with education, respondent's residency time and number of herders. Moreover, gender was correlated with education, respondent's residency time, household size (adult equivalents) and number of owned livestock (livestock units). In addition, the number of owned livestock (livestock units) was correlated with household size (adult equivalents) and respondent's residency time (Table S3, Supporting information). However, there was no strong collinearity detected among these predictor variables (all r s < 0.6), suggesting that any collinearity among variables was unlikely to affect statistical inference (Zuur et al. 2010). In contrast, boma type small stock was highly correlated (r s = 0.76) with boma type cattle (Table S3, Supporting information) and we therefore excluded boma type small stock from subsequent analyses. Eventually, a total of nine variables were included in the models [age, gender, education level, respondent's residency time, household size (adult equivalents), livestock units, number of herders, distance to park boundary and boma type cattle] (Table S1, Supporting information). We ranked candidate models in order of parsimony based on the Akaike Information Criterion (AIC) corrected for small sample sizes (AICc) and model weights (ω i ) (Burnham & Anderson 2002). We computed model-averaged coefficients of predictor variables based on top-ranked models. We considered all models with ΔAICc < 2 to be equally plausible (Burnham & Anderson 2002).
We chose GLMMs with a Poisson error, the most appropriate distribution for count data because they take into account both fixed and random effects in a single model and deal with non-normal response variables (Zuur et al. 2010). We had six Poisson distributed  target/response variables (i.e. number of livestock killed by all predators and number killed by lion, leopard, cheetah, spotted hyena and wild dog). We used the Pearson's chi-squared analyses to test the observed frequency of predation on various types of livestock, contexts of livestock attack events by the five carnivores and the nature of the relationships among independent variables. All statistical analyses were performed using SPSS v. 22.0 (SPSS Inc., Chicago, IL, USA) and the significance level was measured at p < 0.05.

Results
Respondents' demographic, livestock holdings and socio-economic characteristics Overall, the majority of respondents were the Maasai (96%, n = 288) and the rest were the Waarusha (4%, n = 12  Figure 2). There was a slightly higher mean TLU per household in Loibor Siret and Sukuro due to a relatively high number of cattle and small stock in these villages ( Figure 2). On average, respondents from Terat had less livestock (14.16 TLU) compared to other villages.
Reported causes of livestock losses during a 1-month preceding the survey showed that disease was the main cause of livestock loss, followed by depredation, theft and other causes (snake bites, accidents and buffalo assaults). Overall, disease accounted for 90.8% of all stock losses initially reported, depredation 7.1%, theft 1.3% and all other losses 0.8%. On average, a significantly higher proportion of livestock were reportedly lost to diseases compared to other causes of livestock loss (χ 2 = 4205.70, df = 3, p < 0.001). The percentage of stock reportedly lost to depredation over a 1-month period showed that Sukuro (11.8%) had the highest rates of reported depredation, followed by Loibor Siret (6.5%) and Loiborsoit (5.8%), while the remaining villages constituted 4.4% in total. The average annual loss of the total herd to predators was 1.4% when considering the stocking rate in 2014 (ca. 93,382 total stock) and a total of 1906 depredated livestock.

Contexts of livestock depredation
The contexts of attacks on livestock varied among predators. Spotted hyenas were reported to attack livestock at bomas more often than when grazing at pasture (χ 2 = 1016.34, df = 1, p < 0.001), whereas the attacks by cheetahs (n = 55) and wild dogs (n = 150) were reported to occur during the day (in the grazing areas). In contrast, lions and leopards were reported to attack livestock held in boma enclosures during the night as well as the grazing livestock during the day. However, attacks by lions (χ 2 = 0.12, df = 1, p = 0.724) and leopards (χ 2 = 0.52, df = 1, p = 0.469) did not vary significantly between the two contexts. Overall, 75.7% of predator attacks on livestock were reported to occur during the night at bomas, while 24.3% occurred during the day at pasture.

Spatial patterns of livestock depredation
We found that the spatial patterns of livestock depredation were unevenly distributed across the studied villages. The frequency of livestock reportedly lost to different predators differed significantly between villages (χ 2 = 657.51, df = 16, p < 0.001). During the 19month study period, the highest depredation levels were reported in Sukuro (23.2%), Emboret (22.9%) and Loiborsoit (21.9%), with slightly lower levels reported in Terat (17.0%) and Loibor Siret (14.9%). The mean annual livestock loss as reported for all predators was 8.51 head of stock per household (of those that reported loss) ( Table 2). Wild dogs and leopards were reported to cause more attacks in Loibor Siret (67.3%, n = 101 and 39.2%, n = 91, respectively). Spotted hyenas were more often reported to have killed livestock than any other predator in all villages, but less frequent in Loibor Siret (3.5%, n = 47). In contrast, lions were reported to have killed more cattle in Loiborsoit (35.7%, n = 46).

