Modeling democratic erosion in Ecuador: narco-capitalism, state reconfiguration, and criminal governance through PLS-SEM and ANN

1 Introduction

Between 2021 and 2025, Ecuador underwent one of the most abrupt security transformations recorded in Latin America in recent decades. A country historically described as a transit buffer between the world’s two largest cocaine producers became a key node for processing, storage, and export, registering one of the sharpest increases in homicide rates in the region (Pontón Cevallos and Rivera Velez, 2024; Dressler and Wolff, 2024). This transformation has been read through multiple competing lenses—criminological, securitarian, geopolitical—but studies that empirically test the structural articulation between illicit economic dynamics, state reconfiguration, and their effects on democratic life remain comparatively scarce in the Ecuadorian context (Freeman, 2024; Rivera-Rhon and Bravo-Grijalva, 2020).

This article addresses that gap by proposing and testing a structural model that links four theoretically grounded dimensions: narco-capitalism, coercive state reconfiguration, criminal territorial governance, and democratic erosion. Narco-capitalism is understood as a mode of accumulation in which drug-related illicit economies become functionally integrated into circuits of production, laundering, and institutional collusion, reshaping how, where, and over whom the state exercises authority (Ghiabi, 2022; Paley, 2014). Rather than treating “narco-state” as a closed descriptive category—which would require simultaneous evidence of structural institutional penetration, capture of coercive agencies, loss of the violence monopoly, and stable criminal territorial governance (Solís González, 2013)—the study treats state reconfiguration and criminal governance as analytically distinct and empirically testable processes, consistent with recent calls in the comparative literature (Garay Salamanca et al., 2008; Lessing, 2018; Campana et al., 2025). This distinction is central to the article’s theoretical contribution: rather than assuming a direct and totalizing transition from drug trafficking to a narco-state, the model tests whether citizens perceive a sequential and differentiated process in which narco-capitalist accumulation first modifies state and territorial governance arrangements, which then contribute to democratic erosion through identifiable institutional and community-level mechanisms.

The model is estimated through a two-stage hybrid strategy combining PLS-SEM and Artificial Neural Networks (ANN), applied to a cross-sectional survey of 400 adult residents across Ecuador’s four natural regions. PLS-SEM is used to validate the measurement model and assess the significance of structural relationships; ANN captures non-linear patterns and determines the relative importance of each predictor through sensitivity analysis, enhancing the predictive robustness of the structural model beyond what linear estimation alone provides (Hair et al., 2021).

The study makes three contributions. First, it offers a theoretically specified empirical test of how perceived narco-capitalism is associated with state and territorial transformations in Ecuador, avoiding the unsupported use of the narco-state label. Second, it separates two mechanisms that are frequently conflated in public debate: coercive state reconfiguration, referring to securitization, differentiated state violence, and emergency governance; and criminal territorial governance, referring to the local regulation of social life by non-state criminal actors. Third, it demonstrates the analytical value of a PLS-SEM + ANN hybrid strategy for assessing both theoretically specified linear paths and non-linear predictive patterns in governance research under institutional fragility.

2 Literature review

2.1 Political economy of illegalities and narco-capitalism

Illicit economies associated with drug markets constitute a multi-scalar phenomenon whose understanding has shifted from crime-centered readings toward structural approaches (Paley, 2014). In this vein, they are understood as a transnational economy that articulates production, circulation, and accumulation under conditions of prohibition (Santana, 2004; Goodhand et al., 2024). This framework makes it possible to observe their insertion into global dynamics of markets, territorial control, and risk management (Snyder and Duran-Martinez, 2009; Paley, 2014).

A central debate in this literature concerns whether the relationship between illegal and legal capital should be analyzed through the lens of state capture. Durand (2019) emphasizes durable institutional arrangements in which private economic interests systematically shape public decision-making; applying this framework to criminal actors raises the question of whether drug-trafficking organizations penetrate state institutions in the same structural sense as licit economic elites, or whether their relationship to the state is better described as transactional, localized, and unevenly distributed across agencies and territories (Jessop, 2008; Lipari, 2025). Jessop’s (2008) concept of strategic selectivity is particularly useful here: state institutions are not uniformly available for capture but selectively permeable depending on the configuration of forces acting upon them, which helps explain why narco-capitalist penetration of the Ecuadorian state is likely uneven across agencies. Journalistic and policy analysis has pointed to customs, ports, and the prison system as particularly exposed nodes (Freeman, 2024), though a rigorous empirical mapping of this institutional variation remains scarce and falls outside the scope of the present study.

Critical political economy scholarship argues that illegality does not operate as an externality of capitalism but as a selectively managed field (Ghiabi, 2022). Licit and illicit spheres are articulated through financial, institutional, and territorial gray zones that facilitate the convergence of circuits and actors, and their persistence can be interpreted in terms of the accumulation functions they acquire under specific structural conditions (Ghiabi, 2022; Madrid and Tapia, 2024). In Latin America, these dynamics tend to intensify under conditions of state restructuring, persistent inequality, and neoliberal reforms associated with weakening regulatory capacities (Márquez Covarrubias and Autonomous University of Zacatecas, 2015; Freeman, 2024). In Ecuador, strategic geographic location and cumulative institutional fragility have favored the country’s repositioning within regional circuits of drug trafficking, with particular concentration around the port infrastructure of Guayaquil and adjacent coastal corridors (Castillo Rojas and Lévy, 2025; Pontón Cevallos and Rivera Velez, 2024).

2.2 Narco-states, criminal governance, and state reconfiguration

Rather than assuming “narco-state” as a settled analytical category, this study situates Ecuador within a broader comparative discussion that distinguishes several related but non-equivalent concepts. Solís González (2013) traces the emergence of the narco-state concept in the Mexican context, arguing that neoliberal restructuring created the institutional conditions under which organized crime could penetrate segments of the state over time. Garay Salamanca et al. (2008) develops the closely related notion of captured and reconfigured statehood in the Colombian case, in which criminal actors do not merely corrupt isolated officials but progressively reshape institutional design to serve illicit interests—a structural and durable process distinct from individual-level corruption.

Lessing (2018) cautions against the “absent state” thesis in his comparative work on Latin American drug wars, demonstrating instead that states and criminal organizations interact through complex, often tacit arrangements of confrontation, accommodation, and negotiated coexistence—where crackdowns and truces are themselves strategic state choices rather than simply evidence of state weakness. This insight is extended by the emerging literature on criminal governance, which examines how non-state armed actors regulate territories, resolve disputes, and substitute state functions—provision of order, informal employment, conflict resolution—often without seeking control of the central state apparatus (Campana et al., 2025; Duncan, 2006). Campana et al. (2025), drawing on the most systematic cross-national estimate of criminal governance prevalence to date, find that tens of millions of people in Latin America live under governance arrangements in which local order depends simultaneously on state institutions and armed criminal groups—a configuration with profound implications for democratic consolidation. Feldmann and Luna (2023) frame these dynamics as part of a broader governance crisis in which criminal actors exploit institutional gaps rather than necessarily seeking outright capture.

