Explaining human conflict remains an important scientific initiative because it can elucidate causal relationships between environmental change and harmful societal outcomes such as homicide and war (Allen and Jones Reference Allen and Jones2014; Burke et al. Reference Burke, Hsiang and Miguel2015; Hsiang et al. Reference Hsiang, Burke and Miguel2013; Kennett et al. Reference Kennett, Masson, Lope, Serafin, George, Spencer and Hoggarth2022; Zhang et al. Reference Zhang, Brecke, Lee, Yuan-Qing and Zhang2007). As modern anthropogenic climate change threatens to destabilize global societies and ecosystems (Masson-Delmotte et al. Reference Masson-Delmotte, Zhai, Pirani, Connors, Péan, Chen and Goldfarb2021), research is needed to investigate links between socioenvironmental dynamics and violent conflict (Burke et al. Reference Burke, Hsiang and Miguel2015; Hsiang et al. Reference Hsiang, Burke and Miguel2013). In arid regions of the world, future environmental stress will likely result from increasing droughts driven by higher temperatures and decreasing precipitation (Cook et al. Reference Cook, Ault and Smerdon2015; Hoylman et al. Reference Hoylman, Bocinsky and Jencso2022; Lisonbee et al. Reference Lisonbee, Parker, Fleishman, Ford, Bocinsky, Follingstad and Frazier2025; MacDonald Reference MacDonald2010). Given the importance of drought as an environmental limitation in the southwestern United States, several studies have evaluated the response of past societies to changes in moisture (Codding et al. Reference Codding, Coltrain, Louderback, Vernon, Magargal, Yaworsky, Robinson, Brewer and Spangler2022; Finley et al. Reference Finley, Erick Robinson and Hora2020; Schwindt et al. Reference Schwindt, Bocinsky, Ortman, Glowacki, Varien and Kohler2016; Vernon et al. Reference Vernon, Yaworsky, Spangler, Brewer and Codding2022; Yaworsky et al. Reference Yaworsky, Vernon, McCool, Hart, Spangler, Codding, Kurt and Weston2024) and have shown that past droughts often correlate with social upheaval (e.g., Anderies and Hegmon Reference Anderies and Hegmon2011; Benson and Berry Reference Benson and Berry2009; Thomson and MacDonald Reference Thomson and MacDonald2020; Varien Reference Varien, Kohler, Varien and Wright2010). However, the often-tenuous links between macroclimatic trends and human conflict are increasingly debated (Kintigh and Ingram Reference Kintigh and Ingram2018; Masson-Delmotte et al. Reference Masson-Delmotte, Zhai, Pirani, Connors, Péan, Chen and Goldfarb2021), and there is a need to (1) quantify the role of direct limiting factors, (2) evaluate the moderating effects of endogenous social and demographic processes (Kohler et al. Reference Kohler, Ortman, Grundtisch, Fitzpatrick and Cole2014; Mach et al. Reference Mach, Kraan, Adger, Buhaug, Burke, Fearon and Field2019; Masson-Delmotte et al. Reference Masson-Delmotte, Zhai, Pirani, Connors, Péan, Chen and Goldfarb2021), and (3) test general theory to parse spurious correlations from causal relationships. By doing so, researchers can leverage known links between socioenvironmental conditions and violence to inform models predicting outcomes under various, future climate projections.
In the North American Southwest, archaeologists have put forth a concerted effort to describe violent conflict (Coltrain et al. Reference Coltrain, Janetski and Lewis2012; Kuckelman et al. Reference Kuckelman, Lightfoot and Martin2002; Lambert Reference Lambert2002; LeBlanc Reference LeBlanc1999; Martin et al. Reference Martin, Harrod and Fields2010; Snead and Allen Reference Snead and Allen (editors)2011; Solometo Reference Solometo, Arkush and Mark2006; Varien Reference Varien, Kohler, Varien and Wright2010; Wilcox and Haas Reference Wilcox and Haas1994) and formally test its causes (Cole Reference Cole, Timothy and Mark2012; Kohler et al. Reference Kohler, Ortman, Grundtisch, Fitzpatrick and Cole2014; Lekson Reference Lekson2002). Indeed, Southwest archaeologists have played a central role in developing an archaeology of warfare (Arkush and Allen Reference Arkush and Allen2006) and in increasing recognition of its importance for both elucidating patterns and formally testing theories. Regional scholars have also long recognized the significance of human–environment interactions and climate downturns in influencing past conflict, often in reference to broader phenomena such as the depopulation of the Colorado Plateau in the late thirteenth century AD (e.g., Coltrain et al. Reference Coltrain, Janetski and Lewis2012; Kohler et al. Reference Kohler, Varien and Wright (editors)2010; Lambert Reference Lambert2002; Lipe Reference Lipe1995; Matson et al. Reference Matson, William and Curewitz2015; Schwindt et al. Reference Schwindt, Bocinsky, Ortman, Glowacki, Varien and Kohler2016; Varien Reference Varien, Kohler, Varien and Wright2010). Yet, despite a growing compendium of valuable conflict studies, the causal links between exogenous climate change, endogenous socioecological systems, and variation in violent conflict remain poorly understood. To this end, we evaluate the socioclimatological drivers of intergroup conflict using time-series data from precontact farming populations in southeastern Utah.