Financial valuation of livestock losses
The total estimated financial loss for those people interviewed corresponding to 1906 depredated livestock was US$141,847 (equivalent to 235,324,173 Tanzanian shillings) ( Table 2). Spotted hyenas accounted for 70.3% (US$81,905) of the total herd and 57.7% financial loss, while lions accounted for 6.8% (US$38,705) of the total herd and 27.3% of financial loss ( Table 2). The financial loss due to other predators was comparatively low (Table 2). On average, the annual financial loss per household was estimated to be US$633 (of those that reported loss) and US$473 (considering all the respondents). On average, the financial loss per household (of those that reported stock losses) was estimated to be US $464 during the wet season and US$168 during the dry season. The greatest proportion of stock and financial losses were reported on small stock in proportion to their relative abundance (US$82,189, n = 1718) (Tables 1 and 2). There was a significant difference in terms of financial valuation of livestock species losses (χ 2 = 951, df = 12, p < 0.001, n = 1906) and in terms of financial impact among the predators (χ 2 = 78,020, df = 12, p < 0.001). On average, the financial loss due to various livestock diseases for the month preceding the survey ranked highest US $147,235 (US$491 per household), followed by depredation US$7968 (US$27 per household) and theft US $1695 (US$6 per household) ( Table 3). The total financial loss of livestock to large carnivores was relatively higher in Sukuro (US$339,559.30) and lowest in Loiborsoit (US$103,522.60) ( Table 3).

Ecological and socio-economic factors associated with livestock depredation
For all predators combined, model selection using AICc identified two models (ΔAICc < 2; Table S4). In the top model, the reported frequency of livestock depredation (all predators combined) was positively associated with distance to park boundary, respondent's residency time and livestock units, but negatively associated with education level, boma type cattle and number of herders in the grazing area (Table 4). Nevertheless, the reported frequency of livestock depredation by lions at bomas was best explained by the global model containing age, gender, education level, respondent's residency time, household size (AEU), distance to park boundary, boma type cattle and livestock units as significant factors (Table S5, Supporting information). The reported frequency of livestock depredation by lions declined significantly with increasing level of education, distance from the park boundary and household size (AEU) ( Table 5). Moreover, reported depredation frequency was lower among female interviewees compared to males. In addition, reported depredation frequency declined significantly with age and improved boma for cattle but declined (non-significantly) with number of herders in the grazing area (Table 5). Conversely, reported depredation frequency increased significantly with increasing livestock units and respondent's residency time at their households. On the other hand, the top model for spotted hyenas contained education level, respondent's residency time, livestock units, distance to park boundary and boma type cattle at bomas, while in the grazing area, contained livestock units and number of herders as significant factors (Table S5, Supporting information). The reported frequency of livestock depredation by spotted hyenas increased significantly with increasing distance from the park boundary, respondent's residency time,   (Table 5). At bomas, the reported frequency of livestock depredation by leopard was best explained by household size (AEU) (statistically significant) (trend: reported depredation frequency declined with increasing household size), respondent's residency time [trend: reported depredation frequency increased (non-significantly) with residency time] and distance to park boundary [trend: reported depredation frequency declined (non-significantly) with increasing distance from the park boundary] Table 4. Summary statistics of model-averaged estimates of coefficients (β) derived from the top model, standard error (SE), t-statistic and its 95% confidence interval (CI) from generalized linear mixed models (GLMMs) explaining the reported frequency of livestock depredation by all predators and by leopard in the Tarangire ecosystem, Tanzania 2014.  Species-specific models were computed separately for (a) depredation at bomas and (b) depredation in the grazing area. All models consisted of village ID as a random effect. The significance level was set at p < 0.05.
( Table 4). The best-fitting model for wild dog contained two variables; livestock units and number of herders, all showing a negative trend (Table 6,  Table S6 Supporting information). However, only livestock unit reached a statistical significance. In the case of cheetahs, the reported frequency of livestock depredation decreased (non-significantly) with increasing livestock units and increased (non-significantly) with number of herders (Table 6).

