Perceptions and realities of elephant crop raiding and mitigation methods

Crop raiding by African elephants (Loxodonta africana) jeopardizes human livelihoods and undermines conservation efforts. Addressing this issue is particularly important in subsistence farms adjacent to protected areas and requires assessing the perceived and actual scale of the problem and the benefits, limitations and adoption potential of mitigation techniques. To achieve these objectives, we assessed the effectiveness of chili and beehive fences relative to control plots, using a daily farm monitoring protocol implemented on 20 farms bordering the Ngorongoro Conservation Area (Tanzania). Prior to the field study, we interviewed 65 farmers about human–elephant interactions and contrasted interview findings with those of daily farm monitoring. Farmer perception of crop raiding frequency declined with increasing distance from the protected area and was, on average, eight times greater than daily farm monitoring data indicated. The majority of interviewees expressed a willingness to try chili or beehive fences, though chili fences were preferred. Generalized‐linear‐mixed models indicated that neither elephant farm intrusions nor damages were significantly reduced by either chili or beehive fences relative to the control sites. Losses per month and hectare did not differ significantly by fence type. However, farm plots with chili fences did not experience massive damages which occasionally occurred in beehive or control plots. This partial effectiveness of chili fences was further confirmed by contrasting crop losses from a subset of farms that were subject to a cross‐over experimental design. Our multidimensional case study suggests that chili fences have greater adoption potential than beehive fences. Nevertheless, additional efforts are required to increase effectiveness and to realize adoption potential. By providing insights into the context and circumstances that presented challenges to the effectiveness, sustainability and scalability of beehive and chili fences, this study can serve as reference for areas where crop raiding by elephants is a key conservation conflict.

. Combined with a recall study, such assessments can be used to assess the objectivity of selfreporting (Gillingham & Lee, 2003). Moreover, possible disjunctions between perceived and real intensity of crop raiding could hint to deeper-rooted conservation conflicts (Gillingham & Lee, 2003), which may pose challenges beyond implementing technical solutions to reduce crop damages by elephants (Zimmermann, McQuinn, & Macdonald, 2020).
If the aim is a wider adoption of a mitigation method, such assessments not only evaluate the technical efficacy, but also assess the acceptance, sustainability and scalability of a given method (Denninger Snyder & Rentsch, 2020). Providing such multidimensional evidence to inform conservation management decisions has been underpinned by research that suggests conservation decisions are often made based on anecdotes, rather than evidence-based practices, and recommends the use of cross-disciplinary evidence in order to best predict the outcome of a particular intervention (Game et al., 2018;Salafsky et al., 2019).
To address these interrelated issues, we carried out an interdisciplinary study with the aims of assessing and comparing (a) real and perceived patterns of crop raiding by elephants, (b) the willingness of people to adopt beehive and chili fences, and (c) the effectiveness of chili and beehive fences to reduce crop raiding by elephants and. Finally, we discuss our findings with a focus on the wider adoption of the tested mitigation methods.

| Study area
We conducted this study in Tloma village, in the Karatu district of northern Tanzania. The village directly borders the forest section of the Ngorongoro Conservation Area (NCA); the forest is occupied by several wildlife species, including a resident elephant population (Homewood & Rodgers, 1991). Near the study site (c. 3 km; within the NCA) is a large salt lick, which attracts wildlife to the area due to the high concentrations of micronutrients (Mills & Milewski, 2006). Long distance elephant movement does not regularly occur through the agricultural matrix of the Karatu district, and is largely restricted to the Upper Kitete corridor where the NCA forest connects to the Selela forest in the Tarangire ecosystem (Caro, Jones, & Davenport, 2009).
The landscape of the area is undulating and at an elevation of approximately 1700 m asl. The area experiences a bimodal rainfall pattern with precipitation (1,000 mm/ year) typically occurring from late October to December (short rains) and from March to May (long rains). January and February are typically relatively dry and the main dry season is from June to mid-October (Prins & Loth, 1988).
Tloma village hosts several tourist lodges, but commercial coffee farms (protected with electric fences), settlements, and subsistence farming are the primary land-use types. In the Karatu district, the vast majority of people earn their livelihood through subsistence agriculture (Owenya, Mariki, Kienzle, Friedrich, & Kassam, 2011). Common crops in the study area are pigeon peas, beans, and maize. Almost all farmers in this area practice the same sequential cropping scheme in their plots. Pigeon peas are planted in January/February and primarily harvested in September/October. Maize is usually planted in February/ March (in between the peas) and harvested in July and August. Beans are planted in November and harvested in January/February (pers. observation).