Theoretical Framework
Applying theory from evolutionary ecology, we suggest that individuals should only participate in collective violence when the perceived benefits outweigh the anticipated costs (Allen et al. Reference Allen, Bettinger, Codding, Jones and Schwitalla2016; Glowacki et al. Reference Glowacki, Wilson and Wrangham2020). These costs and benefits are defined by how they are expected to influence an individual’s survivorship and access to social rewards (Ember and Ember Reference Ember and Ember1992; Glowacki et al. Reference Glowacki, Wilson and Wrangham2020; McCool, Codding, et al. Reference McCool, Codding, Vernon, Wilson, Yaworsky, Marwan and Kennett2022) and should not be confused with benefits to society. Intergroup conflict is then the aggregate outcome of individuals deciding that participation is worth the costs and risks. Consequently, we focus our theory on costs and benefits experienced by individuals and the way these scale up to group-level patterning visible in the archaeological record (see Bird and O’Connell Reference Bird and O’Connell2006, Reference Bird, O’Connell and Hodder2012; Codding and Bird Reference Codding and Bird2015; Winterhalder and Kennett Reference Winterhalder and Kennett2006). Synthesizing from the literature, we propose three resource attributes (distribution, predictability, and abundance) and one social attribute (inequality) that are likely to influence the costs and benefits of violent conflict (Allen et al. Reference Allen, Bettinger, Codding, Jones and Schwitalla2016; Dyson-Hudson and Smith Reference Dyson-Hudson and Smith1978; Ember and Ember Reference Ember and Ember1992; Glowacki and Wrangham Reference Glowacki and Wrangham2013; Glowacki et al. Reference Glowacki, Wilson and Wrangham2020; Hackman and Hruschka Reference Hackman and Hruschka2013; McCool and Codding Reference McCool and Codding2024; Smith and Price Reference Smith and Price1973; Winterhalder et al. Reference Winterhalder, Flora and Tucker1999).
Distribution and Predictability
The relative individual payoffs for resource defense or capture should be low when resources are homogenously distributed across the landscape, because defense costs are prohibitively high and the benefits are low given that alternatives abound (Dyson-Hudson and Smith Reference Dyson-Hudson and Smith1978; Wilson et al. Reference Wilson, Cole and Codding2023). Under these conditions, individuals should benefit from cooperative interactions, and defenses and conflict should be rare (Field Reference Field2008). In contrast, heterogeneously distributed resources that are clumped in time and space reduce the costs of defense or capture (Dyson-Hudson and Smith Reference Dyson-Hudson and Smith1978). The costs and benefits also vary, depending on the predictability of a resource. When individuals are unable to predict where or when resources are available, it limits the ability for contests to emerge. When resources are dense as well as highly predictable, however, there should be higher payoffs for resource defense (Dyson-Hudson and Smith Reference Dyson-Hudson and Smith1978; Smith and Codding Reference Smith and Codding2021; Wilson et al. Reference Wilson, Cole and Codding2023). Although distribution and predictability structure which resource types offer high payoffs for defense or capture, environments or economies with clustered and predictable resources vary considerably in frequencies of intergroup violence (Allen et al. Reference Allen, Bettinger, Codding, Jones and Schwitalla2016; McCool, Codding, et al. Reference McCool, Codding, Vernon, Wilson, Yaworsky, Marwan and Kennett2022), suggesting that these variables alone are insufficient explanatory factors.
Abundance
The benefits of resource capture that influence aggressive behavior should be structured by the relative value of a resource (Allen et al. Reference Allen, Bettinger, Codding, Jones and Schwitalla2016; Homer-Dixon Reference Homer-Dixon1994; Smith and Price Reference Smith and Price1973). Assuming the relationship between resource value and abundance follows a diminishing returns curve (Blurton Jones Reference Jones and Nicholas1987; Charnov Reference Charnov1976), then as key resources become increasingly scarce, the relative benefits of defending or capturing that resource increase. This also helps explain economic risk preferences, suggesting that individuals with low resource holdings will benefit from risk-prone economic strategies—including violence—whereas those with abundant resources should avoid high-risk conflicts and seek more cooperative interactions (Hackman and Hruschka Reference Hackman and Hruschka2013; McCool and Codding Reference McCool and Codding2024; Winterhalder et al. Reference Winterhalder, Flora and Tucker1999). Resource scarcity can arise on a population level owing to climate downturns or population pressure, for example (Kahl Reference Kahl1998; Kohler et al. Reference Kohler, Ortman, Grundtisch, Fitzpatrick and Cole2014; McCool, Vernon, et al. Reference McCool, Vernon, Yaworsky and Codding2022), or among individuals in subgroups due to inequality (Homer-Dixon Reference Homer-Dixon1994; McCool and Codding Reference McCool and Codding2024).
Inequality
From this framework, we predict that intergroup violence will be more likely when resources are scarce, clustered, and predictable in time and space. Because delayed-return economies produce clustered and predictable resources (e.g., crops), we assume that these conditions are satisfied and that the payoffs for violent conflict should vary depending on absolute and relative resource scarcity. In other words, the payoffs to violence should be mediated through the distribution of resources within a community. When inequality is present, resources may be unevenly distributed, which can induce deficits among individuals in disadvantaged groups or exacerbate existing shortfalls (Homer-Dixon Reference Homer-Dixon1994; McCool and Codding Reference McCool and Codding2024).
Synthesis
Although some case studies show straightforward links between these theoretical expectations and violent conflict (e.g., Ember and Ember Reference Ember and Ember1992; Hackman and Hruschka Reference Hackman and Hruschka2013; McCool and Codding Reference McCool and Codding2024), others highlight the complications that can arise (McCool, Vernon, et al. Reference McCool, Vernon, Yaworsky and Codding2022) given that individuals may respond adaptively to resource stress through strategies such as migrating to more productive areas (Eriksson et al. Reference Eriksson, Betti, Friend, Lycett, Singarayer, von Cramon-Taubadel, Valdes, Balloux and Manica2012; Kaczan and Orgill-Meyer Reference Kaczan and Orgill-Meyer2020) or in situ economic intensification (sensu Boserup Reference Boserup2014; Morgan Reference Morgan2015). Consequently, when socioenvironmental stressors increase but violence does not, it may be due to effective buffering mechanisms that prevent resource scarcity and keep the payoffs for violence low. Nonetheless, an exogenous shock may exceed the adaptive capacity of a socioecological system, and short-term adaptations such as migration and intensification may compromise long-term resilience as populations continue to grow, landscapes become circumscribed, and resource diversity declines (McCool et al. Reference McCool, Anderson, Ja’net Baide, Gonzalez, Codding, Kurt and Weston2024). In either case, supply may fall below a population’s starvation threshold or imperil social standing or economic mobility.