Livestock losses due to depredation in relationship to other causes
Our results showed that 75% of people reported carnivore attacks in their households, which is equivalent to an average of 1.4% of the total herd loss per annum. This figure is within the range of 0.02-2.6% worldwide losses to large carnivores reported by Graham et al. (2004) and is far more than 0.26% of the total herd reported in Ruaha National Park (Dickman 2008), and much less compared to 12% of the total herd reported in Loliondo and Ngorongoro buffer zones (Maddox 2003). Interestingly, we found that carnivore species preyed selectively upon different livestock species corresponding to the size of the predator and in accordance with the size of their prey, prey preference and abundance. Cheetah, spotted hyena, leopard and African wild dog were the predominant predators of smaller prey species (small stock), while lions preyed mostly on larger prey species (cattle and donkey) and rarely on small-sized prey (small stock). This result confirms the preference of lions for larger prey species as reported in various studies (Hayward & Kerley 2005) and the preference of cheetah, spotted hyena, leopard and African wild dog for smaller prey species (Hayward 2006;Hayward, Henschel et al. 2006;. Our results are consistent with previous findings that livestock species selection corresponds to the size of the predator (Patterson et al. 2004) and in accordance with the size of their prey (Hayward 2006). However, small stock were the most preferred prey by cheetah, spotted hyena, leopard and wild dog, probably related to their relative abundance in comparison to other livestock. In addition, spotted hyenas also preyed upon larger livestock such as cattle and donkey that are larger than their own body mass probably due to the fact that spotted hyenas do not have distinct prey species preference (Hayward 2006). On the other hand, wild dogs occasionally preyed upon calves, i.e. prey sizes which are significantly larger than their own body mass due to their group hunting strategy .
Overall, spotted hyenas and leopards accounted for more small stock attacks compared to other carnivore species. The high plasticity of spotted hyenas (Kruuk 1972;Boydston et al. 2003) and leopards (Nowell & Jackson 1996) in habitat use and diets may explain their predominance as small stock predators compared to other carnivore species. Similar studies have also reported spotted hyenas and leopards being responsible for most of the small stock depredation, e.g. around the Serengeti National Park in Tanzania (Holmern et al. 2007) and in the Maasai Steppe of northern Tanzania (Kissui 2008;Mponzi et al. 2014). However, unlike Tarangire ecosystem, wild dogs and spotted hyenas were considered the most problematic species in the eastern part of the Serengeti ecosystem Hampson et al. 2015) and lions in the Ruaha landscape (Dickman et al. 2014).
The total financial loss of livestock depredation by spotted hyenas, lions and leopards reported in this study was much higher compared with other studies (Holmern et al. 2007). In this case, the financial costs of reported losses might be perceived as significant by households experiencing such losses. In Maasai culture, livestock act as a social capital, a sign of status and wealth (Hampson et al. 2015) such that a single depredation event may be devastating for households owning very few animals, hence posing a significant economic impact on rural communities (Hazzah 2006). Perception of costs may also be higher than actual Table 6. Summary statistics of model averaged coefficient estimates (β) associated with their standard errors (SE), t-statistic and 95% confidence interval from generalized linear mixed models (GLMMs) explaining the reported frequency of livestock depredation by wild dogs and cheetah in the grazing area in the Tarangire  costs because, for example, predators are blamed for livestock loss when the cause of livestock death may be due to other factors, such as disease, and a scavenger is blamed for killing livestock when in reality, it is scavenging an animal already dead (Rasmussen 1999). We observed that the total reported livestock loss due to leopard and wild dogs was slightly higher than that by lion, but lion killed cattle contributing to higher financial loss than leopard and wild dogs. Cattle have economic and cultural values placed on them by the Maasai; therefore, loss of cattle is likely to have serious economic and social consequences (Spear & Waller 1993). Our results suggest that diseases were responsible for higher livestock losses than any other cause within and among villages. Our findings concur with other studies conducted in Tanzania, in which disease was found to be the leading cause of livestock loss (Holmern et al. 2007;Dickman 2008;Kissui 2008;Nyahongo & Røskaft 2012). Diseases are particularly known to be responsible for high loss in livestock production (3-6 times higher other than livestock depredation) in sub-Saharan Africa (Gifford-Gonzalez 2000;Frank et al. 2005). Generally, the impact of theft was very low compared to the impacts of livestock depredation and disease contrary to Nyahongo and Røskaft (2012) and Ogada et al. (2003).

