Individual behavior varies for many reasons, but how early in life is such variability apparent, and is it under selection? We investigated variation in early-life behavior in a wild eastern grey kangaroo (Macropus giganteus) population, and quantified associations of this behavior with early survival. Behavior of young was measured while still in the pouch and also as subadults, and we also monitored survival to weaning. We found consistent variation between offspring of different mothers in levels of activity at the pouch stage, in flight initiation distance as subadults, and in subadult survival, indicating similarity between siblings. There was no evidence of covariance between the measures of behavior at the pouch young vs subadult stages, nor of the early-life behavioral traits with subadult survival. However, there was a strong covariance between offspring and mothers’ flight initiation distance (FID) tested at different times. Further, of the total repeatability of subadult FID (55.3%), more than two-thirds could be attributed to differences between offspring of different mothers. Our results indicate that (i) behavioral variation is apparent at a very early stage of development (still in the pouch in the case of this marsupial); (ii) between-mother differences can make up much of the repeatability of juvenile behavior (or ‘personality’); and (iii) mothers and offspring exhibit similar behavioral responses to stimuli, potentially indicating heritability of behavioral responses. However, (iv) we found no evidence of selection via covariance between early-life or maternal behavioral traits and juvenile survival in this wild marsupial. Keywords: animal personality, maternal variance, early-life behaviour and survival, macropods, multivariate Bayesian statistics.

Study population

We studied a wild population of marked eastern grey kangaroos at Wilsons Promontory National Park, VIC, Australia (38º57’S, 146º17’E). In this population, adult females usually reproduce every year, producing at most one offspring per year.  Offspring stay in the pouch for approximately 10 months, then continue to be nursed until approximately 19 months of age (Poole et al. 1982; King and Goldizen 2016). Most offspring are born from November to May; for this analysis, cohorts are named after the calendar year for January of a given breeding season (e.g. an offspring born in either December 2017 or January 2018 would both be assigned to the 2018 cohort). Survival rates of juveniles vary widely across years (range 8 – 89%), possibly due to variation both in density of likely predators, in particular foxes (Vulpes vulpes), and in weather conditions (Bergeron et al. 2023; Plaisir et al. 2022). Approximately 75% to 85% of kangaroos in the study area were individually marked and monitored for survival and reproduction each year, with a resighting probability of 99% for females and 92% for males (Bergeron et al. 2023). Captured individuals were ear-tagged with a unique color combination at first capture; this includes any pouch young weighing >900 g (King et al. 2011). We considered individuals aged three years and older as adults.

This study involved individuals born in four cohorts (2016 to 2019) and their mothers, with behavioral observations and pouch young measures made in 2017, 2018, and 2019. In these years, there were 276-336 marked individuals in the study population, comprising 115-140 adult females, about 55 adult males, 50-85 subadults, and 30-55 pouch young. We estimated the birth dates of pouch young based on body size measurements (Poole et al. 1982). We classified the developmental stage of young as: small pouch young (unfurred, small distension of pouch, aged < 3 months); medium pouch young (not completely furred, head sometimes out of medium-sized pouch, aged 4-6 months); large pouch young (completely furred, often with head outside pouch, aged 7 months or older); and 1-year-old offspring (out of the pouch but unweaned) (Jaremovic and Croft 1991).

Mothers were typically captured from August to November, when pouch young were aged about 7-9 months and were making only occasional exits from the pouch (Poole 1975). Mothers were sedated at capture, but their pouch young were not (King et al. 2011). While we collected measurements of the mother, her young remained inside the pouch.  We then extracted the young from the pouch for measurement and tagging, and whilst we weighed the mother.  We returned the young to the pouch after an average of 8 minutes. As described below, we scored the behavior of each young during its short removal from the pouch. We also did not measure behavior of the mother within a few days following capture.

Definitions of traits

We recorded data on four traits: (1) movement of pouch young during handling at capture; (2) flight initiation distance (FID), of subadults aged 13-35 months; (3) FID of adult females; and (4) survival of subadults to weaning at 21 months. We describe these in turn below; summary statistics and samples sizes for each trait are given in Table 1.

