Depression is a significant pediatric public health concern in the USA as over three million adolescents experience a depressive episode annually (Weinberger et al., Reference Weinberger, Gbedemah, Martinez, Nash, Galea and Goodwin2017). Yet, millions of depressed youth go unrecognized by health providers, and consequentially, remain untreated for this chronic, impairing, and potentially fatal mental illness (Kessler et al., Reference Kessler, Avenevoli and Merikangas2001). In response, several governmental [e.g. the United States Preventive Services Task Force (USPSTF); Siu, Reference Siu2016] and professional (e.g. the American Academy of Pediatrics; Zuckerbrot et al., Reference Zuckerbrot, Cheung, Jensen, Stein and Laraque2018) organizations now recommend universal depression screening starting at age 12. According to the USPSTF, universal depression screening has two aims: (1) identify current distress and impairment, and (2) estimate prospective depression risk (Siu, Reference Siu2016). To date, most research focuses on how well depressive symptom measures classify current states of clinical depression, with few studies examining if these same protocols predict future depression outcomes (see Stockings et al., Reference Stockings, Degenhardt, Lee, Mihalopoulos, Liu, Hobbs and Patton2015). Thus, translational research demonstrating the clinical utility of methods for predicting future incidents of depression is needed.
Developing a screening protocol for future depression onset (i.e. a first lifetime episode of depression) may be particularly important. Depression is a recurrent disorder, and prevention efforts demonstrate better long-term outcomes compared to depression interventions (Weersing et al., Reference Weersing, Jeffreys, Do, Schwartz and Bolano2017). Primary prevention screening can therefore help reduce the burden of depression by preventing its onset. Conceptual models (e.g. the kindling hypothesis) posit that risk factors for first lifetime and recurrent depression episodes may vary (Monroe and Harkness, Reference Monroe and Harkness2011). However, the relatively infrequent use of diagnostic assessments in large, multi-wave studies makes it challenging to disentangle which risk factors are best to preselect for prospectively predicting first depression onset in adolescents.
In addition to differentiating between recurrent and first lifetime episodes of depression, additional translational barriers exist. First, few studies simultaneously examine risk factors, making it challenging to perform formal tests of incremental validity (Johnston and Murray, Reference Johnston and Murray2003). Second, most studies rely on prediction models that use regression-based, structural equation modeling, or machine learning analytic plans. While all of these broad analytic frameworks have significant, and unique, strengths, they often do not speak to all aspects of predictive validity when used in isolation. Specifically, (a) predictive accuracy (e.g. the ability to differentiate between those who will and will not develop depression; also known as discrimination) and (b) precise likelihood estimates (e.g. correctly estimating the risk of depression onset for a range of scores; also known as calibration) should be explicitly tested to demonstrate the clinical usefulness of a prediction model (Hanson, Reference Hanson2016; Youngstrom et al., Reference Youngstrom, Meter, Frazier, Hunsley, Prinstein, Ong and Youngstrom2017; Lindhiem et al., Reference Lindhiem, Petersen, Mentch and Youngstrom2018). Finally, the lack of replication studies is a critical issue for psychological research, including mental health screening. Without replication, prediction models will likely (unintentionally) exaggerate a given screening protocol's ability to identify the risk (Steyerberg, Reference Steyerberg2009).
Beyond statistical prediction, there are other factors to consider when selecting screening measures. For instance, early childhood adversities predict first lifetime episodes (Monroe and Harkness, Reference Monroe and Harkness2011), but focusing on historical events could be stigmatizing and does not help explain the dynamic reason why certain adolescents may be at increased risk (Finkelhor, Reference Finkelhor2018). Instead, psychosocial individual differences, such as cognitive vulnerabilities (e.g. rumination) and emotional predispositions (e.g. negative affect), may be preferable risk factors to target in screening batteries because they can be altered by evidence-based therapeutic approaches (Garber et al., Reference Garber, Korelitz and Samanez-Larkin2012; Nehmy and Wade, Reference Nehmy and Wade2015). Similarly, academic and social difficulties, two forms of impairment associated with adolescent depression (Jaycox et al., Reference Jaycox, Stein, Paddock, Miles, Chandra, Meredith, Tanielian, Hickey and Burnam2009), can also be attenuated via improved problem-solving skills often taught in depression prevention protocols. By focusing on dynamic personal factors that are targeted in existing prevention interventions, the screening process can select youth based on these proposed mechanisms of risk.
In the present study, we sought to develop and validate an algorithm for prospectively predicting first depression onset in youth. To do so, we analyzed data from two independent, multi-wave longitudinal studies with shared methodological characteristics and multiple psychosocial risk factors (e.g. cognitive vulnerabilities, emotional predispositions, impairment, maternal depression). The primary aims for this research are to (a) bolster pediatric depression screening initiatives and (b) advance clinical research. Many randomized clinical trials for depression prevention programs use subthreshold depressive symptoms to select vulnerable youth (Weersing et al., Reference Weersing, Jeffreys, Do, Schwartz and Bolano2017). The current study is among the first to comprehensively test whether this is the best approach. Moreover, our study is well-positioned to quantify the incremental validity and clinical utility of using multiple predictors to forecast depression onset (i.e. a multi-indicator screening approach). Across screening initiatives for pediatric conditions (e.g. obesity; Proctor et al., Reference Proctor, Moore, Gao, Cupples, Bradlee, Hood and Ellison2003), it is well-accepted that multiple indices are necessary for predicting chronic health outcomes. Screening batteries that quantify the risk across several predictors provide flexibility in the decision model and help protect against the error associated with depending on a single cutoff (Cohen et al., Reference Cohen, So, Hankin and Young2018a). Thus, we seek to identify a screening algorithm to predict future first depression onset that is both accurate and clinically useful.
