We utilized machine learning techniques to better understand the relationship between self-reported depressive symptoms and new mental health diagnosis among 2SLGBTQ + youth. Using random forest, both predictive models shared nine of their top ten variables in predicting their respective outcomes. These shared top ten variables were: SRMH, ACE, depressive symptoms, age, internalized homophobia, outness, community connectedness, enacted stigma, and perceived sexual stigma. Two other factors were not shared by the two outcomes, suicidal thought was unique in predicting depressive symptoms, while stressful life was unique in predicting new mental health diagnosis. These results were complemented by the results from the analysis performed using the LASSO algorithm, where similar sets of variables were selected. Of note, seeking help from an HCP at baseline was often selected by the LASSO to be associated with new mental health diagnosis, but not by the random forest model.
The results of help seeking behaviours being an important predictor of new diagnosis is noteworthy. As shown in Appendix 1, the youths were asked about their intention to seek help at baseline, which they may or may not have done between the baseline to the 1-year follow-up (see Appendix 1 for details). In order to receive a mental health diagnosis, they will have to have sought help from a healthcare provider. Hence, it is important to contextualize this result within the framework of access to care; individuals who did not have new diagnoses may not be interpreted as individuals who were not facing mental health problems. Instead, this might be because they do not wish to seek help or face barriers to care, which is common in more marginalized communities, albeit their higher rates of help-seeking behaviours relative to heterosexual counterparts.(22, 38) Previous work from this same data also identified predictors of mental health seeking behaviours, some of which overlap with the current findings, but also identified the importance of sociodemographic variables such as sexual orientation and gender identity.(10)
Model performances varied based on techniques and iterations of imputations, however, the selected variables generally remained consistent. The LASSO selected other predictors that were not selected by the random forest models, but generally results were consistent. While one’s identity—for example, across intersections of sexual orientation, ethnicity, age—were hypothesized to be important, these variables were often not selected to be the most important predictors by the random forest model. A few of the interaction terms with these demographics appeared significant, albeit not strong. These analyses may suggest that social processes that shape psychological well-being, such as internalized homophobia or outness, were more important in predicting one’s future mental health than their socio-demographic identities. This aligns with the minority stress model and underscores the need to address sexual stigma, and build community connectedness, to promote mental health and wellbeing among 2SLGBTQ + persons.(39, 40)
Our machine learning techniques uncovered some unique non-linear associations between the predictors and outcomes (e.g., ACE and depressive symptoms), however the patterns of some of these complex associations are still not well understood. Certain measures with a score range of 5 (e.g., SRMH and stressful life) were treated as continuous; hence, we were unable to examine the range of variations that exist in these patterns. Furthermore, a few of the unique non-linear relationships were still difficult to interpret, and it might have been attributable to the smaller sample size in the second analysis. For example, Fig. 2 may suggest that high levels of outness can also increase depressive symptoms, which might be counterintuitive, but might have also as a result of overfitting due to the smaller sample size among those with the highest overall outness score. Theoretically, it could be that higher outness is linked with increased stigma exposure, or differential treatment across social dimensions of education, employment and relationships, which in turn could be linked with depressive symptoms. This might be improved in future studies by using a wider range of scales and larger sample sizes to fully understand the complex non-linear relationships that may exists in these associations.
Our study was also unable to detect significant interactions that were expected from the literature, particularly with the social identities. Unlike some previous studies, we preserved socio-demographic categories as much as possible in our analyses whenever possible (see Appendix 1 and Table 1). For example, rather than classifying everyone who are not bisexual or gay as “other”, we put them all in their self-identified categories. This may have made it more difficult to detect interactions, but it is important to distinguish these groups as an increasing number of youths are choosing these non-traditional identities (such as queer).(40, 41) In addition, as the initial survey was designed for a smoking cessation program, we might be missing important predictors for mental health (e.g., family support) which were not asked in the survey. Yet our findings did identify social processes as important predictors—including both marginalization (stigma) and protective factors (community connectedness)—which aligns with research on the salience of socio-political contexts and processes that shape 2SLGBTQ wellbeing.(39) To further improve model performance and detecting potential interactions, future studies will need larger sample sizes, as well as using quality data that have more mental health related predictor variables.
While our findings have identified some important predictors for new diagnosis of mental health problems and changes in depressive symptoms among 2SLGBTQ + youth that were consistent across different methods, future studies will need to validate these findings and further disentangle some of the non-linear associations observed results to fully understand the mechanism that leads to mental health challenges in this community. With a few exceptions, results match the existing literature on the known protective and risk factors against mental health problems among 2SLGBTQ + individuals from cross-sectional data. Clinicians and other stakeholders who work with 2SLGBTQ + youth may want to take these findings into account, especially since majority of the predictors identified are similar for both outcomes. With further validation of results in future studies, clinicians and schools may want to focus on the eight most important predictors shared by both outcomes, which will help reduce depressive symptoms and in turn reduce overall mental health diagnosis. Finally, some of the complex non-linear associations that remain not well understood will need to be further studied to find the optimal balance in protective and risk factors for these important mental health outcomes.