(1)   Pouch Young (PY) Movement

We assessed pouch young behavior by scoring body movements across eight stages of handling at the point of capture of the mother: extraction from the mother’s pouch, placement of the offspring inside a cloth bag, weighing, collection of three body measurements (foot, hind leg and head length), tagging, and being held briefly inside an observer’s jacket (for warmth) while the mother was weighed. We recorded 1/0 for any presence/absence of movement at each stage, summed these values across all points to a maximum of 8, then divided by 8 to estimate an individual score ranging from 0 to 1. For practicality, in 2019 we combined responses for foot and hind-leg length measurements and for extraction from the pouch and placement inside a cloth bag, to give a total of only six stages of handling, but again deriving a 0-1 score. The mean total duration of the handling process across all years was 7.9 (±1.9 SD) minutes. The pouch young data (PY Movement) were assessed for the 2017, 2018, and 2019 cohorts, at 6.4 – 10.8 months of age.  The average score of PY Movement was 0.47 (±0.24 SD, range 0 - 1). All analyses of PY Movement included age at capture (mean 8.0 ± 0.8 SD months) as a covariate in the model, to account for any changes in behaviour with age.

(2)   Subadult Flight Initiation Distance (FID) and (3) Adult Female Flight Initiation Distance

We measured flight initiation distance of subadult offspring (age range 13-35 months) and of adult females, as the distance to the nearest meter at which a kangaroo moved away when approached by a human (Strong et al. 2017). High values of FID indicated a response to the approaching human at a greater distance, whereas low values indicated that the individual allowed the human to come closer before it moved, implying a ‘bolder’ response. The observer started at 30 m from a target kangaroo that was positioned alone or in the periphery of a group, walked directly towards that individual at a constant pace, and stopped when the individual took flight. The terrain was flat and there were no obstructions that could limit the view of the approach. We tested one individual per group per approach, in group sizes of 1 to 6. In groups, the target individual was always the closest to the observer. We always tested mothers and their offspring on different dates, and we also always avoided tests when they were within 3 m of each other.  FID measures were made in 2017, 2018 and 2019. Juveniles from the 2016 and 2017 cohorts thus had their responses measured at ages 1 and 2 years, whereas juveniles from the 2018 cohort were measured at age 1 year only. The 2019 cohort was not tested for FID. We measured the FID of 156 adult females between 2017 and 2019; of these 156 females, we also had measures of PY Movement or Subadult FID measures of 87 (Table 1).

For all measures of FID, we also noted the date, time, geographic position, group size and the presence of the mother or an offspring in the same group, when applicable. Geographic positions were recorded with a Garmin^® GPSMAP 64 device, as an east-west coordinate in meters. To determine membership of groups, we followed a ‘10-m chain rule’ (Jarman 1987; King et al. 2015a), so individuals were considered to be in the same group if they were within 10 m of at least one other group member. ‘Group size’ was the number of individuals aged 1+ years in a group. For each individual being tested for FID, we also classified the presence of its mother or of an offspring within the group as a factor: alone, i.e. neither a mother or offspring present; mother present (for 1-year-olds and 2-year-olds); small pouch young present (for adult females); medium pouch young present (for adult females); large pouch young present (for adult females); and 1-year-old offspring (for adult females). To avoid an extra parameter in the models, the rare cases of a female with both a pouch young and a yearling were considered within the class of their pouch young if carrying a large pouch young, but within the class of 1-year-old offspring if carrying a small or medium-stage pouch young. Subadults were tested with their mother in the same group in 31.4% of tests, but never at the same time. For adult females, 20.9% of the tests were with a small pouch young, 6.6% with a medium pouch young, 18.7% with a large pouch young, 9.4% were with a 1-year-old in the same group, and 44.3% were without an offspring in the same group.

Finally, to account for any habituation via multiple testing of FIDs, we recorded the test number and the number of days since the previous test for that individual, and included these variables in our models. Test number indicated whether it was the first, second, or n-th time an individual was being tested for FID. The number of days since the previous test accounted for the variation in time between trials that resulted from repeats both within and between years. As this could not be defined for the first test, the first value was set to an arbitrary value larger than the largest possible number of days since the previous test (500 days; use of a different value did not affect the results).  

FID measures were available on more adult females than just the mothers of the offspring considered in this analysis. We therefore analysed all available data to provide the best estimates of variation in adult female FID. Ninety-nine percent of subadult FID trials were carried out by one observer, WMC, however 9% percent of FID trials on adult females were carried out by another trained observer, so we included observer ID as a two-level factor in the model of adult female data.

(4) Subadult Survival      

We monitored subadult survival from 6 to 21 months for each individual, by which age nearly all young are weaned and have passed their second winter (King et al. 2017). Survival was measured for all four cohorts (2016-2019).

See Table 1 for sample sizes and summary statistics for all traits.

87

1.4 

(±0.6)

0.47 

(±0.24)

86

1.2 

(±0.5)

6.6

(±4.4)

Table 1. Means, standard deviations and sample sizes for all traits. Pouch Young Movement is a score from 0-1 (see Methods); Flight Initiation Distance is in meters (m); Subadults consists of individuals aged 13-35 months. Subadult Survival scores survival from 6 to 21 months (0 died, 1 survived).  Subadult Survival data includes but is not limited to individuals in the PY Movement and Subadult FID datasets.  The mean difference in age between siblings of the same mother was 1.7 (±0.7) years (range 1-3 years).

