Contests over resources are widespread in nature. To optimize outcomes, animals assess fighting abilities, deciding to escalate conflicts based on their own strength (self-assessment) or comparing their own strength with that of their rival (mutual assessment). While most research focuses on one-on-one (dyadic) contests, the assessment strategies employed by groups remain poorly understood, even though animal groups from ants to humans engage in intergroup conflict. Mutual assessment is frequently assumed, as more information is thought to improve decision-making; however, this assumption has rarely been tested. Here, we used a dataset spanning 21 years and 633 intergroup contests in a banded mongoose (Mungos mungo) population in Queen Elizabeth National Park, Uganda. Our results support a model of self-assessment: groups with many males tend to escalate conflicts regardless of the rival group's strength, thus contrasting the commonly held assumption that decisions during intergroup contests are made by mutual assessment. We suggest that assessing rival group strength during conflict could be disproportionately costly, compared with assessing own group strength, which can be done over longer time periods and is easier to obtain. Greater understanding of these dynamics can shed light on the drivers and escalation patterns of intergroup conflict across social species, including humans.

Study Population and Data Collection

All data were collected as part of the Banded Mongoose Research Project, a long-term study on a population of banded mongooses in and around Mweya Peninsula, Queen Elizabeth National Park, Uganda (0°12′S, 29°54′E). This study includes data collected from 19^th February 2000 to 1^st April 2021, encompassing 633 intergroup contests from 39 different groups and 76 unique pairings of groups. In general, at any given time, there are approximately 250 individuals present within the population, making up 10-12 groups consisting of around 10-30 adults each (Cant, 2000; Cant et al., 2013). Every 1-3 days, researchers visited groups to record data on life-history (e.g., births, deaths), composition of each group, and details about their intergroup interactions (below).

Scoring Intergroup Contest Intensity and Escalation

Intergroup contests were recorded opportunistically as they occurred. Intergroup contests were defined as any time when at least two groups directed agonistic behavior towards each other (Fig. 2). Data on contest duration were not available. Intensity and escalation of intergroup contests were used instead as a proxy for cost and scored using comments recorded by the Banded Mongoose Research Project team. As an intergroup contest often involves a series of behaviors in which intensity escalates until one or both groups retreat (at which point the intergroup contest ends), the highest point of escalation was used as the level of intensity for each contest. The group that retreated was defined as the loser group. For ease of analysis, the range of intensities possible during an intergroup contest was divided into two categories (Fig. 2). The lowest level of intensity was “non-physical,” which was defined as an instance where two groups directed non-physical agonistic behavior (e.g., vocal and visual displays; Cant et al., 2016) towards each other and/or fled upon sighting. If this then escalated to fighting between the two groups in which there was aggressive physical contact (e.g., biting, scratching, wrestling) (Cant et al., 2016), this was defined as “physical”. Intergroup contests that escalate to physical combat are expected to have higher contest costs (such as energy used and injury risk) compared to non-physical contests (Green & Patek, 2018; Lane & Briffa, 2017; McGinley et al., 2015). Because the original dataset did not include enough information on whether only one group attempted to escalate, escalation was assumed to reflect a decision (or at least participation) by both groups, regardless of which group initiated the interaction. This approach follows the logic that even losing groups must have engaged physically to some degree and therefore incurred the associated energetic and injury risks. A more detailed behavioral analysis (e.g., whether only a subset of individuals from each group participated at any stage) was unavailable, but notably, some dyadic contest studies have used overall metrics of escalation to test assessment in a similar fashion (e.g., Green & Patek, 2018). Our scoring approach was then translated into a binomial escalation metric with 0 representing non-physical and 1 representing physical. Injurious and lethal violence in which aggressive physical contact results in severe injury or death of one or more individuals was included in the “physical” category, rather than a category of its own. This is because, firstly, injury and mortality were data-poor (9.4% of all intergroup interactions; N = 60), but more importantly, whether individuals suffer an injury is not so much a decision as it is a (potentially random) outcome of such physical combat. By contrast, whether violence escalates from non-physical to physical combat is a decision the group may make. Therefore, our use of the escalation metric likely reflects the cost each group was willing to pay.

For each intergroup contest, a qualitative comment (a description of the events of the contest) was recorded by observers in the field. These comments were then assessed by three researchers (CR, FM, DS) to evaluate whether the contest escalated into physical violence (with 0 representing non-physical and 1 representing physical), or whether the comments did not allow us to determine whether or not the contest escalated (termed: “undeterminable”). Second, all researchers assigned a confidence score (1–3) to each of their categorizations of intensity score, reflecting their certainty in the intensity score. The confidence score was based on criteria such as the clarity and detail of the recorded comment, as well as contextual factors that might influence interpretation (e.g., visibility conditions or proximity to the event). A score of 1 indicated high confidence, 2 indicated moderate confidence, and 3 indicated low confidence. We ended up removing contests that scored with low confidence (3) from the dataset because they were missing data on other variables included as fixed effects in our analysis. The three researchers discussed any ambiguous comments in detail to assign an escalation and confidence score if possible. The discussion was guided by FM, who has over 27 years of experience observing and collecting field data on banded mongooses and managing the Banded Mongoose Research Project.

We report models using both highly confident and moderately confident scores (1,2), and models including only highly confident scores (1). 

Statistical Analysis

All statistical analyses were carried out using R 4.01 (Team, 2021). Binomial escalation response data were analyzed using generalized linear mixed effect models (GLMMs) with a binomial error structure and a logit link function (Bolker et al., 2009) with the ‘lme4’ package (Bates et al., 2015). For both losing groups and winning groups, we tested the effect of two RHP predictor variables—number of adult males (>6 months old) and age of the oldest male (days) (Green et al., 2022). Both variables were scaled with a mean of zero and unit variance using the scale function (Becker et al., 1988).

In total, we ran two GLMM models, one model using data from high and moderate confidence scores, and another using data from only high confidence scores. The model for each confidence category is as follows (“~” represents “as predicted by”):

escalation ~ loser number of adult males + loser age of oldest male + winner number of adult males + winner age of oldest male

Year, winner group ID, loser group ID, and unique pairings of groups (winner group ID + loser group ID) were included as random intercepts in each model to account for repeated measures of intergroup contests between the same groups and group dyads. Test statistics were obtained using the ANOVA function and confidence intervals using bootstrapping. We present the p-values, chi-squared values, parameter estimates (β) on the logit-scale, standard errors, and confidence intervals of each GLMM model, and compare p-values and direction of parameter estimates to the assumptions of pure self-assessment, and mutual assessment (Fig. 1). The collective self-assessment hypothesis would be supported if there was a positive and statistically significant relationship between RHP predictor variables and escalation for losing groups (β>0; p<0.05) and a weaker relationship for winning groups (β>0; no significance threshold) (Fig. 1A-B). The collective mutual assessment hypothesis would be supported if these same tests indicated a significant positive relationship for losing groups (β>0; p<0.05) and a significant negative relationship for winning groups (β<0; p<0.05) (Fig. 1C-D).

Changes after Sep 25, 2025: 

Removed the ‘D:/‘ from ‘data_final <- read_csv("D:/2025_08_20_data_final.csv") in the previous script.

Removed unnecessary columns from the data_final.csv file.

Added raw data used to calculate summary statistics.