Female monopolizability promotes within-community killing in chimpanzees

Male chimpanzees (Pan troglodytes) defend group territories and sometimes kill members of rival communities — a pattern often compared to human warfare1-3. Male chimpanzees also sometimes kill grown males from their own community4-9. Such within-community killings are puzzling, as they reduce the coalition strength needed to win inter-community contests5,10. Here we examine the contexts of within-community killing using data from two neighboring communities at Gombe National Park, Tanzania, as well as published data from other long-term chimpanzee study sites. At Gombe, more killing occurred within the smaller Mitumba community, where fertile females were more monopolizable. Attackers increased their share of mating and paternity following known and inferred killings. Other factors proposed to explain such killings, including the degree of intercommunity threat, male-biased sex ratios, high population density, or generalized aggression, did not explain the high rates of killing in Mitumba. Comparing across study sites, the best predictor of within-community killing was an index of the degree to which fertile females can be monopolized by the highest ranking male11. Our ndings therefore support the hypothesis that within-community killing is a strategy to eliminate reproductive rivals that is more likely to pay off when fertile females are more easily monopolized.

We tested these hypotheses using data from the Mitumba (median = 3 males, 9 females ≥12 years old) and Kasekela (median = 12 males, 23 females ≥12 years old) communities at Gombe (1997Gombe ( -2018. During this study period, we infer that ghting among males led to 5 deaths in Mitumba and 1 in Kasekela (Table S1). The killings in Mitumba began in 2004, when alpha male Vincent was injured in a fall. During the subsequent three months, Vincent avoided the other two adult males of Mitumba, Rudi and Edgar, but when these two males encountered Vincent in December 2004, they attacked and killed him. In January 2005, Edgar's 8. 2 year-old brother Ebony was found dead with injuries consistent with attack by chimpanzees. Circumstantial evidence indicates Ebony was likely killed by a Mitumba male, as most of the Kasekela males were under observation 2.5 km away at the likely time of Ebony's death. Rudi then became alpha male, but Edgar challenged Rudi repeatedly, severely injuring Rudi in 2007. Edgar became alpha male in 2008. Rudi disappeared in 2013; as he had not recently been injured, and also was positive for SIVcpz, we infer that he died from disease. Edgar injured three other males in severe attacks: Forest (2012), Apple (2015), and Fansi (2017). Forest and Apple disappeared soon after the attacks; we infer they died from injuries in icted by Edgar. In 2017, observers found Edgar displaying near Fansi's freshly killed body; we infer Edgar was the killer. In Kasekela, one male, Kris, disappeared after an attack by the alpha male, Ferdinand; we infer Kris died from the resulting injuries. Controlling for population size in the manner commonly used for homicide data (killings per year per 100,000 individuals 18,19 ) yields rates of 2859 (Mitumba) and 242 (Kasekela) killings per 100,000 weaned males per year, much higher than the median rate across long-term chimpanzee studies, which is 0 (Table S2).

Ranging and male coalition strength
We examined both the potential costs and bene ts of killing males from one's own community. We began by exploring the impact of male coalition size on effectiveness in competition for territory. Mitumba had a smaller home range than Kasekela (1997-2016, n=20 years; Mitumba: median=5.13 km 2 , range=3.51-8.03; Kasekela: median=16.5 km 2 , range= 10. 9-19.4). The Mitumba community's range size increased when they had more adult males relative to the number of males in the neighboring community (Pearson correlation; t=2.6, df=18, n=20, p=0.02, r=0.52, 95% CI=0.11-0.79) but Kasekela's did not (t=1.42, df=18, n=20, p=0.17, r=0.32, 95% CI=-0.15-0.67; Figure S1), probably because when Kasekela chimpanzees lost range to Mitumba, they were able to expand their range to the south at the expense of the declining Kalande community 20 . Kasekela's annual range center shifted south by 585 m during the study period range centers were positively correlated (Pearson Correlation, t=2.74, df=18, n=20, p=0.01, r=0.54, 95% CI=0.13-0.79), indicating that Kasekela shifted southward to avoid Mitumba chimpanzees as the number of mature males in Mitumba increased.