The Mexican and Colombian cases illustrate the diversity of these trajectories. In Mexico, criminal organizations have at times exercised forms of territorial authority that coexist with, rather than fully displace, state institutions at the subnational level (Maldonado Aranda and Vera Cabrera, 2025). In Colombia, paramilitary and narco-trafficking networks have historically intersected with regional political elites through what Garay Salamanca et al. (2008) and Duncan (2006) describe as a cooptation of subnational state structures—a process distinct from but related to national-level capture. Muggah (2026) synthesizes these regional trajectories in terms of three mechanisms through which criminal governance erodes democratic functioning: coercion, which deploys threats and selective violence to shape electoral competition and candidate selection; suppression, through which endemic fear deters civic and political participation over time; and substitution, which narrows the perceived policy menu by pressuring mainstream actors to harden their positions to avoid appearing permissive. These mechanisms are not mutually exclusive and may operate simultaneously, as the Ecuadorian case between 2021 and 2025 illustrates.

For Ecuador, this comparative literature suggests that the relevant empirical question is not whether the country “is” a narco-state in any totalizing sense, but which of several non-exclusive processes—institutional collusion, criminal governance, territorial cooptation, or more durable structural capture—best characterizes citizens’ perceptions of how illicit economies and state institutions interact (Freeman, 2024; Maldonado Aranda and Vera Cabrera, 2025). The present study’s narco-capitalism construct, focused on perceived institutional collusion, legal–illegal gray zones, and the reconfiguration of how state authority is exercised in affected territories, is designed to operationalize this question empirically rather than to presuppose its answer.

2.3 Coercive state reconfiguration and securitization

The expansion of illicit economies in Ecuador has coincided with a pronounced intensification of securitized state responses. Following the declaration of an “internal armed conflict” in January 2024, Ecuador entered a sustained cycle of states of exception, military deployment in public security tasks, and the formal designation of criminal organizations as terrorist actors (Dressler and Wolff, 2024). This pattern is consistent with regional trends documented in the literature, in which the perceived threat posed by organized crime is used to justify the suspension or attenuation of ordinary legal procedures, the expansion of military jurisdiction over public order, and the normalization of emergency governance as a quasi-permanent tool rather than a genuinely temporary measure (Escobar-Jiménez, 2024; Freeman, 2024; Freire, 2025). As Muggah (2026) notes, in low-trust settings this shift is reinforced by the expectation that only exceptional tools can deliver order, which further raises the political cost of institutional restraint and accelerates the normalization of coercive state reconfiguration.

Mbembe (2020) offers an interpretive lens for understanding the social distribution of this securitization through his necropolitical framework: sovereignty is expressed not only through direct force but through decisions and omissions that allocate protection, risk, and abandonment unevenly across populations and territories. In the Ecuadorian context, this unevenness is visible in the disproportionate exposure of young people, peripheral neighborhoods, and historically marginalized populations to both criminal violence and coercive state response, as well as in documented selectivities in the sanctioning of larger economic and political networks linked to organized crime (Carrión Mena, 2024; Madrid and Tapia, 2024). Rather than treating necropolitics as an independent latent construct, this study uses it as an interpretive dimension for understanding the differentiated state violence and extralegal governance items that operationalize the coercive state reconfiguration construct.

A further dimension of this reconfiguration concerns the production of security discourses that construct organized crime—and at times entire social groups or territories—as an “internal enemy” against which exceptional measures are legitimized (Escobar-Jiménez, 2024). This discursive frame, associated with analytical traditions in critical discourse analysis (van Dijk, 1993; Verón, 1987), is treated here not as an independent variable but as a contextual mechanism that contributes to explaining why citizens may perceive militarization and exceptionality as normalized or acceptable—a perception that the items NE1 and NE2 of the survey instrument capture indirectly.

2.4 Criminal territorial governance and democratic erosion

A substantial and growing body of literature documents the specific pathways through which criminal governance translates into democratic erosion at the citizen and community levels. Campana et al. (2025) establish that criminal governance—understood as the regulation of everyday life by non-state armed actors in territories where state presence is weak, absent, or coexisting with criminal authority—is far more prevalent across Latin America than previously estimated, with direct implications for how citizens experience political rights, personal security, and institutional trust. Albarracín et al. (2025) extend this analysis by showing that the territorialization of criminal economies is associated with the assassination of civil society actors, activists, and journalists, which progressively narrows the space for democratic participation in affected areas.

At the individual level, Masullo et al. (2025) find, through a large-scale study in Brazil, that crime exposure is associated with a process they term democratic decoupling—a pattern in which citizens retain support for democracy as an abstract system while simultaneously reducing their commitment to the rule of law, civil liberties, and legal constraints on state power. This finding carries particular relevance for the Ecuadorian context: it suggests that securitized state responses may gain public support precisely because crime exposure weakens citizens’ commitment to the legal guardrails that constrain those responses, creating a self-reinforcing dynamic between criminal governance, coercive state reconfiguration, and democratic erosion.

Muggah (2026) synthesizes these micro- and meso-level dynamics into a framework in which the three mechanisms described in Section 2.2—coercion, suppression, and substitution—produce structurally interconnected effects on democratic life: coercion narrows who can participate in political competition; suppression reduces the civic engagement of ordinary citizens through the normalization of fear; and substitution delegitimizes mainstream political actors by demonstrating state incapacity to guarantee basic order. Together, these mechanisms are associated with the deterioration of the conditions that sustain democratic functioning—perceived insecurity, loss of institutional trust, weakening of fundamental rights, and the fragmentation of community bonds—which constitutes the substantive content of the democratic erosion construct operationalized in the present study.

The four analytical dimensions discussed above—narco-capitalism, coercive state reconfiguration, criminal territorial governance, and democratic erosion—constitute the theoretical basis for the structural model proposed and tested in the following section. The model is built on a mechanism-based logic: narco-capitalism is expected to be associated with institutional pressure through collusion, gray-zone accumulation, and selective permeability of state agencies; this pressure can be expressed either through state-level securitization and emergency governance or through local criminal regulation of territory; both mechanisms may be associated with the erosion of democratic conditions by weakening rule-of-law expectations, rights protection, institutional trust, civic participation, and community cohesion.

3 Theoretical model and hypotheses

The present study proposes a structural model in which narco-capitalism operates as the exogenous variable, coercive state reconfiguration and criminal territorial governance function as parallel mediating constructs, and democratic erosion constitutes the final endogenous variable. Figure 1 illustrates the proposed model, which includes five directional hypotheses: two direct paths from narco-capitalism to each mediator (H1, H2), two paths from each mediator to democratic erosion (H3, H4), and one direct path from narco-capitalism to democratic erosion (H5). The rationale for each hypothesis is developed below. Each hypothesis is derived from a concept-mechanism-evidence sequence in order to avoid treating statistical paths as merely exploratory associations: the theoretical claim specifies how a given social process is expected to operate, the empirical literature identifies comparable mechanisms, and the hypothesis translates that mechanism into a testable association among latent constructs.

3.1 Narco-capitalism and coercive state reconfiguration (H1)

The relationship between illicit economies and the reconfiguration of state power has received sustained theoretical and empirical attention in recent years. Falconí (2025), in a systematic review of the literature on narco-capitalism’s effects on formal institutions, documents that drug-related illicit economies are consistently associated with the erosion of institutional integrity, the co-optation of public officials, and the emergence of what some authors describe as “parallel state” structures in which formal institutions are progressively compromised by organized crime interests. This institutional erosion is not uniform but selective, affecting those agencies most directly implicated in territorial control, resource allocation, and enforcement—precisely the agencies whose reconfiguration is most visible to citizens. Ranestad and Pouillard (2026) further show that smuggling and illegal market dynamics are associated with state responses based on intensified repression and militarization, while simultaneously generating corruption and institutional strain that reshape how public power is exercised in affected territories. Complementarily, Lipari (2025) demonstrates that the gray zones between the licit and the illicit are sustained through coordination among economic actors, public officials, and criminal organizations, which is associated with rent capture, the neutralization of institutional controls, and the blurring of the boundary between public authority and criminal interests.

