Lifespan is a key life history trait that shapes ecological strategies and colony fitness in social insects, yet its drivers remain unclear. We evaluated whether intrinsic traits—especially worker body size, a principal pace-of-life axis, and phylogeny—predict survival across 18 ant species from five subfamilies monitored in the field and in captivity. Mark–recapture data analysed with hierarchical models accounting for imperfect detection showed that larger workers lived longer in both environments, highlighting intrinsic physiological properties as the strongest determinant of longevity. Worker lifespans ranged from 23–394 days, and field and lab survival were positively correlated. Body mass scaled negatively with colony size: small colonies invested in large, long-lived workers, whereas large colonies relied on smaller, short-lived workers. In captivity, survival was modestly higher for some clades and comparable for others, indicating mixed lab–field differences consistent with hazard reduction in certain lineages; Dolichoderinae and Myrmicinae showed little to no improvement. Phylogenetic signal was weak, indicating that lineage imposes little constraint beyond body size. These results confirm body size as a strong—and phylogeny as a weak—predictor of worker longevity. Laboratory assays capture relative field lifespans despite environmental complexity, based on the first comparative survival dataset for a local ant assemblage spanning multiple lineages.

Data collection

We conducted our study at the La Trobe Wildlife Sanctuary, a 28-hectare reserve on La Trobe University’s Melbourne campus in southeastern Australia (Fig. S1a). The sanctuary has a mesic, temperate climate, with mean annual minimum and maximum temperatures of 9.7 °C and 20.2 °C, respectively (Bureau of Meteorology, 1973–2017), and average rainfall of ~656 mm. During sampling, near-ground temperatures ranged from 5 °C to 40.5 °C. The site includes revegetated and remnant river red gum (Eucalyptus camaldulensis⁶²) woodland. Daily temperature and precipitation data were obtained from the nearby Bundoora weather station (BOM #086351, 2024). Ants typically inhabited low-productivity, sparsely vegetated areas with open ground, under large rocks or dead red gums. We sampled ants across microhabitats and nests, recording activity from 8 am to 7 pm throughout the study.

Field study design

Nest sampling

We sampled 18 species from 5 subfamilies and 10 genera (Table S1), covering 39 colonies. For each colony we selected 2-5 cohorts for survival estimation. Each cohort was selected at least one month after the previous selection (average 30 days). Multiple cohorts were used to minimize the effects of colony maturation. Each cohort consisted of 30–100 worker ants, ensuring uniformity in worker size and foraging tempo to minimize the influence of polymorphism. In genera with discrete major and minor worker castes genera such as Camponotus, only minor workers (typically ≤ 9 mg dry mass) —the primary foragers—were chosen to maintain consistency^{63, 64}. While this caste-foraging link is well-documented in some taxa, we acknowledge that foraging roles may vary in other polymorphic species, which could introduce bias.  Although this approach excludes other castes (e.g., majors involved in nest defence), we focused on foragers to enable cross-species comparisons. Cohort sizes reflected typical foraging activity, ranging from ~30–50 workers in some species to 100–500 in others (Supplementary File 1, Table S1). Workers were collected at the nest, ensuring they were active foragers. As in fire ants, foragers are typically older; majors begin foraging only in the final 25 % of adult life⁶⁶. Our estimates therefore reflect forager longevity, not full lifespan from eclosion.

Field workers were marked for identification as illustrated in Figure S1(b), with colours (green, purple, blue) assigned sequentially to new cohorts every three months. Laboratory workers were left unmarked because survival was recorded by directly counting all live individuals in closed containers, and preliminary trials indicated that applying paint caused unnecessary stress. To evaluate potential mark loss, we conducted a paint‑retention trial on 30 workers (10 each of Papyrius A, Camponotus consobrinus, and Meranoplus fenestratus) kept under laboratory conditions; no paint loss or colour fading was observed over their entire laboratory lifespans (≤ 4 months), indicating that mark loss is unlikely to bias field estimates. Same-species colonies were intentionally spaced (min. 58.5 m; all other minima ≥ 120 m; Fig. S1a; Table S6) to reduce between-colony movement during the field window.

Colonies were categorized by foraging rate: low (<0.5 workers/min; 30 ants), intermediate (0.5–5; 50 ants), and high (>5; 100 ants). Within cohorts, individuals were indistinguishable by markings. Biweekly 1-hour observations recorded forager counts and marked workers. Observation length varied by cohort and ended after 30 consecutive days without sightings or at day 400 (right-censored). Colonies that relocated were excluded for consistency.

