Prediction of major depressive disorder onset in college students.

BACKGROUND: Major depressive disorder (MDD) in college students is associated with substantial burden. AIMS: To assess 1-year incidence of MDD among incoming freshmen and predictors of MDD-incidence in a representative sample of students.
METHOD: Prospective cohort study of first-year college students (baseline: n = 2,519, 1-year follow-up: n = 958)
RESULTS: The incidence of MDD within the first year of college was 6.9% (SE = 0.8). The most important individual-level predictors of onset were prior suicide plans and/or attempts (OR = 9.5). The strongest population-level baseline predictors were history of childhood-adolescent trauma, stressful experience in the past 12 months, parental psychopathology, and other 12-month mental disorder. Multivariate cross-validated prediction (cross-validated AUC = 0.73) suggest that 36.1% of incident MDD cases in a replication sample would occur among the 10% of students at highest predicted risk (24.5% predicted incidence in this highest-risk subgroup).
CONCLUSIONS: Screening at college entrance is a promising strategy to identify students at risk of MDD onset, which may improve the development and deployment of targeted preventive interventions.

INTRODUCTION

Major depressive disorder (MDD) is one of the leading causes of disability worldwide (Vos et al., 2012) and also one of the most common mental disorders among college students (Auerbach et al., 2016, 2018; Farabaugh et al., 2012). Depression in college students is associated with lower academic performance (Hysenbegasi, Hass, & Rowland, 2005), substantial role impairment (Alonso et al., 2018), increased risk for college dropout (Arria et al., 2013), increased levels of anxiety (Rawson, Bloomer, & Kendall, 1994), physical illness, decreased physical activity, unsafe sexual behavior, increased levels of smoking (Cranford, Eisenberg, & Serras, 2009), alcohol and drug dependency, poorer quality of life, self‐harming behaviors (Serras, Saules, Cranford, & Eisenberg, 2010), and an increased risk of suicide (Eisenberg, Hunt, & Speer, 2013). Together, this underscores the importance of developing tools that identify students at greatest risk to develop MDD during this critical period of development.

Early identification of students at risk for MDD may allow to effectively deploy preventive interventions during college and thereby reduce the incidence, prevalence, severity, duration, and consequences of future depressive episodes as well as of other mental disorders (van Zoonen et al., 2014). To support clinical decision‐making and resource allocation, universities need tools that accurately identify students at high risk of developing depression in the near future.

Although there is a fair amount of studies that estimate the prevalence of MDD among college students, studies on the incidence of MDD among representative incoming students are much scarcer. Zivin, Eisenberg, Gollust, & Golberstein (2009) estimated the incidence of depression in college students at approximately 5% per year, but they did not identify risk parameters that predict MDD incidence.

Several studies of risk indicators for MDD among college students have been carried out, but most of them were limited by being based on cross‐sectional rather than prospective data and thus, cannot disentangle the cause from the effect (Brandy, Penckofer, Solari‐Twadell, & Velsor‐Friedrich, 2015; Leino & Kisch, 2005). Moreover all prior studies focused only on the coefficients of individual predictors rather than developing composite risk measures (Mahmoud, Staten, Hall, & Lennie, 2012). Furthermore, individual‐level effect sizes merely identify specific risk indicators for individuals. However, for prevention purposes, it is important to select risk indicators that are associated with the largest potential health gain at population level. Risk indicators that will lead to the largest population health gain should not only be linked to heightened risk of developing a depression on an individual‐level, but should also be prevalent in the target population. By estimating the population attributable risk proportion (PARP), it is possible to identify the number of cases that would not occur in the population, if a specific risk indicator were eliminated.

Key Points

We investigated whether MDD during the first year of college can be predicted using baseline data.

Strongest predictors on the individual level were prior suicidal behaviors, but when taking also the prevalence of the risk factor into account (OR ≥ 2/PARP ≥ 15%), preventive approaches should focus on students with traumatic experiences, a recent break‐up with a romantic partner, serious ongoing arguments with people close to them, and those with recent stressful life events. A multivariate risk prediction algorithm was able to predict the incidence of MDD, with 36.1% of all cases occuring in the 10% of students at the highest predicted risk.