| Experimental plots and field sampling
To test two crop prevention techniques (chili fences and beehive fences), we selected 20 small-scale farms (average 0.7 ha, range: 0.1-2.3 ha) bordering the NCA forest. The boundary to the NCA is demarcated by a c. 4 m wide fire break. In addition, a thorny hedgerow (ca. 3 m tall and 2-3 m wide) separates the farmland from the fire break and thus serves as an additional physical barrier between the NCA forest and the farmland.
In close cooperation with the 20 farmers (farmers expressed their willingness to support the project and we agreed upon equal benefit sharing of bee products), we distributed two beehive fences and 10 chili fences to the plots; eight sites were left untreated as a control ( Figure 1). Ideally, experiments to assess conservation evidence distribute an equal number of treatments and controls (Khorozyan, 2020), yet funding constraints limited us to two beehive treatments only. As we worked in collaboration with the farmers and some farmers were not willing to host a beehive fence, a random allocation of treatments was not possible. At one plot, the allocated beehive fence remained active for the entire duration of the experiment. However, we had to shift one beehive fence to a different site following complaints by neighbors. To counteract the non-random placement of treatments, and to distribute potential crop-protection benefits of the project in a just way, we swapped the treatment of 16 sites, interchanging chili fences with control sites. This partial cross-over was implemented approximately half way through the experiment so that these paired sites were subject to 11-13 months of chili fencing and 11-12 months without additional fencing.
Construction and maintenance of the fences was conducted by a trained person, in cooperation with the corresponding farmer of each plot. At the beginning of the project, a field crew of PAMS foundation also provided a workshop to demonstrate and practice chili fence construction. In line with Chang'a et al. (2016), we constructed chili fences by soaking sisal strings and rags in a solution of used engine oil and ground chili peppers (Capsicum frutescens). Wooden poles were placed about 5-10 m apart around the perimeter of a field. We tied the soaked sisal rope between the poles and tied the soaked rags along the string (2-3 rags in between poles). The chili fences required periodic refreshing to maintain their offensive odor; we renewed the fences once or twice a month (depending on precipitation) and daily monitored and repaired potential damages (Chang'a et al., 2016). We constructed beehive fences by creating a fence around a field plot with nine foot posts connected by a strong wire. The posts were coated in insecticide to prevent termite infestation. We strung beehives along the wire with small holes in the sides of the hive and 10 m between hives. Beehives consisted of an 80 cm long Kenyan top-bar hive constructed out of industrial plywood, a rainproof roof made from corrugated iron sheet, and a flat-thatched roof to protect the hive from the sun. Hives were locally sourced and cost c. US $50 unit −1 . We aimed to facilitate bee occupation by twig (Ocimum kilimandscharicum) and beeswax baiting. If elephants tried to enter the farm, the wire would be stretched or broken, disturbing the hives and causing bees to become agitated and prone to attacking intruders (King et al., 2011). We maintained and cleaned beehives periodically to prevent degradation and repaired damages to the wire and hives within 24 hours from when the damage was observed.
A trained enumerator monitored fields daily from March 2016 through October 2016, December 2017 through April 2017, and December 2018 through October 2019. Daily assessments included recording the crop type, identifying which species invaded the farm (based on animal signs), quantifying the area (m 2 ) or crop quantity (number of cobs or plants) damaged by wildlife, and assessing the direction from which elephants invaded the farm, the time of day of the intrusion (day or night, based on age of elephant signs and conversation with local farmers), and the number of invading individuals (based on inspection of elephant tracks and estimating how many individuals caused the tracks). Inherent to the monitoring method, time of intrusion and herd sizes had to be based on indirect evidence and is thus subject to some level of uncertainty. In sum, our monitoring stretched over four calendar years and covers a total of 14,528 plot visits.