Taken together, these contributions suggest that the consolidation of narco-capitalist dynamics is positively associated with the intensification of coercive, selective, and militarized forms of state power—precisely the dimensions captured by the coercive state reconfiguration construct. Based on the above, the following hypothesis is proposed: Mechanistically, this relationship is expected because narco-capitalist accumulation increases the perceived permeability of state agencies to illicit interests, which in turn strengthens citizens’ perception that public authority is exercised through militarized, selective, or exceptional practices rather than through ordinary democratic legality.

H1: Perceived narco-capitalism is positively and significantly associated with perceived coercive state reconfiguration in Ecuador.

3.2 Narco-capitalism and criminal territorial governance (H2)

The literature reviewed in Section 2 establishes that narco-capitalism does not only reshape state institutions but also creates the conditions under which non-state armed actors can assume governance functions at the local level. Campana et al. (2025) show that the prevalence of criminal governance across Latin America is structurally linked to the expansion of illicit economies and the institutional gaps they generate or exploit, with criminal organizations regulating everyday life, resolving disputes, and providing services in territories where state presence is weak, contested, or co-opted. Albarracín et al. (2025) extend this analysis by demonstrating that the territorialization of criminal economies is associated with the consolidation of criminal authority, including interference in local political decision-making and the elimination of civil society actors who challenge that authority. Muggah (2026) further shows that substitution—the process through which criminal actors displace state functions — is directly linked to the economic density of illicit markets in affected territories.

These findings jointly suggest that where narco-capitalism is more strongly perceived, the conditions for criminal territorial governance are also more likely to be present. On this basis, the following hypothesis is formulated: The mechanism is territorial: illicit accumulation requires routes, logistics, protection, labor, silence, and dispute regulation; where these conditions become normalized, criminal actors are more likely to be perceived as governing everyday life, replacing or coexisting with public authority at the local level.

H2: Perceived narco-capitalism is positively and significantly associated with perceived criminal territorial governance in Ecuador.

3.3 Coercive state reconfiguration and democratic erosion (H3)

A securitized reconfiguration of the state—expressed through militarization, differential violence, and the normalization of extralegal governance—tends to be associated with the deterioration of the conditions that sustain democratic functioning. Masullo et al. (2025) find that crime exposure in contexts of intense securitization is associated with a process of democratic decoupling, in which citizens progressively reduce their commitment to the rule of law, civil liberties, and the legal constraints on state power, even while maintaining abstract support for democracy as a system. Cunha et al. (2026) complement this finding by showing that the militarization of state agencies in contexts of democratic backsliding is associated with direct repercussions on citizens’ perceived security and rights. Berlin (2025) further shows that when human rights protection mechanisms operate in a limited or selective manner—a defining feature of coercive state reconfiguration—the weakening of rights and the loss of institutional trust follow systematically. The theoretical expectation is therefore not that coercion automatically destroys democracy, but that repeated reliance on militarized presence, differentiated state violence, and emergency measures can normalize weakened legal safeguards, reduce trust in rights protection, and make democratic constraints appear less effective or less necessary.

The following hypothesis is therefore proposed:

H3: Perceived coercive state reconfiguration is positively and significantly associated with perceived democratic erosion in Ecuador.

3.4 Criminal territorial governance and democratic erosion (H4)

The literature documents multiple pathways through which criminal territorial governance translates into democratic erosion at the community level. Muggah (2026) identifies coercion and suppression as the two most direct mechanisms: criminal organizations use selective violence and threats to discipline communities and deter civic participation, while the normalization of fear progressively reduces citizens’ willingness to engage in public life. Campana et al. (2025) establish that where criminal governance is most entrenched, citizens experience the most severe deficits in institutional trust, perceived rights, and community cohesion—the core dimensions of democratic erosion. Albarracín et al. (2025) further show that the territorialization of criminal economies is associated with a contraction of democratic space through the elimination of civil society actors. Bandauko (2026) complements these findings by documenting that governance arrangements characterized by abandonment and selective coercion are linked to the erosion of dignity, rights, and social cohesion in affected communities. This hypothesis is formulated as a territorial mechanism: where criminal actors regulate mobility, speech, commerce, dispute resolution, or local political behavior, democratic erosion is expected to operate through fear, civic withdrawal, reduced community cooperation, and the perceived contraction of public space.

Based on these contributions, the following hypothesis is proposed:

H4: Perceived criminal territorial governance is positively and significantly associated with perceived democratic erosion in Ecuador.

3.5 Narco-capitalism and democratic erosion: direct association and dual mediation (H5)

Beyond the indirect paths running through the two mediating constructs, the model tests whether narco-capitalism retains a direct association with democratic erosion after controlling for both mediators. Falconí (2025) documents that narco-capitalism is associated with the erosion of public confidence and state legitimacy independently of the specific governance arrangements it generates. Ghiabi (2022) further shows that gray zones between legality and illegality, when they become normalized in everyday economic life, are associated with the erosion of citizens’ expectations of rule-of-law protection and impartial institutional treatment—effects that constitute direct dimensions of democratic erosion. The direct path is theoretically important because democratic erosion may arise not only from visible coercion or local criminal rule, but also from the broader normalization of institutional collusion and legal-illegal gray zones, which undermines expectations of impartial legality even among citizens not directly exposed to territorial criminal governance.

The full structural model therefore posits a dual mediation in which narco-capitalism is associated with democratic erosion both directly and indirectly through two parallel channels: coercive state reconfiguration (H1 → H3) and criminal territorial governance (H2 → H4).

H5: Perceived narco-capitalism is positively and significantly associated with perceived democratic erosion in Ecuador, both directly and indirectly through the parallel mediation of perceived coercive state reconfiguration and perceived criminal territorial governance.

4 Methodology

4.1 Research design

The study adopts a quantitative, non-experimental, cross-sectional design aimed at analyzing the structural relationships among four latent constructs: narco-capitalism, coercive state reconfiguration, criminal territorial governance, and democratic erosion. This design is appropriate when the objective is to examine patterns of association among unobservable variables measured through reflective indicators, including theoretically specified direct and indirect relationships, without deliberate manipulation of independent variables (Hair et al., 2019).

The study employs a two-stage hybrid analytical strategy. In the first stage, Partial Least Squares Structural Equation Modeling (PLS-SEM) is used to validate the measurement model and estimate the significance and magnitude of structural relationships among latent constructs. In the second stage, an Artificial Neural Network (ANN) is applied to the latent variable scores extracted from PLS-SEM to capture non-linear patterns of association and determine the relative importance of each predictor through sensitivity analysis. This hybrid PLS-SEM + ANN approach compensates for the assumption of linearity inherent in SEM-based methods and has been increasingly adopted in social science research to enhance the predictive robustness of structural models (Hair et al., 2021).