Ethical Note

Population monitoring and captures were undertaken with ethics approval from the Université de Sherbrooke (permit no.s MFB2016-01, MFB2020-01), the University of Melbourne (no. AEC 1312902.1) and the Australian National University (no. A2018/02) and research permits from the Victorian Department of Environment, Land, Water & Planning (no.s 10007062, 10008630).  Behavioral experiments and observations were conducted with animal ethics approval from the Australian National University (no.s A2017/17, A2018/02).

Statistical analysis

We fitted two multivariate generalized linear mixed models (GLMMs) using a Bayesian Monte Carlo Markov Chain framework in the statistical package MCMCglmm (Hadfield 2010). Code for the MCMCglmm models is provided in the Supplementary Material.  We used Model I to address question  1 (do siblings behave similarly to each other?) and question 2 (are either offspring early-life behavior or maternal behavior associated with offspring survival?). We used Model II for question 3 (what is the contribution of variation between mothers to the repeatability of offspring behavior?) and question 4 (are maternal and offspring behaviors correlated?).

Model I was a four-trait GLMM with response variables of PY Movement, Subadult FID, Adult Female FID (including mothers), and Subadult Survival. Although we had repeated measures of Subadult FID, to facilitate fitting of this complex four-trait model we averaged the multiple measures on each offspring into a single mean value. Each offspring was therefore represented by a single observation for each trait. However, each mother could be represented by multiple offspring, and also by multiple observations for her FID (average 5.6 FID observations per individual, range 1-14; Table 1). 

For all four traits, we fitted Mother ID as a random effect to estimate between-mother variances (or ‘maternal repeatability’) and covariances for offspring traits, and individual-level repeatability for Adult Female FID. Maternal repeatability measures the level of consistent differences between non-siblings’ behavior and indicates similarities amongst siblings: variance between mothers indicates covariance within mothers. For the offspring traits (PY Movement, Subadult FID, and Subadult Survival), the residual variance in the model represents within-mother between-offspring variance, or differences between siblings. We acknowledge that this approach may lead to some overestimation of maternal repeatability of Subadult FID in this model, and that an explicit model of repeated measures with error would have been preferable (see Ponzi et al 2018, Dingemanse et al 2021), but repeated measures of Subadult FID are considered in more detail in Model II below.

For PY Movement, we fitted fixed effects of age at capture (in months), sex, and year (a 3-level factor: 2017, 2018 and 2019). For Subadult FID and Subadult Survival, we fitted fixed effects of sex and year (year was a 3-level factor for FID and a 4-level factor for Subadult Survival). For Adult Female FID, we also included fixed effects of GPS location along an east-west axis to capture the spatial variation in the study area (Menário Costa 2021), group size, presence of an offspring of different stages (as defined above), test number, number of days since the previous test, year, and a two-level observer effect.

The MCMCglmm runs generated posterior distributions of parameter estimates, from which we could estimate posterior means (for fixed effect parameters), posterior modes (for variance-covariance parameters and ratios) and 95% highest posterior density credible intervals (CIs) for each parameter and also for any derived estimates. For each offspring trait, we estimated maternal repeatability by dividing the estimate of the variance between mothers by the total variance in the offspring traits (defined as the sum of all the variance components). For Adult Female FID, we estimated individual repeatability by dividing the between-individual variance by the total variance. Calculations were made on each sample of the posterior distribution to generate posterior distributions of the two repeatability estimates, from which we estimated posterior modes and 95% CIs. As Subadult Survival was a binary trait (0/1), we fitted it specifying the family ‘categorical’ in MCMCglmm. In a binomial model, residual effects have a variance fixed at 1, but can still covary with other random effects. Because parameter estimates from the Subadult Survival model were on the latent (logit) scale, we used the QGicc function in the package QGglmm (de Villemereuil et al. 2016; de Villemereuil 2018) to back-transform the latent-scale variance estimates and also to calculate repeatability on the original data scale as well as on the latent scale (see equations below).

We estimated the latent-scale repeatability in Subadult Survival as: (Equation 1), where ‘var_Residual’=1                    

The data-scale repeatability was estimated as: (Equation 2), where ‘var_Residual’ =1, and ‘var_Binomial sampling’ is the variance related to binomial sampling).

