Personality traits, rank attainment, and siring success throughout the lives of male chimpanzees of Gombe National Park

Study site and subjects

Researchers have collected data during daily focal follows (Altmann, 1974) on adult chimpanzees in the Kasekela community of Gombe National Park, Tanzania since 1973. During these follows, pairs of researchers record party composition via scan samples at 15-minute intervals, as well as the timing of all individual arrivals and departures from the focal individual’s ranging party. In addition, researchers record all-occurrences data on aggression, other dominant and submissive behaviors, mating, hunting, and other behaviors in narrative notes. Finally, researchers record the focal subject’s location every 15 min on paper maps of the park (Goodall, 1986; Wilson, 2012) and, since 2018, with portable GPS units. The research team has conducted a mean of 297 (SD = 58) days of observation each year since 1973.

Subjects included 34 males. We included males that survived until at least 12 years of age, and that survived beyond the start of 1978 when dominance rank data are available (see below). Males begin to travel with adult males and challenge other males for position in the hierarchy around age 12 (Pusey, 1990). The youngest age of first reproduction among males at Gombe was 12.0 years, or 11.4 on the date of siring (Foster et al., 2009; Feldblum et al., 2021).

Dominance rank

To operationalize dominance rank, we used submissive pant-grunt vocalizations with an unambiguous actor and recipient to generate measures of male dominance rank where both actor and recipient were at least 12 years old. This is a well-established and standard method of estimating dominance rank in chimpanzees because pant-grunts are only given from the subordinate to the dominant in any pair (Marler, 1976; Goodall, 1986). The data set included 6,153 total dyadic pant-grunt events (range = 57 to 2,082 per chimpanzee) from 1978 to 2015.

To generate a dynamic longitudinal record of dominance rank, we used a maximum likelihood implementation of the Elo rating method (Elo, 1978; Foerster et al., 2016) in the package “EloOptimized” in R version 4.1.1 (Feldblum, Foerster & Franz, 2019; R Core Team, 2020). To ensure that we accurately captured relationships between personality and rank among young males, we estimated both initial Elo scores for each male upon entry into the hierarchy, as well as the scaling parameter k (Model 3 in Foerster et al., 2016). Because absolute Elo scores could be influenced by group composition, we transformed Elo scores to represent, for each individual on each day, the sum of the pairwise likelihoods of winning an agonistic encounter against all other males in the hierarchy standardized by the total number of individuals present in the community (see Foerster et al., 2016 for details). This measure varies between 0 and 1 while preserving the daily distribution of Elo scores. This resulted in 626 to 11,198 daily observations of Elo score per individual (median = 5,255.5), totaling 149,005 observations.

Personality trait ratings

A discussion of the advantages and disadvantages of using ratings and other measures, including behavioral measures, were the subjects of a target article by Uher (2008a), the Open Peer Commentary and Uher’s 2008b reply, a review by Freeman & Gosling (2010), and a chapter by Weiss & Adams (2013). These sources concurred that ratings of nonhuman primates are reliable and valid, and so produce useful information. This is not surprising for the evidence is considerable. The interrater reliabilities of ratings of chimpanzee personality traits (Crawford, 1938; Buirski, Plutchik & Kellerman, 1978; King & Figueredo, 1997; King, Weiss & Farmer, 2005; Weiss, King & Hopkins, 2007; Dutton, 2008; Uher & Asendorpf, 2008; Weiss et al., 2009; Freeman et al., 2013) are comparable to the interrater reliabilities of ratings of human personality traits (Mõttus et al., 2017). Moreover, studies that use ratings have revealed evidence for mean-level changes in chimpanzee personality traits, some of which resemble those found in humans, and evidence for the stability (the retest reliability) of these traits (Dutton, 2008; King, Weiss & Sisco, 2008; Uher & Asendorpf, 2008; Rawlings et al., 2020). There is also evidence that personality ratings measure behavioral, affective, and cognitive characteristics of chimpanzees. Personality traits measured using ratings are prospectively related to measured behavior (Buirski, Plutchik & Kellerman, 1978; Pederson, King & Landau, 2005; Vazire et al., 2007; Uher & Asendorpf, 2008; Freeman et al., 2013; Brosnan et al., 2015) and heritable (Weiss, King & Figueredo, 2000; Latzman et al., 2015; Wilson et al., 2017). Moreover, higher Dominance and Conscientiousness are positively associated with percentage of gray matter in the subgenual cingulate cortex (Blatchley & Hopkins, 2010), Agreeableness is associated with longevity in males (Altschul et al., 2018), and higher Conscientiousness, Openness, and Dominance have been associated with better performance on cognitive tasks (Hopper et al., 2013; Altschul et al., 2017; Padrell et al., 2020). Finally, rating-derived traits are not the products of anthropomorphic projections (Weiss et al., 2012).