Method
Participants and procedure
The Gene-Environment and Mood (GEM) and Montreal-Chicago (MTL-CHI) studies, two independent, multi-site samples were used to develop our algorithms. For both studies, youth were recruited by partnering with schools and sending letters home to parents, and via other indirect methods (e.g. advertisements placed in local newspapers). Youth were excluded if the parent reported that the youth (a) had an intellectual disability that prevented them from understanding the questionnaires, (b) suffered from psychosis, or (c) did not speak English. For the GEM study, 663 parent–child dyads enrolled, while the MTL-CHI study consisted of 382 dyads. At baseline, youth and a caregiver completed self- and parent-reported depressive symptoms, adversities, maternal depression, impairment, and individual differences (cognitive vulnerabilities, emotional predisposition) inventories. In addition, four diagnostic assessments in a 2-year span (i.e. 6-month follow-up interviews) were completed to assess for depression onset. The GEM study took place from 2008 to 2013 and the MTL-CHI took place from 2003 to 2007.
Only youth who, as determined via our multi-informant diagnostic interview, had not experienced a major depressive episode (MDE) in their lifetime were used in the current study. This resulted in a subsample of 591 youth in GEM and 348 in MTL-CHI. The distribution of age (in years; GEM: M = 11.74, range = 7–17; MTL-CHI: M = 12.56, range = 10–15), sex (GEM: 55.2% female; MTL-CHI: 56.6% female), and race (GEM: White = 62.2%, African-American = 12.3%; MTL-CHI: White = 70.6%, African-American = 12.3%) was similar between the studies (see Hankin et al. (Reference Hankin, Young, Abela, Smolen, Jenness, Gulley, Technow, Gottlieb, Cohen and Oppenheimer2015) and Abela and Hankin (Reference Abela and Hankin2011) for more details on the GEM and MTL-CHI studies). Identical measures were used to assess predictors and depression onset in both studies.
Measures
Depression diagnoses
Trained clinical psychology doctoral students administered the Mood Disorders section of the Schedule for Affective Disorders and Schizophrenia for School Age Children (K-SADS-PL) (Kaufman et al., Reference Kaufman, Birmaher, Brent, Rao, Flynn, Moreci, Williamson and Ryan1997) to youth and a caretaker at baseline and each follow-up. Interviewers were trained and supervised by licensed clinical psychologists. Both interviews were used to determine youths' diagnostic status using best estimate procedures (Klein et al., Reference Klein, Dougherty and Olino2005). Interviewers were blinded to all other study procedures when making a diagnosis. Inter-rater reliability for the KSADs was excellent for both GEM (κ = 0.91) and MTL-CHI (κ = 0.87) based on approximately 20% of interviews reviewed for reliability. Youth were diagnosed with depression if they met DSM criteria for at least five symptoms of an MDE. In total, 5.25% (N = 31) and 8.30% (N = 29) of youth experienced depression onset in the GEM and MTL-CHI studies, respectively. This rate of depression onset in a 2-year period is comparable to past research in a similar age range (Hankin et al., Reference Hankin, Abramson, Moffitt, Silva, Mcgee and Angell1998). The KSADs is a common diagnostic assessment of youth psychiatric disorders and has demonstrated excellent reliability and validity for depression outcomes (Hankin and Cohen, Reference Hankin, Cohen, Prinstein and Youngstromin press).
Depression symptoms
The Children's Depression Inventory (CDI) (Kovacs, Reference Kovacs1992) assessed both self- and parent-reported symptoms in youth. The youth (CDI-Y) and parent (CDI-P) report on the CDI are identical except parents answer with regard to how they believe their child feels. Scores on the CDI-Y and CDI-P are reliable and valid predictors of depression in youth (Garber, Reference Garber1984; Kovacs, Reference Kovacs1992). The CDI is a commonly used screening inventory for depression, with similar predictive validity as other depression inventories (Hankin and Cohen, Reference Hankin, Cohen, Prinstein and Youngstromin press). The CDI-Y (α = 0.88 in GEM; α = 0.87 in MTL-CHI) and CDI-P (α = 0.86 in GEM; α = 0.85 in MTL-CHI) had acceptable levels of reliability in the current study.
Adversities
Items from the Adolescent Life Event Questionnaire (ALEQ) (Hankin and Abramson, Reference Hankin and Abramson2002) were used to assess recent adversity exposure. Proximal adversity exposure (past 3 months) was intentionally queried because recent adversity exposure may be uniquely predictive for depression onset (Monroe and Harkness, Reference Monroe and Harkness2011). The subscale asked about the youth's direct and/or indirect exposure to a number of potentially traumatic events (i.e. any hospitalizations, deaths, arrests, significant medical or emotional problems, parental separation/divorce, job loss, and emotional neglect exposure).
