In the highly climate- and ecology-sensitive eastern margin of the Qinghai–Tibet Plateau (QTP), understanding the adaptation behaviors of agro-pastoralists is crucial for reconfiguring human–climate–ecosystem interactions. However, existing studies often overlook the bounded rationality of micro-level decision-makers. Based on behavioral decision theory, this study constructs a “Perception–Capital–Adaptation” analytical framework. Utilizing micro-survey data from 890 agro-pastoralist households and employing Multinomial Logit (MNL) and Ordinary Least Squares (OLS) models, this paper systematically explores how environmental risk perception and livelihood capital are jointly associated with livelihood strategy choices and land-use behaviors. The findings reveal that (1) risk perception exhibits significant heterogeneity: sudden risks show a strong association with risk-avoidance transitions, whereas gradual risks often manifest as traditional livelihood lock-in due to “cognitive lag.” (2) The moderation of capital exhibits non-linear characteristics: physical and natural capital correspond to path dependence on existing production, while financial capital may be associated with expansionary behaviors, reflecting “maladaptation” in specific contexts. (3) Their interaction corresponds to a duality of adaptation pathways: a coordinated pathway (balancing ecological conservation and livestock reduction) and a conflictive pathway (maintaining production scale). Accordingly, a “risk-capital” trade-off matrix is constructed to identify four typical adaptation patterns: risk-avoidance transformation, path-dependent persistence, resilience-driven expansion, and fragile maintenance. This study demonstrates that climate adaptation essentially reflects the dynamic trade-offs made by boundedly rational actors between cognitive constraints and capital structures, providing a novel micro-behavioral perspective for avoiding maladaptation. 1. Introduction However, as climate and ecosystem uncertainties intensify, analytical paradigms that rely solely on objective environments or capital endowments struggle to fully explain the complexity of adaptation decisions. In the agro-pastoral ecotone on the eastern edge of the QTP, subtle fluctuations in climate and grassland conditions often trigger adjustments in livelihood strategies and the reconfiguration of land-use practices [ 17, 18]. In broader high-altitude regions, such as the Andes and the Alps, local communities’ interpretations of ecological signals like glacial melt are similarly profoundly influenced by socio-cultural boundaries and local knowledge [ 19, 20]; in the Koshi River basin of Nepal and the Ladakh region of India, altitudinal gradients not only affect resource accessibility but also shape households’ exposure to and perception of climate risks [ 21, 22]. This cross-regional evidence suggests that climate adaptation is not a purely technical “adjustment” [ 23, 24], but a contextualized behavioral process shaped jointly by cognitive and resource constraints. Furthermore, adaptive behaviors do not necessarily lead to positive outcomes. Existing studies have pointed out that under complex constraints, some coping strategies may result in “maladaptation,” which alleviates short-term shocks while increasing long-term system vulnerability or leads to risk transfer [ 24, 25]. For instance, in the mountain regions of Nepal, labor migration may relieve economic pressure but exacerbate the social vulnerability of left-behind groups [ 26], while in the sub-Himalayan region of India, the behavior of hedging risks through the over-input of production factors may trigger ecological degradation [ 27, 28]. These phenomena further highlight the necessity of understanding the adaptation process from a behavioral decision-making perspective. Consequently, this study constructs a “Perception–Capital–Adaptation” analytical framework based on behavioral decision theory, aiming to examine the heterogeneous adaptation behaviors of agro-pastoralists in response to climate and environmental risks. By integrating subjective risk perception and livelihood capital endowments, this framework systematically explores their intrinsic associations with livelihood strategy choices and land-use intensity (serving as a concrete manifestation of adaptive behavior). Utilizing micro-survey data from 890 households in the agro-pastoral ecotone on the eastern margin of the QTP, this paper employs Multinomial Logit (MNL) and Ordinary Least Squares (OLS) models to analyze how different types of risk perception, moderated by livelihood capital, are associated with and reconfigure the adaptation pathways of agro-pastoralists. It should be clarified that, due to the constraints of cross-sectional data, the focus of this study is on identifying the correlations and behavioral logic among variables rather than strict causal inference. The objectives of this study are first, to identify differences in how different types of environmental risk perceptions relate to agro-pastoralists’ livelihood strategy choices; second, to elucidate the moderating mechanisms of livelihood capital between risk perception and specific adaptation decisions; and third, to explore how the micro-level adaptive behaviors of agro-pastoralists reconfigure human–climate–ecosystem interactions through the adjustment of land-use patterns. 