Rates of male aggression
If within-community killings occurred as a byproduct of overall higher rates of aggression, then these rates should be higher in Mitumba than in Kasekela. However, controlling for individual identity and observation time, Kasekela and Mitumba males did not differ in their rate of aggression, considering either all aggressive acts (mixed Poisson regression, β=0.64, 95% CI=-0.31-1.55, z=1.43, p=0.152), or only instances of higher-cost aggression (chasing and/or physical contact: mixed Poisson regression, β=-0.30, 95% CI=-0.98-0.42, z=-0.90, p=0.37; Table S3). Even Edgar, who was observed attacking four of the ve victims in Mitumba, nonetheless did not have an exceptionally high rate of aggressive behavior overall ( Figure S5). These ndings indicate that the higher rate of killings in Mitumba did not result from simply from a higher rate of aggression in that community. Moreover, in the observed case, Edgar and Rudi continued attacking Vincent until he was dead, suggesting that killing was intentional. The many wounds found on Fansi's freshly killed body also appear consistent with intentional rather than accidental killing.
Attackers are those listed in Table S1 as known or inferred attackers in fatal cases. Non-attackers are all other mature males (i.e., those that were observed mating in periods before and after violent events). Error bars represent one standard deviation of uncertainty.

Reproductive consequences for attackers
Known and suspected attackers from both communities increased their share of observed mating following severe within-community attacks ( Figure 2, Figure S9; Binomial glmm, Attack Order*Attacker, β=0.62, 95% CI=0.42-0.81, z=6.2, p<0.01). Paternity data also indicated that community size affected the extent to which attackers gained bene ts from intra-community killing ( Figure 3). In the small Mitumba community, the alpha male sired 80% of infants born in Mitumba, but only 18% of infants born in Kasekela (1997Kasekela ( -2018 Figure 3A). In 2009, Kasekela's alpha-male, Ferdinand, similarly increased his share of conceptions after attacking Kris, an adult male who subsequently disappeared. Ferdinand's share of paternities increased from before his attack on Kris (6.3%, n=2) to after (15.4%, n=2); however, this change was not statistically signi cant ( 2 =0.16, df=1, p=0.70, Φ=0.15; Figure 3B). Both Mitumba attackers fathered more infants than all but three Gombe males ( Figure 3C-D), making them exceptionally successful even after taking into account the death of Andromeda (sired by Edgar), whose killing by Kasekela males might have been averted if Vincent had survived. After Rudi died, Edgar obtained 72% of observed matings ( Figure S9A). Edgar likely has sired even more offspring; 14 Mitumba infants remain un-genotyped, with only one conceived before Rudi's death.

Cross-site analysis
Using the Gombe data presented here, together with published data from long-term chimpanzee eld sites 8 (Table S5), we employed an information-theoretic model selection approach to test which of three factors best predicted rates of within-community killing: operational sex-ratio, population density, and monopolizability. The best model included only monopolizability (model weight w = 0.70), and only for monopolizability did the 95% con dence interval of the model-averaged parameter value exclude zero (Table 1). These results support the view that within-community killing of male chimpanzees results from male reproductive competition [4][5]9 , and that monopolizability provides a better predictor than sex ratios. High monopolizability of fertile females can make it worthwhile for males to eliminate rivals.

Conclusions
In summary, we found that within-group killings of male chimpanzees occurred most frequently when fertile females were most readily monopolized. Because killing male group members reduces ability to hold territory, such killings must be offset with substantial reproductive bene ts to be adaptive. Our ndings indicate that this is most likely to be the case in small communities. Nonetheless, such killings of males have been observed in larger communities with many females, such as Ngogo 5 , Mahale 7 , and Budongo 4 . Other factors likely affect both the bene ts and the costs of killing, such as relative ghting ability. Attackers often target young or injured individuals, suggesting that chimpanzees attack when they can do so at low risk to themselves.
That within-community killing occurs even during times of intense inter-community threat challenges a widespread explanation for the evolution of cooperation in humans. Following Darwin 23 , Alexander 24 and others have proposed that within-group cooperation in humans co-evolved with between-group competition. Bowles 25 developed this idea further in the parochial altruism hypothesis: intense competition between groups selects for increased cooperation within groups. However, chimpanzees and human hunter-gatherers experience similar rates of between-group killing 19 . High rates of between-group killing thus appear insu cient to promote the high levels of within-group cooperation observed in human societies.

Animal Subjects
We report data generated from non-invasive observational studies of chimpanzees at Gombe National Park, Tanzania (4°40′S 29°38′E). Long-term researchers at Gombe Stream Research Centre, employed by the Jane Goodall Institute, collected behavioral data. Gombe encompasses a 35.69 km 2 strip of land along the shore of Lake Tanganyika that is home to 3 communities of chimpanzees. Researchers conducted nearly daily focal follows of individual chimpanzees, recording party composition and location at 15-minute intervals while maintaining a continuous record of feeding behavior 15 . Researchers recorded all observed occurrences of selected behaviors, including aggression, mating, and intercommunity interactions. Observers documented the reproductive state of all adult females seen each day, based on the size of ano-genital sexual swellings.