Although each of these factors may have independent effects on rates of violence, we expect that collective violence will peak when conditions interact (Homer-Dixon Reference Homer-Dixon1994). Furthermore, we may expect cyclical effects, given that violence itself (1) further reduces resource access and exacerbates inequality and mistrust (positive feedback; Ember and Ember Reference Ember and Ember1992; Glowacki and Wrangham Reference Glowacki and Wrangham2013) or (2) increases population mortality, which may dampen population pressure on local resources (negative feedback; Kohler et al. Reference Kohler, Ortman, Grundtisch, Fitzpatrick and Cole2014). As outlined in the next section, the key factors influencing the absolute and relative availability of resources in the study area are precipitation, temperature, climate shocks, population density, and inequality.
Socioecological Background
The study area is in the Bears Ears National Monument (BENM) in southeastern Utah (Figure 1). The environment is arid, high desert, with lower elevations defined by relatively warmer temperatures that favor pinyon-juniper woodlands and cooler upland settings that favor conifer forests. Canyon systems can support riparian zones along streams and springs, whereas upland mesa and montane areas are periodically dry, with moisture deriving from winter snowpack and the summer monsoon. This study area was selected because (1) there has been extensive prior archaeological research (Coltrain and Janetski Reference Coltrain and Janetski2019; Lipe and Matson Reference Lipe and Matson2007; Spangler et al. Reference Spangler, Yentsch and Green2010) that provides an existing comprehensive database, (2) the area offers long-term data on subsistence maize farmers living in a marginal and stochastic environment that underwent drastic and volatile climate change, and (3) this context may provide parallels for the millions of contemporary marginalized subsistence farmers living in challenging environments who lack regular access to markets and modern technology (Galani et al. Reference Galani, Orfila and Gong2022; Rapsomanikis Reference Rapsomanikis2015).
Map of the Bears Ears National Monument, Utah, with a digital elevation model (DEM), habitation sites plotted as red points, and study areas labeled. I made all of the figures in R, which is a free, open license software program that needs no accrediation for figures generated using it. (Color online)
Maize farming in and around the BENM is attributed to three archaeological complexes: Basketmaker (ca. 450 BC–AD 750), Fremont (ca. AD 350–1300), and Ancestral Puebloan (ca. AD 750–1300; Benson and Berry Reference Benson and Berry2009; Coltrain and Janetski Reference Coltrain and Janetski2019; Spangler et al. Reference Spangler, Yentsch and Green2010). Although maize was introduced to the region as early as 1150 BC, reliance on maize farming did not emerge until around 450 BC, at the onset of the Basketmaker complex in the Four Corners region (Benson and Berry Reference Benson and Berry2009; Codding et al. Reference Codding, Coltrain, Louderback, Vernon, Magargal, Yaworsky, Robinson, Brewer and Spangler2022; Coltrain and Janetski Reference Coltrain and Janetski2019). The subsequent Ancestral Puebloan complex is thought to have derived from the Basketmaker complex, and the transition is marked by important changes in material culture, including aboveground masonry blockhouses (Pueblos), kiva ceremonial structures, large storage units, greater sedentism, and investment in agricultural infrastructure (Lekson Reference Lekson2009; Spangler et al. Reference Spangler, Yentsch and Green2010). Maize farming was the principal subsistence strategy for roughly 2,400 years, until the thirteenth century AD, when Ancestral Puebloan and Fremont populations ceased farming and migrated out of the area (Benson and Berry Reference Benson and Berry2009).
The “Fremont” is a loose umbrella term for individuals in the western Colorado Plateau and eastern Great Basin who farmed and constructed extensive storage features (Madsen and Simms Reference Madsen and Simms1998). Unlike many Puebloan groups, the Fremont persisted in the use of seasonal foraging and dispersed pithouse residences, rarely concentrating in large aggregate villages. Fremont peoples manufactured a variety of ceramics and ground stone artifacts to process maize and wild plant resources (Metcalfe Reference Metcalfe1984). Although archaeologists have come to realize that Fremont adaptations are better defined by their variability rather than cross-regional patterning (Simms Reference Simms, Hemphill and Larsen1999), the Fremont people living just outside the study area can be roughly defined as a more dispersed, less politically complex population of farmers in comparison to their Puebloan neighbors. Given that the study area is composed of Basketmaker and Ancestral Puebloan groups, we focus on these archaeological complexes.
Resource abundance in the study area was largely a function of maize agricultural productivity (Benson and Berry Reference Benson and Berry2009; Codding et al. Reference Codding, Coltrain, Louderback, Vernon, Magargal, Yaworsky, Robinson, Brewer and Spangler2022; Coltrain and Janetski Reference Coltrain and Janetski2019; Matson Reference Matson2016), which is linked to precipitation and growing-season length, often measured in cumulative growing degree days (GDD; Schwindt et al. Reference Schwindt, Bocinsky, Ortman, Glowacki, Varien and Kohler2016; Yaworsky and Codding Reference Yaworsky and Codding2018). Variation in maize productivity results from interconnected spatiotemporal factors that affect moisture and temperature (Thomson et al. Reference Thomson, Balkovič, Krisztin and MacDonald2019). From a spatial perspective, elevation is the major determinate, with effective moisture increasing with elevation and growing-season length decreasing (Vernon et al. Reference Vernon, Yaworsky, McCool, Spangler, Brewer and Codding2024). Because of this, farmers can maximize returns by either (1) settling in low-elevation drainages with a long growing season and access to surface water for irrigation (Bocinsky and Kohler Reference Bocinsky2014; Boomgarden et al. Reference Boomgarden, Metcalfe and Simons2019), or (2) residing on patches that receive sufficient moisture while allowing for adequate GDD—often referred to as the dry farming niche (Bocinsky and Kohler Reference Bocinsky2014; Spangler et al. Reference Spangler, Yentsch and Green2010). The former strategy was predominant in northern latitudes where rainfall was limited, whereas the latter strategy predominated in more well-watered southern latitudes.