| Interviews
In November 2015, and thus prior to the implementation of the experiment, we conducted 65 structured interviews F I G U R E 1 Map of the study area, indicating the location of the experimental farm plots (treatment distribution according to the last year of the experiment), locations of interviews, and the Ngorongoro Conservation area. The inset in the lower left corner indicates the approximate location of the study area within northern Tanzania with residents of Tloma village. We walked along transects and approached households at c. 200 m intervals. With the aid of translators conversant in Swahili and the local Iraqw language, we asked people for their consent to participate in an interview. We asked predetermined questions about the frequency at which elephant crop-raiding occurred in a year, the typical group size of crop raiding elephants, the methods currently in use to mitigate cropraiding, and willingness to try chili fences or beehive fences in the future (the principles of both beehive and chili fences were briefly explained before asking about willingness to adopt either of the methods; Appendix S1).
Because interviews were distributed across T'loma village, we first tested if reported elephant farm invasion frequency was related to distance from the protected area using a logistic regression model. As dependent variable, we created a two column matrix, where the first column includes the number of reported farm invasions in a year (i.e., successes) and the second column represents the number of days without farm invasions by elephants (failures). As explanatory variable, we used the Euclidian distance between interview location and border to the NCA. To assess patterns of perceived and daily enumerated patterns of crop raiding by elephants, we compared perceived (from the interview data; calculated as reported number of farm invasions per year / 365 days) and enumerated (from the farm monitoring; calculated as F I G U R E 2 (a) Perceived daily farm intrusion likelihood of elephants (assessed via 65 interviews in T'loma village, Tanzania) in relation to Euclidian distance between interviewed household and border of the Ngorongoro Conservation Area. (b) Compares the reported (based on interview data from 2015) and observed (based on daily farm monitoring from 2016-2019) frequency of elephant intrusions on farms. C contrasts reported (based on interview data from 2015) and estimated (based on counting number of different elephant tracks) elephant herd sizes observed number of farm invasions / number of monitoring days) invasions by elephants per plot and day, as well as perceived and estimated elephant herd sizes. Given nonnormal distribution of the data, we tested for differences between perceived and enumerated farm intrusion frequencies and elephant herd sizes using a Mann-Whitney U test.
To assess the willingness of interviewees to adopt either beehive or chili fences, we visually compared the proportion of interviewees willing to adopt the mitigation method.
To assess and compare the effectiveness of chili and beehive fences in reducing the daily likelihood of elephant intrusions and of causing crop damage, we fitted two separate mixed effects logistic regression models (response variable for Model 1: elephant intrusion observed on this day yes [1] or no [0]; response variable for Model 2: crop damage caused by elephants observed on this day yes [1] or no [0]) using the lme4 package (Bates, Maechler, Bolker, & Walker, 2015). We considered all farm intrusions as independent, although the same elephant group may have entered multiple farm plots in one night. However, as treatments were designed as perimeter fences, possible non-independence of the response variable is unlikely to have a major impact. Importantly, if elephants invaded farms during one night, they mostly targeted a single farm: 62% of invasions were recorded on unique plot-day combinations. Of note, 21% of invasions occurred on two plot-unique day combinations (average of five plots in between invasions) and only few nights (17% of nights which experienced at least one intrusion) were subject to more than three elephant intrusions.
As explanatory variables, we considered the fence treatment (three level factor: chili, beehive or no fence) and "month" (11 level factor, no monitoring during November) as fixed effects. Month was entered as fixed effect because crop raiding by elephants is associated with crop maturity (Denninger Snyder et al., 2021), which in our system is primarily a function of the season. To account for the repeated measure design (Zuur, Ieno, Walker, Saveliev, & Smith, 2009), we included "plot id" as random effect. To account for the possibility of elephant habituation to fences (Hoare, 2012;Shaffer et al., 2019), we included the year as random effect in our models. In addition, we included the farm area (in ha) as explanatory variable to control for potential area effects on invasion and damage likelihood.
As a third measure to test the efficacy of treatments, we estimated the monetary value of crop damage by collecting data on damaged area (m 2 ) and number of damaged plants and cobs for maize and pigeon peas (we did not record any damage in beans). To standardize the data, the damaged area was first converted into number of plants. Based on local farming practice, we estimated that six plants made up one m 2 of maize and four plants covered one m 2 of pigeon peas. Vegetables were excluded due to lack of specific market value information and damage occurrence only made up 1% of all crop damages. Assuming that maize plants yield on average two cobs (each weighing 0.22 kg), one maize plant was estimated to produce 0.44 kg of maize. In 2019, the market price for maize was 600 TZS kg −1 . Therefore, one maize plant was valued at 264 TZS. For pigeon peas, we assumed a harvest of 2 kg of peas per plant and a 2019 market price of 800 TZS kg −1 , and thus a value of 1,600 TZS plant −1 . Because farm sizes differed, we standardized the losses to 1 ha. For each plot, we divided the total losses by the corresponding number of months in which the specific treatment was associated with the specific plot to derive an estimate for monthly losses per ha for each farm. To assess whether these economic losses differed by treatment, we carried out a Kruskal-Wallis ANOVA.
Because eight of the experimental plots were subject to a cross-over design, we tested for differences in plots that served both as control sites without additional treatment and with chili fences using a paired Wilcoxon test.