Data Killings
We examined evidence for all reported cases of within-community killing at Gombe for the Mitumba and Kasekela communities (1997-2018). Following Wilson et al. 8 , we rated killings to be 'observed' if observers directly witnessed a within-community attack leading to death, and 'inferred' if the victim disappeared following one or more severe within-community attacks, or if compelling evidence indicated that the individual was killed by chimpanzees (such as a body found with multiple bite wounds, and/or skeletal trauma consistent with a chimpanzee attack). For this analysis, we did not include cases that were only suspected to be killings. For cases that were not directly observed, we considered them to be within-community killings only if circumstantial evidence su ciently ruled out intercommunity attacks or other causes of injury.
To investigate whether chimpanzees who kill community members bene t following lethal attacks, we identi ed which individuals were involved in each attack. For all cases, we used documentation of attack records to identify individuals involved in each killing. For observed cases, we classi ed all individuals observed attacking the killed chimpanzee as 'attackers'. For disappearances preceded by withincommunity attacks, we considered individuals observed attacking the victim leading up to their disappearance as attackers. For other inferred cases, we used circumstantial evidence to identify likely attackers given sightings surrounding the incident. For all cases, we categorized all remaining adult males in the community as non-participants.
Demographics: As part of the long-term study, researchers documented demographic events, including births, deaths, immigrations and emigrations. Using these data, we calculated annual community size and composition using all individuals known to be alive on 01 Jan of each year . For this study, we considered all individuals 12 years of age and older as 'adults' (old enough to participate in intercommunity ghting and reproductive competition). For each year for each community, we calculated the 'male ratio': the number of males in the focal community divided by the number of males in the neighboring community.

Ranging
During focal observations at Gombe, observers collected the focal target's location at 15-minute intervals.
Prior to 2005, Gombe researchers recorded locations manually on maps. These were later digitized into the UTM35S coordinate system. We used these locations to calculate annual home range size for each community, using 99% minimum convex polygons (R package adehabitatHR) 27 for the years 1997-2016, during which location data had been digitized and entered into the Gombe Database. From these ranging data, we calculated the latitudinal range center for each community. Gombe is bordered to the west by Lake Tanganyika and to the east by village land. The boundary between the Mitumba and Kasekela territories runs roughly east-west, along a ridge dividing the Mitumba and Busindi stream valleys. The north-south dimension thus provides a proxy of movement relative to community boundaries. We de ned the 'contested area' between the two communities as the 99% minimum convex polygon enclosing points used by both communities (1997-2016). We calculated the annual use of this contested area as the percentage of each year's location points located within the contested area. To quantify variation in annual inter-community pressure experienced by Mitumba chimpanzees, we calculated the number of days each year that Kasekela males were recorded traveling north of 9485400 N, UTM 35S, a point along the ridge dividing the two communities' ranges.

Aggression
We analyzed aggression data from focal follows of male chimpanzees conducted in Mitumba (2000-2008) and Kasekela (1997-2011) for all years within the study period for which aggression data have been extracted, cleaned, and entered into the Gombe Database. We limited analysis to aggressive acts conducted by the focal target. We classi ed each observed aggressive act into 4 categories of increasing severity (bristle, display, chase, contact) following published previous studies [28][29] . 'Bristle' involved piloerection (raised hair), 'display' involved exaggerated or demonstrative movement directed at another chimpanzee, 'chase' as one individual actively pursuing another, and 'contact' as hitting, kicking, or biting from the focal target. We also combined chase and contact together as 'severe aggression' for analysis.

Density
Using annual range estimates and demographic data described above, we calculated the annual density for each community, de ned as the number of chimpanzees per square kilometer. For cross-site analysis we used published density data 8 .

Mating
We analyzed all cases of observed mating between adults collected during focal follows (1997-2016).
To assess whether participants in killing increased their share of matings, we calculated each male's proportion of mating relative to each within-community killing. In Mitumba, we inferred that 5 withincommunity killings occurred during the study period, so we identi ed the proportion of mating events each male participated in for 6 time periods surrounding these events. In Kasekela, only a single within-community killing took place over the study period. To identify a comparable comparison of change in mating behavior we analyzed mating that took place 3 years before and after the within-community killing in Kasekela to match the time periods between killings in Mitumba (mean = 3.18).

Paternity
Paternity was assigned using genetic material from fecal samples. Using PCR, 11 microsatellite loci were ampli ed and used to match fathers with offspring (see Wroblewski et al. 30 ). Genetic sampling of offspring usually takes place after the infant has survived for two years 30 . Over the study period, the Kasekela community had 45 assigned paternities (1997-2016), while Mitumba had 15 (1997-2015). When assessing lifetime reproductive success, we considered all paternity assignments from Gombe (n = 80).