For this study, we computed chimpanzees’ scores on six personality traits using ratings data collected in 2010 (https://osf.io/s7d9d/). Weiss et al. (2017) detail how these data were collected and on the reliability, stability, and validity of all six traits. Briefly, 18 Tanzanian researchers provided 494 ratings for 141 chimpanzees that they followed during their tenure (up to 35 years; see Weiss et al., 2017 for details). Each of these researchers completed ratings for between 21 and 43 chimpanzees on a 24-item Swahili-language version of the Hominoid Personality Questionnaire (Weiss et al., 2009; Weiss et al., 2017; Weiss, 2017). Each questionnaire item comprises an adjective and a brief description that sets the adjective in the context of behavior. For example, the item “sociable” was “SOCIABLE: Subject seeks and enjoys the company of other chimpanzees and engages in amicable, affable, interactions with them.” (boldface in original). The questionnaire instructs raters to assign each item a score from 1 (Displays either total absence or negligible amounts of the trait) to 7 (Displays extremely large amounts of the trait) based on the “... chimpanzee’s own behaviors and interactions with other chimpanzees” (King & Figueredo, 1997). The questionnaire also instructs raters to not discuss their ratings with one another. These items were used to calculate a score for each of the six traits (see Text S1: Personality Trait Calculations in Weiss et al., 2017). It is important to note that, although rater impressions originated from observations of how individuals behave and how others behave towards those individuals, these impressions were based on a range of contexts. As such, items and trait scores capture behavioral, affective, and other tendencies that are consistent across many contexts.

In the full sample of 106 chimpanzees in the Kasekela community that were rated by at least two research assistants, 19 of the 24 questionnaire items showed evidence of adequate levels of interrater reliability for further analysis (Weiss et al., 2017). The interrater reliabilities of the mean scores (Shrout & Fleiss, 1979) suggested that 11–66% (median = 38%) of the variation between chimpanzees in these 19 traits was attributable to the chimpanzees as opposed to sources of non-systematic variance, that is, error variance (Weiss et al., 2017). These 19 items were used to create six trait scores for the 106 chimpanzees from the Kasekela community (Weiss et al., 2017). These scores were based on previous chimpanzee personality research (King & Figueredo, 1997; Weiss et al., 2009). The interrater reliabilities of these trait scores based on the average of raters’ ratings were 0.51 for Dominance, 0.40 for Conscientiousness, and 0.39, 0.40, 0.35, and 0.53, for Extraversion, Agreeableness, Neuroticism, and Openness, respectively (see Table 3 in Weiss et al., 2017). Thus, around 40–50% of the variance in these scores was attributable to the chimpanzees with the remaining variance being non-systematic. These interrater reliabilities are comparable to repeatabilities (Boake, 1989; Hayes & Jenkins, 1997) of behavioral measures of personality traits across multiple taxa, which Bell, Hankison & Laskowski (2009) reported to be 0.37 (p. 774). Thus, just under two-fifths of the variance in measures, such as behavioral observations or responses to behavioral tests, such as the novel-object test were attributable to between-animal effects; the remaining three-fifths of the variance was error variance.

There is evidence for the long-time stability (retest reliability) of personality ratings among chimpanzees in the Kasekela community. In 1973, seven students and post-doctoral researchers rated 24 members of this community using the Emotions Profile Index (Buirski, Plutchik & Kellerman, 1978; Buirski & Plutchik, 1991). An examination in these chimpanzees of the correlations between the eight traits measured by the Emotions Profile Index and the six traits examined in this study found that the traits from these scales were significantly correlated in ways consistent with the meanings of the traits, e.g., the Emotions Profile Index trait “Trustful” only had a significant (positive) correlation with “Agreeableness” (Weiss et al., 2017).