| Integrated assessment of mitigation methods
To systematically and holistically evaluate the effectiveness and the adoption potential of beehive and chili fences F I G U R E 3 Proportion of interviewees (n = 65) in Tloma village expressing their willingness to try beehive or chili fences (this had been assessed in 2015, prior to field tests of the fences) to prevent elephant crop raiding we adopted a recent framework (Denninger Snyder & Rentsch, 2020) and employed an integrated assessment scheme. Based on quantitative evidence from the interviews and the farm monitoring as well as qualitative aspects, we rated attitudes, effectiveness, sustainability, and scalability of beehive and chili fences using a 3-point scale (3 points: very effective/conducive for adoption; 2 points: effective/conducive for adoption; 1 point: less effective/ conducive for adoption). We present the integrated assessment in the discussion.   Figure 2c).
In addition, our farm monitoring provided insights into crop raiding behavior of elephants. Farm intrusions by elephants apparently exclusively occurred during nighttime. Elephants primarily (71%; 227/318 occasions in which intrusion direction was assessed) entered from the NCA forest, occasionally entered from adjacent lateral farms (37%; 85/318) and rarely entered from the southern side that borders the village (2%; 6/318).

| Attitudes toward mitigation methods
Based on the interviews prior to the experiment, the majority of interviewees were willing to try chili fences (87.5%), and to a slightly lower degree (64%), beehive fences ( Figure 3). Interviewees who did not want to adopt beehive fences frequently justified this by expressing fear of bees.

| Effectiveness of mitigation methods
The odds for elephants to intrude farms with beehive fences were approx. 1.9 times greater compared to farms without additional fences; however this association did not reach statistical significance (Table 1A). Compared to control farms, the odds for elephants to enter farms with chili fences was lower, but this relationship was also not statistically significant. Farm size did affect the likelihood of elephant intrusion. Elephant intrusions were most likely to occur during the months of April, September, and October and least likely in December (all months compared to January; Table 1A). The elephant damage model revealed a similar pattern (Table 1B). The odds for elephant damage occurring in a farm were highest in farms with beehive fences, and this association was statistically significant (odds ratio 95% CI [1.0, 6.5]). Farms with chili fences experienced lower odds of elephant damage compared to farms without additional fences, but this effect was not strong (indicated by p-value = .199 and confidence intervals of odds ratios overlapping with 1). Similar to the elephant intrusion model, crop damages predominantly occurred during the months of April and October. Regression coefficients for the months May-July also suggested greater damage likelihoods compared to the reference month (January), but the estimates were associated with relatively wide margins of error (Table 1B). Farm sizes were neither associated with intrusion nor damage likelihood.
In both models, unexplained variation from farm to farm was evident (SD of the random intercepts: 0.66 The third indicator to test effectiveness of chili and beehive fences, monetary value of raided crops, yielded similar results as frequency based measures (Figure 5a). Fences with beehives suffered the greatest amount of monthly crop losses (xc = 44,835 TZS month −1 ha −1 ; 95% CI [0, 126,631]), followed by farms with no additional treatment (xc = 26,235 TZS month −1 ha −1 ; 95% CI [0, 52,506]) and farms with chili fences (xc = 16,245 TZS month −1 ha −1 ; 95% CI [455, 32,035]). Although differences in mean losses per treatment appear substantial, monetary losses did not differ significantly by fence type (Kruskal-Wallis Χ 2 = 0.797; df = 2, p = .67). Comparing economic losses of the eight plots that were subject to a cross-over design, highlights that economic losses were typically smaller when the plots where fenced with chili rags although one farm plot showed the opposite pattern (Figure 5b). However, the reduction in economic losses was not statistically significant (Paired Wilcoxon test V = 7; p = .27).