2.1.1. Conceptual Definitions Amidst intensifying climate change and escalating ecological uncertainty, the livelihood strategies of agro-pastoralists frequently deviate from the principle of absolute rational profit maximization. Instead, they manifest as boundedly rational choices constrained by cognitive capacity, information availability, and environmental conditions. Based on bounded rationality theory [ 38] and behavioral decision theory [ 29], individuals facing uncertainty do not directly make optimal decisions based on objective environmental changes; rather, they screen, interpret, and reconfigure environmental signals through subjective cognition, thereby forming differentiated behavioral responses. This study posits that agro-pastoralists’ responses to environmental changes are shaped not only by objective environmental conditions but also by the joint influence of subjective risk perception and livelihood capital constraints. In the complex environment where climate change and ecological degradation intertwine, agro-pastoralists’ livelihood decisions are not perfectly rational optimal choices, but rather a non-linear process dually constrained by cognitive biases and resource endowments. To thoroughly analyze the underlying mechanisms of this “Perception–Capital–Adaptation” framework, this paper introduces three key conceptual perspectives: First, status quo bias. In behavioral decision theory, status quo bias refers to individuals’ tendency to maintain existing livelihood paths when facing uncertainty, as they are more sensitive to the potential losses associated with changing the status quo, even if the expected returns of the current path are no longer optimal [ 29]. In this study, this mechanism primarily explains the “path dependence” of QTP agro-pastoralists on traditional animal husbandry. Driven by long-standing cultural identity, asset specificity (e.g., large herds and fixed shelters), and psychological uncertainty regarding non-pastoral transitions, agro-pastoralists exhibit pronounced livelihood stickiness, leading them to maintain the status quo even when environmental risks escalate. Second, Cognitive traps. Cognitive traps refer to the systematic misjudgments formed by individuals when processing complex environmental information, caused by incomplete information or psychological desensitization [ 32]. This study focuses this concept on the heterogeneity of agro-pastoralists’ risk perception, meaning that individuals often exhibit different perceptions toward sudden risks (e.g., extreme cold and snow disasters) versus long-term gradual risks (e.g., pasture degradation and phenological shifts). This perceptual asymmetry constitutes the core psychological barrier in our analysis of “divergent strategies under identical risks.” Third, Maladaptation. Maladaptation refers to the phenomenon where coping actions originally intended to improve short-term resilience inadvertently increase long-term system vulnerability [ 24]. It must be emphasized that this study treats “maladaptation” as a theoretical inference based on the existing literature and the ecological context. For instance, when households with abundant financial capital choose risk hedging, it may secure family livelihoods in the short term; theoretically, however, this behavior might exacerbate the carrying capacity pressure on the fragile alpine ecosystem, thereby evolving into negative externalities for overall ecological security over the long term. Accordingly, this paper proposes the concept of the “duality” of adaptation behavior to characterize how, under the joint influence of bounded rationality and resource constraints, adaptation behaviors may simultaneously exhibit trajectories toward both “proactive adaptation” and “maladaptation.” In other words, identical risk perceptions, under varying capital endowment conditions, may lead agro-pastoralists to adopt entirely different livelihood strategies. This duality constitutes a crucial theoretical perspective for understanding the heterogeneity of agro-pastoralists’ adaptation pathways. 2.1.2. Research Hypotheses In this framework, risk perception is regarded as a crucial antecedent driving behavioral adjustments. Existing research demonstrates that when individuals perceive higher environmental risks, they are more inclined to adjust existing production methods or seek alternative livelihood pathways [ 35]. Therefore, hypothesis H 1 is proposed: Risk perception is a crucial determinant influencing agro-pastoralists’ livelihood strategy choices, and different types of environmental risk perception significantly alter their response tendencies toward livelihood transition. Livelihood capital constitutes a major resource constraint in agro-pastoralists’ adaptation decisions, and capital accumulation not only dictates the capacity to cope with shocks but also influences the likelihood of breaking away from path dependence [ 13]. Therefore, hypothesis H 2 is proposed: Livelihood capital constitutes a major resource constraint in agro-pastoralists’ adaptation decisions, and varying capital endowments dictate their capacity to break away from traditional livelihood path dependence. Furthermore, based on the “duality” of adaptation behaviors, livelihood capital may exert oppositely directed moderating effects between risk perception and adaptive behavior, with its impact contingent on the capital structure. On the one hand, generalized capital (e.g., human and financial capital) can enhance information acquisition and risk-coping capacities, thereby facilitating proactive adaptation. On the other hand, specialized physical capital (e.g., large