Parental depression
One risk factor that may be particularly important to assess for depression screening in youth is parental depression (Siu, Reference Siu2016). In response, the Beck Depression Inventory (Beck et al., Reference Beck, Steer and Brown1996) was used to assess current depressive symptoms and the Schedule of Clinical Diagnoses (SCID) (Lobbestael et al., Reference Lobbestael, Leurgans and Arntz2011) was used to assess lifetime history of parental depression. Both the BDI and SCID are valid and reliable measures of adult depression (Pettersson et al., Reference Pettersson, Boström, Gustavsson and Ekselius2015). Overall, 27% (N = 160) and 32% (N = 112) of participating caretakers reported a history of major depression in the GEM and MTL-CHI studies, respectively. BDI and SCID scores were used as independent indices of parental depression.
Social and academic impairment
Another ALEQ (Hankin and Abramson, Reference Hankin and Abramson2002) subscale measured academic and social impairment. Examples of academic impairment included receiving a bad report card and having a bad teacher/class. Social impairment reflected social isolation (e.g. not having as many friends as you'd like to) and conflict (e.g. arguments with parents) with peers and family members. Impairment was conceptualized as a multi-dimensional measure that assessed both social and academic impairment (rather than social or academic impairment) based on recommendations within the field (Fabiano and Pelham, Reference Fabiano, Pelham, Goldstein and Naglieri2016). The ALEQ is a reliable and valid measure of items related to adolescent impairment (Hankin et al., Reference Hankin, Stone and Wright2010).
Individual differences
Several psychosocial risk factors were assessed via self-report questionnaires. These include the Children's Dysfunctional Attitudes Scale (Abela and Sullivan, Reference Abela and Sullivan2003), the Adolescent Cognitive Style Questionnaire (Hankin and Abramson, Reference Hankin and Abramson2002) (to assess inferential styles), and the Children's Response Style Questionnaire (Abela et al., Reference Abela, Vanderbilt and Rochon2004) (to assess rumination). These inventories collectively measure cognitive styles articulated in preeminent theoretical models that explain depression onset in adolescents (Hankin et al., Reference Hankin, Snyder, Gulley and Cicchetti2016). Additionally, the Positive and Negative Affect Scale for Children (Laurent et al., Reference Laurent, Catanzaro, Joiner, Rudolph, Potter, Lambert, Osborne and Gathright1999) was included as a measure of positive and negative affect, both of which are temperamental risk factors for depression onset in youth (Muris and Ollendick, Reference Muris and Ollendick2005; Farchione et al., Reference Farchione, Fairholme, Ellard, Boisseau, Thompson-Hollands, Carl, Gallagher and Barlow2012). In the current study, Cronbach α for dysfunctional attitudes (α = 0.85 in GEM; α = 0.67 in MTL-CHI), inferential style (α = 0.93 in GEM; α = 0.90 in MTL-CHI), rumination (α = 0.87 in GEM; α = 0.89 in MTL-CHI), positive affect (α = 0.83 in GEM; α = 0.90 in MTL-CHI), and negative affect (α = 0.89 in GEM; α = 0.88 in MTL-CHI) were acceptable and consistent with past research (Laurent et al., Reference Laurent, Catanzaro, Joiner, Rudolph, Potter, Lambert, Osborne and Gathright1999; Muris and Ollendick, Reference Muris and Ollendick2005; Abela and Scheffler, Reference Abela and Scheffler2008; Hankin et al., Reference Hankin, Snyder, Gulley and Cicchetti2016).
Data analytic plan
As the GEM study had a larger sample size, it represented our ‘test’ study, while MTL-CHI was our ‘validation’ study. Analyses first examined the properties of each measure dimensionally to establish the predictive validity of a given index test (Youngstrom et al., Reference Youngstrom, Meter, Frazier, Hunsley, Prinstein, Ong and Youngstrom2017). Specifically, because few translational studies have used a risk factor approach to depression screening, our initial aim was to test, and subsequently validate, each index test (i.e. a pre-selection approach; Steyerberg, Reference Steyerberg2009). Once a collective set of valid predictors was identified, we subsequently conducted multivariate analyses to test and validate if the multi-indicator screening algorithm was incrementally valid across both datasets. To minimize potential bias in our parameter estimates, bias-corrected accelerated bootstrap methods were used when appropriate.
There are a variety of analytic approaches that can quantify the utility of a prediction model. In the extant literature, various machine learning and related data mining approaches have become common exploratory steps in identifying a subset of novel predictors. However, given that we were focused on the clinical utility of a finite number of relatively well-known risk factors, the advantages of a machine learning approach are mitigated in the current context. Instead, we began with a regression-based approach as it has shown to be more robust compared to alternative methods (e.g. machine learning) within studies that have comparable methodological characteristics (i.e. <25 predictors and sample sizes 1000) (Steyerberg, Reference Steyerberg2009). Specifically, discrete-time survival models (Singer and Willett, Reference Singer, Willett and Willett2003) were first conducted to examine which measures predicted time to first episode onset.