Ratings data were collected on 106 chimpanzees from the Kasekela community in 2010, and 28 males from that sample served as subjects in the present study. Twelve researchers completed 88 questionnaires for these males: 25 males were rated by three of these researchers, two were rated by four, and one was rated by five. The researchers knew the chimpanzees that they rated for between 2 and 29 years (mean = 15.7, SD = 5.8). When the researchers began to collect behavioral data on the 28 subjects, the chimpanzees ranged in age from less than one year old to 26 years old (mean = 11.3, SD = 6.3). When the researchers last saw the 28 males, the chimpanzees ranged in age from 9 to 41 years (mean = 27.0, SD = 8.8).

Preliminary analyses: validating personality data

We first tested for the validity of Dominance and Conscientiousness by using bivariate regressions to examine their associations with elements of dominance rank trajectories. These tests were based on restricted sample sizes created using arbitrary criteria to include/exclude subjects that were appropriate to each measure (e.g., minimum age of survival for some analyses).

We delineated six elements of dominance rank trajectories, defined as follows: (1) Elo score at entry (restricted to the 21 males that entered the hierarchy at age 12); (2) slope of initial rise in Elo score (change in Elo score between entry and the first date of achievement of highest ordinal rank, divided by the time between entry and first achievement of highest ordinal rank; restricted to the 21 males that entered the hierarchy at age 12); (3) highest Elo score achieved (restricted to the 19 males for which we have rank data until at least age 26); (4) highest ordinal rank achieved (restricted to the 19 males for which we have rank data until at least age 26); (5) length of time at high Elo (the time difference between the last day of highest ordinal rank achieved and the first day of the same; restricted to the 11 males that rose to at least an ordinal rank of 3; (cf. Gilby et al., 2013) entered the hierarchy before age 16, and for which we have rank data until at least age 26); and (6) slope of rank decline (change in cardinal Elo score over two years following the last date of highest ordinal rank, which was restricted to 10 males that survived at least two years beyond their last recorded date at their highest ordinal rank).

We re-ran these analyses excluding Pax, a male that sustained severe testicular injuries as a juvenile and remained low ranking and peripheral throughout his life (Goodall, 1986). These additional analyses were conducted to ensure that his unusual rank trajectory did not bias results of these analyses.

Study 1: personality and lifetime rank trajectories

We sought to avoid problems related to sample size and inclusion criteria. We therefore used generalized additive mixed models (GAMMs; Wood, 2006) with daily Elo scores for all 28 males that were at least 12 years of age, and for whom we had personality data and rank data between 1978 and 2015. Because GAMMs are flexible enough to accommodate a wide range of trajectories, we did not exclude Pax for these analyses. These males were thus present for a mean of 14.1 years beyond the age of 12 (range 1.7–30.7 years).

In our GAMMs, we examined associations of age, personality, and cardinal Elo scores. These models were implemented in the R package mgcv (Wood, 2011). We chose to use GAMMs because unlike linear models, additive models are suited to modeling non-linear trajectories, such as male chimpanzee rank (Foerster et al., 2016). In our models, we used all Elo outcome data available (149,031 observations).

Different types of splines must be used for fixed and random effects. For fixed effects, we used tensor product interaction (ti) smooths because they allow separation of main and interaction effects as both can be specified using distinct ti formula terms. We opted to use penalized cubic regression splines with shrinkage for these main effects and interactions as these splines are more conservative and can shrink a parameter estimate to zero, thus removing it from the equation.

We fit six models. The first included only the main effect of age. The second and third added z-transformed Dominance and Conscientiousness, respectively, along with their interaction with age. The fourth included both Dominance and Conscientiousness, and their interactions with age. The fifth included all six chimpanzee personality dimensions and their interactions with age. The sixth was added at the request of a reviewer, to deal with potential biases of autocorrelation and includes the same fixed effects as model five, but an additional random effect—see below. For random effects, to reduce overfitting, we specified smooths with ridge penalties for each individual variable. The model included random intercepts and random slopes for each age and personality main effect, with random effect grouping determined at the level of individual chimpanzees, and a random effect for date in the sixth model.