herds and fixed infrastructure) may reinforce path dependence and induce cognitive traps due to sunk costs and status quo bias, potentially leading to maladaptation. Therefore, competing hypotheses are proposed. H 3a: Livelihood capital (particularly generalized capital) exerts a positive moderating effect between risk perception and adaptive behavior; higher endowments facilitate the translation of risk perception into proactive adaptation strategies. H 3b: Livelihood capital (particularly specialized physical capital) may also exert a negative moderating effect between risk perception and adaptive behavior; higher endowments might reinforce the inclination to maintain the status quo due to sunk costs and path dependence. Compared to existing studies, this paper primarily presents three contributions: First, at the theoretical level, by constructing the “Perception–Capital–Adaptation” framework and introducing behavioral decision theory into sustainable livelihood research, it broadens the analytical paradigm predominantly driven by the objective environment. Second, at the empirical level, based on large-sample micro-data from an alpine ecologically fragile region, it identifies the differentiated impact characteristics of risk perception. Third, at the analytical level, it proposes a “risk–capital” trade-off mechanism and reveals the heterogeneity of adaptation pathways under different capital endowment conditions. These results provide a novel behavioral perspective for understanding the evolution of human–environment systems in fragile alpine regions and offer a theoretical foundation for formulating differentiated adaptation policies. Based on the above theories and conceptual definitions, this study constructs a Perception–Capital–Adaptation analytical framework. To visually map out the research hypotheses and theoretical derivation process, we present this framework as a Pathway Selection Diagram of Adaptation Behavior (). Rooted in Behavioral Decision Theory, this framework systematically elucidates an intrinsic decision-making mechanism characterized by “cognitive filtering—resource constraints—behavioral response.” Within this mechanism, objective environmental signals—such as climate change and grassland ecological evolution—do not directly trigger behavioral changes. Instead, they first enter the micro-level decision-making system through the subjective risk perception of agro-pastoralists, thereby completing the “cognitive filtering” process. Subsequently, livelihood capital endowments—comprising natural, human, physical, social, and financial capital—establish the resource constraint boundaries that define the feasible set of adaptation options available to these households. Ultimately, the interaction between risk perception signals and livelihood capital endowments shapes the heterogeneous adaptation behaviors of agro-pastoralists, which are specifically manifested in their adjustment of livelihood strategies and the reconfiguration of land-use intensity. By integrating psychological cognition with a material foundation, this framework aims to reveal how agro-pastoralists navigate complex decision-making trade-offs under the dual influence of bounded rationality and resource constraints when responding to climate risks. 2.2. Variable Construction To better explore the adaptation pathways of agro-pastoralists under risk perception and capital constraints, this paper constructs a system of dependent variables, core explanatory variables, and control variables centered around livelihood strategy choices and land-use behaviors. The definitions and measurement methods for each variable are as follows. 2.2.1. Independent Variables (1) Risk Perception Regarding variable processing, preliminary tests yielded relatively low Cronbach’s Alpha coefficients for the predefined dimensions (0.302 and 0.241, respectively). This indicates that risk perceptions among agro-pastoralists on the QTP do not behave as “reflective” indicators driven by a single underlying psychological trait; rather, they form a multidimensional cognitive structure characterized by high heterogeneity and independence. Consequently, rather than employing a simple equal-weight aggregation method, this study applies Principal Component Analysis (PCA) combined with Varimax Rotation. Addressing the borderline Kaiser-Meyer-Olkin (KMO) value of 0.508, and acknowledging the objective weak correlations among the indicators, this study adopts a more prudent methodological position. Specifically, PCA is strictly characterized as a pragmatic dimensionality reduction tool, rather than a method for identifying underlying latent variables. It should be noted that compared to traditional socio-economic variables, agro-pastoralists’ risk perceptions exhibit stronger context dependence and typological heterogeneity, meaning different risk dimensions do not necessarily demonstrate high correlations. The relatively low KMO value herein reflects the inherent “weak correlation feature” of the risk perception structure, which aligns with the reality of fragile alpine ecosystems where compound risks coexist and are mutually independent. Therefore, this study explicitly excludes any theoretical interpretation based on latent variables; the PCA procedure is utilized solely to mathematically transform the multidimensional perception indicators into several uncorrelated composite indices. This approach simplifies the complexity of the empirical models while identifying key structural features of