Screening at college entrance is promising to identify students at high risk for MDD onset, which may improve the development and deployment of targeted preventive interventions.

The aims of the current study were to: (1) estimate the 1‐year incidence of MDD among college freshmen who were without MDD in the 12 months prior to college enrollment, and further, among these individuals to, (2) examine individual‐level, and (3) population‐level predictors of 1‐year incidence of MDD. Finally, we aimed (4) to evaluate the prediction accuracy of a baseline multivariate risk prediction model aimed at identifying college freshmen at highest risk for MDD onset over the subsequent 12 months.

METHOD

Longitudinal data were obtained from the Leuven College Surveys (LCS), which are part of the WHO World Mental Health Surveys International College Student project (WMH‐ICS). Full procedures of the LCS have been reported elsewhere (Mortier et al., 2017). In the academic year 2014–2015, all 4,130 Dutch‐speaking incoming freshmen aged 18 years or older were invited to participate in the baseline survey. The inclusion of the baseline sample consisted of three consecutive stages, with different refusal conversion strategies to increase final response rates: In the first stage, the baseline survey was part of a routine psychomedical check‐up. All incoming freshmen (i.e., complete enumeration or census sampling) were sent a standard invitation letter for this check‐up. This means that all units of the freshmen population were eligible to complete the survey on a desktop computer in the waiting room of the students’ mental health center. One reminder letter for the medical check‐up was sent by mail by the students’ mental health center. In a second stage, nonrespondents to the first stage were personally contacted using customized emails containing unique electronic links to the survey. Two reminder emails were sent with a 1‐week interval. By implementing this stage, we removed the physical barrier between the initial nonrespondents and the mental health center, since the survey could then be completed on a personal computer at home. The third stage was identical to the second stage, but additionally included an incentive, that is, a raffle for store credit coupons. Two reminder emails were sent with a 1‐week interval. When including the reminder (e)mails used in each stage, the maximum amount of contacts was set to eight attempts. A total of 2,519 students completed the baseline survey, equivalent to a baseline response rate (RR) of 61%. Students were contacted for a follow‐up survey 12 months after the baseline assessment, using a similar sampling design to the one used at baseline. Personalized emails with unique electronic links to the survey were sent, including up to seven reminder emails. Beginning with the fifth reminder email, emphasis was put on a store credit coupon raffle. The survey was delivered fully digital and all survey sections were presented in a precise predetermined order.

Of the students who responded at baseline, 958 (38%) responded to the follow‐up survey (57.5% response rate after adjusting for nonparticipation due to college dropout). Earlier reports of the WMH‐ICS initiative have shown that lifetime‐ (22.4%, 95%CI: 21.2–23.7) and 12‐month mental disorders (19.1%, 95%CI: 17.9–20.2) are somewhat lower in the LCS sample, than average prevalence rates of the countries included in the first prevalence estimates assessed in eight participating WMH‐ICS countries (Auerbach et al., 2018). The study's protocol was approved by the University Hospital Leuven Biomedical Ethical Board (B322201215611). Informed consent was obtained from all subjects who participated in the study. Students who reported any past year STB or nonsuicidal self‐injury were presented with links to local mental health resources.

MEASURES

The WMH‐ICS survey instrument includes multiple screening instruments measuring a wide range of mental health outcomes. The included assessments are described below.

Socio‐demographic variables

Socio‐demographic characteristics of freshmen were obtained from the KU Leuven student administration office, including gender, age, nationality, parent financial situation, parent education, familial composition, university group membership, and secondary school educational type.

Parental psychopathology and traumatic experiences in childhood–adolescence

Traumatic experiences in childhood and adolescence (i.e., prior to the age of 17) were assessed using 19 items adapted from the Composite International Diagnostic Interview (CIDI) (Kessler & Üstün, 2004), the Adverse Childhood Experience Scale (Felitti et al., 1998), and the Bully Survey (Swearer & Cary, 2003). Items assessed parental psychopathology (i.e., any serious mental or emotional problems, substance abuse, suicidal thoughts and behaviors or death by suicide, criminal activities or interpersonal violence), physical abuse, emotional abuse, sexual abuse, neglect, bully victimization (i.e., direct verbal or physical bullying, indirect bullying or cyberbullying), and dating violence. Items were rated on a 5‐point Likert scale (“never,” “rarely,” “sometimes,” “often,” and “very often”). Confirmatory factor analysis using our data showed a strong unidimensional structure of responses (Comparative Fit Index = 0.991; Tucker–Lewis = 0.988; root mean square of approximation = 0.019). To obtain dichotomously coded variables, “rarely” was used for all items, except bully victimization which had a cut‐off of “sometimes” (Nansel et al., 2001).