Irrespective of farming locations, diachronic changes in precipitation and temperature also impact maize productivity (Scheffer et al. Reference Scheffer, Egbert, Darcy Bird and Timothy2021; Thomson et al. Reference Thomson, Balkovič, Krisztin and MacDonald2019). Given that this article focuses on diachronic variation in conflict, we rely on spatially resolved temporal climate estimates as a proxy for resource abundance (see Materials and Methods section, below). Reduced precipitation coupled with increased evaporation rates due to high temperatures can result in water scarcity for irrigation and dry farming, thereby limiting the amount of water that can be allocated to crops and affecting growth, yield, and quality. Aridity can also render previously arable patches no longer viable. Cool temperatures can cause crop failure due to late spring frosts and can also reduce the amount of arable high-elevation land. Overall, as precipitation decreases, so does maize productivity, while with temperature productivity declines during hot or cold intervals.
Although precipitation and temperature influence annual productivity, farmers rely on past climate to anticipate future conditions and plan subsistence strategies (Kennett and Marwan Reference Kennett and Marwan2015). When local environmental conditions are highly variable, as they are in the Southwest, maize farmers may try to mitigate risk by adopting a suite of strategies to reduce crop harvest variance (Winterhalder et al. Reference Winterhalder, Flora and Tucker1999). However, when climatic trends are unpredictable it can be impossible to evaluate the relative costs and benefits of alternative strategies with productivity rapidly declining when there is a mismatch between farming strategies and local climate. Therefore, shocks in precipitation or temperature may negatively impact resource abundance even when long-term trends are favorable.
It is also necessary to differentiate the abundance of resources from their distribution. Population growth will increase demand from local resources (Boserup Reference Boserup2014; Weitzel and Codding Reference Weitzel and Codding2022), and inequality may lead to resource scarcity among subordinated groups, even when resources are abundant overall (Wilson and Codding Reference Wilson and Codding2020; Wilson et al. Reference Wilson, Cole and Codding2023). Although sociopolitical complexity was relatively muted in the study area, we present evidence of diachronic changes in inequality in response to variability in precipitation.
To summarize, absolute and relative resource scarcity and therefore violent conflict are expected to increase when local conditions are defined by (1) reduced rainfall, (2) especially hot or cold temperatures, (3) precipitation or temperature shocks, (4) population pressure, and (5) resource inequality. As noted earlier, the interaction of these factors may play a more important role than direct effects. If these conditions are present, but no markers of violent conflict are observed in the archaeological record, this may reveal effective resource buffering.
Materials and Methods
Archaeological Site Database
To capture variation in socioenvironmental conditions, site attribute data (site location and accessibility, number of rooms, room type, and habitation room area) was recorded throughout the BENM (Figure 1), including at (1) Elk Ridge to sample high-elevation locations, (2) Comb Wash to sample well-watered open basins, (3) multiple incised canyon complexes, and (4) Cedar Mesa to sample open mesas. Our database includes 216 habitation sites that range between one and 41 rooms (mean = 5.16), each with a relative chronological-period assignment that follows local conventions (Spangler et al. Reference Spangler, Yentsch and Green2010): Basketmaker II (450 BC–AD 500, BM2), Basketmaker III (AD 500–750, BM3), Pueblo I (AD 750–900, P1), Pueblo II (AD 900–1150, P2), and Pueblo III (AD 1150–1300, P3; see Supplementary Material 1 for database).
Site Defensibility
When landscapes offer naturally defensible localities, small groups tend to rely on these features to restrict access to habitation areas rather than more costly fortifications and other artificial constructions (Field Reference Field2008; LeBlanc Reference LeBlanc1999; Martindale and Supernant Reference Martindale and Supernant2009; McCool and Yaworsky Reference McCool, Yaworsky, Codding, Whitaker and Stevens2019; Solometo Reference Solometo, Arkush and Mark2006). Prior research shows that site accessibility tracks levels of conflict whereby communities relocate to naturally defensible locations as conflict intensifies (LeBlanc Reference LeBlanc1999; Martindale and Supernant Reference Martindale and Supernant2009; McCool and Yaworsky Reference McCool, Yaworsky, Codding, Whitaker and Stevens2019). Settlement defensibility can be a reliable proxy for conflict for several reasons. First, living defensively is costly, given that naturally defensible locations are rarely located in areas that minimize travel and transportation costs to and from resource areas such as farm plots or fresh water. Second, living in defensive locations can come with its own risks given that families may be exposed to dangerous natural features such as cliffs, as is common with defensive sites in the study area (Figure 2). Third, these locations often require households to cluster densely, which can increase disease transmission and reduce space for housing domestic animals or storing food. As a result of these difficulties, individuals will have major incentives to relocate to more accessible locations when the risk of violent encounters declines. Although the defensive location of a single habitation site may not be sufficient to infer region-wide conflict, an analysis of a large sample of habitation sites spanning a broad spatiotemporal context can allow inferences to be made about population-level rates of intergroup conflict. These arguments are further bolstered by studies showing that defensive settlement patterns predictably covary with high rates of violent trauma in bioarchaeological assemblages and with other markers of warfare (Kohler et al. Reference Kohler, Ortman, Grundtisch, Fitzpatrick and Cole2014; LeBlanc Reference LeBlanc1999; Lipe Reference Lipe1995; McCool, Vernon, et al. Reference McCool, Vernon, Yaworsky and Codding2022).