Additive models do not produce standard regression coefficients. Smoothing splines parsimoniously fit curves, or surfaces in the case of interactions, to the data. The effect of every predictor variable can be shrunk to a linear relationship, or even to zero, if doing so generates optimal fit, minimizing overfitting. Shrinking the effect of a variable to zero is a special case where the model suggests that there is no relationship between the predictor and outcome variables, and the variable can effectively be removed from the regression equation. Therefore, although overall significance for each spline can be assessed through quantitative statistics, the effect of variables must be interpreted graphically by plotting the splines. We determined the approximate overall significance as well as specific regions of significance for two-dimensional interaction surfaces to understand the relationships among age, personality, and rank.

The modeling approach and concomitant fit statistics in GAMM are not equivalent to those of linear models. Therefore, although AIC statistics can be obtained from GAMM models, they are not reliable in the same way as are the fit statistics derived from linear models. This is because GAMM AICs cannot be scaled appropriately for weighting, so the model with the best fit always looks like it fits the data perfectly, which is not accurate. We therefore assessed model fit with deviance explained, adjusted R², and robust cross-validation measures (see Validation Analyses section in Text S1). Where appropriate, models were compared using χ² tests of twice the difference in minimized smoothing parameter selection score (in this case restricted maximum likelihood). The section of the plots that correspond to individuals in their 30s and 40s are based on fewer individuals because fewer males survive to these ages. These sections of the surfaces are thus not as reliable, and so we exercised caution when interpreting them. For more on how to fit and interpret GAMMs, we recommend Wood (2006).

Study 2: personality and reproductive success

For the second study, we analyzed siring events with known paternity between 1986 and 2014, using personality and other traits to predict likelihood of male siring success. We included males as potential sires if they were at least 11 years old on the siring date, as the youngest male known to sire an offspring in Gombe was 11.4 years old at the time of siring (and 12.0 years old at birth, as noted above), and we had both genetic and personality data for the males. This resulted in a data set that included 55 siring events, with 24 unique mothers and 22 unique males. These males had a mean age of 22.7 (range 11 to 40.8) years, and were present in the data set for a mean of 27 (range 1–55) siring events. The dataset had a median 11 (range 8–14) potential sires per siring event (see Inclusion Criteria in Text S1).

For each siring event, we estimated siring date by subtracting the Gombe-specific mean gestation length for singleton births of 228 days (Feldblum et al., 2021) from each offspring’s date of birth. We then recorded each potential sire’s cardinal Elo score and age on each siring date. Because closely related male–female dyads are less likely to mate and reproduce in this population (Feldblum et al., 2014; Walker et al., 2017), we also included pairwise genetic relatedness values (R; Queller & Goodnight, 1989) for each potential male sire with each mother in the sample. Genetic relatedness data were missing for some dyads because genetic sampling began after the death of some subjects. We excluded those dyads from the analysis (see above). To quantify male reproductive success, we used paternity data based on genetic material collected from non-invasive fecal sampling using a microsatellite-based exclusion method (Wroblewski et al., 2009; Walker et al., 2017). All paternities in the current analysis were reported in earlier publications (Constable et al., 2001; Wroblewski et al., 2009; Gilby et al., 2013; Feldblum et al., 2014; Walker et al., 2017).

For the analyses, we used information-theoretical model comparisons (Burnham & Anderson, 2002) to determine (1) whether Dominance and Conscientiousness were important predictors of male siring success controlling for dominance rank score, and (2) whether different levels of these personality traits were associated with male siring success at different times in the life course.

Following Feldblum et al. (2021), we first constructed a base model against which to compare to more complex models. The base model was a binomial generalized linear mixed model with siring success as a binary outcome variable and a logit link function. The base model included male age, male Elo score on the estimated date of siring, and male relatedness to the mother as predictor terms, as well as random intercepts for male identity. We then constructed additional models to compare with the base model to determine if including Dominance and Conscientiousness, as well as interactions between personality trait scores and other predictor terms such as Elo score and age, would improve fit. We also included models testing potential nonlinear longitudinal relationships between personality and male siring success; this was to test the prediction that alternative personality traits (e.g., high and low Dominance) should be associated with higher reproductive success at different stages in the life course. These models included both linear and squared personality trait scores, and interactions between those terms and male age. Additionally, as with the analysis of male Elo scores from Study 1, we included a model that included all six traits to reduce the likelihood of omitted variable bias and incorrectly classifying personality traits.