4.2 Stage 1—PLS-SEM

In plain terms, PLS-SEM estimates relationships among unobserved (latent) constructs—such as narco-capitalism or democratic erosion—based on how citizens responded to the survey items, and tests whether the hypothesized paths between these constructs are statistically significant. PLS-SEM was implemented in SmartPLS 4. The PLS-SEM algorithm used the path weighting scheme, a maximum of 300 iterations, and a stop criterion of 10⁻⁷. This estimator was selected because it is appropriate for reflective latent constructs, dual-mediation structures, and predictive aims without requiring multivariate normality (Hair et al., 2014, 2019; Reinartz et al., 2009). Bootstrapping with 5,000 subsamples was used to estimate the statistical significance of path coefficients and indirect effects (Hair et al., 2013).

Model evaluation followed the two-stage sequence recommended in the literature (Hair et al., 2019). The measurement model was assessed for internal consistency (Cronbach’s α and composite reliability ρ_c), convergent validity (AVE ≥ 0.50; Fornell and Larcker, 1981), and discriminant validity via the Fornell–Larcker criterion and the HTMT ratio (Henseler et al., 2015). The structural model was then evaluated through path coefficient significance, explained variance (R²), effect sizes (f²), predictive out-of-sample relevance (assessed via PLSpredict with k = 10 folds and 10 repetitions; Shmueli et al., 2016, 2019), and internal collinearity (VIF < 3.3; Kock, 2015).

Unlike covariance-based SEM, PLS-SEM is a variance-based, prediction-oriented technique and does not have an equivalent, well-established set of global fit indices (e.g., χ², CFI, RMSEA). We nonetheless report the approximate fit indicators available in SmartPLS 4 as a supplementary check: the Standardized Root Mean Square Residual (SRMR) was 0.038 for the saturated model and 0.083 for the estimated model, and the Normed Fit Index (NFI) was 0.919 (saturated) and 0.908 (estimated). The saturated-model SRMR—the value originally proposed by Henseler et al. (2014) as the primary PLS-SEM fit diagnostic—falls comfortably below the conventional 0.08 threshold; the estimated-model SRMR (0.083) is modestly above that threshold, which is common in PLS-SEM structural models and reflects the additional constraints imposed by the hypothesized path structure rather than a measurement-model problem. Given ongoing debate over the use of SRMR as a strict goodness-of-fit test in PLS-SEM (Hair et al., 2019; Henseler et al., 2014), we treat these values as supplementary evidence rather than as a substitute for the primary evaluation criteria already reported: reliability, convergent and discriminant validity, explained variance (R²), effect sizes (f²), and predictive relevance (Q²_predict via PLSpredict).

4.3 Stage 2—artificial neural network (ANN)

In plain terms, an artificial neural network (ANN) is a flexible, data-driven model that can capture non-linear patterns among the same three predictor constructs that a linear structural model might miss. Here it is used as a complementary robustness check on the PLS-SEM results, not as a replacement for the theory-driven structural model. Latent variable scores extracted from PLS-SEM were used to build the ANN model. The three predictor constructs—narco-capitalism, coercive state reconfiguration, and criminal territorial governance—were entered as inputs, while democratic erosion was specified as the target output. The analytical pipeline employed Pandas and NumPy for data structuring and matrix manipulation of the latent variable scores (McKinney, 2010; Harris et al., 2020); Scikit-learn for preprocessing—specifically, hold-out splitting (80% training/20% test) and standardization of inputs via StandardScaler—and for computing regression performance metrics (R², MAE, RMSE) (Pedregosa et al., 2011); TensorFlow and Keras as the deep learning framework for designing, compiling, and training a Multilayer Perceptron (MLP) with democratic erosion as the target variable (Abadi et al., 2016; Chollet, 2015); and Matplotlib and Seaborn for topological mapping of the network architecture and visual analysis of loss-function convergence across training epochs (Hunter, 2007; Waskom, 2021).

The MLP architecture consisted of an input layer (three predictor constructs: narco-capitalism, coercive state reconfiguration, and criminal territorial governance), two hidden layers with ReLU activation (8 and 4 units, respectively), and a single linear output neuron predicting democratic erosion scores. Predictor variables were standardized (zero mean, unit variance) prior to training. The model was compiled with the Adam optimizer (default learning rate = 0.001) and mean squared error (MSE) as the loss function (mean absolute error tracked as a secondary metric), trained for 100 epochs with a batch size of 16, holding out 10% of the training data as an internal validation set to monitor convergence (no early-stopping criterion was applied). The 80/20 train/test split used a fixed random seed (random_state = 42) to ensure reproducibility. To assess the stability of model performance beyond a single hold-out split, we additionally report results across 20 independently seeded training runs (mean R² = 0.6945, SD = 0.0922; mean RMSE = 0.5886, SD = 0.0824; mean MAE = 0.4064, SD = 0.0446).

Sensitivity analysis—computing the relative importance of each input construct in predicting democratic erosion—was conducted using permutation importance (Breiman, 2001; as implemented in scikit-learn following Fisher et al., 2019): each predictor construct’s values were independently and randomly permuted across the test set (10 repetitions), and the resulting mean increase in prediction error (MSE) was taken as that construct’s relative importance, normalized to sum to 1 across the three predictors. This analysis provides a non-linear, model-agnostic complement to the standardized path coefficients obtained in Stage 1.

Relative weight analysis (RWA) is a simpler, regression-based alternative for apportioning explained variance among correlated predictors, and would likely produce a similar ranking of predictor importance here, since it—like PLS-SEM—assumes linear, additive relationships. The specific value the ANN adds is its capacity to capture non-linear and interactive relationships among the three predictor constructs that neither PLS-SEM path coefficients nor RWA can detect by construction. In this study, the close correspondence between the ANN sensitivity-analysis ranking and the PLS-SEM ranking is itself informative: it indicates that the relationships in the data are predominantly linear, rather than that the ANN uncovered a substantially different structure.

4.4 Participants and sampling

The survey was administered to a non-probabilistic convenience sample of 400 adult participants (18 years of age or older) residing across Ecuador’s four natural regions. Recruitment was conducted digitally through email, social media, and messaging applications during 2023–2024, using the Dillman (2011) tailored design method to reduce nonresponse bias. Participation was voluntary and anonymous.

The final sample of 400 was obtained after excluding approximately 5% of initial responses for incompleteness or invariant response patterns (Hair et al., 2019). The sample covers Ecuador’s four natural regions: Costa (coast; 52.50%), Sierra (highlands; 26.25%), Amazonía (Amazon basin; 15.50%), and Insular (the Galápagos Islands; 5.75%), and includes urban (42.75%), peripheral (31.25%), and rural (26.00%) areas. Women represent 58.75% and men 41.25%. Full sociodemographic detail is presented in Table 1. Because recruitment was non-probabilistic, findings cannot be statistically generalized to the Ecuadorian population (Evans and Mathur, 2005).

Table 1. Sociodemographic profile of the sample.

Frequency distributions and percentages reported for sociodemographic variables in the sample.

4.5 Instrument and operationalization

The instrument consists of 24 items across four reflective constructs (six items each), scored on a 5-point Likert scale (1 = strongly disagree, 5 = strongly agree). Items were adapted from published empirical literature and operationalize citizen-level perceptions of recognizable social and institutional phenomena. Table 2 summarizes the constructs, abbreviations, analytical dimensions, dimension abbreviations, and the main theoretical and measurement sources used to guide item adaptation and construct specification. The complete item wording is provided in Supplementary Table S1, together with the mapping of each item to its corresponding construct and dimension.

Table 2. Construct operationalization, analytical dimensions, and main theoretical and measurement sources.