Next, we used a ‘best practice’ receiver operator characteristic (ROC) approach (Youngstrom, Reference Youngstrom2014) to compute the AUC for each measure. Across both machine learning and regression-based approaches, ROCs are useful in determining whether a predictor is not only statistically significant, but also clinically useful (Youngstrom et al., Reference Youngstrom, Meter, Frazier, Hunsley, Prinstein, Ong and Youngstrom2017). If a given predictor was significant in both the survival analyses and ROC, it was deemed to have satisfactory predictive accuracy. Subsequently, Diagnostic Likelihood Ratios (DLRs) were computed to estimate the likelihood of developing depression based on minimal, subthreshold, and threshold levels of risk. Consistent with an evidence-based medicine approach (Straus et al., Reference Straus, Glasziou, Richardson and Haynes2011), these groups were formed by using cutoffs that generated three equal groups of adolescents. Finally, the Expected-Observed (E/O) Index was then used to determine whether the likelihood estimates generated by the DLRs were valid (Hanson, Reference Hanson2016). Specifically, the DLRs for each risk profile (e.g. threshold rumination) in the GEM study were multiplied by the pre-test odds for depression onset in the MTL-CHI study. This represented the expected number of episodes given the number of youth with that risk profile in the MTL-CHI study. We then divided this number by the observed number of episodes for that profile in the MTL-CHI study. The Poisson variance for the logarithm of the observed number of cases was used to create the confidence interval (CI) for the E/O Index. A CI that included 1 indicates strong calibration (Hanson, Reference Hanson2016).
Several tests of incremental validity were next conducted with our valid predictors. First, discrete-time survival analyses using backward selection methods based on changes in the maximum likelihood estimates of the likelihood ratio were examined (Steyerberg, Reference Steyerberg2009). If a predictor was eliminated across both studies, it was dropped from further analyses. Further, AUCs of remaining predictors were compared using the DeLong test for paired ROC curves (Youngstrom et al., Reference Youngstrom, Meter, Frazier, Hunsley, Prinstein, Ong and Youngstrom2017). A significant difference in the DeLong test suggests that certain index tests should be prioritized in prediction tools.
Finally, an additive approach was used to form a Cumulative Risk score. Specifically, profiles were formed based on whether youth were at minimal risk (0), subthreshold risk (1), or threshold risk (2) on each measure. This categorical approach is similar to how cumulative risk is typically conceptualized within basic research (e.g. Evans, Reference Evans2003) and in multi-indicator risk algorithms for pediatric health screening (e.g. Proctor et al., Reference Proctor, Moore, Gao, Cupples, Bradlee, Hood and Ellison2003). We then replicated the analytic plan described above to test whether (a) the Cumulative Risk score (comprised of the categorical risk score on each index test) predicted depression onset above and beyond the individual index tests and (b) the Cumulative Risk score prospectively predicted depression onset above and beyond current approaches that rely on depressive symptoms. Additional calibration plots were created to test whether the predicted probabilities for the Cumulative Risk score mirrored the observed probabilities within each dataset (Lindhiem et al., Reference Lindhiem, Petersen, Mentch and Youngstrom2018). Online Supplementary Table S1 summarizes our analytic approach. The Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) guidelines were followed in reporting this study's methods and results (Collins et al., Reference Collins, Reitsma, Altman and Moons2015).
Results
Missing data and preliminary analyses
As is common in longitudinal studies, there were missing data across the follow-ups. Little's Missing Completely at Random (MCAR) test suggested that these data were missing completely at random in both the GEM [χ2(35) = 21.76, p = 0.96] and MTL-CHI [χ2 (173) = 201.10, p = 0.07] studies. Thus, missing baseline data were imputed using expectation maximization algorithms for the predictor, and follow-up data were censored after the last completed follow-up. Descriptive statistics for our dimensional predictors can be found in Table 1. Online Supplementary Fig. S1 displays Kaplan–Meier curves to illustrate the rate of depression onset in our study.
Descriptive statistics for multi-site study
CDI-Y, Children's Depression Inventory (Kovacs, Reference Kovacs1992) (Youth Report); CDI-P, Children's Depression Inventory (Kovacs, Reference Kovacs1992) (Parent Report); Adversities, Adversity subscale of the Adolescent Life Event Questionnaire (Hankin and Abramson, Reference Hankin and Abramson2002); BDI, Beck Depression Inventory (Beck et al., Reference Beck, Steer and Brown1996); Impairment, Impairment subscale of the Adolescent Life Event Questionnaire (Hankin and Abramson, Reference Hankin and Abramson2002); Dysfunctional Attitudes, Children's Dysfunctional Attitudes Scale (Abela and Sullivan, Reference Abela and Sullivan2003); Attributional style, Attributional Cognitive Style Questionnaire (Hankin and Abramson, Reference Hankin and Abramson2002); Rumination, Children's Rumination Scale Questionnaire (Abela et al., Reference Abela, Vanderbilt and Rochon2004); Positive affect, Positive and Negative Affect Scale for Children (Laurent et al., Reference Laurent, Catanzaro, Joiner, Rudolph, Potter, Lambert, Osborne and Gathright1999) (Positive Affect Subscale); Negative affect, PANAS (Laurent et al., Reference Laurent, Catanzaro, Joiner, Rudolph, Potter, Lambert, Osborne and Gathright1999) (Negative Affect Subscale).
