Suicidal behavior is one of the most critical and complex concepts associated with mental disorders. Approximately 800,000 deaths by suicide occur worldwide each year (World Health Organization, 2021). A substantially larger number of individuals experience suicide attempts, and an even greater proportion report suicidal ideation, a common yet difficult-to-predict phenomenon. Suicidal behavior encompasses a range of experiences, including passive and active ideation, attempts, and potential fatality as outcomes (Gvion and Apter, 2012). This complex phenomenon is prevalent across different demographic groups and is characterized by a multitude of risk and protective factors (Beghi et al., 2021; Evans et al., 2017; Nock et al., 2008). Despite ongoing efforts to understand the profiles of those at risk for suicide, accurately predicting suicide remains challenging (Belsher et al., 2019), due to the multitude of factors that influence risk. Previous suicide attempts and suicidal ideation and are key risk factors for suicide behavior and potential fatality, making them central targets in psychology research (Baptista et al., 2022; Harmer et al., 2022; May and Victor, 2018; Turecki et al., 2019).

There is still an ongoing debate concerning the latent structure of suicidal ideation, as evidence from prior research remains inconsistent regarding whether this phenomenon is best conceptualized as categorical or dimensional. Some studies have reported findings suggestive of a categorical latent structure (e.g., Orlando et al., 2015), whereas others have provided evidence in favor of dimensional models (e.g., Liu et al., 2015). Despite numerous investigations, this question remains unresolved. The present study aims to conduct a taxometric investigation of the latent structure of suicidal ideation in two non-clinical adult samples to contribute to this debate.

Classification systems are essential tools that help health professionals interpret the existence of symptoms and signs of various mental disorders and communicate in a universal technical language (Kendler, 2009). Since the first version of the Diagnostic and Statistical Manual of Mental Disorders (DSM; American Psychiatric Association, 1952), there has been a concerted effort to categorize mental health conditions. This initiative aims to facilitate the collection of reliable information about individuals in psychological assessment contexts and facilitate global exchange by establishing organized criteria for diagnoses, which were – and still are – primarily categorical, focusing on the presence or absence of a disorder.

However, more recent approaches have emphasized the need to rethink the comprehension of mental disorders based on dimensional models (Conway et al., 2019; Hopwood et al., 2018). These models guide the understanding of disorders through the quantitative gradation of symptoms, which can vary in intensity on a continuum that ranges from healthy to pathological (American Psychiatric Association, 2022; Kotov et al., 2017). Several researchers favor the dimensional perspective on the taxonomy of mental disorders (Brown and Barlow, 2005; Krueger and Bezdjian, 2009; Widiger and Samuel, 2005). Achenbach and Edelbrock (1983) were precursors of dimensional approaches, indicating that assessment scales consistent with dimensional proposals could favor a more complete and qualitative view of human phenomena. That is, they could provide enhanced clinical information (Achenbach et al., 2016). More recently, dimensional proposals gained prominence with recent representatives as the National Institute of Mental Health’s Research Domain Criteria (RDoC; Insel et al., 2010), the Hierarchical Taxonomy of Psychopathology (HiTOP; Kotov et al., 2017), the International Statistical Classification of Diseases and Related Health Problems (ICD-11; World Health Organization, 2018), and the DSM-5 alternative model for personality disorders (American Psychiatric Association, 2013).

Researchers have used taxometric procedures to investigate the latent structure of several psychopathologies. Understanding whether the latent structure of psychopathology is more suitable for categorical or dimensional models, can guide research and favor clinicians in their evaluations and interventions. Meta-analytical studies demonstrate the latent structures of most mental disorders are better explained by dimensional models (Haslam et al., 2012). In a more recent meta-analysis, Haslam et al. (2020) similarly found that the majority of the investigated constructs and psychopathologies show evidence of dimensionality. However, in this meta-analysis, suicide risk was among the few psychopathologies in which results suggested plausible taxonicity (i.e., categorical latent structure). Despite this, the authors noted that the plausibility level was not very high, as the taxonic results were replicated only a single time. Moreover, in that review, suicide risk was operationalized broadly, encompassing heterogeneous indicators, including suicidal ideation, suicidal self-injury, and suicide risk under the same umbrella. A more specific meta-analysis focused exclusively on suicidal ideation reported a mean CCFI value of 0.48, indicating an ambiguous latent structure (Siddaway et al., 2021). This meta-analysis revealed mixed findings across studies, with some suggesting dimensionality and others taxonicity. The authors further emphasized the limited number of available studies and the inconsistency of the results, concluding that it is premature to draw definitive conclusions regarding the latent structure of suicidal thoughts and that additional research is needed to clarify this issue.