risk cognition. Following the extraction principle of eigenvalues greater than 1 and the rotated factor loading matrix, four common factors were extracted: winter cold and snow disaster perception, climate warming and drying trend perception, extreme disaster frequency perception, and ecological phenology change perception. These four factors cumulatively explain 64.16% of the variance (see Table 2). Specifically, winter disaster perception reflects agro-pastoralists’ awareness of extreme winter cold and snow disaster risks, which constitute the most fundamental and traditional survival risks in pastoral areas. Warming and drying trend perception captures individuals’ holistic understanding of long-term climate warming. Disaster frequency perception reflects the awareness of the occurrence frequency of natural and geological disasters. Lastly, ecological phenology perception reveals individuals’ acuity towards crucial ecological timing events and physical functional changes within the ecosystem. (2) Livelihood Capital As a key moderating variable, livelihood capital was constructed based on the Sustainable Livelihoods Approach (SLA), comprising five dimensions: natural, human, physical, social, and financial capital (see Table A1). Each dimension is measured by several specific indicators (e.g., pasture area, labor force size, livestock inventory, social network, and income levels), which are standardized and aggregated to form composite indices. In the baseline model, the equal weighting method was adopted to construct the livelihood capital index to maintain theoretical consistency with the DFID framework, which assumes the equality and systemic nature of the five capitals [ 13, 42]. The calculation procedures are as follows. First, following the standardization of tertiary indicator data, the secondary indicator index is calculated using the arithmetic mean method: M d = ∑ i = 1 n i n d e x s d i n (1) Second, upon obtaining the secondary indicator indices, the identical aggregation method is applied to compute the indices for the five major types of livelihood capital: C d = ∑ i = 1 N M d i N (2) where C d represents the primary livelihood capital index for a specific dimension (e.g., human capital index), M d i denotes the index value of the i -th secondary indicator under that dimension, and N represents the total number of secondary indicators encompassed by that dimension. In the baseline regressions, this study utilized the equal weighting method to construct the comprehensive livelihood capital and dimensional capital indices of agro-pastoralists, aiming to preserve the intuitive economic interpretability of each indicator. Simultaneously, to avoid potential biases introduced by a single subjective weighting method, this study further employed PCA during the robustness check phase to objectively reduce dimensionality and re-weight the internal indicators of the five livelihood capitals. Through min–max normalization, these were mapped onto a 0–1 interval and utilized as alternative core independent variables to verify the reliability of the model’s conclusions. (3) Interaction Terms To test the bidirectional moderating role of livelihood capital between risk perception and adaptation behavior, this study, drawing upon the existing literature and insights from field surveys, purposefully selects and constructs two representative interaction terms. This aims to precisely validate the resource constraint mechanisms under risks of different natures. First, the interaction between sudden risk perception and financial capital. Winter cold and snow disasters represent typical sudden and severe risks. Agro-pastoralists’ immediate coping strategies for such risks (e.g., emergency purchase of external forage, reinforcement of disaster-prevention facilities, and post-disaster livelihood turnover) are highly dependent on the household’s financial liquidity and risk-sharing instruments (such as agricultural and pastoral insurance). Therefore, financial capital in this context represents not merely economic accumulation, but more importantly, the household’s “disaster-avoidance capital” and immediate buffering capacity. The introduction of this interaction term aims to verify how financial resources alter agro-pastoralists’ adaptation pathways under risk shocks. Second, the interaction between gradual risk perception and natural capital. Grassland phenological changes and degradation are long-term gradual risks. The core impact of such risks on livelihoods lies in the prolonged erosion of the quality and scale of production carriers (i.e., grasslands and cultivated lands). Agro-pastoralists’ path dependence on traditional livelihood models, or their momentum to transition toward non-pastoral sectors, is constrained by the structural lock-in of their land resource endowments (i.e., natural capital). This interaction term is introduced to examine how the resource base functions as a “livelihood baseline” in exerting a moderating effect during long-term ecological evolution. In summary, by comprehensively considering the heterogeneity of risk perception (sudden vs. gradual) and capital functions (immediate disaster avoidance vs. production baseline) in the design of the interaction terms, this study not only echoes the preceding theoretical framework but also effectively avoids statistical redundancy caused by indiscriminate variable combinations. 