Abbreviations identify the latent constructs and their analytical dimensions. The table follows a construct-level operationalization format; the complete item wording can be provided as Supplementary material during submission.

4.6 Data collection and procedure

Items were presented in four fixed blocks: RE1–RE6, NC1–NC6, ED1–ED6, CG1–CG6. This order deliberately did not reproduce the theoretical causal sequence (NC → RE/CG → ED). The purpose was to reduce priming effects and the risk that respondents would narratively align their answers with an implied directional model (Podsakoff et al., 2003). Thus, the order of presentation functioned as a procedural control for common method bias rather than as a representation of the hypothesized structural model. The dataset contains 400 observations across 31 variables: 7 sociodemographic and 24 item-level variables (RE1–RE6, NC1–NC6, ED1–ED6, CG1–CG6).

4.7 Ethical considerations

The study complied with the principles of voluntariness, anonymity, and confidentiality applicable to social research with adult participants. Formal ethical approval was not required given the anonymous, non-sensitive nature of the data (Winter and Gundur, 2024). Informed consent was obtained implicitly through voluntary questionnaire completion.

4.8 Common method bias assessment

Given the single-source, cross-sectional design, common method bias (CMB) was assessed through procedural and statistical controls (Podsakoff et al., 2003). Procedurally, items were worded in concrete, non-technical language; anonymity was assured; and construct blocks were presented in an order that does not match the theoretical model sequence. Statistically, internal VIF values for all predictor–endogenous pairs fell between 1.000 and 2.039—well below the threshold of 3.3 (Kock, 2015)—and Harman’s single-factor test explained less than 50% of total item variance, providing no indication of dominant method variance. Results should nonetheless be interpreted as citizens’ perceived co-occurrence of these phenomena rather than as structural causal evidence. These tests do not prove the absence of method bias; rather, they indicate that no single statistical diagnostic suggested a level of method variance strong enough to invalidate the interpretation of the structural paths.

5 Results

5.1 Sample description

The sample consisted of 400 participants (Table 1), predominantly women, in the mestizo ethnic category, and concentrated in the 18–24 age group and in urban areas. Full sociodemographic frequencies and percentages are reported in Table 1.

5.2 Measurement model: reliability and validity

Prior to interpreting the structural model, the measurement model was evaluated to verify the reliability and validity of the four reflective constructs. Results are presented in Tables 3, 4. Internal consistency was high across all constructs (Table 3). Cronbach’s α ranged from 0.940 (Coercive State Reconfiguration and Democratic Erosion) to 0.963 (Narco-Capitalism), ρA from 0.940 to 0.966, and composite reliability (CR) from 0.952 to 0.970, all well above the recommended threshold of 0.70 (Hair et al., 2019). Convergent validity was also supported: AVE values ranged from 0.769 (Coercive State Reconfiguration) to 0.845 (Narco-Capitalism), all exceeding the minimum criterion of 0.50 (Fornell and Larcker, 1981), indicating that each construct explains a substantial proportion of the variance in its indicators.

Table 3. Construct reliability and convergent validity.

α, ρA, and CR assess internal consistency; AVE assesses convergent validity. Thresholds: α, ρA, CR > 0.70; AVE > 0.50 (Hair et al., 2019; Fornell and Larcker, 1981).

Table 4. Discriminant validity—Fornell–Larcker criterion.

Diagonal values (bold) = square root of AVE. Off-diagonal values = inter-construct correlations. Discriminant validity is supported when diagonal values exceed all off-diagonal values in the same row and column.

All individual item loadings exceeded 0.83 (Figure 2), surpassing the recommended minimum of 0.70 (Hair et al., 2019). The lowest loading was RE2 (0.835) and the highest NC2 (0.943), indicating that all items contribute substantially to their respective constructs.

Discriminant validity was assessed using two complementary criteria (Table 4 and Table 5). The Fornell–Larcker criterion (Table 4) was satisfied: the square root of AVE for each construct (diagonal) exceeded all inter-construct correlations in the corresponding row and column. The highest inter-construct correlation was observed between Narco-Capitalism and Democratic Erosion (0.818), which nonetheless remained below the Narco-Capitalism AVE root (0.919) and the Democratic Erosion AVE root (0.878).

Table 5. Discriminant validity—HTMT ratio.

The Heterotrait-Monotrait (HTMT) ratio estimates the true correlation between two constructs, correcting for measurement unreliability; off-diagonal values represent the HTMT ratio for each construct pair. Values below the conservative threshold of 0.90 (Henseler et al., 2015) support discriminant validity between those constructs; all values here fall below that threshold.

The HTMT ratio (Table 5) provided additional confirmation: all values fell below the conservative threshold of 0.90 (Henseler et al., 2015), with the highest value observed between Narco-Capitalism and Democratic Erosion (0.856). These results jointly support the empirical differentiation of the four constructs.

5.3 Structural model: path coefficients, explanatory power, and effect sizes

With the measurement model supported, the structural model was estimated through bootstrapping with 5,000 subsamples. Results for path coefficients, R², f², and indirect effects are presented in Tables 6–9.

Table 6. Path coefficients—bootstrapping results (B = 5,000).

β (O) = original sample coefficient; M, bootstrap mean; SD, standard deviation; CI, bias-corrected confidence interval. Significance assessed at α = 0.05.

Table 7. Explained variance (R²).

Table 8. Effect sizes (f²).

f² interpretation: < 0.02 negligible, 0.02–0.14 small, 0.15–0.34 medium, > 0.35 large (Cohen, 1988, as adapted by Hair et al., 2019).

Table 9. Specific indirect effects (B = 5,000 bootstrap subsamples).

Significance assessed at α = 0.05 based on bootstrapped t-statistics.

5.3.1 Path coefficients

Table 6 presents the standardized path coefficients with their bootstrapped statistics and 95% bias-corrected confidence intervals. Three of the five hypothesized paths were statistically significant at p < 0.001. Narco-Capitalism was positively and significantly associated with Coercive State Reconfiguration [β = 0.716; t = 17.685; p < 0.001; 95% CI (0.627, 0.787)], supporting H1. Narco-Capitalism was also positively and significantly associated with Criminal Territorial Governance [β = 0.741; t = 18.105; p < 0.001; 95% CI (0.652, 0.814)], supporting H2. The direct path from Narco-Capitalism to Democratic Erosion was likewise significant [β = 0.584; t = 9.748; p < 0.001; 95% CI (0.462, 0.698)], providing support for the direct component of H5.

Coercive State Reconfiguration was positively and significantly associated with Democratic Erosion [β = 0.238; t = 4.269; p < 0.001; 95% CI (0.129, 0.348)], supporting H3. By contrast, the path from Criminal Territorial Governance to Democratic Erosion did not reach statistical significance [β = 0.085; t = 1.584; p = 0.113; 95% CI (−0.017, 0.194)], leading to the rejection of H4.

5.3.2 Explanatory power

Table 7 presents R² and adjusted R² for the three endogenous constructs. The model explains 51.2% of the variance in Coercive State Reconfiguration (R² = 0.512; adjusted R² = 0.511), 54.9% in Criminal Territorial Governance (R² = 0.549; adjusted R² = 0.548), and 71.0% in Democratic Erosion (R² = 0.710; adjusted R² = 0.708). Following Cohen's (1988) benchmarks as adapted for PLS-SEM (Hair et al., 2019), these values indicate substantial explanatory power across all endogenous constructs, with Democratic Erosion showing particularly high explained variance.

