As a former practitioner who worked with adolescents in acute psychological distress, I read with great interest the study recently published by Bitar et al. (Reference Bitar, Mendes, Lopes Ferreira, Samuel and Catunda2026). The authors argue that “traditional approaches to understanding youth suicide risk have often focused on individual risk behaviors (…) in isolation,” whereas adolescents’ habits “often form interconnected patterns, known as health lifestyles, which reflect both their personal choices and the social environments they navigate” (p. 1). This framing, combined with the authors’ strong emphasis on prevention and intervention, naturally raises expectations regarding both theoretical innovation and practical relevance.
Unfortunately, the study does not fully deliver on its stated aims. While the analyses are presented confidently, the prevention implications are stronger than what their results can support. Moreover, key methodological decisions likely undermine the stability and interpretability of the latent class analysis (LCA) on which the applied recommendations rest. I elaborate on these points below.
Overstated prevention implications
Throughout the manuscript, the authors repeatedly invoke prevention and intervention, presenting their findings as offering “practical guidance for prioritizing prevention resources” and enabling clinicians to “focus attention on adolescents most likely to benefit from targeted support” (p. 8). The authors also state that “screening tools, health education programs and counseling services can be more effective when tailored to the specific behavior profiles we identified” (p. 1).
Clinicians and other practitioners are generally attached to optimized and pragmatic approaches. They would like to know, for instance, whether the five “health profiles” identified by the authors represent more efficient screening indicators of suicidality than depression or parental psychiatric history. In that context, it is unclear why the authors excluded anxio-depressive symptoms from the regression models, despite their availability in the dataset, as evidenced by other publications using the same data (Brisson et al., Reference Brisson, Mendes and Catunda2023). The authors adjusted for several factors “to account for potential confounding” (p. 3). However, depression and anxiety are among the strongest and most proximal predictors of adolescent suicidality (Orri et al., Reference Orri, Galera, Turecki, Forte, Renaud, Boivin, Tremblay, Côté and Geoffroy2018), making their omission difficult to justify within the authors’ own adjustment strategy. Moreover, the behaviors included in the LCA are well-established correlates or potential consequences of internalizing distress. Without accounting for anxio-depressive symptoms, it remains unclear whether the reported associations between lifestyle classes and suicidality reflect independent behavioral configurations or simply underlying psychological vulnerability. This omission introduces substantial risk of omitted-variable bias and renders interpretation of the class “effects” ambiguous. Including anxio-depressive symptoms could meaningfully alter the results – either attenuating the observed associations or strongly confirming their independence. Because the authors did not demonstrate that lifestyle is independently associated with suicidality beyond established psychological predictors, claims regarding screening and prevention remain ill-grounded.
Premature prevention inference
Although the cross-sectional design is acknowledged as a limitation (p. 8), its implications for causal direction and prevention inference are not fully integrated into the authors’ applied claims. Using the profiles identified by the authors in prevention frames would be premature, as the study provides no evidence of predictive validity. The latent classes were not validated longitudinally, nor were sensitivity, specificity or positive predictive values assessed. There is also no demonstration that the identified classes are stable across time, contexts or plausible alternative model specifications.
Given the extensive dichotomization of indicators (see below) and the exploratory nature of the latent class solution, relatively minor analytical modifications could plausibly yield a different class structure, further limiting the robustness of the proposed “profiles” as targets for prevention. For instance, several relevant variables available in the dataset have been excluded, including cannabis use and dietary indicators such as the consumption of sweets and sugar-sweetened beverages. This omission is particularly puzzling given that the authors explicitly refer in their theoretical framing to associations between high consumption of sugary drinks and suicidality (p. 2). The rationale for excluding these variables is not provided, nor are sensitivity analyses reported to assess whether the latent class structure is robust to alternative indicator sets. Furthermore, because temporality cannot be established, the study does not demonstrate that the identified behavioral configurations represent antecedent risk patterns rather than consequences or concurrent manifestations of psychological distress, rendering prevention recommendations potentially unwarranted.