Field assumptions

1. Ant cohorts not observed for a month were considered deceased.
2. Workers within marked cohorts were of similar age and were mature enough for foraging; cohorts were marked simultaneously to ensure age similarity.
3. Individuals within a colony were indistinguishable, as each cohort was marked with the same color.
4. Each colony was treated as a 'closed system' during the observation period, with no emigration or immigration, allowing changes in cohort size to be attributed solely to mortality or non-detection.

*Colony closure and spacing
*To minimise potential movement among same-species colonies during the recapture window, we deliberately spaced same-species colonies during site selection. GPS coordinates (Fig. S1a) show a minimum within-species inter-colony distance of 58.5 m (Camponotus claripes gp. A), with all other minima ≥ 120 m (Table S6). We did not observe cross-marked individuals at non-source colonies during field checks. This design reduces the likelihood that immigration/emigration among same-species colonies biased daily survival (φ) estimates.

Ant trait measurements

We examined how body mass and colony size affected survival. For each colony, we measured 10 ants and used mean dry mass for analysis. Foraging activity—used to stratify cohorts and model survival—was quantified by counting workers joining a trail over 60s (Table S1). To estimate colony size for each species, we employed Chapman's estimator⁶⁷, a refinement of the Lincoln-Petersen estimator, using mark-recapture data collected over multiple days of field observations. We selected Chapman’s estimator because it corrects for small-sample bias inherent in the classic Lincoln–Petersen method under closed-population, sampling-without-replacement assumptions^{68, 69}. Initially, ants were marked on Day 0, and the total number of ants marked was recorded. Recapture data were subsequently collected on the following days (Days > 0), documenting both the total number of ants observed and the number of marked ants recaptured.

Chapman's estimator was calculated for each observation day using the formula:

Chapman= (M + 1)(C + 1)/ (R + 1) - 1

where M is the total number of initially marked ants, C is the total number of ants observed on a recapture day, and R is the number of marked ants recaptured on a recapture day. Given our design of single-marking on Day 0 with independent daily recaptures, applying Chapman separately to each valid day and averaging results mitigates bias from variable and sometimes low daily recapture numbers. Recent simulations confirm Chapman’s superior bias and variance performance among two-visit estimators across a range of sample sizes⁷⁰. We included all observation days with ant activity (C > 0), excluding days with no activity (C = 0), as valid capture-recapture estimates require at least one observed individual. Colonies contained multiple cohorts. We calculated the Chapman estimator separately for each cohort using recapture data from active days, then averaged cohort estimates to produce a single colony size. This approach accounted for daily variation and improved robustness across species.

Laboratory study design

We conducted laboratory observations using the same colonies as the field study (Figure S1(c)). On the field marking day, one laboratory fragment of 30 workers was created from each source colony, yielding 38 fragments in total (1–3 per species, median = 2; Table S5). All fragments were maintained at 25 °C under a reduced light regime simulating nest conditions. Lights remained off except during feeding and hydration to minimize stress. Each species had two replicates of 30 workers, housed in 3.75-liter plastic containers (13 × 13 × 9.5 cm) coated with fluon and covered with mesh lids. Containers included an egg carton nest on a Plaster of Paris™ base⁷¹. Colony fragments were randomly positioned. Workers received ad libitum honey–water (50:50), dried insects, and saturated cotton for hydration. Laboratory room humidity tracked ambient outdoor conditions (Melbourne average daily RH 56–76%; Bureau of Meteorology, 1973–2017). Each 3.75 L container provided a humidity gradient: ants could occupy drier open areas away from the cotton-wool water source or seek more humid microclimates beneath the egg-carton nest. Water was replenished every 48 h. While this standardized setup was designed to mitigate extreme humidity stress and allow some behavioural regulation of moisture exposure, we acknowledge that it may not equally match the natural humidity or thermal preferences of all species, even though they were all collected from the same local site. Species-level differences in microhabitat use may still introduce survival bias (see Supplementary Table S1 for summary of nest types and temporal activity niches across species). Mortality was recorded daily, with deceased ants removed; to account for collection stress, death counts began two days after transfer, providing exact estimates of days to mortality for all 30 workers in each colony fragment. We therefore regard laboratory survivorship as an index or bioassay of intrinsic longevity under this standardized abiotic regime, not an absolute measure of each species’ maximal lifespan.