2.2.2. Dependent Variables The dependent variables in this paper primarily reflect agro-pastoralists’ adaptation behavior, measured across two dimensions: livelihood strategy choice and land-use behavior. Second, in the land-use dimension, given the reality of pastureland transfer and communal grazing on the eastern edge of the QTP, the sample includes 398 households that do not hold contracted pastureland under their names but still maintain substantial livestock scales. If “grazing intensity” (stocking rate per unit area) were utilized as the sole dependent variable, it would not only result in significant sample attrition due to a zero denominator but also potentially trigger Sample Selection Bias. Therefore, a dual dependent variable system comprising grazing intensity and livestock scale was selected to reflect the actual impact of agro-pastoralists’ production behaviors on the ecosystem, utilized for OLS model analysis. These indicators can characterize differences in adaptation pathways at the behavioral outcome level, providing empirical evidence for identifying potential “maladaptation”; that is, when production inputs continuously intensify while ecological pressure rises, it may reflect a negative adaptation outcome. 2.2.3. Control Variables To mitigate omitted variable bias, this paper introduces multiple control variables at both the individual and household levels. At the individual level, variables such as the household head’s age and highest education level are selected to control for the impact of human capital differences on decision-making behavior. At the household level, variables such as household labor force size and household income level (logarithm) are incorporated to reflect the household’s resource base. These control variables help enhance the robustness of the model estimates, allowing the impacts of core variables to be identified on a relatively consistent comparative basis. 2.3.1. Multinomial Logit Model A MNL model was constructed to examine the direction and intensity of the impact of risk perception on agro-pastoralists’ livelihood strategy choices. Because livelihood strategy types (agro-pastoral type, mixed type, off-farm type) are multi-categorical discrete variables with no natural ordinal relationship, the MNL model is suitable for this decision analysis. Unlike traditional methods that treat risk perception as a single aggregated indicator, this paper splits risk perception into multiple independent dimensions based on PCA results and incorporates them into the model separately. This approach enables the identification of the differentiated impacts of various risk cognitions on livelihood strategy choices, thereby circumventing information loss caused by indicator aggregation. Concurrently, interaction terms between risk perception and livelihood capital are introduced to characterize the “cognition–resource” coupling mechanism. Under this framework, risk perception not only directly influences livelihood strategy choices but also has its effect strength constrained or amplified by capital endowments, thereby shaping agro-pastoralists’ adaptation pathways. Let the probability of a household choosing the j-th livelihood strategy be: P i j = e x p ( U i j ) ∑ k = 0 J e x p ( U i k ) (3) where V i j is the utility function for an individual selecting strategy j. With the agro-pastoral type (j = 0) as the reference category, the relative utility function is expressed as: V i j = α j + ∑ k = 1 4 β k j R P k i + ∑ m γ m j C a p i t a l m i + ∑ k = 1 4 δ k j ( R P k i ୍ଠ C i ) + θ j Z i + ε i j (4) where R P k i ( k = 1 , 2 , 3 , 4 ) represents the four risk perception factors extracted via PCA; C a p i t a l m i represents the five types of livelihood capital variables; C i represents the comprehensive livelihood capital index (used to construct the interaction term); R P k i ୍ଠ C i is the interaction term; Z i represents the control variables; and v a r e p s i l o n i j is the random disturbance term. 2.3.2. Ordinary Least Squares (OLS) Model To analyze the specific production response behaviors of agro-pastoralists to environmental risks, this study constructs regression models utilizing the OLS method. To ensure the robustness and comprehensiveness of the conclusions, this study constructs a dual dependent variable system parallelizing grazing intensity and livestock scale: (1) Grazing Intensity Model (for the sub-sample with pastureland) When analyzing production responses, grazing intensity—measured as stocking rate per unit area—is utilized to represent the land intensification pressure of agro-pastoralists. The model is specified as follows: G I i = β 0 + β 1 R P i + β 2 C a p i t a l i + β 3 R P i ୍ଠ C a p i t a l i + θ m C o n t r o l i m + ϵ i (5) (2) Livestock Scale Model (for the full sample) The logarithm of the household’s total Standard Sheep Units (SSUs) is utilized as the dependent variable. This specification incorporates households without contracted pastureland, elucidating the overall scale expansion logic of groups relying on landless grazing or rented pastureland. The model is specified as follows: l n S c a l e i = γ 0 + γ 1 R P i + γ 2 C a p i t a l i + γ 3 R P i ୍ଠ C a p i t a l i + δ m C o n t r o l i m + μ i (6) where G I i represents the grazing intensity of household i ; l n S c a l e i represents the total livestock scale (logarithm of SSU); R P i denotes the risk perception variables; LCi denotes the livelihood capital variables; ( R P i ୍ଠ C a p i t a l i ) represents the interaction term, employed to identify the moderating role of capital endowment in risk responses; C o n t r o l i m represents the control variables; ϵ i and μ i denote the random error terms. Statistical significance is evaluated using conventional thresholds at the 10%, 5%, and 1% levels. Specifically, p-values below 0.10, 0.05, and 0.01 indicate statistical significance at the marginal, standard, and stringent levels, respectively. These thresholds are widely adopted in applied econometric analysis [ 44, 45], and are therefore used in this study. 