Practical and temporal limitations of the proposed screening strategy
The authors’ proposal of repeated universal prevention – specifically, the suggestion that adolescents be screened on an annual basis (p. 8) – raises important questions regarding feasibility and resource allocation that are not addressed in the manuscript. In many countries and school systems, practitioners simply do not have the resources (staff, time and funding) to interview or screen every adolescent in a school, city or region once per year. Presenting such an approach as a concrete recommendation based on the study findings disregards the practical constraints under which prevention and mental health services operate.
Moreover, the temporal dynamics of suicidality further complicate the authors’ recommendations. Suicidal ideation and attempts can emerge rapidly, sometimes within a few months following acute stressors such as severe bullying or sexual victimization. Annual universal screening based on relatively static “health lifestyles” may therefore fail to identify adolescents who transition into acute risk between assessment points.
Finally, clinical experience consistently shows that a substantial proportion of adolescents in distress are reluctant to disclose sensitive information and may misreport health behaviors (Mirichlis et al., Reference Mirichlis, Burke, Bettis, Dayer and Fox2025).
Massive dichotomization of ordinal variables
Perhaps the most damaging methodological decision in the study is the systematic dichotomization of ordinal variables. This practice, long criticized in the methodological literature (e.g., MacCallum et al., Reference MacCallum, Zhang, Preacher and Rucker2002), entails substantial loss of information, reduced variability and an increased risk of misclassification.
Here, dichotomization is particularly problematic because there is no theoretical foundation for considering as equivalent: (1) daily smokers and adolescents who smoked two cigarettes at a party in the past month; (2) weekly binge drinking with friends and the consumption of a single glass of wine at a family event in the past month; and (3) adolescents reporting no physical activity at all and those reporting a level just below the cutoff used by the authors. Treating such markedly different behaviors as interchangeable collapses meaningful distinctions that are central to the concept of “health lifestyles.”
This issue is further compounded in the case of physical activity. The cutoff adopted by the authors reflects a strict interpretation of earlier recommendations, whereas current guidance from the World Health Organization emphasizes that “children and adolescents should do at least an average of 60 min per day of moderate- to vigorous-intensity, mostly aerobic, physical activity, across the week” (WHO, n.d.). Because the authors measured daily activity rather than weekly averages, their operationalization may be overly restrictive and of questionable validity.
The consequences of such aggressive dichotomization are far-reaching. First, it obscures meaningful variation within the sample, masking important differences between subgroups of adolescents. Second, it can inflate or attenuate associations between variables, producing misleading impressions of which behaviors cluster together in so-called “health lifestyles.” Third, it increases the risk of misclassification bias, whereby adolescents with substantially different behavioral profiles are treated as equivalent. Fourth, it questions the interpretation of the classes “high substance use” and “high alcohol use,” as, for instance, having one alcoholic drink over the past month is not reflective of such a high use, especially compared to weekly or daily drinking. Finally, because these dichotomized indicators serve as the foundation for all subsequent analyses, any bias introduced at this stage propagates throughout the study, calling into question the reliability of the identified classes and the conclusions drawn from them.
Latent class model misspecification and overextraction
The local (conditional) independence assumption – a core premise of LCA (Lee et al., Reference Lee, Jung and Park2020) – is neither tested nor discussed, despite the inclusion of multiple closely related behaviors that are likely to remain correlated within classes. When this assumption is violated, LCA commonly compensates by extracting additional classes to absorb residual associations, leading to overextraction and distorted class profiles.
Moreover, class enumeration decisions raise red flags. Although the Bayesian Information Criterion favored a three-class solution, the authors selected a five-class model, citing lower AIC values and “substantive interpretability.” Entropy and posterior probabilities were also invoked, despite the fact that these indices do not justify the number of classes or diagnose model misspecification. Notably, the average posterior classification probabilities were higher for the three-class solution than for the retained five-class model, offering no empirical support for the added complexity. In sum, no evidence is provided that the additional classes reflect qualitatively distinct subgroups rather than artifacts of dichotomization and unmodeled dependence.