Univariate analyses
Results from univariate survival analyses and ROC are presented in Table 2. Based on the Wald statistic and AUC, only three significant predictors from the GEM study were also significant in the MTL-CHI study: Rumination, Impairment, and Negative Affect. Of note, both self- and parent-reported depressive symptoms demonstrated inconsistent findings across the two studies. DLRs and the accompanying E/O indices for Rumination, Impairment, and Negative Affect are presented in Table 3, along with the corresponding cutoffs for minimal, subthreshold, and threshold levels of risk. Importantly, youth were approximately 1.5–2 times more likely than the sample average to be diagnosed with a first lifetime episode of depression if they scored in the upper third on any one of the three measures. CIs of the E/O Index suggest that each predictor was able to reliably estimate the likelihood of depression onset for minimal, subthreshold, and threshold risk levels.
Univariate discrete-time survival analyses and receiver operating characteristics
CDI-Y, Children's Depression Inventory (Kovacs, Reference Kovacs1992) (Youth Report); CDI-P, Children's Depression Inventory (Kovacs, Reference Kovacs1992) (Parent Report); subscale Adversities, Adversities subscale of the Adolescent Life Event Questionnaire (Hankin and Abramson, Reference Hankin and Abramson2002); BDI, Beck Depression Inventory (Beck et al., Reference Beck, Steer and Brown1996); II Dysfunctional attitudes, Children's Dysfunctional Attitudes Scale31; Attributional style, Attributional Cognitive Style Questionnaire (Hankin and Abramson, Reference Hankin and Abramson2002); Rumination, Children's Rumination Scale Questionnaire (Abela et al., Reference Abela, Vanderbilt and Rochon2004); Positive affect, Positive and Negative Affect Scale for Children (Laurent et al., Reference Laurent, Catanzaro, Joiner, Rudolph, Potter, Lambert, Osborne and Gathright1999) (1999) – Positive Affect Subscale; Negative affect, PANAS (Laurent et al., Reference Laurent, Catanzaro, Joiner, Rudolph, Potter, Lambert, Osborne and Gathright1999) – Negative Affect Subscale. Impairment, Impairment subscale of the Adolescent Life Event Questionnaire (Hankin and Abramson, Reference Hankin and Abramson2002); Parent depression Dx, Lifetime depression diagnosis from the depression module of the SCID.
All estimates from both the survival and receiver operating characteristic analyses are bootstrapped.
**p ⩽ 0.01.
Diagnostic Likelihood Ratio and expected/observed indices
Rumination, Children's Rumination Scale Questionnaire (Abela et al., Reference Abela, Vanderbilt and Rochon2004); Negative affect, PANAS – Negative Affect Subscale (Laurent et al., Reference Laurent, Catanzaro, Joiner, Rudolph, Potter, Lambert, Osborne and Gathright1999); Impairment, Impairment subscale of the Adolescent Life Event Questionnaire (Hankin and Abramson, Reference Hankin and Abramson2002); DLR, Diagnostic Likelihood Ratios (Straus et al., Reference Straus, Glasziou, Richardson and Haynes2011); E/O Index, Expected number of cases in the MTL-CHI study based on posterior probabilities for a given risk profile derived from the GEM study/the number of actual observed cases for a given risk profile in the MTL-CHI study. 95% Confidence Interval, The confidence interval for the E/O Index (if the CI includes 1, it is significant) (Hanson, Reference Hanson2016).
Multivariate analyses
When entered simultaneously into multivariate survival analytic models, neither Rumination, Impairment, nor Negative Affect consistently emerged as incrementally significant predictors using backward selection methods. Further, DeLong tests indicated that there were no significant differences between the AUCs for our predictors (p > 0.05). We therefore included, and equally weighted, Rumination, Impairment, and Negative Affect to form the Cumulative Risk score. This created a range of scores from 0 (minimal risk across each predictor) to 6 (elevated risk for each predictor) for each youth-based one's level of risk (0–2) on the three predictors.
We next tested the incremental validity of our Cumulative Risk score (see Table 4). First, Cumulative Risk was entered into a model with Rumination, Negative Affect, and Impairment as individual predictors, as well as the CDI-Y and CDI-P, two traditional approaches to depression screening (Siu, Reference Siu2016). Cumulative Risk was selected as the only significant prospective predictor of time until first depression onset in the GEM study, and these findings were validated in the MTL-CHI study. Next, to provide a more rigorous test of incremental validity, we created a Multi-Informant Risk score (summing minimal, subthreshold, and threshold risk across the CDI-Y and CDI-P). The purpose of these analyses was to examine our Cumulative Risk score against a ‘gold standard’ multi-informant approach (Klein et al., Reference Klein, Dougherty and Olino2005). As there were only two predictors entered, we examined boostrapped estimates of the Cumulative Risk and Multi-Informant scores simultaneously. Findings suggested that both predictors were unique in predicting time to first onset in the GEM study, but only Cumulative Risk predicted time to first onset in the MTL-CHI study. AUCs for Cumulative Risk were 0.73 and 0.68, respectively, the highest AUCs of any predictor for each study.