Photos of habitation sites in the study area exhibiting (top) low to (bottom) high levels of natural defensibility. Site trinomials from top left to bottom right: 42SA11767, 42SA5271, 42SA4295, 42SA256.
The location of habitation sites in the study area varies considerably from sites on flat and open terrain to totally inaccessible sites (Figure 2). Following previously established methods, we implement two quantitative measures of accessibility. The first, inspired by Martindale and Supernant (Reference Martindale and Supernant2009), measures the radial degrees (0–360) of access for each habitation site. For this method, a point of entry for each radial degree was deemed inaccessible if (1) the slope was >80° (nearly vertical) for >2 m and would therefore require climbing, or (2) there was a fortification feature that blocked a point of access. We then calculated the radial degrees of access from 0 (no points of access) to 360 (all points of entry are accessible) for all sites. The second measure, following McCool and Yaworsky (Reference McCool, Yaworsky, Codding, Whitaker and Stevens2019), estimates site defensibility using a computationally derived slope measure. To calculate this measure, we used an Arrow 100 Global Navigation Satellite System (GNSS) receiver to record the submeter location of each site. When sites were in locations that block satellite reception, a laser range finder was used to shoot in precise coordinates. The Terra package in R (Hijmans et al. Reference Hijmans2020) was used to generate 10 m buffers—capturing habitation site approaches—around each site and to calculate the mean slope for each buffer. Given the small area of sites in the study area, a 10 m buffer was sufficient to cover the landscape immediately surrounding the site core. When a habitation site contained occupations that span multiple landform types that differ in their natural defensive characteristics, multiple loci and corresponding GPS coordinates were created (Figure 3a).
Time series plots of (a) site accesability, (b) Gini index of inequality, (c) population estimate, and (d) climate estimates.
A Pearson’s r correlation test shows that the computational and ground-surveyed measures are tightly correlated (see Supplementary Material 1); therefore, we use the computational measure because of its reproducibility and applicability to other regions.
Variation in baseline topographic characteristics can bias natural defensibility estimates (Bocinsky Reference Bocinsky2014). This potential bias can be controlled for by including the topographic feature where sites are located into statistical models. However, this would exclude a critically important decision variable: when violent conflict increases, individuals living in highly accessible landscapes may opt to move to more naturally defensible landscapes rather than simply repositioning onto the most defensible landform in the area. Consequently, we evaluate model results without controlling for topographic variation due to the expectation that as conflict increases, more individuals will choose to move to geographic locations that provide superior natural defensibility. A plot of site frequency supports this decision, which shows considerable variation over time, suggesting large population shifts from more to less accessible landscapes (see Supplementary Material 1 figures ).
Finally, it is well recognized that groups in the Southwest aggregated in times of conflict to provide mutual protection and may have relied on large viewsheds or site intervisibility (LeBlanc Reference LeBlanc1999). Aggregate sites can bias accessibility measures of defense when larger, more formidable groups can afford to live in more convenient, less defensible locations (LeBlanc Reference LeBlanc1999; McCool and Yaworsky Reference McCool, Yaworsky, Codding, Whitaker and Stevens2019). However, our computationally derived accessibility measure strongly correlates with the total number of structures per site (see Supplementary Material 1), showing that large aggregate sites still relied on natural defensibility for protection. Site intervisibility has also been argued to be an important form of defense (Allen Reference Allen, Jim and Reycraft2008; Mullins Reference Mullins2016). However, intervisibility alone is an inadequate defensive strategy unless paired with aggregation or mechanisms that restrict access to a settlement. For instance, a site with a broad viewshed might allow early detection of approaching groups, but without natural or artificial defenses, such visibility is only beneficial when protection in numbers renders that group safe from confrontation or because the early warning enables flight to a defensible refuge. As stated above, aggregate sites in our study area occupy highly defensible locations, suggesting that intervisibility and aggregation alone are insufficient indicators of defense. In addition, many sites in the study area are situated in topographically restricted settings (e.g., deep, narrow canyons) that limit viewsheds and intervisibility but provide significant barriers to access. This pattern suggests that limiting access, rather than maximizing visibility, is the primary defensive concern.
Inequality
We measured the length and width in meters of the interior of each habitation feature for all sites in our database (nonhabitation structures are excluded). Surface area for each residential structure was then calculated in meters squared. Following Kohler and Higgins (Reference Kohler and Higgins2016), residential floor size was used to calculate the Gini coefficient, a number between 0 and 1 that demonstrates the distribution of wealth inequality within a group, where 0 represents total equality and 1 represents absolute inequality (Figure 3b). Gini coefficients are produced separately for each time period and therefore do not compare room areas between time periods when residential architectural traditions changed. Gini indexes offer the advantage of providing a single, standardized measure with broad spatiotemporal coverage that reduces the probability of sampling bias.
Population Density
Local population changes were estimated using the dates as data approach (Crema Reference Crema2022; Rick Reference Rick1987), which assumes that larger populations will deposit more datable materials into the archaeological, which will result in a greater frequency of dates. Consequently, a summation of all dates should result in lower summed probabilities during intervals of low population size, and in higher summed probabilities during periods of large population size. Dates were collected from the Canadian Archaeological Radiocarbon Database, including dates from Kelly et alia (Reference Kelly, Mackie, Robinson, Meyer, Berry, Boulanger and Codding2022), and the tree-ring-dates dataset compiled by Kohler and Bocinsky (Reference Kohler and Bocinsky2016). This involves generating 1,000 unique kernel density estimates (KDEs) for both radiocarbon and tree-ring dates. Dates were included if the age fell between 0 and AD 1300 and if the error range was 100 years or less. To account for intersite research bias, we thinned each sample, limiting the number of radiocarbon dates per site to 20 and the number of tree-ring dates per site to three. The bandwidth for each of the KDEs was set to 25 years to better display general trends that correspond roughly to human generation lengths (Figure 3c).