Survival analyses

Natural colonies: We employed a hierarchical modeling approach to estimate survival of individuals within natural ant colonies while controlling for imperfect detection. Survival probabilities, as unitless measures, represent the proportion of individuals expected to survive over a given interval, enabling approximate comparison across field and laboratory conditions while acknowledging differences in colony structure (e.g., queen presence) and detection methods. Our approach is a modification of the standard N-mixture model⁷², following Lyons et al. (2019), which contains two binomial models representing the apparent survival of an ant (state process) and the ability to detect an individual (observation process).

State Process:  S_ij ~ Binomial(H_i, Φ_i^days)

Observation Process: C_{ij ~~ Binomial(S}ij_{, d}ij~)

The number of ants surviving for the ith colony at time j (S_ij) was considered as having a binomial probability where the number of trials equates to the initial number of marked individuals (H_i) and the probability of success equates to the probability of survival to a given number of days (Φ_i)^days,. The number of marked individuals observed (C_ij) at each cohort (i) and at each observation period (j) was also considered a binomial process with the number of trials equating to the true, but unobservable number of individuals surviving for the ith cohort (S_ij) and the detection probability (d_ij) of an ant within the ith cohort during the jth visit. Because we followed multiple cohorts per colony and multiple colonies per species, we modeled species and cohort-within-species as random effects on daily survival. This lets species differ in their average survival while also capturing extra among-cohort variation within species. For each species we summarized the cohort effect size as a standard deviation (reported as cohortsig[i]) and also as an intra-class correlation (ICC), which expresses the share of (logit-scale) variance attributable to cohorts. We used weakly informative uniform (0–2) priors for the random-effect standard deviations. Detection was modeled as described above using the same environmental covariates.

Survival probabilities were assumed constant over time and modeled on a daily scale. We investigated differences in survival and detection probability among subfamilies, environmental, and behavioural variables (foraging activity, rain, minimum temperature) using a logit link, where log transformed colony size and mass were included as fixed effects on survival, and foraging activity, rain, and minimum temperature were included as fixed effects describing the detection process.

Our approach assumes that worker ants are exchangeable (survival and detection are the same) and not double counted during an observation period, and that survival and detection probability are homogenous among colonies or are accounted for via colony (species, families, etc.) or observation period (temperature, precipitation, etc.) specific covariates. 

Laboratory Colonies: We estimated survival (Φ_i) of marked individuals using a geometric model, which is a special case of the negative binomial distribution. In our analysis, the number of failures for the ith individual in colony j (C_ij) refers to the number of days an individual survived.  There is only 1 success event (mortality), and the probability of success (a mortality event) equaled the probability of mortality (1 - Φ_i).

C~ij ~~ Negative Binomial(1- Φ_i, 1)

We investigated differences in survival among species using the logit link. Specifically, we included each species as a fixed effect, along with the log transformed estimate of size of the colony they were collected from along with the log transformed estimate of mass. We controlled for multiple colonies within species using a normally distributed random effect.

Lab assumptions

1. It is assumed that resources such as food and water are consistently available and provided equally across all laboratory colonies, thus not influencing survival differentially.
2. The absence of a queen or brood has a limited effect on survival, although queenlessness can extend worker lifespan in some species^{74, 75}. This represents an important limitation when comparing laboratory and field estimates.

Model implementation and statistical framework

We produced separate subfamily‐level and species‐level models within a Bayesian hierarchical framework to estimate field and laboratory survival via JAGS (version 4.3.0, Plummer) interfaced with R (version 4.0) through the jagsUI package⁷⁷. Survival and detection probabilities were modeled using a logit link function. Covariates for survival and detection were fitted with vague, normally distributed priors (mean = 0, precision = 0.25) on the logit scale, following Northrup and Gerber (2018). The species random effect was modeled as normally distributed (mean = 0; precision equal to the inverse square of an uninformative uniform distribution on [0, 2]). Cohorts, nested within species, had random effects modeled similarly, with cohort variances drawn from uniform priors. In the lab survival model, colony was treated as a random effect, modeled as normal (mean = 0) with precision from a uniform prior on [0, 2]. Derived estimates of 10‐day survival were calculated for both field and lab settings, as this period adequately captured the uniform decline in survival rates, simplifying analysis and presentation. Derived quantities. In addition to daily and 10-day survival, we reported two cohort-level summaries per species: (i) the cohort random-effect SD (Table S8), and (ii) the cohort intra-class correlation (ICC) on the logit scale,     where is the logistic residual variance (Table S7). We visualized with a forest plot (Fig. S6).Posterior survival estimates and 95% credible intervals were used to assess survival and model coefficients. We ran three parallel MCMC chains of 1,000,000 iterations each, discarding the first 100,000 samples as burn‐in after a 5,000‐iteration adaptation phase. Every 10th sample was retained, yielding 270,000 posterior estimates across chains. Convergence and mixing were confirmed by visual inspection of trace and density plots and via the Gelman–Rubin diagnostic (all R̂ < 1.1). The species‐level model differed from the subfamily‐level model by using identity parameterization for 18 individual species rather than a species random effect, allowing species‐specific survival and detection effects. Derived survival estimates (daily) were calculated separately for each species, providing finer resolution than broader subfamily groupings. The species‐level model used identity parameterization for all 18 species, allowing species-specific survival and detection estimates, while the subfamily model pooled colonies into five taxonomic groups. To quantify information loss from pooling, we computed a two-way, absolute-agreement intraclass correlation coefficient (ICC[A,1]) on posterior mean daily survival values, treating each species as having two ratings: its own estimate and its subfamily mean. Concordance was low (ICC = 0.154, 95% CI = –0.341 to 0.574; F₁₇,₁₇.₂ = 1.35, p = 0.269; package irr⁷⁹. Despite this, subfamily pooling was retained to increase sample size tenfold and enable stable estimation of the continuous body mass effect. Including both species-level intercepts and body mass in the same model led to overdispersed posteriors and unstable trace plots. Full species-level survival estimates and 95% credible intervals are provided in Supplementary Table S3.

The subfamily model included five subfamily indicators and a continuous body mass effect under the fast–slow life history hypothesis. The species model used identity parameterisation (18 indicator variables) without mass to yield unadjusted species-level estimates for rank testing, phylogenetic analyses, and visualization. Due to strong collinearity (Spearman’s ρ = −0.73) between body mass and colony size, we retained mass alone in final survival models. In field and lab, mass was a continuous logit‑scale predictor entered additively with subfamily indicators (β₁–β₅) and random effects for species and cohorts. The β coefficients received vague normal priors (mean 0, precision 0.25), and posterior samples quantified effect size. Full model details are in Supplementary File 1.

Rank order

We used a Bayesian framework to evaluate the relationship between lab and field survival at the species level. Specifically, we obtained posterior samples of each species’ survival estimates from a Markov chain Monte Carlo (MCMC) analysis. At each iteration, we fit a linear model of field survival (y) as a function of lab survival (x):
Field survivalᵢ ∼ α + β × lab survivalᵢ
We extracted the slope (β), intercept (α), and R² at each iteration to generate posterior distributions. To test subfamily-level deviations from this relationship, we grouped species-level residuals by subfamily and performed a one-way ANOVA at each iteration, yielding posterior F-statistics and p-values. Analyses were conducted in R (v4.4.2), primarily using base functions and ggplot2.

Phylogenetic signal

We tested for phylogenetic signal in survival using Blomberg's K⁸⁰, defined relative to a Brownian motion model, which assumes trait variation accumulates with branch length. Although ecological traits may be shaped by adaptation, Brownian motion serves as a null for detecting deviations due to convergence or non-neutral dynamics⁸⁰–⁸². We used a time‐calibrated phylogeny from Economo et al. (2019), retaining one representative per genus. Study species were added as soft polytomies, and trimmed taxa were removed. Chelaner, absent from the original tree, was added as a tip, sister to Monomorium⁸⁴.
We extracted MCMC samples of species-level survival from the Bayesian model and calculated Blomberg's K at each iteration to account for uncertainty and obtain posterior distributions. We summarized these distributions by their mean and credible intervals, and used a likelihood ratio test at each iteration to assess departure from zero (no phylogenetic signal), with p-values summarized across iterations.
Given the limited power of our 18-species sample—below the recommended minimum of 21 tips (Blomberg et al. 2003)—we used a significance threshold of p < 0.10. The ‘phylosig’ function in the phytools package (v2.3)⁸⁵ was used for all K calculations. We also applied Blomberg’s K to species-mean body mass and to residuals from the field–lab survival relationship to test for phylogenetic structure in each.

Exploratory trait–mismatch check
To evaluate whether species’ traits could explain laboratory–field survival differences, we computed Spearman rank correlations between lab–field contrasts (daily and 10-day derived from daily) and colony size, foraging rate, worker mass, and thermal niche, applying Benjamini–Hochberg correction across tests (n = 18 species). Full outputs are reported in Table S5; these analyses were interpretive only and did not affect the primary model estimates.