5.3.3 Effect sizes

Effect sizes (f²) are presented in Table 8. Narco-Capitalism showed a large effect on Coercive State Reconfiguration (f² = 1.049) and on Criminal Territorial Governance (f² = 1.216), indicating that Narco-Capitalism is the dominant predictor of both mediating constructs. In explaining Democratic Erosion, Narco-Capitalism showed a medium-to-large effect (f² = 0.481), Coercive State Reconfiguration a small effect (f² = 0.065), and Criminal Territorial Governance a negligible effect (f² = 0.008), consistent with the non-significant path coefficient observed for H4.

5.3.4 Indirect effects and mediation

Table 9 presents the specific indirect effects tested for the mediation component of H5. The indirect path through Coercive State Reconfiguration was positive and statistically significant (β = 0.171; t = 4.067; p < 0.001), confirming partial mediation of Coercive State Reconfiguration in the relationship between Narco-Capitalism and Democratic Erosion. The indirect path through Criminal Territorial Governance was not statistically significant (β = 0.063; t = 1.540; p = 0.124), consistent with the non-significant direct path from CG to ED (H4). The full mediation component of H5 is therefore partially supported: the indirect channel through Coercive State Reconfiguration is confirmed, while the channel through Criminal Territorial Governance is not.

5.3.5 Internal collinearity

Internal VIF values for predictor–endogenous variable pairs in the structural model are reported in Table 10. All values fell below the conservative threshold of 3.3 recommended by Kock (2015) for detecting common method bias-related collinearity, ranging from 1.049 to 1.216 among the predictors of Democratic Erosion. These results indicate no problematic multicollinearity in the structural model and provide additional reassurance against common method bias, complementing the procedural controls described in Section 4.8.

Table 10. Internal collinearity statistics (VIF)—structural model.

Internal VIF < 3.3 indicates absence of common method bias-related collinearity (Kock, 2015).

5.4 Predictive relevance (Q²)

The Q²_predict metric reported here builds on the original cross-validated redundancy concept of Q² introduced by Stone (1974) and Geisser (1974), later extended by Shmueli et al. (2016, 2019) into PLSpredict, an indicator-level, out-of-sample prediction procedure. Predictive out-of-sample relevance was assessed using PLSpredict (Shmueli et al., 2016, 2019), which estimates construct- and indicator-level Q²_predict through a k-fold holdout procedure, comparing the PLS-SEM model’s prediction error against a naïve linear model (LM) benchmark. Results are presented in Table 11. All three endogenous constructs showed positive Q²_predict values: Coercive State Reconfiguration (RMSE = 0.707; MAE = 0.522; Q²_predict = 0.509), Criminal Territorial Governance (RMSE = 0.680; MAE = 0.467; Q²_predict = 0.546), and Democratic Erosion (RMSE = 0.582; MAE = 0.399; Q²_predict = 0.667). At the indicator level, PLS-SEM outperformed the LM benchmark (lower RMSE) for 9 of 18 indicators, underperformed for 7, and tied for 2—a mixed pattern indicating medium predictive power in the sense of Shmueli et al. (2019), rather than the uniformly high predictive power that would be implied by PLS-SEM outperforming LM across all indicators. We report this more precisely in the revised text rather than characterizing predictive relevance as uniformly high.

Table 11. Predictive relevance—PLS predict.

Q²_predict > 0 indicates predictive relevance beyond the naïve mean baseline; comparison against a linear model (LM) benchmark (Shmueli et al., 2019) determines whether predictive power is high, medium, or low. RMSE, root mean square error; MAE, mean absolute error.

5.5 Stage 2—artificial neural network results

Before presenting results, it is useful to clarify what each part of the network represents. The input layer consists of the three predictor constructs (Narco-Capitalism, Coercive State Reconfiguration, Criminal Territorial Governance). The hidden layers are not mediating constructs in a theoretical sense—unlike the mediators specified in the PLS-SEM model—but intermediate computational layers that allow the network to combine the three inputs in flexible, non-linear ways. The output layer produces the predicted Democratic Erosion score. The ANN is used here purely as a predictive, data-driven complement to the theory-driven PLS-SEM model, not as an alternative theoretical structure. The latent variable scores for Narco-Capitalism (NC), Coercive State Reconfiguration (RE), and Criminal Territorial Governance (CG), extracted from PLS-SEM, were used as inputs to a Multilayer Perceptron (MLP) with Democratic Erosion (ED) as the target variable. The architecture consisted of an input layer (three constructs), two hidden layers with ReLU activation, and a single linear output neuron, compiled with the Adam optimizer and MSE as the loss function. The dataset was split into training (80%, n = 320) and test (20%, n = 80) sets via hold-out, with inputs standardized using StandardScaler prior to training.

5.5.1 Model performance

On the held-out test set, the ANN achieved a mean R² = 0.6945 (SD = 0.0922), mean MAE = 0.4064 (SD = 0.0446), and mean RMSE = 0.5886 (SD = 0.0824) across 20 independently seeded training runs (Table 12). This level of predictive accuracy is comparable to, rather than exceeding, the R² = 0.710 obtained in PLS-SEM for Democratic Erosion, suggesting that the two approaches perform similarly and that the relationships among the three predictor constructs are predominantly linear (see Section 6, point 12). Rather than adding predictive power beyond the structural model, the value of the hybrid two-stage strategy here lies in cross-validating the PLS-SEM findings through an independent, non-linear estimation technique (Hair et al., 2021)—the convergence of both approaches on the same predictor ranking strengthens confidence in that ranking, even though it does not indicate meaningful non-linearity in the underlying relationships. The loss convergence curve (Figure 3) shows stable and monotonic convergence for both the training and validation sets from approximately epoch 20 onward, with no evidence of overfitting—the validation loss stabilizes at a value close to but above the training loss, consistent with normal generalization behavior.

Table 12. ANN performance metrics—test set (20%, n = 80), mean ± SD across 20 independently seeded training runs.

Metrics computed on the held-out test set using Scikit-learn, averaged across 20 independently seeded training runs (fixed data split, re-initialized network weights) to account for non-deterministic GPU-based training. R², coefficient of determination; MAE, mean absolute error; RMSE, root mean square error.

5.5.2 Sensitivity analysis: relative importance of predictors

Sensitivity analysis was conducted using permutation importance (Breiman, 2001; Fisher et al., 2019): each input construct was randomly permuted across the test set while the remaining predictors were left unchanged, and the resulting increase in MSE was recorded. Results are presented in Figure 4 and summarized in Table 13.

Table 13. Relative importance of predictor constructs—ANN permutation importance, mean ± SD across 20 independently seeded training runs.

Relative importance estimated via permutation importance (Breiman, 2001; Fisher et al., 2019), mean ± SD across 20 independently seeded training runs. Values represent the normalized mean increase in prediction error (MSE) when each construct is randomly permuted.

Narco-Capitalism was the dominant predictor of Democratic Erosion in the non-linear model, with a mean relative importance of 0.6751 (SD = 0.0958) across 20 independently seeded training runs — roughly 2.8 times that of Coercive State Reconfiguration (M = 0.2402, SD = 0.0789) and 8 times that of Criminal Territorial Governance (M = 0.0846, SD = 0.0467); this ranking (NC > RE > CG) was consistent across all runs.