2.4. Study Area and Data Sources This study selects Zoigê County in the Aba Tibetan and Qiang Autonomous Prefecture of Sichuan Province as the empirical study area. Located on the northeastern edge of the QTP at an average elevation of approximately 3500 m, the county features a typical alpine semi-humid monsoon climate and represents a quintessential agro-pastoral ecotone transitioning from the QTP to inland regions. Natural grasslands account for approximately 76.1% of the county’s total land area, and alpine grass-fed animal husbandry—primarily rearing yaks and Tibetan sheep—constitutes the core traditional livelihood for local agro-pastoralists. In recent years, tourism has expanded rapidly in the region, attracting a growing number of visitors due to its distinctive natural landscapes. However, given its unique ecological niche and relatively monolithic industrial structure, the local livelihood system remains highly sensitive and vulnerable to fluctuations in climate conditions and land-cover changes. Therefore, Zoigê County provides a representative research context for exploring the risk perception and livelihood adaptation behaviors of agro-pastoralists in alpine ecologically fragile regions. The data utilized in this study were derived from the “Zoigê Rural Livelihood and Household Development Survey,” conducted by the research team between July and August 2022 in Zoigê County, Aba Prefecture, Sichuan Province. A multi-stage stratified sampling method was employed. Households were stratified based on the locational conditions and economic development levels of the townships, and household surveys were carried out in 17 villages across 7 townships. Within each sampled village, random sampling was conducted taking into account village size and the distribution of agro-pastoralist households. This yielded 913 questionnaires, of which 890 valid samples were retained, corresponding to an effective response rate of 97.5%. The sample covers households with various livelihood types and resource endowments, demonstrating good representativeness. Prior to the survey, the research team conducted field visits and interviews to gain a comprehensive understanding of local socioeconomic conditions and livelihood characteristics. The questionnaire design drew upon established survey instruments, including the China Family Panel Studies (CFPS) and the Chinese Social Survey (CSS), while fully integrating the regional culture and livelihood characteristics specific to the Zoigê pastoral areas. The questionnaire encompassed five primary modules: basic household demographic characteristics, household economic conditions, social participation, health status, and public services. Comprising approximately 60 sets of questions, the survey ensured a comprehensive assessment of agro-pastoralists’ livelihood capital and risk perception, thereby providing a reliable data foundation for the measurement of livelihood capital and risk perception variables. 3.1. Descriptive Statistics The overall sample exhibited livelihood and household characteristics typical of Zoigê County ( Table 3). In terms of educational attainment, this study employs the highest educational level within the household (i.e., the family member with the most years of schooling) as a proxy variable to more comprehensively reflect the household’s human capital endowment, rather than relying solely on the household head or an average level. Descriptive statistics indicate that the surveyed agro-pastoralists were relatively well-educated overall, with 47.4% having attained a college degree or above, while 18.1% and 19.7% had completed junior high school and senior high school education, respectively. With respect to household demographic structure, the majority of households consisted of 4–7 members, and the dependency burden was generally high: approximately 66% of households had a dependency ratio exceeding 60%, indicating a relatively heavy labor burden on working household members. Regarding land resources, approximately 55.3% of agro-pastoralist households possessed cultivated land or pasture resources, whereas 44.7% lacked contracted pastureland or cropland but maintained livestock production through rented grazing land or landless rearing. In terms of income structure, agro-pastoral income remained a vital source of household income, with approximately 51.1% of households deriving more than 60% of their total income from agriculture and animal husbandry. Furthermore, the survey revealed that livestock production was dominated by small- to medium-scale operations, with about 72% of households raising fewer than 600 Standard Sheep Units (SSUs). To further identify the heterogeneity in livelihood capital among agro-pastoralists, this study applied a K-means cluster analysis based on the livelihood capital index, categorizing the sample into three groups: low, medium, and high livelihood capital ( Table 4). Overall, households with medium livelihood capital constituted the largest proportion (50.56%), while the low and high livelihood capital groups