| DISCUSSION
This interdisciplinary case study suggests that stakeholders may have an exaggerated perception of the frequency (and possibly also herd sizes) of elephant crop raids. Importantly, neither beehive nor chili fences significantly reduced the occurrence and economic magnitude of crop damages by elephants. Because these two methods are widely implemented to prevent elephant crop raiding in East Africa and beyond, reporting negative results is important. Below, we discuss the context and circumstances that possibly impaired the effectiveness and wider adoption of beehive and chili fences in this case study.

| The perception-reality gap associated with HECs
The difference between perceived and independently enumerated frequency or intensity of wildlife-caused damages is a widespread reported phenomenon observed in studies on human-wildlife interactions such as livestock depredation (Kissui, Kiffner, König, & Montgomery, 2019) and crop raiding by great apes (Campbell-Smith, Sembiring, & Linkie, 2012) or multiple species (Gillingham & Lee, 2003;Hill, 2004). In this study, the magnitude of difference was c. 7.6 fold. This may even be an underestimation, as we considered perceived invasion frequency from the entire village, including data from areas further away from the protected area where elephant intrusions are less frequent (Figure 2a). We acknowledge a slight temporal mismatch between the interview data (interview data: recall for 2014-2015; farm monitoring: 2016-2019). However, given the magnitude of the difference and the relatively small variation in the yearly effects (Table 1), we assume that such disjunctions are unlikely to be explained by year-to-year variation in elephant movement.
Interestingly, interviewees apparently also perceived that elephant groups were larger than observed based on tracks. Certainly, accurate herd size estimation is difficult if based solely on tracks, yet the emerging pattern ( Figure 2c) suggests that this difference may indeed be disjointed (as differentiating tracks from single vs. large elephant herds can be done with some confidence).
F I G U R E 5 (a) Estimated value of crops lost per month (expressed in Tanzanian shillings; exchange rate: 1 US $ 2,300 TZS as of December 2019) caused by elephants in subsistence farms with beehive, chili, or no additional fences averaged over the treatment-specific study period and standardized by plot size. We removed one data point for the beehive treatment as this value was identified as outlier (1,256,178 TZS month −1 ha −1 ). (b) Pairwise comparison of monthly crop losses per ha in plots subject to both chili and no additional fence treatment (cross-over design; identical plots are connected by grey lines) In this case study, most crop raiding likely occurred by single elephants which aligns with other research that has found elephant bulls to be responsible for the majority of crop raiding events (Hoare, 2001;Jackson et al., 2008).
Multiple hypotheses have been suggested to explain this apparent disjunction between perception and reality, including extreme damage events (which occasionally occurred in our study as indicated by the wide margin of crop damages Figure 5a), opportunity costs associated with guarding farm plots, a feeling of powerlessness to deal with wildlife (Gillingham & Lee, 2003), or underlying deep-rooted conflicts such as broader disagreements with protected area management (Zimmermann et al., 2020). Additional studies could be insightful to substantiate these hypotheses and may help elucidating whether local attitudes toward elephants improve if elephant damage is effectively reduced (Davies et al., 2011).

| Integrated assessment of mitigation methods
The work presented by Denninger Snyder and Rentsch (2020) provides a suitable framing for a multidimensional assessment of the two mitigation methods ( Table 2).
The general attitudes toward the tested mitigation methods suggested openness among the interviews toward beehive and chili fences, whereas more interviewees were willing to implement chili fences. While the majority of interviewed subjects were open to try beehive fences (Figure 3), those who did not expressed their concerns that bees may sting them while working on farms. Indeed, we had to shift one of the beehive fences to another plot because neighbors did not tolerate the bees. In addition, we received complaints that bees caused problems when livestock grazed on the farms (livestock grazing is common once crops have been harvested), when farmers worked on their farms, or when children played in the vicinity of the fences. In terms of attitudes, we thus evaluated chili fences with 3 and beehive fences with 2 points (Table 2).
Surprisingly, none of the three indicators employed (likelihood of farm intrusion, crop damage by elephant, or the extent of monetary losses) provided strong evidence for the effectiveness of either beehive or chili fences in reducing crop damages caused by elephants. In part, nonsignificance of treatment effects may be the result of limited spatial replication and unequal sample sizes of treatments (especially considering few beehive treatments), possibly resulting in low test power (Khorozyan, 2020). Nevertheless, nonsignificant reduction of crop damage (and occasional greater elephant intrusion and damage likelihood in treatments relative to control sites) is surprising because previous studies reported strong evidence for the effectiveness of beehive (King et al., 2011(King et al., , 2017Scheijen et al., 2019), and T A B L E 2 Qualitative assessment of beehive and chili fences as a means to mitigate crop raiding by elephants in the Tloma case study, Tanzania 2 (can easily be scaled up)