For the radiocarbon dates, the KDE is based on sampled years from the radiocarbon probability distributions, producing a composite KDE. For tree-ring dates, we generated KDEs using two methods. The first involves the use of the absolute date without errors, and the second randomly samples from a Poisson distribution (lambda = 12.5 years) to add a way of accounting for uncertainty between cutting and construction dates. Given that both approaches yield very similar results, we use the population estimate without a Poisson offset.
The KDE estimates produce two matrices, in which rows represent calendar years and columns represent KDEs. The matrices are then summed, yielding 1,000 population estimates for each calendar year. The resulting matrix provides an envelope of demographic trends that displays temporal intervals of high or low uncertainty. During model analyses, each row (or year) of the matrix is sampled to produce a single estimate of the relative population that year. The estimates for all years are then incorporated into the statistical model fit during each simulation as an estimate of change in population through time (see Supplementary Material 1).
Climate Change
Paleoclimate is measured using fine-grained climate reconstructions based on a method developed by (Bocinsky Reference Bocinsky2015; Bocinsky and Kohler Reference Bocinsky2014; Bocinsky et al. Reference Bocinsky, Rush, Keith and Timothy2016) and implemented using the SKOPE interface (Bocinsky et al. Reference Bocinsky, Rush, Keith and Timothy2016). This method regresses modern climate data on the complete tree-ring record for North America provided by the International Tree-Ring Data Bank using a high-dimensional regression and variable selection approach (Bocinsky and Kohler Reference Bocinsky2014; Bocinsky et al. Reference Bocinsky, Rush, Keith and Timothy2016) based on the well-known correlation between annual tree-ring width and annual temperature and precipitation. We use the resulting model to hindcast summer (May–September) maize-growing degree days (GDD, °F) and annual water-year (October–September) precipitation over the temporal sequence (see Supplementary Material 1 for details). Given that the SKOPE interface contains 800 m spatial block data, we select separate climate datasets for the two major study regions—Cedar Mesa and the surrounding canyons, and the Abajo Mountains—so that climate data are spatiotemporally resolved. Climate values are averaged across each spatial extent (Figure 3d). Climate shocks calculate the difference between estimates in precipitation (mm) and maize GDD (°F) for each calendar year and the average for the previous five years; we then calculate the standard deviation for the previous five years and divide the difference by the standard deviation to establish the five-year binned z-score. This tells us whether the precipitation and temperature during any given calendar year is higher or lower than the prior five-year average and what the magnitude of that change is (see Supplementary Material 1; Figure 3d).
Statistical Models
Our methods follow Wilson et alia (Reference Wilson, McCool, Brewer, Zamora-Wilson, Schryver, Lamson, Huggard, Coltrain, Contreras and Codding2022, Reference Wilson, McCool, Coltrain, Kurt and Weston2024) by coupling Monte Carlo resampling with a random forest (RF) machine-learning model to account for uncertainty in age estimates and interactions between predictor variables. Each simulation has six steps:
1. A random calendar date is drawn from a uniform distribution covering the assigned culture period under the assumption that the habitation site was equally likely to be occupied during any year in that range.
2. The precipitation, maize GDD, and shock values for the region in which the site occurs are pulled, and the values that correspond to the sampled occupation year are paired with the site.
3. A Gini coefficient is paired with the site using the site’s cultural time period.
4. We find the row in the population matrix associated with the sampled occupation year and then sample one of the 1,000 relative population estimates for that year and associate it with the site.
5. Steps 1 through 4 are repeated for each site during each iteration.
6. Fit the RF model.
An RF regression model of the accessibility estimates is built with all covariates using the randomForest package in R (Liaw and Wiener Reference Liaw and Wiener2002; R Core Team 2024). Given the sampling strategy used to fit individual decision trees (bootstrapping), the ensemble or aggregate of those trees—the “random forest”—will provide estimates of both direct and interaction effects. The iterations are repeated 1,000 times. We then combine predictions of all 1,000 model runs by calculating the median 50%, 2.5%, and 97.5% quantiles to estimate the central tendency with 95% confidence intervals. We then plot the partial dependence from each model along with interaction terms and variable importance (% increase in Mean Squared Error when the variable is permuted out of the model; see Supplementary Material 1).
Results
Our RF model compares the effects of precipitation, maize GDD, population density, and a Gini index of wealth inequality to our response variable—site defensibility—which is our proxy for conflict. Additional information about our data and measures can be found in the Materials and Methods section above. The results from the 1,000 RF models show several clear trends. First, maize GDD has a positive covariance with site defensibility—our proxy for conflict—indicating conflict was more prevalent during intervals when resources would be less abundant owing to excessively high temperatures (Figure 4a). Second, precipitation has a negative relationship with defensibility (Figure 4b), with conflict increasing during periods of reduced rainfall, which would have hindered maize productivity. Third, neither precipitation nor maize GDD shocks (defined as unpredictable precipitation or GDD) have strong effects on conflict. Indeed, a variable importance plot shows that permuting the climate shock variables out of the model improves model fit in over 50% of simulations (see Supplementary Material 1). Consequently, climate shocks are dropped from our final RF model. Fourth, population positively covaries with site defensibility (Figure 4c), with conflict rising during intervals of population growth and resultant competition. Fifth, a Gini index of inequality has a positive association with site defensibility (Figure 4d), with conflict increasing when resource inequality is higher.