accounted for 24.83% and 24.61%, respectively. This indicates substantial heterogeneity in livelihood capital endowments across the study area. The agro-pastoral dominant type comprised 448 households, accounting for 50.34% ( Table 5), indicating that traditional farming and animal husbandry remained the principal livelihood pattern in the area. the off-farm dominant type included 175 households, accounting for 19.66%, a group that has achieved livelihood and income diversification through non-agricultural employment or tourism-related businesses; the mixed type included 267 households, accounting for 30.00%, whose income sources are dispersed, representing a transitional phase from traditional pastoralism to modern diversified livelihoods. Table 6 presents the descriptive statistics of agro-pastoralists’ risk perceptions. Regarding climate change perception, respondents generally perceived that both summer rainfall (mean = 2.753) and winter snowfall (mean = 2.308) had increased over the past decade. For temperature perception, respondents observed a warming trend in winter (mean = 2.469) and a cooling trend in summer (mean = 1.376). In terms of natural disaster perception, the respondents generally leaned towards the view that the frequency of disasters had somewhat decreased (mean = 1.964). Regarding pasture ecological changes, agro-pastoralists observed an advanced green-up period (mean = 3.072) and a delayed senescence period (mean = 2.755) for grassland vegetation, while simultaneously perceiving a reduction in regional wetland areas (mean = 2.211). 3.2. Effects of Risk Perception and Livelihood Capital on Livelihood Strategy Choice A MNL regression model was employed to examine the effects of risk perception and livelihood capital on agro-pastoralists’ livelihood strategy choices. The agro-pastoral dominant strategy was set as the reference category, while the off-farm dominant and mixed strategies were treated as dependent outcome categories. The results are presented in Table 7. The likelihood ratio test results (LR χ 2 (18) = 168.80, p = 0.000) demonstrated that the model possessed robust explanatory power. 3.2.1. Analysis of the Off-Farm Dominant Household Model Among the risk perception variables, winter cold and snow disaster perception exerted a significantly positive effect on the likelihood of agro-pastoralists transitioning towards an off-farm dominant livelihood strategy (β = 0.484, p χ 2 = 0.9419, which is far above conventional significance levels, indicating that the null hypothesis of IIA cannot be rejected. This demonstrates that the substitution relationships among different livelihood strategy options are stable, and the MNL model specification is reasonable. Multicollinearity test. A Variance Inflation Factor (VIF) test was conducted on all explanatory variables within the OLS models. The results showed that the maximum VIF value across the models was 4.84, significantly below the commonly used critical threshold of 10, indicating that the models do not suffer from severe multicollinearity issues. Adjusting the sample structure. This study further conducted robustness checks by adjusting the sample structure, specifically by re-estimating the model after expanding the sample (see Table 8, Model 3). The results demonstrate that the coefficient directions and significance levels of the core explanatory variables remain fundamentally consistent under different sample specifications, indicating that the research conclusions do not rely on a specific sample structure and possess strong universality and robustness. Incorporating village fixed effects. Considering that unobserved village-level environmental factors—such as specific village regulations, the connectivity and quality of pastureland, and infrastructure accessibility—might simultaneously influence both risk perception and behavioral responses, this study re-estimated the models by incorporating village fixed effects. The results indicate that after controlling for village-level heterogeneity, the significance levels and influence directions of the core variables and interaction terms remain highly consistent with the baseline models (see Appendix A Table A4 for details). This confirms the universality of the “perception–capital” interactive logic across different village contexts. 4. Mechanism Analysis and Discussion The findings indicate that the adaptation of agro-pastoralists on the QTP to environmental risks is not a process of purely rational economic calculation. Instead, it corresponds to a boundedly rational process associated with the bidirectional moderating role of cognitive biases and capital endowments. In this process, risk perception reflects significant heterogeneity, while livelihood strategy choices exhibit a pronounced duality of pathways. Against this background, this study applies behavioral decision theory to interpret the livelihood adaptation choices among agro-pastoralists on the QTP. 4.1. Mechanism Analysis: The Interaction Between Risk Perception and Livelihood Capital To further reveal the behavioral logic behind agro-pastoralists’ livelihood decisions, this study systematically analyzes their adaptation mechanisms from two dimensions—risk perception characteristics and livelihood capital constraints—based on the results of the multinomial Logit and OLS regressions. 