Transparency and reporting deficiencies
First, the process of latent class enumeration is insufficiently documented. The authors state that “although the BIC was lowest for the three-class model, we selected the five-class solution based on a combination of statistical (e.g., lower AIC) and substantive interpretability criteria” (p. 3). However, the three-class solution – despite being favored by the BIC and reportedly exhibiting substantially better posterior classification probabilities – is not presented, either in the main text or in a Supplementary Material. The same applies to the other unselected solutions. As a result, readers are unable to assess whether the retained five-class model represents a meaningful qualitative improvement over simpler alternatives.
Second, the authors report that “the distribution of each variable was assessed, and outliers, coding errors, missing data and multicollinearity were checked” (p. 3), yet provide no information on how outliers or coding errors were defined, detected or handled, nor do they report any multicollinearity diagnostics. The number of affected cases is not indicated, and no sensitivity analyses are presented. Such statements, therefore, remain largely declarative and do not allow readers to evaluate the rigor of the data-screening process or its potential impact on the results.
Third, although the authors report the proportion of missing values for each variable and state that missing data were handled using the expectation–maximization algorithm implemented in the poLCA package (p. 3), no justification is provided for the implicit assumption that data are missing at random. It would have been informative to present, in the Supplementary Material, results from analyses using listwise deletion or alternative missing-data strategies to assess the sensitivity of the findings to different assumptions.
Taken together, these reporting limitations do not merely reflect omissions of secondary details; they materially constrain the reader’s ability to assess the validity, robustness and generalizability of the study’s conclusions.
The added value of the “health lifestyle” approach
The authors’ invocation of health lifestyle theory (Cockerham, Reference Cockerham2005) is conceptually problematic. In that tradition, lifestyles are not clusters of co-occurring behaviors but socially structured patterns of practice shaped by social position, stratification, collective norms and enduring dispositions. “Health lifestyle” refers to patterned action emerging from the interplay between structure and agency – not to statistical groupings derived from response similarity.
In the present study, however, “health lifestyles” are operationalized exclusively as latent classes based on seven dichotomized behavioral indicators. No social mechanisms, structural constraints or dispositional processes are modeled. Although the introduction emphasizes socio-ecological frameworks and the role of social environments, these dimensions are not incorporated into the latent class specification. What is ultimately identified are statistical profiles of behaviors – not theoretically grounded, socially embedded lifestyles. Equating these two constructs conflates a sociological theory of patterned action with a data-reduction technique.
Moreover, the added empirical value of this approach remains limited here. A substantial body of literature has already established that substance use and problematic social media engagement are independently associated with adolescent suicidality. The present study, through LCA, primarily reorganizes these well-documented associations into behavioral profiles. Because these risk behaviors do not perfectly overlap – some adolescents engage in substance use without problematic social media use, and others show the reverse – the identification of distinct classes combining these factors in varying configurations is statistically foreseeable. Thus, the findings largely confirm patterns that variable-centered analyses have already identified, without demonstrating that the person-centered reconfiguration generates substantively new insights, clarifies mechanisms or meaningfully extends explanatory understanding. In this context, the “health lifestyle” framing appears more descriptive and terminological than transformative.
Conclusion
These issues undermine the study’s central claims. Rather than offering a breakthrough in understanding adolescent health lifestyles or informing prevention, the study risks reifying statistically unstable groupings and overstating their practical relevance.
The findings should be interpreted with considerable caution. As currently presented, the study does not convincingly demonstrate that it provides a reliable basis for prevention. The authors’ impact statement that “screening tools, health education programs and counseling services can be more effective when tailored to the specific behavior profiles we identified” (p. 1) should be considered as overconfident and exceeding what their data can reasonably support.
I hope this comment provides the authors with an opportunity to clarify several methodological decisions (e.g., item selection and dichotomization) and to present additional information and analyses that would help readers better assess the robustness of the findings (e.g., the three-class model, sensitivity analyses and the definition and handling of outliers).
This is an Open Access article, distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives licence (http://creativecommons.org/licenses/by-nc-nd/4.0), which permits non-commercial re-use, distribution, and reproduction in any medium, provided that no alterations are made and the original article is properly cited. The written permission of Cambridge University Press or the rights holder(s) must be obtained prior to any commercial use and/or adaptation of the article.