Incremental validity and screening properties of risk score algorithm
Parameters for discrete-time survival analyses (DTSA) are derived from two separate models. In the top panel, parameters represent estimates from successive backward elimination models. The middle panel represents parameter estimates from simultaneous DTSA models in the GEM and MTL-CHI studies, respectively. In the top two panels, Risk score represents the sum of whether one was at minimal (0), subthreshold (1), or threshold (2) levels of risk. In the bottom panel, Risk score represents minimal (dimensional scores 0–2), subthreshold (3–4), and threshold levels of risk. The equation for the Diagnostic Likelihood Ratio (DLR) is: (number of episodes in risk profile/number of episodes total)/(number of non-episodes in risk profile/number of non-episodes total) (Straus et al., Reference Straus, Glasziou, Richardson and Haynes2011). DLRs above are combined between the two studies and weighted by the number of individuals in each study. Base rates for depression onset in GEM and MTL-CHI were 5.25% and 8.30%, respectively.
Finally, we examined the E/O Index for our Cumulative Risk score. As can be seen in online Supplementary Fig. S2, the estimated likelihood matched the observed likelihood across the range of scores. Relatedly, the predicted probabilities were similar to the observed probabilities in both studies (online Supplementary Fig. S3). These findings suggest the Cumulative Risk score can be used as a dimensional prospective predictor of first depression onset (Hanson, Reference Hanson2016). Alternatively, it may be useful to have cutoffs that correspond to minimal, subthreshold, and threshold risk to further facilitate clinical decision-making (Straus et al., Reference Straus, Glasziou, Richardson and Haynes2011). The bottom panel of Table 4 provides the DLRs and E/O indices, as well as other complementary statistics commonly used in the screening literature [sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV)], for those at minimal risk, subthreshold risk, and threshold risk. Online Supplementary Fig. S3 shows how many individuals presented at each risk level across both studies. E/O indices at the bottom of Table 4 suggests the categorical version of the algorithm is well-calibrated. Online Supplementary Figure 4 provides the survival curves based on these cutoffs. This figure illustrates that threshold scores on the Cumulative Risk score developed in the GEM study were at an increased risk for depression onset in the MTL-CHI study.
Discussion
Increasingly, agencies and organizations are recommending pediatric depression screening with the aim of identifying current distress and impairment as well as future risk (Siu, Reference Siu2016; Zuckerbrot et al., Reference Zuckerbrot, Cheung, Jensen, Stein and Laraque2018). However, few studies examine the optimal method for screening for prospective first onsets of depression. Overall, our findings suggest that relying on depression symptom inventories may lead to inconsistent findings when predicting future depression onset. Instead, measures of rumination, impairment, and negative affect represent reliable, clinically useful options for prospectively predicting depression onsets in youth. Below, we discuss the implications of this algorithm, and explain how these findings can help bridge the translational gap by providing prevention initiatives with a feasible solution for quantifying depression risk.
A key emphasis within the assessment literature is developing algorithms that accurately predict risk with results that replicate across independent samples (Steyerberg, Reference Steyerberg2009). Whereas youth and parent-reported depressive symptoms predicted depression onset in the test sample, these results did not replicate in the validation sample. Meanwhile, rumination, negative affect, and impairment all emerged as significant predictors that were incrementally valid compared to multi-informant symptom approaches. Our findings are consistent with theoretical models and corresponding basic research that conceptualizes these three risk factors as preceding depression onset (Laurent et al., Reference Laurent, Catanzaro, Joiner, Rudolph, Potter, Lambert, Osborne and Gathright1999; Muris and Ollendick, Reference Muris and Ollendick2005; Abela and Scheffler, Reference Abela and Scheffler2008; Hankin et al., Reference Hankin, Snyder, Gulley and Cicchetti2016). Yet, methodological limitations have inhibited the clinical utility of these measures, despite recent calls to integrate a focus on risk factors into depression prevention efforts (Garber et al., Reference Garber, Korelitz and Samanez-Larkin2012; Siu, Reference Siu2016). By examining these risk factors within the context of other viable predictors of depression onset (e.g. maternal depression), and using a translational analytic plan to facilitate clinical decision-making, our findings provide a foundation for an evidence-based screening approach for prospectively forecasting first depression onset in youth.