Sum 17 24
Note: Criteria of the assessment (attitudes, effectiveness, sustainability, and scalability) have been adopted from a recent framework developed by Denninger Snyder and Rentsch (2020). Points (3 points: very effective/ conducive for adoption; 2 points: intermediate effective/conducive for adoption; 1 point: less effective/conducive for adoption) were assigned by the authors based on interview data (attitudes), daily farm monitoring on experimental plots (effectiveness), and experiences during the case study (sustainability and scalability).
chili (Chang'a et al., 2016;Davies et al., 2011) fences. However, such a finding is consistent within a broader context of assessing evidence for the effectiveness of methods to prevent damage by wildlife, as reviews on this topic typically highlight mixed and often limited effectiveness of different prevention methods (Hoare, 2012(Hoare, , 2015van Eeden et al., 2018). In this case study, chili fences appear as partially effective in reducing crop damage ( Figure 5), but the effect sizes were statistically not significant. While chili fences did not prevent the occurrence of crop damages, the economic crop losses in this treatment were in a narrower range (US $ 0.2-13.9 month −1 ha −1 ) as compared to the control (US $ 0-22.8 month −1 ha −1 ) or the beehive fences (US $ 0-55.1 month −1 ha −1 ). This suggests that chili fences can at least prevent large economic losses in farms. Although elephants occasionally broke the chili fences, elephants possibly spend little time (and thus cause relatively little damage) in chili-fenced farms due to the olfactory deterrence function of chili (Osborn, 2002). The limited effectiveness of beehive fences can partially be explained by the fact that elephants in our study, and indeed other study areas (Jackson et al., 2008;King et al., 2011), primarily raid crops at night, when bees reduce their activity (Kaiser, 1988;Kronenberg & Heller, 1982). Thus, beehive fences may be limited in their ability to reduce nocturnal crop raiding by elephants (Hoare, 2012). While deterring elephants with bees may clearly work (King et al., 2010;Ngama, Korte, Bindelle, Vermeulen, & Poulsen, 2016;Vollrath & Douglas-Hamilton, 2002), the nature of the deterrent may cause unforeseen issues. For example, on one occasion elephants approached a beehive fence upon returning to the forest. The elephants triggered the wire, followed by a subsequent attack by bees. As the fence blocked their way to the forest, the elephants fled from the bees by running back to the village land, causing additional trampling damage to adjacent properties in the process. In summary, we thus assigned 2 points for chili and 1 point for beehive fences for the effectiveness criteria (Table 2).
Likely, the limited effectiveness of both methods in this case study is associated with aspects of sustainability of their implementation. For example, because the deterrence effect of chili is likely closely related to time since application, a more frequent re-application of the chili-oil solution (and including the time since re-application as covariate in assessment models) may provide greater effectiveness, yet requires more manual labor and greater financial investment. Similarly, we hypothesize that the overall poor performance of beehive fences in these trials was related to low bee occupancy. From 2018 to 2019, bee occupancy was relatively low in the two fences (28.6% [6/21]; 34.8% [8/23]) and this may explain their limited effectiveness (Scheijen et al., 2019). Indeed, low bee occupancy and substantial time lags between beehive fence construction and bee occupancy have been observed in other studies as well (King et al., 2011).
Obviously, the investment and maintenance costs are an important consideration even if costs for these trials were mostly covered by external funds. Renewing chili fences costs approx. US $14 per hectare and application (Chang'a et al., 2016) and thus c. US $10 for an average sized farm (0.7 ha) in our area. With c. US $50 per hive and a required spacing of 10 m, a perimeter fence for one hectare would cost c. US $2000 for hives only (and thus excluding additional costs for wire and poles). Even for a 0.7 ha farm (US $1400), these costs are likely prohibitive for subsistence farmers in our study area. Theoretically, beehive fences can produce bee products (honey and wax) and thus may generate money to recuperate the initial investment and generate income (King et al., 2017). Unfortunately, in our case study third parties repeatedly stole honey and wax, and hence effectively reduced the potential of income generation. In addition, beehives were repeatedly stolen by third parties and placed inside the NCA for honey production. Although the stolen beehives were recovered with the help of village and NCA management, such human inference can limit both the effectiveness and economic viability of this method. Moreover, these examples highlight the clandestine human-human conflicts that can hinder the effective solution of human-wildlife conflicts. In sum, and based on the experience of this case study (which may not be representative for other settings) chili fences thus scored more points that beehive fences in terms of sustainability (Table 2).
Similarly, chili fences scored higher in the scalability criteria than beehive fences (Table 2). Actively trying to increase the occupancy of beehives in fences (Ngama et al., 2016), or implementing large-scale beehive fences may not only be constrained by funding and technical expertise but also by the carrying capacity of bees in the landscape. In our case study, this appears limited by the availability of flowering plants and widespread distribution of beehives (operated by subsistence beekeepers and small business) across the landscape. Despite initial demonstrations to train farmers in setting up chili fences (Gunaryadi et al., 2017), no additional farmers set up chili fences; we assume that the associated costs of setting up chili fences may have prevented widespread adoption. If the observed slight damage reductions ( Figure 5) are deemed sufficient enough to upscale the use of chili fences (Chang'a et al., 2016), cost-benefit considerations will be an important aspect. In terms of fairness and benefit distribution, it may be a worthwhile consideration that stakeholders who benefit financially from the presence of elephants (e.g., lodges, administration of the protected area) could bear these costs (Sommerville, Jones, Rahajaharison, & Milner-Gulland, 2010). Possibly, such concentrated and joint efforts could also alleviate parts of the apparent deep-rooted conflict that may underlie the negative perceptions of elephants in this area. If such a funding and implementation scheme could be established, fencing off the border line between the village and the NCA (instead of perimeter fences around single farms) could be a cost-effective solution.
In summary, our multi-dimensional assessment suggests that chili fences appear more effective and score more points in terms of attitudes, sustainability and scalability and thus appear to have greater adoption potential (Table 2). Nevertheless, our analyses highlight substantial potential for improving the effectiveness, sustainability and scalability of both methods and the need to include socio-economic factors when evaluating methods to mitigate human-wildlife conflicts (Denninger Snyder & Rentsch, 2020).