You have
Access
Open access
As a former practitioner who worked with adolescents in acute psychological distress, I read with great interest the study recently published by Bitar et al. (Reference Bitar, Mendes, Lopes Ferreira, Samuel and Catunda2026). The authors argue that “traditional approaches to understanding youth suicide risk have often focused on individual risk behaviors (…) in isolation,” whereas adolescents’ habits “often form interconnected patterns, known as health lifestyles, which reflect both their personal choices and the social environments they navigate” (p. 1). This framing, combined with the authors’ strong emphasis on prevention and intervention, naturally raises expectations regarding both theoretical innovation and practical relevance.
Unfortunately, the study does not fully deliver on its stated aims. While the analyses are presented confidently, the prevention implications are stronger than what their results can support. Moreover, key methodological decisions likely undermine the stability and interpretability of the latent class analysis (LCA) on which the applied recommendations rest. I elaborate on these points below.
Overstated prevention implications
Throughout the manuscript, the authors repeatedly invoke prevention and intervention, presenting their findings as offering “practical guidance for prioritizing prevention resources” and enabling clinicians to “focus attention on adolescents most likely to benefit from targeted support” (p. 8). The authors also state that “screening tools, health education programs and counseling services can be more effective when tailored to the specific behavior profiles we identified” (p. 1).
Clinicians and other practitioners are generally attached to optimized and pragmatic approaches. They would like to know, for instance, whether the five “health profiles” identified by the authors represent more efficient screening indicators of suicidality than depression or parental psychiatric history. In that context, it is unclear why the authors excluded anxio-depressive symptoms from the regression models, despite their availability in the dataset, as evidenced by other publications using the same data (Brisson et al., Reference Brisson, Mendes and Catunda2023). The authors adjusted for several factors “to account for potential confounding” (p. 3). However, depression and anxiety are among the strongest and most proximal predictors of adolescent suicidality (Orri et al., Reference Orri, Galera, Turecki, Forte, Renaud, Boivin, Tremblay, Côté and Geoffroy2018), making their omission difficult to justify within the authors’ own adjustment strategy. Moreover, the behaviors included in the LCA are well-established correlates or potential consequences of internalizing distress. Without accounting for anxio-depressive symptoms, it remains unclear whether the reported associations between lifestyle classes and suicidality reflect independent behavioral configurations or simply underlying psychological vulnerability. This omission introduces substantial risk of omitted-variable bias and renders interpretation of the class “effects” ambiguous. Including anxio-depressive symptoms could meaningfully alter the results – either attenuating the observed associations or strongly confirming their independence. Because the authors did not demonstrate that lifestyle is independently associated with suicidality beyond established psychological predictors, claims regarding screening and prevention remain ill-grounded.
Premature prevention inference
Although the cross-sectional design is acknowledged as a limitation (p. 8), its implications for causal direction and prevention inference are not fully integrated into the authors’ applied claims. Using the profiles identified by the authors in prevention frames would be premature, as the study provides no evidence of predictive validity. The latent classes were not validated longitudinally, nor were sensitivity, specificity or positive predictive values assessed. There is also no demonstration that the identified classes are stable across time, contexts or plausible alternative model specifications.
Given the extensive dichotomization of indicators (see below) and the exploratory nature of the latent class solution, relatively minor analytical modifications could plausibly yield a different class structure, further limiting the robustness of the proposed “profiles” as targets for prevention. For instance, several relevant variables available in the dataset have been excluded, including cannabis use and dietary indicators such as the consumption of sweets and sugar-sweetened beverages. This omission is particularly puzzling given that the authors explicitly refer in their theoretical framing to associations between high consumption of sugary drinks and suicidality (p. 2). The rationale for excluding these variables is not provided, nor are sensitivity analyses reported to assess whether the latent class structure is robust to alternative indicator sets. Furthermore, because temporality cannot be established, the study does not demonstrate that the identified behavioral configurations represent antecedent risk patterns rather than consequences or concurrent manifestations of psychological distress, rendering prevention recommendations potentially unwarranted.