Within translational research, it is not enough for an algorithm to be significant, but it also must be clinically useful. A traditional method for operationalizing clinical utility is by evaluating the positive and NPVs (see the bottom of Table 4) (Trevethan, Reference Trevethan2017). In the present study, the PPVs were modest, suggesting that our algorithm is vulnerable to false-positive results (i.e. identifying youth at-risk for depression who do not go on to experience a first lifetime episode). In response, some may suggest abandoning a risk stratification approach for depression onset, and instead, use less selective preventive intervention strategies for depression or for clinicians to delineate idiographic risk factors via consultation with the patient (Carter et al., Reference Carter, Milner, McGill, Pirkis, Kapur and Spittal2017; Large et al., Reference Large, Ryan, Carter and Kapur2017). Alternatively, because predictive values are correlated with base rates, others have argued that relying on these frequentist metrics may underestimate the utility of algorithms for outcomes with low base rates (e.g. 5–10%; Lavigne et al., Reference Lavigne, Meyers and Feldman2016). Instead, EBM advocates recommend using posterior probabilities stemming from DLRs when evaluating clinical prediction models (Youngstrom, Reference Youngstrom2014). From this Bayesian perspective our algorithm was useful because it identified youth who were at 2–3 times elevated risk compared to the sample average, as well as individuals who were at minimal risk for experiencing depression onset. Thus, considering both perspectives, a reasonable conclusion when comparing our algorithm to current practices of using depressive symptoms is that we identified an incrementally valid approach for recommended preventive screening for depression onset (Siu, Reference Siu2016). However, because of the low PPVs, our solution should not be considered a ‘gold standard’ and identifying an optimal strategy for predicting depression onset remains an important aim for future research. Thus, similar to existing algorithms for low base rate psychological outcomes (e.g. suicide; Fazel et al., Reference Fazel, Wolf, Larrson, Mallett and Fanshawe2019), it may be best to conceptualize our algorithm as a marker of continuous risk for depression onset as opposed to a strict classification tool.
Overall, our findings suggest that using rumination, impairment, or negative affect is a more reliable screening approach for depression onset compared to symptom inventories. Further, we found using the combination of indices provided incremental validity above and beyond using any measure independently. Using a multi-indicator approach facilitates a decision-making process that more closely aligns with the dimensional nature of adolescent depression (Hankin and Cohen, Reference Hankin, Cohen, Prinstein and Youngstromin press) and reduces the errors associated with making referrals based on scores near or at the cutoff on a single measure (Sheldrick et al., Reference Sheldrick, Benneyan, Kiss, Briggs-Gowan, Copeland and Carter2015). In addition, examining converging scores across rumination, impairment, and negative affect can inform evidence-based decision-making by understanding risk across multiple indicators. For instance, low impairment scores conferred DLRs below 0.25, suggesting that the onset of depression is highly unlikely with low scores on this particular index (Straus et al., Reference Straus, Glasziou, Richardson and Haynes2011). Therefore, even with scores that approach, or even exceed, the cutoff for negative affect and rumination, one may not refer for services if these developmental risks are not manifesting in the context of impairment.
Ultimately, this study's algorithm can be tailored to a prevention protocol's resources so that evidence-based decisions can be made across a variety of settings. Table 5 helps to illustrate the clinical utility of our Cumulative Risk score. All three case examples represent actual screening profiles from 14-year-old girls who participated in the GEM study. As can be seen, the probability of prospectively being diagnosed with a first depression episode is approximately 50% higher than the sample average for those with risk scores of 5. Based on this probability, decisions can be made as to whether resources are available to engage these youth or if only those with a score of 6 (an over twofold risk for depression onset compared to the sample average) should be engaged. Alternatively, since each measure within the algorithm demonstrated a reliable and valid forecast for depression onset, qualitative decisions can also be made so that the screening process can best match the available prevention response. For instance, for ‘Girl A’ in Table 5, one of her high scores is rumination, while for ‘Girl B’ her rumination score is in the average range. If a cognitive–behavioral prevention program is offered in the area (e.g. Penn Resiliency Program; Gillham et al., Reference Gillham, Reivich, Freres, Chaplin, Shatté, Samuels, Elkon, Litzinger, Lascher, Gallop and Seligman2007), screening initiatives may prioritize ‘Girl A’ for a referral as it selects youth high in the mechanism of risk targeted by the preventative intervention (i.e. cognitive vulnerability). Thus, the algorithm retains flexibility through its dimensional risk approach, while still ensuring referrals and evidence-based decisions can be made to help connect youth to available services.
Diagnostic Likelihood Ratios and posterior probabilities for example screening cases
Rumination, Children's Response Style Questionnaire (Abela et al., Reference Abela, Vanderbilt and Rochon2004); Negative affect, Positive and Negative Affect Schedule (PANAS) for Children Score (Laurent et al., Reference Laurent, Catanzaro, Joiner, Rudolph, Potter, Lambert, Osborne and Gathright1999); Impairment, 18-item subscale from the Adolescent Life Event Questionnaire (ALEQ); DLR, Diagnostic Likelihood Ratio [(number of episodes in risk profile/number of episodes total)/(number of non-episodes in risk profile/number of non-episodes total)] (Straus et al., Reference Straus, Glasziou, Richardson and Haynes2011); Pre-test probability, The baseline percentage of 14 year-old females in the Gene-Environment Mood study with first lifetime episodes within a year period; Post-test probability [(prevalence/(1–prevalence) × DLR)/((prevalence/(1–prevalence)) + 1] (Straus et al., Reference Straus, Glasziou, Richardson and Haynes2011). Scores following our individual predictors represent the score assigned based on whether their score fell in the average (1) range or elevated (2) range. The Total risk score is derived by summing the scores across these three predictors. DLRs are the combined estimates from the GEM and MTL-CHI studies (weighted by the number of youth in each study). Although the development of the algorithm did not take into account gender and age, clinicians are able to account for these differences through use of an informed pre-test probability (as done above) (Youngstrom et al., Reference Youngstrom, Meter, Frazier, Hunsley, Prinstein, Ong and Youngstrom2017). Scores for Girl C illustrates the importance of undertaking an actuarial approach, in which, compared to Girls A and B, a Total risk score difference of 1 can dictate more intensive follow-up mental health services. Girls A and B, on the other hand, demonstrate an instance in which scoring profiles would call for similar screening responses within an actuarial framework. However, given the difference in risk pathways associated with Girl A and Girl B (e.g. higher rumination v. higher impairment) relying on a qualitative understanding of these cases may best inform subsequent clinical services based on the available resources.