Stressful events experienced in the past 12 months

Stressful events were assessed using 12 items taken from well‐validated screeners (Bray & Hourani, 2007; Brugha & Cragg, 1990; Vogt, Proctor, King, King, & Vasterling, 2008). Items assessed life‐threatening illness, accidents or death of a family member or close friend, interpersonal events (i.e., break‐up with a romantic partner, serious betrayal by someone else than partner), physical or sexual assault, and legal problems (i.e., time spent in jail).

Twelve‐month mental disorder

The CIDI Screening Scales (CIDI‐SC) (Kessler & Üstün, 2004; Kessler et al., 2013) were used to assess two mood disorders (major depressive disorder and [hypo]manic episodes), two anxiety disorders (generalized anxiety disorder and panic disorder), and drug use disorder (abuse or dependence either on cannabis, cocaine, or any other street drug, or on a prescription drug either used without a prescription or used more than prescribed to get high, buzzed, or numbed out). The CIDI‐SC consists of a range of DSM‐IV‐based screening scales containing well‐validated self‐report items that were developed to deliver reliable estimates of mental disorder diagnoses. Concordance with blinded clinical diagnoses in clinical reappraisal studies were in the range AUC = 0.70–0.78 (Kessler et al., 2013). Information on lifetime and 12‐month MDD was assessed by asking respondents about age, age of onset of MDD, and most recent age with MDD.

The Alcohol Use Disorders Identification Test (AUDIT) (Saunders, Aasland, Babor, De la Fuente, & Grant, 1993) is a 10‐item screening tool developed by the WHO to determine alcohol consumption, risk for alcohol dependence, and alcohol‐related harm. The AUDIT is well‐validated in college students (DeMartini & Carey, 2012). Consistent with prior recommendations (Babor, Higgins‐Biddle, Saunders, & Monteiro, 2001), the AUDIT was used to identify students with 12‐month “risky or hazardous drinking” and students with 12‐month “risk for alcohol dependence.”

A modified version of the Columbia Suicidal Severity Rating Scale (Posner et al., 2011) was used to assess 12‐month suicidal thoughts and behaviors (STB), including suicidal ideation, suicide plans, and suicide attempts. After the assessment of outcomes with suicidal intent, 12‐month nonsuicidal self‐injury (NSSI) also was assessed (Nock, Holmberg, Photos, & Michel, 2007).

Analyses

All analyses were performed with SAS version 9.4, Mplus version 7.4, and R version 3.3.2. Nonresponse propensity weights (Rosenbaum & Rubin, 1983) were used to adjust for possible bias caused by final nonresponse. Multiple imputation by chained equations (van Buuren, 2007) was used to adjust for survey attrition and within‐survey item nonresponse. All analyses were conducted in the subsample without 12‐month MDD at baseline. Lifetime MDD was added as a covariate in all analyses. Incidence was reported as a weighted proportion (%) and associated standard error (SE). Logistic regression analysis was used to test the individual‐level strength of the association between baseline risk indicators and the onset of MDD. All analyses were adjusted for lifetime history of MDD. Measures of association were reported as odds ratios (OR) and associated 95% confidence intervals (95%CI). Firth's penalised likelihood estimation was applied to avoid overfitting and inconsistent estimators due to data sparseness (Firth, 1993). The population‐level impact of baseline risk indicators on the onset of MDD was estimated by population attributable risk proportions (PARPs) (Krysinska & Martin, 2009) using the predicted probabilities resulting from the logistic regression equations as a summary predictor (Nock, Borges, & Ono, 2012). PARPs could be interpreted as the proportion of cases that would be prevented if the targeted risk indicator were fully blocked in the population.