Practical and temporal limitations of the proposed screening strategy
The authors’ proposal of repeated universal prevention – specifically, the suggestion that adolescents be screened on an annual basis (p. 8) – raises important questions regarding feasibility and resource allocation that are not addressed in the manuscript. In many countries and school systems, practitioners simply do not have the resources (staff, time and funding) to interview or screen every adolescent in a school, city or region once per year. Presenting such an approach as a concrete recommendation based on the study findings disregards the practical constraints under which prevention and mental health services operate.
Moreover, the temporal dynamics of suicidality further complicate the authors’ recommendations. Suicidal ideation and attempts can emerge rapidly, sometimes within a few months following acute stressors such as severe bullying or sexual victimization. Annual universal screening based on relatively static “health lifestyles” may therefore fail to identify adolescents who transition into acute risk between assessment points.
Finally, clinical experience consistently shows that a substantial proportion of adolescents in distress are reluctant to disclose sensitive information and may misreport health behaviors (Mirichlis et al., Reference Mirichlis, Burke, Bettis, Dayer and Fox2025).
Massive dichotomization of ordinal variables
Perhaps the most damaging methodological decision in the study is the systematic dichotomization of ordinal variables. This practice, long criticized in the methodological literature (e.g., MacCallum et al., Reference MacCallum, Zhang, Preacher and Rucker2002), entails substantial loss of information, reduced variability and an increased risk of misclassification.
Here, dichotomization is particularly problematic because there is no theoretical foundation for considering as equivalent: (1) daily smokers and adolescents who smoked two cigarettes at a party in the past month; (2) weekly binge drinking with friends and the consumption of a single glass of wine at a family event in the past month; and (3) adolescents reporting no physical activity at all and those reporting a level just below the cutoff used by the authors. Treating such markedly different behaviors as interchangeable collapses meaningful distinctions that are central to the concept of “health lifestyles.”
This issue is further compounded in the case of physical activity. The cutoff adopted by the authors reflects a strict interpretation of earlier recommendations, whereas current guidance from the World Health Organization emphasizes that “children and adolescents should do at least an average of 60 min per day of moderate- to vigorous-intensity, mostly aerobic, physical activity, across the week” (WHO, n.d.). Because the authors measured daily activity rather than weekly averages, their operationalization may be overly restrictive and of questionable validity.
The consequences of such aggressive dichotomization are far-reaching. First, it obscures meaningful variation within the sample, masking important differences between subgroups of adolescents. Second, it can inflate or attenuate associations between variables, producing misleading impressions of which behaviors cluster together in so-called “health lifestyles.” Third, it increases the risk of misclassification bias, whereby adolescents with substantially different behavioral profiles are treated as equivalent. Fourth, it questions the interpretation of the classes “high substance use” and “high alcohol use,” as, for instance, having one alcoholic drink over the past month is not reflective of such a high use, especially compared to weekly or daily drinking. Finally, because these dichotomized indicators serve as the foundation for all subsequent analyses, any bias introduced at this stage propagates throughout the study, calling into question the reliability of the identified classes and the conclusions drawn from them.
Latent class model misspecification and overextraction
The local (conditional) independence assumption – a core premise of LCA (Lee et al., Reference Lee, Jung and Park2020) – is neither tested nor discussed, despite the inclusion of multiple closely related behaviors that are likely to remain correlated within classes. When this assumption is violated, LCA commonly compensates by extracting additional classes to absorb residual associations, leading to overextraction and distorted class profiles.
Moreover, class enumeration decisions raise red flags. Although the Bayesian Information Criterion favored a three-class solution, the authors selected a five-class model, citing lower AIC values and “substantive interpretability.” Entropy and posterior probabilities were also invoked, despite the fact that these indices do not justify the number of classes or diagnose model misspecification. Notably, the average posterior classification probabilities were higher for the three-class solution than for the retained five-class model, offering no empirical support for the added complexity. In sum, no evidence is provided that the additional classes reflect qualitatively distinct subgroups rather than artifacts of dichotomization and unmodeled dependence.