Determining the latent structure of suicidal ideation is critical for understanding whether individuals experiencing suicidal thoughts should be conceptualized as belonging to a distinct subgroup (i.e., a taxon) or whether suicidal ideation is better understood as varying continuously in severity across individuals. These alternative conceptualizations have important implications for clinical practice, particularly with respect to the use of diagnostic thresholds, risk stratification, intervention, and prevention decisions. A categorical model emphasizes the identification of cutoffs to distinguish individuals who “have” versus “do not have” suicidal ideation, whereas a dimensional model suggests that suicidal thoughts may be present to varying degrees and warrant clinical attention across the full continuum (Haslam et al., 2020). Moreover, latent structure has implications beyond cross-sectional classification. A dimensional structure would suggest that suicidal ideation reflects a gradual accumulation of risk, supporting models that emphasize early identification and intervention even at lower levels of ideation severity. In contrast, a categorical structure would imply a qualitative shift in psychological functioning, consistent with threshold-based identification of a high-risk subgroup. Clarifying this distinction may also inform efforts to differentiate individuals who experience suicidal ideation but never attempt suicide from those who progress to suicidal behavior (Gvion and Apter, 2012; Klonsky et al., 2018), thereby contributing to more precise risk stratification and targeted prevention strategies.

Before proceeding with the taxometric analyses, we assessed the validity and adequacy of the indicators. Table 1 presents the validity (Cohens d) and descriptive statistics (mean, standard deviation, skewness, and kurtosus) of the indicators in both the BSS and the SIDAS datasets. All indicators of both datasets obtained adequate skewness and kurtosis estimates. Cohen’s d-values were acceptable (≥1.25) for all indicators in the BSS dataset. In contrast, two indicators in the SIDAS dataset (items 1 and 2) fell below the required threshold and were therefore excluded from subsequent analyses. Consequently, taxometric analyses for the SIDAS dataset were conducted using items 3 through 5.

After attesting the adequacy and validity of the indicators, we proceed with the taxometric analyses. Figures 1, 2 depict the taxometric analyses results for the BSS and SIDAS datasets, respectively. Table 2 presents the CCFI values for the three taxometric methods used in this study (MAMBAC, MAXEIG, and L-Mode), as well as the mean CCFI across methods.

For all taxometric methods, the gray lines in Figures 1, 2 represent the shape of the simulated taxonic and dimensional data, whereas the solid black line represents the empirical data. Visual inspection of the BSS results suggests that, in the MAMBAC method, the empirical curve more closely resembles the simulated taxonic curve than the simulated dimensional curve. In contrast, for MAXEIG and L-Mode, the empirical curves appear more similar to the simulated dimensional curves. A comparable pattern was observed for the SIDAS dataset: visual inspection suggests greater similarity to the simulated taxonic curve in MAMBAC, but closer resemblance to the simulated dimensional curves in MAXEIG and L-Mode. However, visual inspection is inherently limited by its subjectivity and the absence of objective thresholds for determining the degree of similarity between empirical and simulated curves. Therefore, we relied primarily on the CCFI values to interpret the results. As shown in Table 2, the CCFI values were generally consistent with the visual impressions, with MAMBAC suggesting a structure closer to dimensionality and MAXEIG and L-Mode suggesting a structure closer to taxonicity across both datasets. Nevertheless, most CCFI values (except for MAXEIG) fell within the ambiguous range (0.40–0.60), impeding definitive conclusions regarding the latent structure. Importantly, the mean CCFI values also fell within this ambiguous interval, with values of 0.46 for the BSS dataset and 0.45 for the SIDAS dataset, further supporting an interpretation of structural ambiguity.

To further examine the latent structure of the data, we generated CCFI profiles, which display CCFI values derived from MAMBAC, MAXEIG, and L-Mode across a range of putative taxon base rates. Figures 3, 4 present the CCFI profiles for the BSS and SIDAS datasets, respectively.

In both cases, visual inspection of the profiles indicates that CCFI values across most putative base rates predominantly fall within the ambiguous range for both datasets. More importantly, the aggregate CCFI values also remained within this ambiguous interval. For the BSS dataset, the mean aggregate CCFI profile was 0.57, with method-specific values of 0.58 for MAMBAC, 0.57 for MAXEIG, and 0.57 for L-Mode. For the SIDAS dataset, the mean aggregate CCFI profile was 0.43, with corresponding values of 0.40 for MAMBAC, 0.40 for MAXEIG, and 0.51 for L-Mode.

We aimed to investigate the nature of the latent structure of suicidal ideation. To contribute with empirical evidence to the ongoing debate about this topic, we leveraged data from two psychological instruments assessing suicidal ideation and two distinct nonclinical adult samples to taxometrically examine the latent structure of suicidal ideation. However, the taxometric ambiguity that sustains discussions about suicidal ideation also permeated our results, which were inconclusive. For both the BSS and SIDAS datasets, results consistently fell within the ambiguous interval. This pattern also persisted when testing the robustness of the findings through sensitivity analyses involving alternative methodological choices. The uncertain taxonicity or dimensionality of suicidal ideation poses two venues of discussion, one about the suicidal ideation itself, and the other about the applicability of the taxometric analysis. We prompt researchers to exercise caution in drawing robust inferences regarding our findings, as it was not possible to conclusively discriminate between categorical and dimensional constructs (Ruscio et al., 2011).