Finally, a multivariate model was estimated, including socio‐demographic variables, childhood–adolescent traumatic events, 12‐month stressful experiences, 12‐month risk for mental disorders, and lifetime history of MDD. Nagelkerkes pseudo‐R2 was used as a measure of total effect size. Based on the multivariate equation, individual‐level predicted probabilities were created, receiver operating characteristic (ROC) curves were generated, and to evaluate prediction accuracy area under the curve (AUC) values were calculated. Predicted probabilities were then discretized into deciles (10 groups of equal size ordered by percentiles) and cross‐classified with observed cases to visualize the concentration of risk associated with high composite predicted probabilities. We defined sensitivity as the proportion of cases found among predefined proportions of respondents (e.g., 10%) with the highest predicted probabilities. Positive predictive value (PPV) was defined as the probability of actually developing an MDD when estimated among the 10% of respondents with the highest predicted probabilities. We used the method of leave‐one‐out cross‐validation (Efron & Gong, 1983) to correct for the over‐estimation of prediction accuracy when both estimating and evaluating model fit in a single sample. Although leave‐one‐out cross‐validation shows a downward bias of true prediction accuracy compared to other cross‐validation techniques (Smith, Seaman, Wood, Royston, & White, 2014), this method was preferred as it allows for the straightforward cross‐validation of multiple imputed datasets.

RESULTS

Sample description

Descriptive characteristics of the sample can be found in Table 1. The majority of the sample was female (54.5%), only few participants (6%) were of non‐Belgian nationality and 15.3% of the students indicated that they were raised in households, in which their parents financial situation was difficult. Parental education was high for both parents for the majority of the students (63.4%), only few students (14.8%) indicated that neither of their parents had a high education level. As can be seen in Table 3, the burden of mental disorders in the sample was quite high. Approximately, one‐third of all students experienced at least one 12‐month disorder (35.2%), and 28.6% reported exactly one, 5.3% exactly two, and 1.3% three or more 12‐month mental disorders. Approximately, half of the sample (52.5%) reported at least one traumatic experience before the age of 17; with 28.4% experiencing parental psychopathology as the most reported type of traumatic experience, followed by bully victimization (25.4%). Every second student also reported at least one 12‐month stressful life event (52.5%, Table 2).

Table 1

Socio‐demographic Variables as Baseline Predictors for Depression Onset during Follow‐up

				Bivariate Modela
	Prevalence			Subsample no. 12‐m MDD
	n (w)	% (w)	(SE)	OR	95%CI	PARP (%)
I. Socio‐demographic variables
Being male	1,021	45.5	0.9	0.68	(0.41–1.12)	−14.4
Age > 18 years	495	22.1	0.8	1.40	(0.83–2.37)	7.0
Non‐Belgian nationality	134	6.0	0.4	1.48	(0.63–3.45)	3.2
Parents’ financial situation difficult	344	15.3	0.7	1.05	(0.53–2.07)	1.0
Parental educationb
Both parents high	1,422	63.4	1.0	(ref)	–	–
Only one parent high	487	21.7	0.8	0.95	(0.54–1.68)	−0.8
None of parents high	333	14.8	0.7	1.03	(0.51–2.10)	0.7
Nonintact familial compositionc	472	21.0	0.8	1.23	(0.68–2.20)	4.3
College‐related socio‐demographics
University Group membership
Human Sciences	1,171	52.2	0.9	(ref)	–	–
Science & Technology	623	27.8	0.8	0.61	(0.35–1.07)	−12.7
Biomedical Sciences	448	20.0	0.7	0.61	(0.31–1.17)	−9.3
Non‐GSE pre‐educational level	131	5.8	0.4	1.33	(0.49–3.57)	1.9

Note. Significant odds ratios/PARPs are shown in bold (α = 0.05); OR = odds ratio; PARP = population attributable risk proportion; GSE = general secondary education.

The bivariate associations are based on a separate model for each row, with the variable in the row as the only predictor in the model, adjusted for lifetime MDD at baseline.

High degree of parental education defined as holding a college bachelor degree or more.