We compared model fits using corrected Akaike’s Information Criterion (AICc) scores (Burnham & Anderson, 2002). Models with lower AICc scores are more likely to minimize information loss when predicting siring success. We also calculated Akaike weights for each model, which can be interpreted as conditional probabilities for each model (Burnham & Anderson, 2002). All models were fitted using the R package lme4, version 1.1–27.1 (Bates et al., 2014). Model comparisons were conducted using the MuMIn package, version 1.43.17 (Bartoń, 2015).

Preliminary analyses: validating personality ratings data

Dominance and Conscientiousness were inversely correlated in the full sample of 46 males for whom personality ratings were available (r =  − 0.596, t =  − 4.918, df = 44, p < 0.001). Due to the small sample sizes in the validity analyses, we report below associations where effect size and confidence intervals suggest an association may be present even if the 95% confidence interval does not exclude 0. Note that cardinal Elo scores can range from 0 to 1. Dominance was positively associated with Elo score at entry (β = 0.027, 95% confidence interval (95% CI) = [−0.016–0.071]). This represents about 7% of the total range of Elo scores at entry (which ranged from 0.08 to 0.48). This association remained after excluding Pax (β = 0.036, 95% CI [−0.009–0.080]). Dominance was also positively associated with the slope of initial Elo rise (β = 0.020, 95% CI [0.005–0.036]). It was not possible to include Pax in this analysis because he never increased in ordinal rank after entering the hierarchy. Among the 19 males that survived until at least age 26, Dominance was positively associated with the highest cardinal Elo score achieved (β = 0.099, CI [0.008–0.190]) and highest ordinal rank achieved (β = −0.986, 95% CI [−1.934 to −0.038]), although these effects were attenuated after excluding Pax (cardinal Elo score: β = 0.052, 95% CI [−0.030–0.134]; ordinal rank: β = −0.359, 95% CI [−1.034–0.317]).

Although a smaller effect, there was evidence that Conscientiousness was inversely associated with Elo score at entry (β = −0.035, 95% CI [−0.082–0.011]). This association remained after excluding Pax (Elo score at entry: β = −0.039, 95% CI [−0.086–0.008]). Slope of initial Elo rise was also inversely associated with Conscientiousness (β = −0.013, 95% CI [−0.031–0.006]). As before, Pax was not included in the latter analysis because his Elo score never rose beyond his score at entry. Finally, among the 12 males (excluding Pax) for whom we have Elo score data by age 16, and until at least age 26, and that rose to at least the third position in the dominance hierarchy, there was an inverse association between time at their highest ordinal rank and Conscientiousness: males that were lower in this trait had longer tenures at the highest ordinal rank they achieved (β = −0.223, 95% CI [−0.393 to −0.054]).

The remaining components of rank trajectories were not associated with Dominance or Conscientiousness. This may reflect the fact that calculating measures of these latter components required further restricting sample sizes and choosing arbitrary age cutoffs for inclusion.

Study 1: personality and lifetime rank trajectories

In the 28 males in the sample, rank scores tended to rise gradually, experience a peak, and then fall gradually (Fig. 1). Yet the shape of rank score trajectories still varied considerably between individuals: some individuals changed dramatically over short intervals, other individuals never achieved marked gains in rank score, and the trajectories of others were cut short by death or because the study period ended while these individuals were young (Fig. 2).

To test whether Dominance or Conscientiousness play different roles in rank attainment at different times in a male chimpanzee’s life, we modeled overall and time-varying relationships between individual personality traits and rank. GAMM fit was strong: for the simplest age model, 82% of the variance (adjusted R²) and 78% of the deviance (the likelihood-based goodness-of-fit statistic D = 12905.53) in rank scores could be explained by the model. The fifth model, which contained all six personality dimensions and their interactions with age, fit the data well: 91% of the variance (adjusted) and 89% of the deviance (D = 6342.319) in rank scores could be explained (χ² = 140617.6, effective df = 28, p < 0.001). At a reviewer’s request, we fit a more complex model with the addition of a random effect for date. This more complex model (Table S1, Figs. S1–S5) accounted for 91% of the variance (adjusted R²) and 89% of the deviance (D = 6375.037), but it differed in no substantive way from the simpler model, so subsequent analyses interpreted the less complex version of the GAMMs, the fifth model.