We used Model I (the four-trait model) to estimate the covariances between the four traits. In particular, we estimated maternal-level covariance between the Mother ID effects, for all behavioral responses, to test for maternal-level associations across offspring traits (e.g. do mothers whose offspring have high PY Movement also have offspring with high FID?), and for associations between maternal FID and offspring average behavior and Subadult Survival (e.g. do mothers with a high FID have offspring with high PY Movement, and do they have offspring with high survival rates?). We also fitted residual covariances for offspring traits to test for within-mother offspring-level associations (e.g. for an individual offspring, is high PY Movement associated with high FID?). For the random effect Mother ID, a covariance matrix between traits was set using  an unstructured  covariance structure in MCMCglmm (Hadfield 2017). We also fitted the covariance for the residual variance-covariance of the three offspring traits. We could not estimate residual covariance with the Adult Female FID because this was the only trait measured on adults, so we fixed this covariance to zero.

Model II was a bivariate model for Subadult FID and Adult Female FID, and was used to estimate the contribution of between-mother variation to the repeatability of juvenile behavior, and also the covariance between mother and offspring FID. This model included multiple measures on each offspring for Subadult FID (average 4.6 observations per individual, range 1-10; Table 1). For the response variable Subadult FID, we fitted random effects of Mother ID and Offspring ID, and for the response variable Adult Female FID, we fitted a random effect of Mother ID. For both traits, we fitted the fixed effects of GPS location along the east-west axis, group size, presence of the mother or offspring, test number, the number of days from the previous test, and year. Further, for Subadult FID, we fitted fixed effects of age at observation (in months, as a covariate) and sex, and for Adult Female FID, a two-level effect of observer.

 As in Model I, we fitted the covariance between the Mother ID effects across the behavioral responses. We fixed the covariance to zero for the residual effects, as there could be no covariance across the two traits in within-individual effects.

Both Models I and II were run for 1300 x 10³ iterations, with a burn-in of 300 x 10³, a thinning interval of 1 x 10³ and an inverse-Wishart prior distribution. We report the posterior distribution mode and 95% CIs for each parameter, considering there to be statistical support for the covariance of random effects if the 95% CIs did not overlap 0 and, for fixed effects, if pMCMC (Probability of Markov Chain Monte Carlo) was <0.05.

For Subadult FID in Model II, we estimated maternal repeatability by dividing the variance between mothers (Mother ID variance) by the total variance, and individual repeatability by dividing the variance between offspring (Offspring ID variance) by the total variance as above. For Adult Female FID, we estimated individual repeatability by dividing the variance between individuals (Mother ID variance) by the total variance. Model II only had Gaussian-family traits, so unlike Model I, we did not need to convert repeatability to the data scale in Model II.

In the Subadult FID response, we then estimated total repeatable variance between individuals from the variance between mothers plus the additional variance between offspring, and hence a total repeatability by dividing this sum by the total variance: (Equation 3)

Finally, we estimated the proportion of total repeatability (offspring plus maternal variance, as in a model without Mother ID the between-mother variance would be attributed to between-offspring variance) that could be ascribed to maternal-level effects as: (Equation 4)

Mother-offspring regression for FID

We did not have a sufficiently informative multi-generational pedigree to estimate heritability of FID from a quantitative genetic ‘animal model’ (Kruuk, 2004). Instead, we used the estimates of the phenotypic covariance between mothers’ and offspring’s FID to calculate the mother-offspring regression slope, and hence (having doubled the estimate; Falconer and Mackay 1996) to provide an upper limit on the heritability  (Falconer and Mackay 1996), noting that this is very likely to be inflated by maternal or shared environmental effects (Kruuk and Hadfield 2007; Lande and Price 1989): (Equation 5)

Note that both Models I and II provided estimates of the covariance between mothers and offspring in FID and the variance between mothers, and hence an estimate of the regression slope. We opted to use the estimates from Model II only to calculate , as it was a more detailed FID model, using the repeated measures for both mothers and offspring. We estimated the , as described above, for each sample of the posterior distribution, and report the resulting posterior mode and 95% CI.

Finally, we made use of the (rare) occurrence of adoption in kangaroos (King et al 2015a), and examined the association between mothers and offspring in 6 cases of adoption in 2016 and 2017: 4 involving reciprocal switches and 2 involving mothers whose biological young disappeared. All adoptions occurred when PYs were aged 8-10 months. The adoptive mother was responsible for all subsequent maternal care until weaning, about 6 – 9 months later (King et al 2015a). We repeated the analysis of the association between Adult Female FID and Subadult FID using observations on adopted individuals for which we had measured FID for the young and both the biological and adoptive mothers. Despite the very small sample size, this analysis could potentially indicate the relative importance of genetic and non-genetic causes of similarity in behavior.

Files can be accessed using Excel and R (R Studio).