Nonintact familial composition defined as parents being divorced or separated.

Table 3

Twelve‐Month Mental Disorders as Baseline Predictors for Depression Onset during Follow‐up

				Bivariate Modela
	Prevalence			Subsample no. 12‐m MDD
	n (w)	% (w)	(SE)	OR	95%CI	PARP (%)
IV. Twelve‐month mental disorders
Generalized anxiety disorder	70	3.1	0.3	4.11	(1.75–9.70)	7.5
Panic disorder	23	1.0	0.2	0.94	(0.07–12.05)	0.2
Broad mania	22	1.0	0.2	2.75	(0.58–12.98)	1.6
Low risk for alcohol use disorder	1,660	74.0	0.8	(ref)	–	–
Risky or hazardous drinking	505	22.5	0.8	1.17	(0.69–1.96)	3.3
Risk for alcohol dependence	77	3.4	0.4	1.34	(0.39–4.63)	1.2
Drug abuse/dependence	25	1.1	0.2	1.93	(0.39–9.41)	1.0
Nonsuicidal self‐injury	161	7.2	0.5	2.53	(1.28–5.02)	8.7
No STB	2,154	96.1	0.4	(ref)	–	–
Suicidal ideation	64	2.9	0.3	3.76	(1.65–8.57)	5.9
Suicide plans and/or attempts	24	1.1	0.2	9.55	(2.96–30.78)	5.0
Any mental disorder	790	35.2	0.9	2.12	(1.34–3.36)	25.6
Number of mental disorder
0	1,452	64.8	0.9	(ref)	–	–
1	641	28.6	0.9	1.76	(1.09–2.86)	14.3
2	119	5.3	0.4	3.34	(1.55–7.19)	7.8
3+	30	1.3	0.2	6.27	(1.85–21.33)	4.0
F‐test (p‐value)c				F = 4.16 (0.006)

Note. Significant odds ratios/PARPs are shown in bold (α = 0.05); STB = suicidal thoughts and behaviors; OR = odds ratio; PARP = population attributable risk proportion.

The bivariate associations are based on a separate model for each row, with the variable in the row as the only predictor in the model, adjusted for lifetime MDD at baseline.

Cochran–Armitage trend test. The F‐test evaluates significance (α = 0.05) of 200 pooled Cochran–Armitage χ2 (3) linear trend tests.

Table 2

Childhood–Adolescent Traumatic Experiences and 12‐Month Stressful Experiences as Baseline Predictors for Depression Onset during Follow‐up

				Bivariate Modela
	Prevalence			Subsample no. 12‐m MDD
	n (w)	% (w)	(SE)	OR	95%CI	PARP (%)
II. Twelve‐month stressful experiences
Life‐threatening illness or injury of a friend or family member	481	21.5	1.1	1.18	(0.60–2.34)	3.9
Death of a friend or family member	437	19.5	1.0	1.07	(0.52–2.20)	1.7
Break‐up with a romantic partner	394	17.7	1.0	2.63	(1.37–5.09)	20.3
Romantic partner cheated	87	3.9	0.5	3.81	(1.11–13.07)	8.5
Serious betrayal someone else than partner	210	9.4	0.7	2.35	(1.07–5.17)	10.4
Serious ongoing arguments or break‐up with friend or family member	284	12.7	0.8	2.78	(1.46–5.31)	16.7
Life‐threatening accident	22	1.0	0.3	2.60	(0.21–32.11)	1.9
Seriously physically assaulted	65	2.9	0.4	1.22	(0.18–8.09)	1.2
Sexually assaulted or raped	8	0.4	0.1	2.26	(0.13–38.25)	0.6
Any serious legal problem	44	1.9	0.3	2.20	(0.31–15.73)	2.6
Any stressful event	1,177	52.5	1.2	2.12	(1.20–3.75)	34.5
Number of stressful experiences
0	1,065	47.5	1.2	(ref)	–	–
1	620	27.6	1.1	1.53	(0.83–2.82)	8.8
2	355	15.9	0.9	2.04	(1.01–4.13)	9.8
3+	202	9.0	0.7	4.29	(1.85–9.96)	16.2
F‐test (p‐value)b				F = 2.69 (0.046)
III. Traumatic experiences (≤ age 17)
Parental psychopathology	637	28.4	1.1	1.96	(1.13–3.39)	19.7
Physical abuse	96	4.3	0.5	2.01	(0.74–5.45)	4.0
Emotional abuse	329	14.7	0.8	2.51	(1.34–4.71)	16.4
Sexual abuse	18	0.8	0.2	8.01	(1.64–39.06)	3.6
Neglect	116	5.2	0.5	1.44	(0.51–4.05)	2.4
Bully victimization	570	25.4	1.1	1.19	(0.68–2.09)	4.5
Dating violence	115	5.1	0.5	2.94	(1.03–8.40)	8.5
Any traumatic experience	1,141	50.9	1.2	2.00	(1.17–3.43)	31.5
Number of traumatic experiences
0	1,099	49.1	1.2	(ref)	–	–
1	688	30.7	1.1	1.46	(0.80–2.66)	9.0
2	272	12.1	0.8	2.57	(1.28–5.15)	11.4
3+	181	8.1	0.7	3.43	(1.52–7.74)	11.7
F‐test (p‐value)b				F = 3.04 (0.029)