This ranking is fully consistent with the pattern of standardized path coefficients and effect sizes obtained in Stage 1: NC showed the largest direct path to ED (β = 0.584) and the largest f² (0.481), while CG showed a negligible f² (0.008) and a non-significant path coefficient. The ANN sensitivity analysis therefore corroborates and extends the PLS-SEM findings by confirming that the dominance of Narco-Capitalism over Democratic Erosion holds under non-linear estimation, and that the contribution of Criminal Territorial Governance remains marginal even when non-linear interaction patterns are accounted for.

6 Discussion

6.1 Narco-capitalism as the dominant predictor: structural over local

The most consistent finding across both analytical stages is the dominant role of perceived narco-capitalism in predicting perceived democratic erosion. In PLS-SEM, the direct path from Narco-Capitalism to Democratic Erosion yielded the largest coefficient in the model (β = 0.584), the largest effect size (f² = 0.481), and the highest explained variance contribution to Democratic Erosion (R² = 0.710). In the ANN sensitivity analysis, Narco-Capitalism accounted for a mean relative importance of 0.6751 (SD = 0.0958) across 20 independently seeded training runs—roughly 2.8 times the contribution of Coercive State Reconfiguration (M = 0.2402) and 8 times that of Criminal Territorial Governance (M = 0.0846). This convergence across linear and non-linear estimation strengthens the substantive interpretation: citizens’ perception of narco-capitalism, as perceived institutional collusion, legal–illegal gray zones, and the reconfiguration of state authority in affected territories, is the construct most strongly associated with citizens’ perception of democratic erosion.

This finding resonates with Terán-Mantovani and Scarpacci's (2024) analysis of intertwined criminal economies and accumulation in Latin America, which documents how the entanglement of licit and illicit circuits—through evasion, contrabanding, money laundering, and corrupt state intermediation—constitutes not a marginal phenomenon but a structural feature of contemporary extractivist capitalism in the region. In Ecuador, the dollarized economy has facilitated precisely these entanglements by providing a stable medium for transnational capital circulation that serves both formal and illicit interests (Maldonado Aranda and Vera Cabrera, 2025). Falconí (2025) similarly finds that narco-capitalism operates as a systemic force that erodes institutional trust and state legitimacy independently of the specific governance arrangements it generates—consistent with the large direct effect observed here.

The implication for democratic theory is significant: democratic erosion in Ecuador does not appear, from the citizen perspective, primarily as a consequence of local criminal governance or even coercive securitization, but as a more diffuse effect of the perceived normalization of institutional collusion and gray-zone economic activity. This is consistent with Ghiabi’s (2022) argument that gray zones, when sufficiently normalized in everyday economic and political life, erode citizens’ expectations of impartial institutional treatment—which is the foundational condition for democratic legitimacy.

6.2 The confirmed mediation: coercive state reconfiguration as institutional channel

The indirect path from Narco-Capitalism through Coercive State Reconfiguration to Democratic Erosion was positive and significant (β = 0.171, p < 0.001), confirming Coercive State Reconfiguration as a partial mediator and supporting H3 and the mediation component of H5 through the state channel. This standardized indirect effect (β = 0.171) provides a meaningful complement to the direct effect (β = 0.584). The 51.2% of variance explained in Coercive State Reconfiguration (R² = 0.512) is attributable entirely to Narco-Capitalism, its sole antecedent in the structural model.

This pattern is consistent with the comparative literature on securitization and democratic backsliding. Masullo et al.'s (2025) finding of democratic decoupling—whereby crime exposure reduces commitment to civil liberties and rule-of-law constraints even among citizens who retain abstract support for democracy—provides a mechanism for understanding how coercive state reconfiguration translates into democratic erosion: the normalization of military presence, exceptional governance, and differentiated state violence conditions citizens to expect less from the institutional protections that democracy is supposed to guarantee. In Ecuador between 2021 and 2025, this dynamic is particularly visible in the rapid public acceptance of military deployment, states of emergency, and the formal designation of criminal organizations as terrorist actors—responses that Dressler and Wolff (2024) document as generating both immediate security effects and longer-term normalization of exceptional governance.

The Colombian and Mexican cases provide relevant comparative context. Garay Salamanca et al.’s (2008) analysis of Colombia’s captured and reconfigured statehood shows that coercive state reconfiguration under narco-capitalism pressure tends to produce not a strong securitizing state, but a selectively permissive one—capable of intense coercion toward some populations and territories while simultaneously accommodating criminal interests in others. Maldonado Aranda and Vera Cabrera (2025) document analogous dynamics in Mexico, where military deployment in drug-war contexts has repeatedly been associated with human rights violations concentrated in peripheral and marginalized communities—precisely the kind of differentiated state violence captured by items RE3 and RE4 in this study’s instrument.

6.3 The rejected hypothesis: why criminal territorial governance did not mediate

The non-significance of the direct effect of Criminal Territorial Governance on Democratic Erosion (β = 0.085, p = 0.113) and the corresponding indirect path (β = 0.063, p = 0.124) constitutes the most substantively interesting finding of the study, and warrants careful interpretation rather than dismissal. Two non-mutually exclusive explanations deserve consideration. Importantly, this result should not be framed as a theoretical failure of the criminal governance literature. It more plausibly indicates that the democratic consequences of criminal territorial governance are conditional on territorial exposure and on dimensions of democracy that the present ED scale only partially measures.

The first is a measurement explanation. Criminal territorial governance, as a newly operationalized construct in the Ecuadorian context, may capture phenomena that are geographically concentrated—port cities, peripheral neighborhoods, frontier zones—and therefore less uniformly perceived across the national convenience sample than the other three constructs. Campana et al.'s (2025) own data show significant variation in criminal governance prevalence across Latin American countries, and within-country variation is likely even larger. A sample that includes respondents from Costa, Sierra, Amazonía, and Insular regions may systematically underrepresent the most affected territories—particularly Guayaquil’s peripheral areas and the northern border corridor—where criminal territorial governance effects on democratic erosion would likely be strongest. A nationally distributed convenience sample is therefore likely to capture the perception of narco-capitalism and securitized state response more consistently than the localized experience of criminal rule, which may be concentrated in specific neighborhoods, ports, prison-linked territories, and border corridors.

The second is a substantive explanation with important theoretical implications: criminal territorial governance may affect democratic erosion through pathways that are not captured by the four ED items in this study—specifically, through electoral manipulation, candidate assassination, and the narrowing of political competition documented by Albarracín et al. (2025) and Muggah (2026). The items ED1–ED6 measure institutional distrust, rights deterioration, insecurity, and social fragmentation, but do not directly capture the contraction of political participation and civic space that criminal governance produces through coercion and suppression. This suggests a partial construct validity limitation for Democratic Erosion as operationalized here: it captures the experiential and institutional dimensions of democratic erosion but not its participatory and competitive dimensions.

Both explanations suggest that H4’s rejection should be interpreted as an invitation to refine the measurement of criminal territorial governance and democratic erosion in future research—including participatory and competitive dimensions in the ED construct, and sampling strategies that over-represent the territories of highest criminal governance concentration.