Spatio-temporal patterns of livestock depredation
We observed variation in the timing and contexts of depredation by different carnivore species similar to other reported findings. Cheetahs and wild dogs are diurnal and typically attack grazing herds by day (Ogada et al. 2003). Leopards, spotted hyenas and lions attack livestock at any time of the day, either in the grazing area or at bomas (Patterson et al. 2004), although other studies found that spotted hyena and leopard attacks prevail at night at bomas (Woodroffe et al. 2007;Kissui 2008). Surprisingly, we found that the mean annual livestock loss due to depredation was relatively lower for households in Loibor Siret (1.29 per household) and Terat (1.45 per household) than expected compared to other villages. We expected that the reported frequency of livestock depredation would be relatively higher for households in Loibor Siret due to the proximity of this village to the TNP. A possible explanation for the lower depredation rates in Loibor Siret could be due to the presence of fortified bomas which reduced incidences of carnivore attacks on livestock. Between 37% and 40% of respondents reported using fortified bomas in Loibor Siret to keep cattle and small stock, respectively, which might have influenced our results in various ways (Mkonyi et al. submitted). In addition, boma type was a good predictor of livestock depredation levels (e.g. boma type for cattle associated negatively with livestock depredation by all predators, lions and spotted hyenas in our boma depredation model). The impact of fortified bomas on large carnivores has also been tested in Loibor Siret where overall depredation rates by lions, leopards and spotted hyenas declined by 90% (Lichtenfeld et al. 2014). Similarly, there were fewer reported incidences of depredation in Terat households. The reason for the lower depredation in this village is uncertain; however, it could possibly reflect the low density of carnivore species around this village. Nevertheless, the highest depredation rates reported in Sukuro, Emboret and Loiborsoit may be related to many factors, including low density of wild prey species, continued human encroachment onto carnivore habitat and poorly constructed night-time enclosures (bomas). Studies show that livestock depredation is more common in areas with low prey abundance (Bagchi & Mishra 2006), high human population, increased encroachment and poor livestock husbandry practices (Treves & Karanth 2003).

Seasonal patterns of livestock depredation
Our study revealed that there was a seasonal variation in livestock depredation by lions, cheetahs, leopards, spotted hyenas and wild dogs, with clear peaks of depredation during the wet season. The depredation peak during the wet season has also been reported by Mponzi et al. (2014) and Koziarski et al. (2016) for the western part of this ecosystem. But in other areas of Africa, depredation mainly occurs during the dry season (e.g. Hemson et al. 2009). Our findings somewhat contradict the 'reduced natural prey hypothesis' which emphasizes that depletion in natural prey abundance promotes attacks on livestock . Our results could reflect the seasonal shifts in wild prey distributions from TNP into the communal village lands (Kahurananga & Silkiluwasha 1977), with more predators (lions, cheetahs, wild dogs and spotted hyenas) following natural prey and hence coming into conflict with humans and livestock. This seasonal variability in depredation patterns has also been reported for hyenas and lions in the Maasai Steppe of northern Tanzania (Kissui 2008;Mponzi et al. 2014) and for lions in Tsavo National Park in Kenya (Patterson et al. 2004). Patterson et al. (2004) found that livestock depredation by lions peaks during the wet season when natural prey are in better condition and more widely dispersed, hence difficult for predators to acquire. Interestingly, we found no support for the prediction that livestock depredation by leopards is independent of season. Hence, it is possible that even leopards could be moving in the same manner with ungulate migrations following the increase in livestock predation by leopards in the wet season.
However, additional research with verified data would be appropriate in explaining this seasonal variation.