Note. Significant odds ratios/PARPs are shown in bold (α = 0.05); OR = odds ratio; PARP = population attributable risk proportion.

The bivariate associations are based on a separate model for each row, with the variable in the row as the only predictor in the model, adjusted for lifetime MDD at baseline.

Cochran–Armitage trend test. The F‐test evaluates significance (α = 0.05) of 200 pooled Cochran–Armitage χ2 (3) linear trend tests.

Twelve‐month incidence of MDD during college freshman year

Twelve‐month prevalence of MDD at baseline was 11.0% (95%CI: 10.0–12.0%, n = 277/2,519). Among the remaining cases without 12‐month MDD (n = 2,242), lifetime prevalence of MDD was only 3.5% (95%CI: 2.9–4.2%; n = 79). All analyses were restricted to the 2,242 students who had no history of MDD during the 12 months prior to the baseline survey. The incidence of depressive disorder in the first year after college matriculation was estimated at 6.9% (95%CI: 5.3–8.4%; n = 154/2,242). Most of these cases were first‐onset incidence cases (94.15%; n = 145/154).

Individual‐ and population‐level predictors of 12‐month MDD incidence

Models adjusting for lifetime MDD at baseline (Tables 1, 2, 3) revealed the following key findings. First, socio‐demographic variables did not significantly predict the onset of MDD in students in their first year of college. Second, the most important predictors of MDD onset at the individual level were 12‐month suicide plans and/or attempts (OR = 9.55), sexual abuse prior to the age of 17 (OR = 8.01), three or more 12‐month mental disorders other than MDD (OR = 6.27), three or more 12‐month stressful events (OR = 4.29), and 12‐month generalized anxiety disorder (OR = 4.11). However, the impact of these predictors at population level were all small (PARP < 12%) due to the low prevalence of these predictors.

Third, large proportions of MDD onset were attributable to any 12‐month mental disorder at baseline (other than MDD, PARP = 25.6%), any childhood–adolescent trauma (PARP = 31.5%), and any stressful experience in the past year (PARP = 34.5%). Specific associations regarding stressful experiences included: break‐up with a romantic partner, romantic partner cheated, serious betrayal by someone else than partner, and serious ongoing arguments or break‐up with a friend or family member (median OR = 2.7; median PARP = 13.5%). In relation to any childhood–adolescent trauma, specifically parental psychopathology, emotional abuse, sexual abuse, and dating violence (median OR = 2.7; median PARP = 12.5%) were significantly associated with MD. With regard to being at risk for comorbid mental health issues, specific associations included generalized anxiety disorder, nonsuicidal self‐injury, suicidal ideation, and suicide plans and/or attempts (median OR = 3.9; median PARP = 6.7%).