6.4 Ecuador and the narco-state debate: partial reconfiguration, not totalizing capture

The findings of this study contribute to the broader theoretical debate on whether Ecuador constitutes, or is becoming, a narco-state. As Solís González (2013) establishes, “narco-state” as an analytical category requires the simultaneous observation of at least six conditions: structural penetration of criminal economies in state institutions; sustained capture of coercive and judicial agencies; significant economic dependence on illicit circuits; partial loss of the state’s monopoly on violence; articulation between political, economic, and criminal elites; and relatively stable criminal territorial governance. The perceptual evidence from this study supports some of these dimensions—particularly institutional collusion (NC1–NC2), gray-zone economic activity (NC3–NC4), and the normalization of exceptional governance (RE5–RE6)—but does not permit structural conclusions about the others.

What the perceptual data suggest is a pattern more consistent with citizens’ experience of hybrid configurations of statehood—partial criminal capture, segmented criminalization of the state, or competitive coexistence between state sovereignty and criminal sovereignties (Garay Salamanca et al., 2008; Lessing, 2018)—than with a totalizing narco-state. This is, of course, an inference from citizen perception rather than a direct structural diagnosis, and should be read accordingly. This aligns with Maldonado Aranda and Vera Cabrera’s (2025) characterization of Ecuador’s security crisis as a product of the “neoliberal machine”: the structural articulation of financial liberalization, institutional retrenchment, prison system collapse, and criminal organization expansion that created the conditions for rapid criminalization of state segments without requiring the full-scale institutional capture that a narco-state label would imply.

This interpretation is reinforced by the role that plutocratic dynamics play in this configuration. The perceived articulation between narco-capitalism and democratic erosion documented here is not simply, in citizens’ accounts, a product of criminal actors operating outside state structures, but appears instead linked in perception to the intersection between criminal economies and existing patterns of elite capture of Ecuadorian institutions—what Durand (2019) analyzes in terms of plutocratic capture of regulatory and judicial agencies, extended here to the specific context of criminal-economic articulation. The ANN finding that Narco-Capitalism is dominant over State Reconfiguration and Criminal Territorial Governance in predicting Democratic Erosion suggests that it is the structural, political-economy dimension of this articulation — rather than its securitized or locally governed expressions—that most directly shapes how citizens perceive the erosion of democratic conditions. This is consistent with Jessop’s (2008) framework of strategic selectivity: plutocratic and criminal interests converge around those state agencies and regulatory functions most consequential for capital accumulation, selectively reconfiguring their operation while leaving others comparatively intact.

6.5 Limitations

Several limitations constrain the interpretation and generalizability of these findings. First, the cross-sectional, single-source design means that all associations are strictly perceptual and contemporaneous; no inference about temporal precedence or causal directionality is warranted. The common method bias controls applied—varied item ordering, anonymity assurance, internal VIF < 3.3, and Harman’s single-factor test—substantially reduce but cannot eliminate the risk of inflated associations due to shared method variance (Podsakoff et al., 2003). Second, the non-probabilistic convenience sample, recruited digitally, underrepresents rural populations, older adults, and communities with limited internet access—precisely those groups most likely to live in territories of high criminal governance concentration. Third, the Democratic Erosion construct, as operationalized here, captures experiential and institutional dimensions of erosion but not its participatory and competitive dimensions—a limitation that partially accounts for the non-significant Criminal Territorial Governance → Democratic Erosion path discussed in Section 6.3. Fourth, the ANN architecture and sensitivity analysis method, while providing a non-linear complement to PLS-SEM, involve design choices—number of hidden layers, activation functions, hold-out split ratio—that affect results and cannot be fully optimized with a sample of n = 400. Future research should replicate the hybrid PLS-SEM + ANN approach with larger, probability-based samples and an expanded Democratic Erosion construct that includes participatory and competitive dimensions.

7 Conclusion

This study proposed and tested a structural model linking narco-capitalism, coercive state reconfiguration, criminal territorial governance, and democratic erosion in Ecuador between 2021 and 2025, using a hybrid two-stage analytical strategy that combined PLS-SEM with an Artificial Neural Network applied to a cross-sectional survey of 400 adult citizens across the country’s four natural regions.

Four findings stand out. First, perceived narco-capitalism emerged as the dominant predictor of perceived democratic erosion across both analytical stages—with the largest direct path coefficient (β = 0.584), the largest effect size (f² = 0.481), and the highest relative importance in the ANN sensitivity analysis (M = 0.6751, SD = 0.0958 across 20 seeded runs). This supports H1 and H2, confirms the direct component of H5, and indicates that citizens’ perception of institutional collusion, legal–illegal gray zones, and the reconfiguration of state authority in territories affected by illicit economies is the construct most strongly associated with citizens’ perception of the deterioration of democratic conditions.

Second, coercive state reconfiguration functioned as a significant partial mediator between narco-capitalism and democratic erosion (indirect β = 0.171, p < 0.001), confirming H3 and the state-channel component of H5. This finding suggests that securitization may operate as a meaningful institutional mechanism through which narco-capitalist dynamics are associated with citizens’ democratic experience, even if its independent contribution is small relative to the direct narco-capitalism path.

Third, criminal territorial governance did not produce a statistically significant direct effect on democratic erosion (β = 0.085, p = 0.113), leading to the non-confirmation of H4 and of the Criminal Territorial Governance mediation channel in H5. Rather than dismissing this result, the study interprets it as a measurement and construct-validity signal: the Democratic Erosion operationalization used here captures experiential and institutional dimensions of erosion but likely underrepresents participatory and competitive dimensions—electoral manipulation, civic space contraction, candidate coercion—through which criminal governance most directly affects democratic functioning.

Fourth, the hybrid PLS-SEM + ANN strategy demonstrated confirmatory rather than incremental analytical value: the ANN’s mean R² of 0.6945 (SD = 0.0922 across 20 independently seeded training runs) on the held-out test set was comparable to, rather than exceeding, the PLS-SEM’s R² of 0.710 — indicating that the relationships among the three predictor constructs are predominantly linear. The sensitivity analysis nonetheless clarified the relative weight of each predictor in a way that path coefficients alone do not, cross-validating the PLS-SEM ranking through an independent, non-linear estimation technique.

Taken together, these findings support an interpretation of citizens’ perceptions of Ecuador’s contemporary security crisis as reflecting a process of partial and contested state reconfiguration under narco-capitalist dynamics—closer, in how citizens experience it, to hybrid statehood, segmented institutional criminalization, and plutocratic-criminal articulation than to the totalizing narco-state category that public and political discourse has sometimes invoked. The citizen perceptual data suggest that what appears most strongly associated with perceived democratic erosion is not the presence of criminal governance in specific territories, nor securitization alone, but the perceived normalization of institutional collusion and gray-zone economic activity—a pattern that, if it reflects underlying dynamics accurately, would call for structural, political-economy responses alongside, and not instead of, security-focused interventions.

Future research should address four specific gaps identified by this study: probability-based sampling strategies that over-represent high-criminal-governance territories; an expanded Democratic Erosion construct that includes participatory and competitive dimensions; longitudinal designs that permit analysis of how these perceptual associations evolve as securitization deepens; and qualitative complementary work that examines the specific mechanisms through which narco-capitalist dynamics translate into the institutional collusion and gray-zone normalization that citizen surveys detect but cannot explain.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fpos.2026.1931291/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Stuart Colin Carr, Massey University, New Zealand

Reviewed by: Katharina Naswall, University of Canterbury, New Zealand

Nilton Marques De Oliveira, Federal University of Tocantins, Brazil

Carlos Alfonso Laverde Rodriguez, Universidad Militar Nueva Granada, Colombia