Transparency and reporting deficiencies
First, the process of latent class enumeration is insufficiently documented. The authors state that “although the BIC was lowest for the three-class model, we selected the five-class solution based on a combination of statistical (e.g., lower AIC) and substantive interpretability criteria” (p. 3). However, the three-class solution – despite being favored by the BIC and reportedly exhibiting substantially better posterior classification probabilities – is not presented, either in the main text or in a Supplementary Material. The same applies to the other unselected solutions. As a result, readers are unable to assess whether the retained five-class model represents a meaningful qualitative improvement over simpler alternatives.
Second, the authors report that “the distribution of each variable was assessed, and outliers, coding errors, missing data and multicollinearity were checked” (p. 3), yet provide no information on how outliers or coding errors were defined, detected or handled, nor do they report any multicollinearity diagnostics. The number of affected cases is not indicated, and no sensitivity analyses are presented. Such statements, therefore, remain largely declarative and do not allow readers to evaluate the rigor of the data-screening process or its potential impact on the results.
Third, although the authors report the proportion of missing values for each variable and state that missing data were handled using the expectation–maximization algorithm implemented in the poLCA package (p. 3), no justification is provided for the implicit assumption that data are missing at random. It would have been informative to present, in the Supplementary Material, results from analyses using listwise deletion or alternative missing-data strategies to assess the sensitivity of the findings to different assumptions.
Taken together, these reporting limitations do not merely reflect omissions of secondary details; they materially constrain the reader’s ability to assess the validity, robustness and generalizability of the study’s conclusions.
The added value of the “health lifestyle” approach
The authors’ invocation of health lifestyle theory (Cockerham, Reference Cockerham2005) is conceptually problematic. In that tradition, lifestyles are not clusters of co-occurring behaviors but socially structured patterns of practice shaped by social position, stratification, collective norms and enduring dispositions. “Health lifestyle” refers to patterned action emerging from the interplay between structure and agency – not to statistical groupings derived from response similarity.
In the present study, however, “health lifestyles” are operationalized exclusively as latent classes based on seven dichotomized behavioral indicators. No social mechanisms, structural constraints or dispositional processes are modeled. Although the introduction emphasizes socio-ecological frameworks and the role of social environments, these dimensions are not incorporated into the latent class specification. What is ultimately identified are statistical profiles of behaviors – not theoretically grounded, socially embedded lifestyles. Equating these two constructs conflates a sociological theory of patterned action with a data-reduction technique.
Moreover, the added empirical value of this approach remains limited here. A substantial body of literature has already established that substance use and problematic social media engagement are independently associated with adolescent suicidality. The present study, through LCA, primarily reorganizes these well-documented associations into behavioral profiles. Because these risk behaviors do not perfectly overlap – some adolescents engage in substance use without problematic social media use, and others show the reverse – the identification of distinct classes combining these factors in varying configurations is statistically foreseeable. Thus, the findings largely confirm patterns that variable-centered analyses have already identified, without demonstrating that the person-centered reconfiguration generates substantively new insights, clarifies mechanisms or meaningfully extends explanatory understanding. In this context, the “health lifestyle” framing appears more descriptive and terminological than transformative.
Conclusion
These issues undermine the study’s central claims. Rather than offering a breakthrough in understanding adolescent health lifestyles or informing prevention, the study risks reifying statistically unstable groupings and overstating their practical relevance.
The findings should be interpreted with considerable caution. As currently presented, the study does not convincingly demonstrate that it provides a reliable basis for prevention. The authors’ impact statement that “screening tools, health education programs and counseling services can be more effective when tailored to the specific behavior profiles we identified” (p. 1) should be considered as overconfident and exceeding what their data can reasonably support.
I hope this comment provides the authors with an opportunity to clarify several methodological decisions (e.g., item selection and dichotomization) and to present additional information and analyses that would help readers better assess the robustness of the findings (e.g., the three-class model, sensitivity analyses and the definition and handling of outliers).
Open peer review
To view the open peer review materials for this article, please visit http://doi.org/10.1017/gmh.2026.10243.
Data availability statement
Not applicable.
Author contributions
Seth A. Hun is the sole author of this article.
Funding statement
None.
Competing interest
The author declares none.