Strengths of the present study include the use of independent studies for development and validation, as well as repeated diagnostic interviews with relatively short recall periods to assess for prospective depression onset. At the same time, there are noteworthy limitations. First, it is important to replicate our findings outside of a research setting (Youngstrom et al., Reference Youngstrom, Meter, Frazier, Hunsley, Prinstein, Ong and Youngstrom2017). These studies will be necessary for demonstrating the robust nature of our predictors across contexts and for identifying optimal cutoffs based on the objectives and resources of the setting, a necessary step for validating a screening algorithm (Youngstrom, Reference Youngstrom2014). Second, the current study used a pre-selection approach based on cross-validated univariate analyses to demonstrate which predictors may be useful for clinical decision-making (Steyerberg, Reference Steyerberg2009). However, other studies should replicate our findings using other analytic approaches (e.g. machine learning models), ideally with larger sample sizes across different length follow-ups, to determine if our findings generalize under different methodological conditions. Third, our study relied on subjective self- and parent-reported predictors for depression onset. Incorporating psychophysiological measures may demonstrate incremental improvements in predicting depression outcomes, particularly first lifetime episodes of depression (Cohen et al., Reference Cohen, Thakur, Burkhouse and Gibb2019). Fourth, the effect sizes and DLRs for our predictors were relatively small compared to screening tools for current outcomes (Youngstrom et al., Reference Youngstrom, Meter, Frazier, Hunsley, Prinstein, Ong and Youngstrom2017). We suspect this reflects the difficulty in predicting prospective diagnostic outcomes compared to current diagnostic status. Although our effect sizes are comparable to other recent findings concerning prospective, depression-related outcomes (e.g. suicide; Fazel et al., Reference Fazel, Wolf, Larrson, Mallett and Fanshawe2019) and first depression onset (Cohen et al., Reference Cohen, Thakur, Burkhouse and Gibb2019), future research should build upon our proposed algorithm to develop more targeted screening solutions for depression onset. Fifth, given the discontinuous nature of depression between childhood and adolescence (Cohen et al., Reference Cohen, Andrews, Davis and Rudolph2018b), it is important to examine if our algorithm can be extended downward to childhood. Alternatively, it should be tested if our findings generalize to detecting depression onset later in adolescence (e.g. ages 16–17), as the rates of depression markedly increase during this time (Twenge et al., Reference Twenge, Cooper, Joiner, Duffy and Binau2019). Ultimately, balancing personalized algorithms based on development and the translational benefits of having a ‘one size fits all’ approach has important clinical implications for depression prevention programming.
Finally, our recommended battery consisted of 58 items. This number is significantly lower than common mental health screening protocols (e.g. ASEBA assessments; Achenbach and Rescorla, Reference Achenbach and Rescorla2001), yet it also comes at a time when briefer screening efforts are being encouraged (Lavigne et al., Reference Lavigne, Meyers and Feldman2016). Therefore, future studies could use Item Response Theory (IRT) to identify shorter versions of our screening battery (Youngstrom et al., Reference Youngstrom, Meter, Frazier, Hunsley, Prinstein, Ong and Youngstrom2017). Continued efforts to include risk factor measures in preventative screening will fulfill the aims of universal pediatric depression screening (Siu, Reference Siu2016), and ultimately, help reduce the prevalence and burden associated with depression onset at a young age (Kessler et al., Reference Kessler, Avenevoli and Merikangas2001; Weinberger et al., Reference Weinberger, Gbedemah, Martinez, Nash, Galea and Goodwin2017).
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291719002691.
Acknowledgements
This research was supported by NIMH grants (5R01MH077195, 5R01MH077178) awarded to Benjamin Hankin and Jami Young. Joseph Cohen's time on this manuscript was supported by the National Institute of Justice (2018-R2-CX-0022). The study protocol was approved by the institutional ethical review boards of the participating universities. All procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation with the Helsinki Declaration of 1975, as revised in 2008. Participants gave informed consent before participating. The authors have no conflicts of interest to report. JRC had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Financial support
This research was supported by the National Institute of Mental Health (NIMH) grants (5R01MH077195, 5R01MH077178) awarded to Benjamin Hankin and Jami Young. Joseph Cohen's time on this manuscript was supported by the National Institute of Justice (2018-R2-CX-0022).