4.1.1. The Duality of Agro-Pastoralists’ Risk Perception: Cognitive Filtering and Temporal Scale Differences The results demonstrate that risk perception functions as a behavioral filter associated with objective environmental changes and agro-pastoralists’ livelihood decisions. Although meteorological observations indicate an overall warming and wetting trend along the eastern margin of the QTP, the subjective perceptions of local agro-pastoralists display a certain degree of misalignment and divergence. This divergence relates primarily to the fact that agro-pastoralists interpret environmental changes through embodied experience and production practices rather than standardized meteorological data, thereby forming contextualized risk judgments at the cognitive level. A similar heterogeneity is also evident in perceptions of pasture ecology. On the one hand, agro-pastoralists observe positive phenological signals, such as an earlier green-up and delayed senescence. On the other hand, they perceive wetland areas to be shrinking. This seemingly contradictory perception is consistent with the complex hydrological feedback mechanisms of alpine ecosystems. Increased precipitation does not necessarily correspond to wetland recovery, as it may be accompanied by accelerate evaporation and weaken water retention capacity. Accordingly, agro-pastoralists tend to simultaneously observe localized ecological improvements while developing deeper concerns regarding overall ecological stability. Further analysis suggests that the heterogeneity of risk perception and its interaction with livelihood capital are associated with the distinct differences in behavioral responses across different temporal scales. Based on the interaction analysis, sudden risks (such as winter cold and snow disasters), characterized by severe potential losses and high uncertainty, exhibit a strong association with agro-pastoralists’ tendency toward livelihood adjustment. Within this process, financial capital reflects a bidirectional moderating role. When confronting sudden disasters, financially endowed households do not necessarily opt for risk-avoidance transitions (e.g., developing off-farm activities); instead, they tend to leverage their capital advantages to purchase forage or improve infrastructure. This “resistance” logic based on resource endowments corresponds to a further path dependence on traditional pastoralism, which, from a long-term perspective, poses a potential risk of maladaptation. By contrast, gradual ecological risks (such as changes in pasture phenology) unfold slowly and may be accompanied by phase-specific signals of improvement, corresponding to a “weakened” perception among agro-pastoralists. This is associated with a diminished sense of urgency regarding risks and a reinforcement of existing livelihood patterns. Due to the systemic nature of such risks, the bidirectional moderating role of household-level natural capital (e.g., pasture area) is not statistically significant. This suggests that in the context of declining overall ecological carrying capacity, relying solely on household-level natural resource endowments may not provide a substantial buffer against livelihood pressure. Consequently, households are more likely to fall into “fragile maintenance” under the combined effects of cognitive lag and resource constraints. These findings suggest that risk perception is not a simple mapping of objective environmental changes, but rather a behavioral variable jointly influenced by cognitive filtering and temporal scales. 4.1.2. The Duality of Livelihood Strategy Choice: Coordinated and Conflictive Pathways Empirical results indicate that under the shock of climatic and ecological risks, agro-pastoralists’ livelihood strategies do not exhibit a single directional adjustment; rather, they exhibit a duality of adaptation pathways. This divergence reflects the heterogeneity of mechanisms through which agro-pastoralists’ risk perception relates to specific livelihood decisions under varying capital endowment conditions. The first pathway corresponds to a coordinated adaptation pathway. Under this pathway, risk perception encourages livelihood diversification and a reduction in dependence on traditional pastoral production, consistent with a synergy between risk avoidance and resource pressure alleviation. Agro-pastoralists endowed with higher levels of human capital exhibit a higher probability of adopting this adaptive strategy. Their educational level, skill structure, and information acquisition capacity are associated with the capacity to move away from traditional pastoral pathways and enter non-farm employment sectors, reflecting risk dispersion through smaller herd sizes or off-farm activities. At the same time, for agro-pastoralists with relatively limited natural and financial capital, environmental risks demonstrate a stronger association with a stronger “push effect.” Lacking sufficient pasture resources and having a weak capacity to withstand shocks, even minor fluctuations in the ecological environment correspond to potential threats to their livelihood security; meanwhile, financial constraints are linked to a limited ability to hedge risks through technological or capital inputs. Against this backdrop, sudden risks (such as snow disasters and cold waves) reflect potential livelihood crises, associated with heightened risk sensitivity and a tendency toward risk-avoidance adjustments, such as proactively reducing livestock or turning to diversified income sources. Furthermore, the “mental accounting” characteristic exhibited by agro-pastoralists during the risk-coping process is further associated with this pa