| Ways toward human-elephant coexistence
Results of this case study mirror previous conclusions that mitigating crop raiding by elephants is challenging (Hoare, 2012(Hoare, , 2015 and that no single prevention method is a panacea to solve this issue. One potential for more effective crop reduction could be to focus interventions on months when crops are mature (Denninger Snyder et al., 2021), and thus elephant damage is most likely (in this study area from April to July and September to October; Table 1). In addition, this study indicated that chili fences are only partially effective, and that over time elephants may become habituated to crop protection methods (Figure 4). Therefore, use of additional deterrent methods during high risk periods may be a worthwhile consideration. If such additional methods are to be implemented, particular emphasis should be directed toward assessing not only technical effectiveness, but also attitudes, sustainability and scalability to maximize the potential for local adoption (Denninger Snyder & Rentsch, 2020).

ACKNOWLEDGMENTS
We dedicate this paper to Wayne Lotter who initiated this project and whose life was tragically taken while he was fighting to save elephants. We sincerely thank the participating farmers in Tloma for their cooperation throughout the study and Clemence Inyasi for diligently collecting the farm data. The study was generously funded by PAMS Foundation, Silent Heroes Foundation and Ian Somerhalder Foundation. We thank two anonymous reviewers for constructive comments.

CONFLICT OF INTEREST
Hayley Adams and Krissie Clark were partly funding this research project, yet were not involved in study design or data analysis.

DATA AVAILABILITY STATEMENT
All data are available on reasonable request from the corresponding author (ckiffne@gwdg.de).

ETHICS STATEMENT
This research was carried out with permission from TAWIRI and COSTECH (permits: 2015-169-ER-2013-191 to 2019-92-NA-2013) and consent of the T'loma village administration. The interview protocol was reviewed and was exempt from further Institutional Review Board examination under Type B, Category 2 of the US federal code 45 section 46 on human subject protections in research. All interviewees were above 18 years of age and expressed their consent, and the data were managed to ensure anonymity.

SUPPORTING INFORMATION
Additional supporting information may be found online in the Supporting Information section at the end of this article.
How to cite this article: