Main

Understanding the range of past responses of human societies to disturbances is a global priority across the social and natural sciences and will support the development of solutions to future crises1,2,3. Numerous case studies have addressed past cultural collapse, transformation and persistence, although how to best characterize these processes is a subject for debate6. A major unresolved issue is the lack of comparability between cases of population resilience in the historical sciences4,5. Few studies explicitly model impacts, recovery and adaptation, or formally account for long-term history, which contains important and previously overlooked variation within and between cultural or environmental settings. Furthermore, a tendency to narrowly focus on responses to extreme events in both natural and social systems7,8 may overemphasize local or short-term adaptive success at the expense of understanding large-scale or long-term vulnerabilities6,9. A well-known example is the shift to a narrow marine diet among the Greenland Norse that initially offset the short-term risk of crop failure yet heightened societal vulnerability to longer-term North Atlantic cooling10. Here, we establish a global comparative approach to long-term resilience to identify the factors that structure the response of prehistoric populations to disturbances through time. The approach measures population capacity to withstand changes, as well as the rate of recovery following a disturbance through the common proxy of radiocarbon time–frequency data11,12. Disturbances are the inferred drivers of relative reductions in population or archaeological activity in prehistory, which are described variously as recessions, downturns, busts, negative deviations or similar12,13,14,15 and form the focus of this study, using summed probability distributions (SPDs) of calibrated radiocarbon dates. SPDs function as an index of relative levels of human activity, or population change, over time16,17. Population downturns are defined as periods when SPDs are significantly below expected growth trajectories in response to disturbances. Our efforts focus on two key questions: (1) how quickly do past populations recover after downturns; and (2) what factors mediate past resistance and resilience to downturns?.

To quantify patterns in population resistance–resilience, we performed a meta-analysis of 16 published study regions that used archaeological radiocarbon data to reconstruct regional palaeodemographic trends (Supplementary Table 1). Our approach trades specificity for a large-scale, comparative perspective that still accounts for variation between cases to focus on the emergent properties of the statistical analysis. Studies were reviewed based on three criteria: evidence for significant downturns, their scope and the inclusion of radiocarbon datasets. A lack of any single criterion resulted in the exclusion of a study. Cases with no reported downturns were not included, nor were those whose scope was restricted to specific activities within a regional radiocarbon dataset, such as flint mining18. Where published data have been superseded by later compilations, dates were added from the People3k database19, a systematic compilation of cleaned radiocarbon dates, based on the geographical area of the original study. Our global sample of regions ranges from the Arctic to the tropics and spans 30,000 years of history (Fig. 1a). Population downturns were reproduced using our protocol (Methods) and resistance–resilience metrics (Fig. 1b) were collected on 154 periods of population downturn, with a median of 8.5 periods in each region (Fig. 1c and Supplementary Tables 1 and 2). The numerical metrics collectively capture the severity, chronology and frequency of periods of statistically significant population downturn. Disturbances were classified into both general categories and specific drivers according to the original studies and expert interpretation of regional social, cultural and environmental history (Extended Data Table 1). The broad category of land use and evidence for adaptive change during, or in the wake of, downturns were also recorded. The focus of this meta-analysis is to identify the factors governing the relative depth of downturns (resistance) and the rate of recovery at their conclusion (resilience). A suite of hierarchical linear mixed models was developed to test for significant associations between parameters while controlling for regional variability (Methods).

Fig. 1: Map and summaries of all world regions and disturbance types included in this study.
figure 1

a, World map of study regions made with Natural Earth. Time ranges in calendar years before present (cal bp) and 14C dates used in this meta-analysis are indicated for each region. b, Conceptual diagram of measuring resistance–resilience on palaeodemographic downturns against expected growth trajectories. Downturn A is longer and faster, with low resistance and low relative rate of recovery (resilience). Downturn B is shorter and slower, displaying higher resistance and higher resilience. Other combinations of high or low resistance and/or resilience are possible (Extended Data Fig. 4). Parameters b, x and e for the equations in Table 1 are indicated in white dots. c, Summary of disturbance types within three broad categories reported in published palaeodemographic studies. Listed proxies are examples drawn from the surveyed literature. ‘Unclear’ downturns (downturns with a lack of clear evidence for a driver) are not shown (Extended Data Table 1).

This approach to past resistance–resilience provides a glimpse into population-level responses to disturbances throughout human history. Results demonstrate that a single factor—the frequency of downturns—increases both the ability to withstand disturbances and to recover from them. Additionally, the frequency of downturns experienced by a given population is influenced by land use: agricultural and agropastoral populations experience significantly more downturns over time than other land-use patterns recorded during downturns (Extended Data Table 1). These findings collectively suggest that the global shift to food-producing economies during the Holocene (starting 11,700 calendar years before present) may have increased population vulnerability to disturbances, yet in the process enhanced their adaptive capacity through repeated exposure. Parallels to our long-term perspective on human population change in macroecology suggest that comparative approaches in the historical sciences have the potential to generate profound insights into past human–environmental relationships on a broad scale.

Cross-sectionally high resilience

The most common high-level driver of downturns, after those with a lack of evidence for a specific cause (unclear, n = 65), is environmental (n = 48, 31%), followed by mixed (cultural–environmental) (n = 33, 21%). The most common disturbance type is aridity in the environmental category (n = 31), followed by mobility (n = 20) in the cultural category. Only a third of recorded downturns have resistance values that drop more than 50% from pre-downturn activity levels (n = 53, 34.4%; Fig. 2a). Most downturns (n = 133, 86%) end before baseline relative population levels are attained; in other words, observed SPD values are lower at the end of most downturns than at the pre-disturbance reference value. Resilience is relatively high across all cases (median = 0.64, n = 154), with 40% (n = 63) still attaining 90% of pre-downturn conditions by their end (Fig. 2b). Full returns to SPD baseline conditions are frequently interrupted by subsequent downturns. Downturns associated with cultural drivers return the highest median resilience (0.74), whereas the median mixed (cultural–environmental) resistance is highest at 0.79. Resistance (0.65) and resilience are lowest among climate-driven downturns (0.57). However, we do not find support for significant differences in either metric across disturbance categories (analysis of variance, resistance: F = 0.541, P = 0.65, d.f. = 3, n = 154; resilience: F = 0.04, P = 0.98, d.f. = 3, n = 154).

Fig. 2: Summaries of downturn numerical metrics showing the relationship between resistance, resilience and the duration and relative pace of downturns.
figure 2

a,b, Distributions of resistance (a) and resilience (b) across disturbance categories (Extended Data Table 1). The lower and upper hinges correspond to the 25th and 75th percentiles The upper and lower whiskers extend from the hinges to 1.5 × the interquartile range (IQR). The thick black line represents the group median. Data beyond the whiskers are individually plotted outlying points. The dashed line represents the combined data median. c, Distribution of the duration of downturns and the time to SPD minima across all recorded downturns. d, Relative pace (overall duration normalized by time to minimum) approximates a normal distribution and enables comparison of the speed of downturns. Modelled downturn durations skew towards multidecadal and centennial timescales.

Global variation in recovery rate

Initial modelling indicates that the geographical location of downturns does not affect resistance, with the exceptions of significantly higher values in the Caribbean archipelago and the Italian peninsula. Conversely, there is support for significantly lower values for resilience in three regions: the Central China Plains, the Caribbean archipelago and the Korean Peninsula, when compared with all other regions (Extended Data Table 2). Further examination of these cases reveals that a large minority of downturns in these three regions return negative values of resilience (n = 11, 39%), which are produced when the population proxy exceeds the SPD value at the start of the downturn by its end (Extended Data Fig. 4, 12). Although the SPD population proxy remains below modelled expectations in all of these cases and therefore they are, in the strictest sense, downturns relative to the null models, the results nevertheless imply that populations in these regions were, on average, able to recover faster than the norm. Owing to the observed range of variation and its potential impact on the metrics, study region was retained as a random effect variable in the mixed-effect models.

Long-term downturns are the norm

The durations of population downturns tend towards centennial (100–500 years, n = 48, 31%) and decadal periods (less than or equal to 50 years, n = 47, 30.5%), with a median of 98 years across the sample. Downturns lasting longer than 500 years are a minority (n = 29). The time taken to reach SPD minima skews further towards decadal timescales (Fig. 2c). Only a single downturn that commences with the 8.2-thousand-years-ago event in the Near East20 has a time to minimum longer than a millennium (2,070 years). Both variables have a strong positive skew (duration = 2.84, time to minimum = 4.23). To control for the distribution and broad range of variation in these variables, the time to minimum was normalized by downturn duration to produce an index of its relative speed, which we term ‘pace’ (Fig. 2d). This transformation enables comparison of the variation between downturns, with higher numbers reflecting slower downturns (σ = 0.55, s.d. = 0.23) and lacking support for non-normally distributed values as in the time to minimum and duration variables (Shapiro–Wilk W = 0.98396, P = 0.07108). Consequently, relative pace was employed as a fixed-effect candidate in the mixed-effect modelling.

Land use mediates resilience

The frequency of downturns over time by region was estimated by normalizing the cumulative number of downturns in a region by their duration (Table 1). We transformed this to the logarithm of events per millennium to account for its strongly skewed distribution. The variable allows us to compare the rate at which downturns occur. It consistently displays the strongest relationship to both resistance and resilience (P < 0.001 in both cases) (Extended Data Table 3) and is the only fixed variable retained by the information criterion-based selection procedure.

Table 1 Metrics and model parameters extracted from global case studies of past population downturns

Collectively, these results indicate that populations experience an enhanced ability to withstand disturbances as frequency increases, as well as to recover in the aftermath (Fig. 3a,b). Further examination of frequency of downturns shows that, among the recorded disturbance types, changes to mobility regimes (median frequency of downturns = 2.26, n = 20) and high environmental variability (median frequency of downturns = 2.19, n = 17) occur at the highest rate per millennium, whereas cooling occurs at the lowest rate (median frequency of downturns = 0.761, n = 4) (Fig. 3c). In terms of regional variation, the highest frequency of downturn is recorded in the South African Greater Cape Floristic Region (median frequency of downturns = 2.41, n = 17 over 9,950 years) and the lowest in the Korean Peninsula (median frequency of downturns = 0.58, n = 3 over 4,000 years).

Fig. 3: Effect sizes of model terms and dominant land use during downturns over time.
figure 3

a,b, Resilience (a) and resistance (b) are strongly influenced by the rate of downturn per millennium, with a stronger effect for resistance. Values are extracted from fitted models I and II. c, Standardized regression coefficients for significant (P < 0.05) mixed-model terms for n = 154 independent samples based on a two-sided test. Hunter-gatherer was set as the reference group for land use. d, Proportions of dominant land-use types during downturns, in 1,000-year time slices. Pleistocene downturns (before 11,000 calendar years before present, n = 24) have been combined. The 95% confidence intervals are indicated in ac by shaded bands and error bars. *P < 0.05, **P < 0.01 and ***P < 0.001.

Treating the frequency of downturns as a response variable (Methods) revealed that agricultural and agropastoralist land-use patterns are associated with significantly higher rates of downturn compared with low-level food production, marine foraging or mixed subsistence. Results from this additional modelling exercise suggest that the frequency of downturns is likely to have an important effect on resistance and recovery, whereas the frequency of downturns itself covaries with the dominant pattern of land use and disturbance type in a given time and place. A larger sample size of downturns would increase the explanatory power of our approach and enhance our characterization of these relationships.

Discussion

This meta-analysis has examined the potential factors influencing resistance and resilience across a broad archaeological sample, and was controlled for regional variation. The frequency of downturns is the main determinant of the observed outcomes among the examined factors. The relationship between resistance and resilience displays variable rates, although these events all tend to unfold at multidecadal to centennial timescales. Well-known historical examples support this finding: the catastrophic depopulation of indigenous groups across the Americas took place over centuries21, and the collapse of imperial power in Western Rome was preceded by a long period of rural depopulation22. Systematic data on independent population downturns in prehistory are less common. However, what data are available23,24 corroborate this result: palaeodemographic downturns resolved in radiocarbon data tend to last at least one human generation but frequently much longer.

The frequency of downturns is associated with both the ability of past populations to withstand downturns and the rate of recovery following them across a broad sample of human populations. The results suggest the existence of a common mechanism among human populations that confers resilience to disturbances. The size of this interaction is greater for resistance (eta-squared (η2) = 0.46, P < 0.001; Fig. 3b) than resilience (η2 = 0.29, P < 0.001; Fig. 3a). In practical terms, this suggests that the ability to withstand downturns is distinct from the ability to recover in their wake. We note that this result does not imply a monocausal or deterministic relationship; there are likely to be several adaptive pathways that increase population resistance and resilience by means of the mechanism of increasing downturn frequency.

Further testing indicates that land-use practices associated with food production, such as farming and herding, are significantly and positively correlated with the frequency of downturns (Fig. 3c and Extended Data Table 3). From the early Holocene, the proportion of land use associated with food production in our sample of downturns also increased, as the aggregate global population became gradually more reliant on domesticated species for meeting subsistence needs (Fig. 3d). Collectively, these trends show that although populations generally increased resistance and resilience over time, the heightened rate of downturn over time is itself likely to be linked to the historical tendency towards food-producing subsistence systems. Current archaeological evidence does not indicate that this process was unidirectional or inevitable; foraging and food production are neither mutually exclusive nor in opposition. Our synthetic findings agree with specific cases showing that the behavioural and social changes that food production entailed had trade-offs in other arenas23.

Traditional agricultural or agropastoral practices include a diverse range of socio-ecological systems that are often highlighted as potential sources of inspiration for solutions to current sustainability, biodiversity and conservation challenges5,10,25,26. The results suggest that population downturns or collapse are an inherent property of these systems and a potential trade-off of promoting their use. Systematic reviews in disturbance ecology indicate that frequent natural disturbances enhance the long-term resilience of key ecosystem services and that localized subsystem declines are an important mechanism through which this occurs27. Our study provides insight into the possible existence of an analogous relationship within our sample of human populations; higher frequencies are strongly correlated with both smaller downturns and closer matches to pre-downturn values in the SPD proxies. We suggest that humanity’s overall constant long-term population growth28 may have been sustained due, in part, to the emergent positive feedback between vulnerability, resistance and recovery documented here.

Population decline has been termed an ‘inevitable’ feature of our species’ demographic dynamics29. In their systematic review of historical collapse and resilience dynamics, Cumming and Peterson1 list depopulation as a key metric or factor in every ancient socio-ecological system studied. We anticipate that this singular status will continue undiminished. Our contribution indicates that downturns play an important role in human population history by enhancing the resilience of survivor populations. We speculate that the creation of biased cultural transmission may be responsible; downturns provide critical opportunities for landscape learning and the strengthening of local-to-regional knowledge networks to propagate through a cultural system30,31. Population downturns have been identified as potential triggers of labour investment in infrastructure, social cohesion and technological advancement15. Together, these mechanisms have the potential to enhance the preferential transmission of knowledge and practices that promote future resistance or resilience10. Raising population thresholds by intensifying land use may also heighten the risk of more serious collapse in return for increasingly marginal benefits1,24,32. Other non-trivial and historically contingent factors that are likely to affect outcomes are the diversity and ecology of domesticated species assemblages, degree and type of political complexity, and settlement patterning in relation to the environment. Indicators such as the cessation of monument construction, loss of literacy or economic turmoil can provide additional insight into the consequences and/or potential drivers of population downturns. These potential causal links must be rigorously tested, however, as they are not easily disentangled. A realistic model for the generative mechanism underlying the resilience of human populations will therefore have to be multiscalar and sensitive to cascading effects to account for how various exogenous impacts unfold and endogenous strategies are developed to solve them. The approach used for our comparative analysis of palaeodemographic resistance–resilience, however, does not distinguish between these elements of the studied populations. Future research may translate between our work and the microscale, the patterns of which are only truly understandable within the kind of generalist, synthetic frame of reference that we provide.

This study finds parallels in macroecology, where analogous resistance–resilience outcomes have been suggested to only fully resolve at centennial timescales or above33,34. Archaeology is uniquely suited to examining past population history, and the dynamics that underlie these trends, at such long-term timescales35,36,37. Understanding past societies’ responses to crises is often explicitly motivated by the goal of applying learnings from the historical sciences to present-day policy and activism, contributing to the ultimate objective of fostering resilient adaptations for the future6. Most archaeological work on past resilience is historically particularistic4,9 and emphasizes the contingencies, decisions and practices that underwrote successful adaptations in specific times and places38,39. This specificity can be illuminating but, if the historical sciences are to play any role in fostering future resilience, improving our understanding of the processes and drivers that influence long-term, centennial-scale resilience is a necessary prerequisite5. Our approach has highlighted the global relationship between population change and frequency of disturbance over millennial timescales and applies across a broad geographical and chronological sample of past populations, including prehistoric examples that have been overlooked in systematic reviews of societal resilience more broadly. Improved clarity on the drivers of exposure frequency and type in the past will help to reveal the mechanism(s) behind the dynamics we describe and their potential limits, which is particularly important as environmental variability is predicted to increase in the future40,41. Archaeological and historical case studies have focused on the frequency of volcanism, warfare42 and hydroclimatic oscillations23,43, as well as the rate of abrupt or extreme events in general8. Comparable evidence on different categories of downturns is necessary. Synthesis of these or similar data in a comparative framework can provide important insights into the causal links between population resilience, risk of exposure and, ultimately, the ability to recover.

Methods

Calibration and aggregation

Archaeological radiocarbon dates were collated from published studies that previously adopted null hypothesis significance testing (NHST) approaches towards prehistoric demography. Our literature search identified 16 studies that collectively span six continents and approximately 30,000 radiocarbon years (Supplementary Tables 1 and 318). We applied a consistent protocol to the calibration of radiocarbon assays. The calibrate function within the R package rcarbon17 was used to convert 14C radiocarbon years to calendar years before present. The IntCal20 (ref. 44) and SHCal20 (ref. 45) curves were applied to dates in the northern and southern hemispheres, respectively. Calibrated ages are reported as the 95.4% (two-sigma, 2σ) age range. Laboratory codes and site identification numbers were appended to each calibrated date range and postcalibration distributions were not normalized. To account for between-site variation in sampling intensity, several dates from a single site that are within 50 radiocarbon years of each other were pooled (‘binned’) before aggregation into regional SPDs.

Bayesian model fitting

Our protocol aimed to replicate the results of the 16 identified case studies to the greatest extent possible. To reproduce statistically significant negative population events (‘downturns’) produced by rcarbon’s ‘modelTest()’ function in the original studies, while simultaneously addressing the well-known limitations of using summed probability distributions in NHST, we adopted an alternative, Bayesian modelling approach. Markov Chain Monte Carlo (MCMC) methods implemented in the nimbleCarbon package46 for radiocarbon data can obtain robust parameter estimates, accounting for radiocarbon measurement errors and sampling error simultaneously. Previously, this has been a major drawback of NHST approaches to aggregated radiocarbon data, with several alternatives proposed in the literature47,48,49. Using the MCMC-derived parameter estimates in posterior predictive checks enabled us to detect periods where expected growth trajectories were lower than the fitted model parameters and which were more robust and accurate than least-squares regression approaches applied directly on SPDs. The protocol produced outputs that are analogous to those in previous NHST studies (Extended Data Fig. 1).

We analysed regional SPDs separately by fitting identical bounded exponential growth models to each dataset. This common model is defined by three parameters: growth rate (r) and boundaries (a and b). A weakly informative exponentially distributed prior (λ = 1/000.4) was used for r to capture a broad range of potential growth rates among the cases. Parameters a and b were adopted from the original studies. Markov chain traceplots (Extended Data Figs. 2 and 3) evaluate chain mixing alongside model convergence (Gelman-Rubin Ȓ) and effective sample size diagnostics (Extended Data Table 5). Three chains of 50,000 iterations were run per region, with a burn-in of 5,000 iterations and a thinning interval of 2. To ensure comparability of results with published studies that used a logistic growth model as a null hypothesis, regional datasets were subset based on expert judgement at documented palaeodemographic transitions and treated as two separate exponential growth models. Subsetting was only performed on the Near East and Italy, Sicily and Sardinia datasets. Downturns adjacent to transition points were removed from the sample to avoid including data points introduced by the subsetting. Posterior predictive checks were executed using samples of parameter estimates obtained by the MCMC approach to simulate and back-calibrate a number of radiocarbon dates equal to the sample size. The procedure was repeated 1,000 times to derive critical envelopes.

Resilience metrics

The resilience metrics target periods when empirical SPD curves are below the 90% confidence envelopes of the fitted models, according to the posterior predictive checks (periods termed ‘downturns’). Extraction was performed using a bespoke R function modified from Riris and De Souza12, which is available at ref. 50. The principal response metrics in our analysis are resistance and resilience (Extended Data Fig. 4). Respectively, these metrics quantify the normalized response to downturns and the rate of recovery relative to baseline conditions. Resistance is measured on SPDs using two parameters: SPD values at the start of a downturn (b) and at downturn minima (x), whereas resilience is measured across the entire period of decline until its end (e) relative to the minimum and baseline (Table 1). Resistance ranges between 0 and 1, indicating a 100% change from baseline to no change. Resilience ranges between −1 and 1, with 1 indicating full recovery by the end of the downturn. Negative values of resilience indicate that the baseline value has been exceeded, although remaining outside the expectations of the null model. Finally, zero indicates no recovery11,51. Variations in the shape of the SPD can result under comparable scenarios (that is, with similar demographic dynamics, archaeological sample sizes and disturbances) and hence can produce different results in the metrics because of calibration effects despite their similarity. Although this issue may contribute to noise and error in measurement, it is nevertheless highly unlikely to be systematic or to correlate with other variables of interest. Finally, we conservatively only consider events greater than ten years in duration for our statistical modelling. These events are at an elevated risk of being artefacts of the null model rather than true downturns with an unclear driver.

We also collected information on the start and ends of downturns, their duration, elapsed time until SPD minima were reached, the cumulative number of a downturn and the frequency of downturns (Supplementary Table 2). The frequency of downturns is calculated on a per downturn basis within study regions, using the cumulative number of the downturn, normalized by its duration. We report frequency of downturns for each downturn as a logarithm per millennium.

Statistical modelling

The target of our comparative analysis is the resistance and resilience of human populations to disturbance as defined in each individual study. Our approach assumed that high values reflect resilient populations that successfully reestablish growth regimes after periods of decline related to disturbance events. We also assumed that downturns are randomly distributed in time and geographical space. To account for the influence of interregional and interevent cultural variation on outcomes, we drew on expert judgement and close readings of the published literature to record the broad category and specific type of disturbance during downturns, as well as the dominant land-use type and the nature of resulting socio-cultural changes, if any (Extended Data Table 1). These variables provided a control on whether a given population within a cultural system retained its identity and function over time, or whether system transformation and adaptive change is archaeologically evident.

Linear mixed-effect models were executed to evaluate the presence and strength of relationships between resistance, resilience, the recorded variables and case study locations. This analysis was performed using the cAIC4 and lme4 R packages52,53; scripts are available at ref. 50. Initial models were defined with resistance and resilience as response variables, with only case identifiers (region) as a random effect. As observed downturns were sequential within each case, the random effect controlled for potential pseudoreplication and avoided the need to weight the data by group size. A stepwise search using Akaike’s information criterion was implemented for investigating the information gain of including fixed effects in each model in turn. These candidate models were sequentially fitted using restricted maximum likelihood. Most fixed covariates (Extended Data Table 3) were left out of the final models. Region was retained as a random effect in all cases, to produce two models:

$${\rm{resistance}}\approx \left(1{\rm{| }}{\rm{region}}\right)+{\rm{frequency}}\,{\rm{of}}\,{\rm{downturn}}$$
$${\rm{resilience}}\approx \left(1{\rm{| }}{\rm{region}}\right)+{\rm{frequency}}\,{\rm{of}}\,{\rm{downturn}}$$

Model output is summarized in Extended Data Table 3 and diagnostics are shown in Extended Data Figs. 5 and 6. We present standardized residuals, by region and in full, as well as leverage and Cook’s distance.

To further explore the relationship between rates of downturns and resistance and resilience, we performed an additional modelling exercise with the same random effect and full suite of fixed effects, with the frequency of downturn as the independent variable (Extended Data Table 4 and Extended Data Fig. 7).

$${\rm{frequency}}\,{\rm{of}}\,{\rm{downturn}}\approx \left(1{\rm{| }}{\rm{region}}\right)+{\rm{land}}\,{\rm{use}}+{\rm{change}}+{\rm{disturbance}}\,{\rm{type}}+{\rm{pace}}$$

The effect sizes (standardized coefficients) of the significant model terms are plotted graphically in Fig. 3c and reported in full in Extended Data Table 4. We report effect sizes in the text as η2, that is, the total variance explained by differences between means.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.