Partial dependence plots from the Random Forest model (a–d): x-axis is the scaled predictor variables (z-scored to facilitate effect size comparisons), and y-axis is site defensibility (mean slope), our proxy for conflict. All the 1,000 random forest model runs are plotted (each as a gray line), with red illustrating the mean fit, and quantiles showing the range of 95% of modeled responses; (e–f) interaction plots (axes are z-scores); (g) variable importance plot showing the percent increase in mean standard error for each predictor variable. (Color online)
There are also important interactive effects (Figure 4e–f), whereby conflict peaks when climate is both dry and hot and during intervals when both population density and inequality are highest. The variable importance plot (Figure 4g) shows inequality to have the largest effect, followed by population density, maize GDD, and precipitation.
Model diagnostics for the 1,000 RF models indicate strong model performance, with a root-mean-square error (RMSE) of 18.175–19.966 (95% CI) and an R2 of 0.1444–0.297 (95% CI; see Supplementary Material 1 for details).
Discussion
Model results strongly support our theoretical framework, showing that conflict peaked during intervals of socioclimatologically driven resource scarcity and heterogeneity. Intergroup conflict resulted from a series of direct and interactive effects that include exogenous climate change and endogenous social, environmental, and demographic processes.
Absolute Resource Scarcity
Results show that conflict escalated during intervals of high temperatures (excessive maize GDD; Benson Reference Benson2011) and low precipitation (Figure 4). Maize GDD and precipitation also strongly interact, with atypical hot and dry conditions producing the highest levels of conflict. The lack of any climate shock effect suggests that climate downturns structured conflict more strongly than stochasticity. Aridity can cause resource scarcity directly through its impact on maize productivity, and indirectly through population reshuffling as groups depart habitats with insufficient moisture (Kaczan and Orgill-Meyer Reference Kaczan and Orgill-Meyer2020; Varien Reference Varien, Kohler, Varien and Wright2010; Vernon et al. Reference Vernon, Yaworsky, McCool, Spangler, Brewer and Codding2024). It can also favor intensification, a process that happened in later periods on the Colorado Plateau (Matson Reference Matson2016; Scheffer et al. Reference Scheffer, Egbert, Darcy Bird and Timothy2021; Spangler et al. Reference Spangler, Yentsch and Green2010). The lack of a pronounced climate shock effect may suggest that Puebloan strategies in BENM were effective for coping with sustained climate uncertainty, so long as conditions overall were adequately productive.
The direct correlation between climate and population growth is weak, though it may have initiated feedback mechanisms that influenced population aggregation during the P3 period, when, as conditions became more xeric, groups clustered into canyon complexes and other locations with available surface water (Matson Reference Matson2016; Schwindt et al. Reference Schwindt, Bocinsky, Ortman, Glowacki, Varien and Kohler2016; Spangler et al. Reference Spangler, Yentsch and Green2010). This pattern reveals interactions and feedbacks of human decisions and environmental realities, given that this transition may have been spurred by a desire to live in more defensible landscapes with access to surface water (LeBlanc Reference LeBlanc1999; Matson Reference Matson2016). Finally, compared to the endogenous demographic and inequality measures, precipitation and maize GDD have relatively weak effects on conflict. This may be due to dynamic relationships between climate and sociodemographic factors, whereby arid conditions did not result in increasing levels of conflict during earlier periods when population density and resource inequality were low.
Conflict peaked when increasing population growth was accompanied by a decreasing extent of occupied land, promoting competition over limited arable land. Although population aggregation can have positive effects via economies of scale, it appears that a threshold was reached when demographic dynamics transitioned from positive-sum to zero-sum interactions as populations pressured the marginal and spatially delimited resource base.
Relative Resource Scarcity
At the same time landscape-level productivity was declining, there was increasing inequality between occupied patches. Variation in inequality is significantly higher between habitation sites than within them (see Supplementary Material 1), suggesting inequality manifested as resource patch imbalance in the heterogenous Southwest landscape, with some individuals cooperating to territorialize productive patches and others living on marginal land. Conflict peaked during intervals when population and inequality were highest, suggesting that densely occupied habitats with pronounced ecological inequality promoted violent competition in a way that population or inequality alone did not. These results tie into our general theory and prior empirical results that show that low average resource holdings combined with inequality predict high rates of violent conflict (Allen et al. Reference Allen, Bettinger, Codding, Jones and Schwitalla2016; Daly Reference Daly2017; Ellyson et al. Reference Ellyson, Kohler and Cameron2019; Hackman and Hruschka Reference Hackman and Hruschka2013; Homer-Dixon Reference Homer-Dixon2010; McCool and Codding Reference McCool and Codding2024; McCool, Codding, et al. Reference McCool, Vernon, Yaworsky and Codding2022) and suggest that population pressure and inequality decrease a population’s resilience to climatic downturns. This result also supports prior research showing a link between inequality and conflict in the Southwest (Ellyson et al. Reference Ellyson, Kohler and Cameron2019; Kohler and Ellyson Reference Kohler, Ellyson, Timothy and Michael2018).
Cascading Effects
As a summary, we propose a plausible explanatory model to elucidate the general process by which socioclimatological conditions fostered intergroup conflict (Figure 5). First, drought conditions encouraged migration to well-watered canyon complexes (Matson Reference Matson2016). In these spatially limited yet water-available locations, people quickly reached carrying capacity and intensified maize production, driving population growth despite overall xeric conditions. These processes promoted population aggregation where, initially, Allee effects yielded payoffs for aggregation that later transitioned into scalar stress and zero-sum competition for scarce land and resources (LeBlanc Reference LeBlanc1999). Competition in heterogeneous habitats combined with population pressure to promote rising ecological inequality, given that some individuals restricted access to productive locations. This led to relative declines in resource availability, with those in marginal patches experiencing chronic stress. Individuals in disadvantaged groups then found violent encounters worth the risks, resulting in intergroup conflict. As a result of increased violence, people moved to more defensive locations that retained water access, which was crucial during hot and dry periods. The spatial limitations on water availability and potential social circumscription from competing groups restricted mobility and exchange, perhaps fostering mistrust (Ember and Ember Reference Ember and Ember1992; Lekson Reference Lekson2002), which further exacerbated resource scarcity—particularly for those in marginal locations.