We also used cross-validation, which compares real outcome values for data not used to fit the model with what the model predicts the outcome values will be when those data are fed to the model as new inputs. We evaluated the cross-validation performance of the fifth model using the mean squared error, which represents the difference between true values and model predicted values. If the model’s prediction was perfect, the mean squared error would equal zero. Cross-validation yielded strong mean squared errors of 0.0078 and 0.0449 for stratified 10-fold and forward chaining average predictive error, respectively. In general, scores seemed to be poorest in the middle, where the training dataset included between six and eight folds. On average, scores were good and compared favorably with the test error of the models (Table S2). While the results of 10-fold cross-validation do not rule out overfitting, they do suggest that the scores are inflated due to temporal autocorrelation, as expected. We also cross-validated the sixth model, and its critical forward chaining score was 0.0452 (Table S3), indicating no improvement on the performance of the simpler fifth model. We also checked for multicollinearity by calculating variance inflation factors (Fox & Monette, 1992) for our best-fitting model and our base model using the car package, version 3.0-11 (Fox & Weisberg, 2019). All variance inflation factors were below 4, indicating no risk of multicollinearity.

Parameter estimates for the fifth model are presented in Table 2. Effects in generalized additive models are represented by effective degrees of freedom (edf), which represent the shape of the association between the independent variables and their interactions, and the dependent variable. An edf of 1 indicates a linear relationship, an edf of 2 indicates a quadratic relationship, and so on. The association between age and rank was statistically significant, approached a sixth-order polynomial function (edf = 5.97; Fig. 2), and was consistent with rank trajectories described elsewhere (Foerster et al., 2016). Another finding consistent with the literature was the presence of individual variation in rank trajectory, that is, the chimpanzee ID × age (random effect) term was significant (edf = 4.32). Dominance was linearly associated with rank (edf = 1.17) such that higher Dominance individuals tended to have higher rank regardless of age (Fig. 3, top). Although the association between Conscientiousness and rank varied across individuals (edf = 10.26), the main effect of Conscientiousness was significant (edf = 3.69). Individuals with the highest Conscientiousness scores, therefore, tended to have lower rank, while those individuals with Conscientiousness scores that were slightly above average had slightly higher rank (Fig. 3, bottom).

The effects of the interactions between age and Dominance indicated that all males tend to start out at lower ranks. Young adults (aged around 20 to 25 years) who scored average or higher on Dominance had higher rank scores, while males in the same age range who had low scores on Dominance had relatively lower rank scores (Fig. 4, top). Conscientiousness showed only a few age-dependent associations with rank, but they are difficult to interpret (Fig. 4, bottom). Young males (12–18) of average Conscientiousness appear to have slightly higher rank, as did high Conscientiousness males aged about 21–26.

Study 2: personality and reproductive success

Six models had lower AICc scores than the base model, three of which were within 2 AICc points of the best model, indicating roughly equivalent support (Table 3). In addition to the terms in the base model, the best model (Akaike weight = 0.269) included a linear term for Dominance and a quadratic term for this trait (i.e., Dominance²). The parameters for the best fitting model are presented in Table 4. Male age was negatively associated with siring success, although the 95% confidence interval did not exclude one (odds ratio (OR) = 0.78, 95% CI [0.55–1.11], p = .171; Fig. 5B). Genetic relatedness was also negatively related to siring success (OR = 0.15, 95% CI [0.03–0.71], p = .017). Male Elo score was positively associated with siring success (OR = 1.44, 95% CI [1.05–1.97], p = .023; Fig. 5A). Finally, even though this model controlled for rank, Dominance was associated with a greater likelihood of siring offspring (OR = 2.59, 95% CI [1.30–5.16], p = .007), and this relationship was not linear (Dominance² OR = 0.72, 95% CI [0.52–1.00], p = .047), such that the most successful males were those with above average, but not the highest, Dominance scores (Fig. 5B).

The two best-supported models included Dominance², and models including this term accounted for 52% of total model weight, indicating substantial support for a quadratic relationship between Dominance and siring success. Nevertheless, several models with better fit than the base model contained only linear effects of Dominance (Table 3). Therefore, although we found strong support for a relationship between Dominance and male siring success, and substantial support for this relationship suggests that it is nonlinear, without a larger sample, it is too soon to determine the precise shape of this relationship. The siring success model that included all six personality traits had a poorer fit than the models with only Dominance.