Ecological and socio-economic factors associated with livestock depredation
Our hypothesis that socio-ecological (distance to park boundary, education level) and economic factors (number of livestock owned) would influence the reported frequency of livestock depredation by all predator species was supported. Surprisingly, distance to park boundary was positively associated with reported frequency of livestock depredation, which was contrary to our prediction. This is a clear indication that the reported frequency of livestock depredation by all predator species was relatively lower in households located closer to PA than further away. Clearly, there is a high variation in this variable because the households ranged from 7 to 52 km from the park boundary. Our findings contradict with findings reported elsewhere that livestock depredation declines significantly with increasing distance from the park boundary (Patterson et al. 2004;Holmern et al. 2007). This unexpected pattern may be explained by improved fortified bomas in households closest to the park (i.e. around 42% of the traditional bomas were fortified), while distant households had few or no fortified bomas. However, this trend varied significantly by species. For instance, reported lion attacks declined with increasing distance from the park boundary, although the reverse was true for spotted hyenas. This discrepancy can be partly explained by the fact that spotted hyenas killed the largest number of small stock in households located further away from the park where fortified bomas were mostly absent. Moreover, our findings suggest that lions were more likely to attack livestock in households that were closer to the PAs due to the fact that lions usually stay close to their natural habitat (Holmern et al. 2007). However, spotted hyenas often move far from PAs and are able to survive well in humandominated landscapes due to their opportunistic feeding patterns and adaptive ranging behaviour (Hofer & East 1993;Kolowski & Holekamp 2011). Our findings showed that respondents with formal education experienced lower depredation rates than those without any formal education, consistent with our prediction and previous studies Holmern et al. 2007).
Consistent with other studies (Zimmermann et al. 2005;Hemson et al. 2009), our results indicate that people who owned large numbers of livestock experienced more livestock losses to large carnivores. This finding contradicts Koziarski et al. (2016) who found that education, psychological and demographic attributes were more influential in wildlife conflict perceptions than economic considerations (livestock ownership). Moreover, consistent with our prediction, the incidences of attacks on livestock by predators declined with increasing number of herders and fortified bomas as it has been demonstrated in other studies (Ogada et al. 2003;Woodroffe et al. 2007;Lichtenfeld et al. 2014). Our prediction that increased exposure to wildlife-related risks (i.e. long-term residency) would be positively associated with livestock depredation was supported, and this may be because long-term residency has been found to be associated with negative attitudes towards large carnivores (see Newmark et al. 1993;Arjunan et al. 2006). Furthermore, we found that men reported more depredation frequency with lions than women in the study area. This may be because men claim ownership of livestock and they come more frequently into contact with lions during livestock herding (Hampson et al. 2015;Koziarski et al. 2016). We also found that reported depredation frequency with lions at bomas was negatively associated with interviewee age, suggesting that as interviewees get older, they perceive lower levels of depredation. It could be that older interviewees are more likely to have reinforced bomas and this corroborates our previous findings (Mkonyi et al. 2017), showing that the longer people are in a place, the more tolerant they are likely to become. In addition, we assume that retaliatory killing and also culturally motivated killing of lions by humans in response to damages caused by lions is negatively affecting lion populations in this ecosystem (Kissui 2008;Lichtenfeld L. Pers. Comm. 2014). Based on this evidence, we can assume that it is also likely for other species.

Conclusion and management implications
Human-carnivore conflict is a complex issue for management, especially where humans live adjacent to or within PAs (Dickman 2010). Our study suggests that conflict due to livestock depredation could be significantly reduced by improving formal and conservation awareness education at all levels (i.e. during primary, secondary and tertiary school education), fortifying boma enclosures, improving herding practices such as increasing the number of herders (particularly adults) per herd. We also suggest the need for finding out high and lowrisk areas where livestock is more or less susceptible while grazing and eventually educating herders to avoid grazing their livestock in high-risk areas (predator hotspots) or always be vigilant while grazing in such areas. Conservation efforts for mitigating conflicts should concentrate more on households that are situated further away from national park by improving boma enclosures for livestock. In addition, local people should receive tangible benefits (through benefit-sharing programmes) from large carnivore presence on village land that could offset costs of livestock losses and increase local people's tolerance for these predators. While increased carnivore attacks on livestock in the study area can engender significant socioeconomic costs to local households, conservation efforts would benefit from combined carnivore conservation initiatives and livestock depredation reduction.
Conservation education and awareness programmes focusing on large carnivore behavioural ecology may also reduce the human-carnivore conflict and increase local people's tolerance for large carnivores (Sillero-Zubiri et al. 2007). In order to control and prevent livestock depredation by large carnivores, there is a need to understand predatorspecific protection measures which can then be integrated into conflict mitigation programmes. As disease was perceived to be a greater cause of livestock losses than depredation in the surveyed villages, interventions would be to control and manage livestock diseases through preventive vaccinations and increase access to veterinary services .
Our study has provided new insights into the complexities of human-carnivore conflicts among the five large African carnivores and determinants of reported conflict with these species in the Tarangire ecosystem of northern Tanzania. We recommend further research along these lines to evaluate the actual frequency of conflict in the study area (using field verification methods) and continued monitoring of conflict situations over time or other causes of mortality. This might help in understanding the 'conflict hotspots' or sites predisposed to livestock depredation across the village land, allowing herders and wildlife managers to concentrate livestock protection and conservation education programs in such areas.