An explanatory systems model illustrating demonstrated and probable links between exogenous and endogenous variables with possible feedback loops. (Color online)
This whole cycle then becomes a feedback loop whereby deteriorating climate conditions provoke even greater aggregation and inequality, which exacerbates resources scarcity and further promotes coalitional violence (e.g., Burke et al. Reference Burke, Hsiang and Miguel2015; Cappelli et al. Reference Cappelli, Conigliani, Costantini, Lelo, Markandya, Paglialunga and Sforna2020). Such a continued loop would then produce conditions that exceed the adaptive capacity of the local population and become one of important reasons why the Ancestral Puebloan groups decided to leave the Colorado Plateau at the end of the thirteenth century AD (Lipe Reference Lipe1995; Matson et al. Reference Matson, William and Curewitz2015; Varien Reference Varien, Kohler, Varien and Wright2010).
This study may have implications for modern subsistence farmers facing anthropogenic climate change. Vulnerable subsistence populations in arid regions of the world will likely face similar hardships as drought reduces the amount of quality arable land, populations continue to grow exponentially, and groups face circumscription due to political or sectarian borders. Although this article represents but one case study and additional tests of model expectations are needed, our theoretical framework and empirical results suggest that socioclimatological cascading effects may provide a model by which we can forecast similar outcomes when contemporary rural subsistence populations in marginal environments experience a comparable set of stressors. Nonetheless, caution is urged when applying these results to modern groups, particularly those who can rely on state-level or international interventions when resources become scarce.
Importantly, the intergroup conflict observed in the precontact BENM was the result of an unfortunate set of conditions, whereby exogenous factors well outside the control of local farmers—such as prolonged drought and habitat heterogeneity—combined with endogenous social, economic, and demographic conditions to create circumstances that were ripe for endemic conflict. Far from unique, these conditions have been shown to promote violence across modern and premodern populations (e.g., Allen et al. Reference Allen, Bettinger, Codding, Jones and Schwitalla2016; Burke et al. Reference Burke, Hsiang and Miguel2015; Kennett et al. Reference Kennett, Masson, Lope, Serafin, George, Spencer and Hoggarth2022; McCool, Vernon, et al. Reference McCool, Vernon, Yaworsky and Codding2022; Zhang et al. Reference Zhang, Brecke, Lee, Yuan-Qing and Zhang2007), including contemporary rates of homicide in post-industrial states (Daly Reference Daly2023; Gobaud et al. Reference Gobaud, Mehranbod, Dong, Dodington and Morrison2022; McCool and Codding Reference McCool and Codding2024). These cross-regional regularities strongly suggest underlying motivational principles, which we theorize to be conditional resource availability structured by interacting socioenvironmental forces.
Nonetheless, difficult conditions do not predetermine deleterious outcomes. During later periods, Puebloan groups living in modern-day Arizona and New Mexico experienced high population density and maize intensification that were not accompanied by spikes in conflict (Kohler et al. Reference Kohler, Ortman, Grundtisch, Fitzpatrick and Cole2014). Although this may be an outcome of divergent ecological and climatological circumstances, there is evidence that later Puebloan socioeconomic systems were developed to curb inequality (Ellyson et al. Reference Ellyson, Kohler and Cameron2019), create adequate resources stores and water management systems to avoid shortfalls after exogenous shocks (Duwe and Anschuetz Reference Duwe and Anschuetz2013), and foster internal political cohesion (Kohler et al. Reference Kohler, Ortman, Grundtisch, Fitzpatrick and Cole2014, Reference Kohler, Crabtree, Bocinsky and Hooper2015; Lipe Reference Lipe, William and Hegmon1989; Lipe and Matson Reference Lipe and Matson2007), all of which may have built resilience into the socioecological system that helped buffer against chronic resource stress and attendant conflict. Consequently, although chronic resource shortfalls strongly affect rates of violence in past and present societies, and not all conditions are within our control, socioeconomic institutions and practices can protect against harmful societal outcomes such as violence and war.
Although there is certainly additional work to be done on the causes of conflict in the Southwest, we hope this article has elucidated several key driving mechanisms. Future work will benefit from the publication of additional local instrumental climate archives; continued refinement of demographic dynamics; a richer understanding of the observed changes in inequality, social institutions, intra- and inter-regional exchange, and the types of violence that took place; and especially, the creation of additional direct dates on habitation sites to further refine spatiotemporal changes.
Acknowledgments
We thank Abby Baka, Ja’net Baide, Daniel Dalmas, Matt LoBiondo, Tyler Ferree, Hugh Radde, Izzy Osmundsen, Kate Magargal, Isaac Hart, August Bress, Dan Contreras, Kelsey Carlston, and Grant Thomas for fieldwork assistance; the BLM staff (Permit #22UT85199 R); members of the University of Utah Archaeological Center; Simon C. Brewer, Jerry D. Spangler, Mark Allen, and Kyle Bocinsky, and one anonymous reviewer. All photographs and figures are courtesy of the authors.
Funding Statement
This research was funded by the National Science Foundation (SBE SPRF grant #2104456).
Data Availability Statement
All data used in our analysis are available in the supplemental materials, with the exception of site location data, which—according to the Archaeological Resource Protection Act—cannot be shared with the public. Site location data will be made available to accredited researchers upon request.
Competing Interests
The authors declare none.
Supplementary Material
The supplementary material for this article can be found at https://doi.org/10.1017/aaq.2025.10107.
Supplemental Material 1. The dataset and R code used for our analysis. As per federal regulations, site location information has not been included in the supplementary file. For access to this data please contact the lead author (wmccool@calpoly.edu).
Open access