Genotyping and paternity
This study included paternity for 69 offspring previously reported [30][31][32][33][34] , and new paternities established for 11 offspring (nine in Kasekela and two in Mitumba) (Table S6). Two offspring were genotyped at 8-11 microsatellite loci following previous PCR and repeat size-based sequencing methodology (described in Walker et al. 34 , Wroblewski et al. 30 ), while the other nine offspring, their mothers, and candidate males were genotyped at 8 of those loci using next generation sequencing-adapted PCR and amplicon sequencing (described in Barbian et al. 35 ). Paternity analysis was done by the exclusion principle and con rmed via likelihood methods using Cervus 3.0 36 , as described by Walker et al. 34 , conservatively including all genotyped males as candidate fathers, regardless of community residence or known dates of birth or death. For the two offspring assessed using the sized-based genotypes, candidate males included all sampled males from all three Gombe communities (Mitumba, Kasekela, and the unhabituated but monitored community of Kalande), while those newly assessed using next-generation sequence-based genotypes included all genotyped males from Mitumba and Kasekela, but not Kalande (as updated genotypes for them were not available). In both cases, however, we also simulated 90% sampling of candidate males for a conservative estimate of paternity con dence in Cervus.
If a parameter had been estimated for a speci c site, we used that estimate, however if a site-speci c estimate was unavailable, we used the average estimate for the subspecies. If no published estimates were available within the subspecies, we used the average estimate across all study sites. The relationship between predicted monopolizability and number of adult females thus varied across subspecies ( Figure S10).

Observed Monopopolizability
To test the validity of the predicted monopolizability estimates, we calculated two measures of observed monopolizability at Gombe, based on (1) mating and (2) sexual swellings. We de ned observed monopolizability as the proportion of days on which the number of females observed mating, or with a maximally tumescent sexual swelling, was zero or one. For each day in each community, we calculated the number of females that were observed mating, and with a full swelling. We compared these annual measures of observed monopolizabilities with predicted values for both communities. We found that predicted monopolizability estimates t monopolizability calculated with swelling observations (linear regression, β = 1.41, 95% CI = 1.01-1.81, t = 7.11, p < 0.01, R 2 = 0.60) better than monopolizability calculated with mating observations (linear regression, β = 0.13, 95% CI=-0.05-0.30, t = 7.11,t = 1.48, p = 0.15, R 2 = 0.06).
Sexual swelling data exhibited more variation than mating data and appeared a better overall measure of monopolizability. This is likely because matings last only a few seconds and can be missed, whereas swellings last many days and are readily observed. Moreover, the discrepancy between monpolizability calculated from matings versus swellings was higher for Kasekela than Mitumba, which is consistent with other evidence that Kasekela females were less monopolizable than Mitumba females.

Reproductive cycles
The calculation of observed monopolizability may overestimate actual monopolizability, because chimpanzees often travel in multiple scattered parties, while researchers generally follow a single focal target each day. Additionally, chimpanzees on consortships often appear to avoid other chimpanzees and researchers directly observe consortship pairs rarely. To capture reproductive state for all females, including those not directly observed, we calculated the daily number of cycling, parous females in Mitumba and Kasekela over a period for which complete data on female reproductive states were inferred in both communities using sexual swellings to identify reproductive cycling (1997-2014; Fig. 1). We limited this analysis to parous females. Male chimpanzees appear to prefer mating with parous females, likely because nulliparous females experience prolonged periods of adolescent subfecundity prior to their rst birth 40 . We excluded females that were likely to be pregnant or lactating, based on records of sexual swellings and the timing of births. We identi ed likely pregnancy start dates based on the end of the menstrual cycle closest to 225 days before birth. For females that had given birth, we classi ed them as having resumed cycling upon their rst observed sexual swelling after birth. We inferred that menstrual cycles took place if females were observed with sexual swellings (at least two of them full swellings) over a 14-day window, followed by at least 5 days without a full swelling. We determined the day of detumescence as the subsequent day without a full swelling or, in the case of a gap in observation, halfway between the nal observation of full swelling and the next observation.

Cross-site Analysis
We used data published by Wilson et al. 8 for cross-site analysis, including occurrences of withincommunity killing, demographics, and range size. We updated numbers for Mitumba and Kasekela using data from other sections of this paper.