Ambiguous results in taxometric analysis indicate that both models, categorical and dimensional, presented a comparable (good or bad) fit to the data, making conclusions about the latent structure of a construct hard to reach (Ruscio et al., 2011). Importantly, such ambiguity should not be interpreted as a lack of empirical information. Rather, when ambiguity is observed consistently across methods, samples, and analytic decisions, it constitutes meaningful evidence that the construct may not conform cleanly to either a categorical or a purely dimensional structure. In the present study, the ambiguous pattern of suicide ideation was consistently observed across different measurement instruments, samples, and sensitivity analyses. Importantly, these findings are unlikely to reflect methodological artifacts, as the analyses followed best-practice recommendations in the taxometric literature (e.g., consistency testing, simulation techniques, and the use of averaged CCFI values; Ruscio et al., 2007, 2010, 2018). This convergence across samples and analytic approaches suggests that the observed ambiguity is unlikely to be attributable to sample-specific characteristics, measurement limitations, or methodological artifacts. Instead, our findings are consistent with prior taxometric studies reporting ambiguous or divergent results for suicidal ideation and related constructs (e.g., Baptista et al., 2019; Haslam et al., 2020; Siddaway et al., 2021), and may reflect the inherent complexity of suicide ideation as a psychological phenomenon.

If our assumption is correct, that is, the ambiguous findings are not due to measurements, samples, and statistical analyses, then it may be an inherent characteristic of the phenomenon, or at least of the way we are defining what suicidal ideation is. Suicidal ideation is a complex phenomenon encompassing multiple distinct characteristics (Reeves et al., 2022), such as passive thoughts of death, active desire for suicide, and preparatory or planning-related thoughts. The heterogeneity of these ideational components suggests that suicidal ideation may not be characterized by a single, uniform latent structure. When different ideational facets are examined jointly, they may reflect varying underlying structures, which could contribute to taxometric ambiguity. For example, some aspects of suicidal ideation, such as frequency or intensity of suicidal thoughts, may vary dimensionally, whereas other components, such as the emergence of active desire or preparatory ideation, may reflect more discontinuous shifts. Consequently, when suicidal ideation is treated as a unified phenomenon comprising multiple ideational components, taxometric analyses may fail to clearly favor either a categorical or dimensional model, resulting in ambiguous findings. Because suicidal ideation was treated as a unified phenomenon in our analyses, this may partly explain why our findings did not yield a clear latent structure.

Given the persistence of ambiguous and non-concordant findings regarding the latent structure of suicidal ideation, an alternative conceptual and analytic approach may be offered by the network perspective (Borsboom and Cramer, 2013; Borsboom et al., 2021). Network models allow discrete and continuous variables to be examined simultaneously and do not require assumptions about an underlying latent variable. From this perspective, suicidal ideation is conceptualized neither as a categorical nor as a dimensional phenomenon, but rather as a dynamic complex system in which symptoms and related characteristics mutually influence one another over time (Blanken et al., 2018). For a detailed discussion of the application of network analysis to suicidal ideation, see de Beurs et al. (2021).

Notwithstanding this study’s contributions, some methodological limitations should be considered when interpreting our findings. First, we performed this study with two convenient non-clinical samples, which can compromise the generalizability of our findings, as such samples may include relatively few individuals at elevated suicide risk. However, it is important to note that a substantial proportion of participants in both samples self-reported suicidal ideation and a history of suicidal behavior. Second, our data collection was conducted online, which can lead to a relatively high number of invalid and careless responses (Choi et al., 2017). Third, we used two specific measures of suicidal ideation (SIDAS and BSS) and a corresponding set of indicators, which were selected based on the researchers’ familiarity with these instruments. However, the choice of these measures was largely driven by convenience, and the use of alternative suicidal ideation instruments might have yielded different results. We suggest that future studies address these limitations and consider investigating individual components of suicidal ideation separately. In addition, researchers may benefit from applying alternative methodological approaches, such as network analysis, to better capture the complexity of this phenomenon.

Despite the highlighted limitations, we believe our study has two main strengths. First, although our findings do not provide conclusive evidence regarding the underlying nature of suicide ideation, they offer valuable insights for future investigations focused on its latent structure. Specifically, suicidal ideation may represent a complex phenomenon encompassing multiple components, each potentially characterized by different latent structures. Alternatively, conceptualizing suicidal ideation strictly as either categorical or dimensional may not be the most appropriate approach, and researchers may benefit from considering alternative frameworks for studying psychopathology. As such, the question of whether suicidal ideation is best understood as dimensional or categorical remains unresolved, underscoring the need for caution against premature conclusions favoring either a taxonic or dimensional interpretation. Second, our study appears to be the first, to our knowledge, to be conducted with adults in a Latin American sample, expanding results typically relying on White, educated, industrialized, rich, and democratic (WEIRD) samples. Most studies comprise samples from North America and Europe; these characteristics do not necessarily represent the world population (Azar, 2010). Therefore, our study also contributes to investigating the suicide ideation phenomenon in samples with different sociodemographic characteristics, demonstrating that the ambiguous results reported in the literature so far may not be related to these sample features.