Statistical Tests
We conducted statistical tests in Rstudio (Version 3.6.1) 41 . To assess the bene ts that males gain by having other male community members, we tested the impact of two measures of male coalition strength at Gombe: the raw number of adult males and the male ratio. We tested the extent to which each of these variables correlated with three measures of range use -range size, the northing of the range center (UTM 35S) and use of the contested area -using Pearson correlation tests. Range size did not correlate with raw number of males ≥ 12 years old in Mitumba (Pearson correlation; t = 1.8, df = 18, n = 20, p = 0.09, r = 0.39, 95% CI=-0.06-0.71) or Kasekela (Pearson correlation, t=-0.27, df = 18, n = 20, p = 0.79, r=-0.06, 95% CI=-0.49-0.39; Figure S1). All other test statistics are reported in the body of the paper.
To address the hypothesis that within-community killing results from generalized aggression within communities, we used GLMMs with Poisson error structure (lme4 R package 42 ) to compare rates of aggression by the targets of focal follows of males in Kasekela and Mitumba. We ran two separate regressions: one for hourly aggression rate using all instances of aggression, and another that only considered contact and chase aggression. In both regressions, we assessed the degree to which individuals differed between Kasekela and Mitumba, using individual identity as a random variable and log(focal follow duration) as an offset.
To assess the possibility that population density drives within-community killing, we used a two-tailed ttest to compare annual population density in Mitumba andKasekela (1997-2016). To further examine the connection between killing rate and population density, we used a Mann-Whitney U test to compare population in density in years during which within-community killing occurred and other years.
To assess the possibility that operational sex-ratios affect rates of within-community killing, we used a two-tailed t-test to compare annual operational sex ratios in Mitumba and Kasekela.
We then assessed the relationship between the monopolizability of mating opportunities and withincommunity killing rate. We used a two-tailed t-test to compare the annual mean number of cycling, parous females present in Mitumba and Kasekela. Using all observed mating events, we calculated the daily number of females observed mating. We tested whether the proportion of days with one female observed mating differed between Kasekela and Mitumba using a chi-squared test.
To assess reproductive gains that males experienced following within-community kills, we used both mating observations and paternity assignments. For Mitumba, we binned these mating events around within-community killings. We conducted a GLMM with a binomial error structure (lme4 R package 42 ) to assess whether attackers increased their share of matings following an attack, controlling for attack identity and individual identity. We classi ed individuals as suspected attackers or non-attackers and tested whether there was a non-zero coe cient for the interaction between attacker and time period relative to the attack. We compared lifetime individual reproductive success across communities, de ned as the sum of all assigned paternities over an individual's lifetime. To test for an impact on reproductive shares surrounding killing, we used a chi-squared test to compare the distribution of reproductive success in Mitumba before and after Rudi and Edgar killed Vincent, the alpha male. We also conducted a chisquared test to compare the share of reproductive success of Ferdinand in Kasekela before and after he attacked Kris, who is inferred to have died from the resulting injuries.
To investigate which factors best explained the number of within-community killings of weaned males per community on a larger scale, we used killing and demographic data published from 18 chimpanzee communities at long-term study sites 8 . Using information-theoretic model selection procedures 43 , we examined a set of a priori speci ed models based on three hypotheses: operational sex-ratio, density, and monopolizability. All models were Poisson regressions with the number of within-community killings observed or inferred in each community as the dependent variable, and the log(weaned male years) as an offset. Here, we consider 'weaned male years' to be the product of the average number of weaned males (≥ 3 years old) multiplied by the years of demographic data reported. We included ve models in our model set: one for each hypothesis, a null model, and a full model. Using the AICcmodavg R package 43 , we identi ed model-averaged parameters weighted by the t of each model to identify non-zero coe cients. Data Availability: Summary data used in this analysis are available in the Dryad data repository with the identi er https://doi.org/10.5061/dryad.bcc2fqzbf. We do not provide access to the raw data used in some of these analyses, as these represent a substantial fraction of the long-term data from Gombe, which are not publicly available at this time because of multiple ongoing studies, but are available from the corresponding author on reasonable request.
Code Availability: R code used for data analysis is available in the Dryad data repository with the identi er https://doi.org/10.5061/dryad.bcc2fqzbf. Figure 1 Timeline of cycling, parous females in Mitumba and Kasekela, showing the daily number of parous, cycling (neither pregnant nor lactating) females in Mitumba and Kasekela. Mitumba consistently had fewer of these high value mates and more instances of within-community killing.

Figure 2
Mating success for attackers and non-attackers before and after 5 known or inferred killings in Mitumba and 1 inferred within-community killing in Kasekela. Attackers are those listed in Table S1 as known or inferred attackers in fatal cases. Non-attackers are all other mature males (i.e., those that were observed mating in periods before and after violent events). Error bars represent one standard deviation of uncertainty.