Fourth, a positive dose–response relationship was found within each domain, with MDD risk among students who had three or more risk indicators substantially elevated both at the individual level (ORs between 3.4 and 6.3) and at the population level (PARPs between 4.0 and 16.2%). Finally, when considering both individual‐ and population‐level effects, the most important risk indicators (i.e., OR ≥ 2 and PARP ≥ 15%) were any traumatic experience prior to the age of 17, break‐up with a romantic partner in the past year, serious ongoing arguments or break‐up with a friend or family member in the past year and three or more stressful life events in the last 12 months.

Multivariate model for MDD onset during freshman year

The total effect size (Nagelkerke pseudo‐R2) of risk indicators was 0.23. The prediction model had a reasonable performance with a cross‐validated AUC of 0.73 (SE = 0.04). The 10% of students at highest predicted risk for subsequent onset of MDD within the first 12 months after college matriculation included 36.1% (SE = 6.1) of all observed MDD cases (Table 4). The probability of MDD onset in this 10% of respondents was 24.7% (SE = 4.9). The only significant predictors in the final model, when adjusted for all other risk domains, were suicidal ideation (OR = 2.88; 95%CI = 1.10–7.56; PARP = 4.4%) and suicide plans and/or attempts (OR = 6.77; 95%CI = 1.55–29.62; PARP = 3.9%). A full overview of the multivariate estimates can be found in the supplementary materials.

Table 4

Concentration of Risk of Depression Cases in Different Proportions of Incoming Freshmen at Highest Predicted Risk based on a Multivariate Modela Including all Risk Factors

	Depression Onset
% at Highest Predicted Risk	Sensitivity (%[SE])b	PPV (%[SE])c
100	100.0 (0.0)	6.9 (0.8)
90	96.0 (2.4)	7.3 (0.9)
80	91.8 (3.4)	7.9 (1.0)
70	87.4 (4.1)	8.6 (1.2)
60	82.7 (4.8)	9.5 (1.3)
50	77.2 (5.2)	10.6 (1.5)
40	71.0 (5.9)	12.2 (1.8)
30	63.1 (6.3)	14.4 (2.3)
20	52.4 (6.5)	17.9 (3.0)
10	36.1 (6.1)	24.7 (4.9)

See the model in the supplementary materials covering multivariate model construction (Supplementary Tables 1, 2). Model‐based AUC values were 0.78 [SE = 0.03] for depression onset. Cross‐validated AUC values were 0.73 [SE = 0.04].

Sensitivity = proportion of depression cases found among the row % of respondents at highest predicted risk, based on cross‐validated predicted probabilities.

Positive predictive value (PPV) = probability of effectively developing a depression when being among the row % at highest predicted risk, based on cross‐validated predicted probabilities.

DISCUSSION

Main findings

This prospective study examined the onset of MDD in a large representative sample of college students. In the first year of college, the incidence of MDD was estimated at 6.9%. Among the 10% of students with the highest predicted risk of MDD onset based on our model, approximately one out of four developed MDD. Suicidal plans and/or attempts were most strongly associated with MDD onset at the individual level. The largest population‐level effects, however, were found for any 12‐month mental disorder at baseline (PARP = 25.6%), a history of any childhood–adolescent trauma (PARP = 31.5%), and stressful experiences in the past 12 months (PARP = 34.5%).

Limitations

Several limitations are noteworthy. First, response rates were moderate (61.0% at baseline; 57.5% at follow‐up). However, these response rates compare favourably to those achieved in other large‐scale prospective college student surveys (39–44%) (Eisenberg et al., 2013; Paul, Tsypes, Eidlitz, Ernhout, & Whitlock, 2015). In addition, state‐of‐the‐art missing data techniques were applied to increase the representativeness of the findings. Nonetheless it is possible that systematic nonresponse might have biased results. Second, baseline risk for mental disorders was not assessed by diagnostic interviews but with self‐report measures and a categorical cut‐off scoring system. The latter measures were well‐validated screening scales used in prior general population surveys that have shown high concordance with with blinded clinical diagnoses in clinical reappraisal studies (Kessler et al., 2010). However, it remains unknown whether screening scale performance is different among college students. Although we plan to carry out clinical reappraisal studies to address this limitation in future iterations of the WMH college surveys, this has not yet been done and caution is consequently needed in interpreting results regarding prevalence estimates. Third, the survey was conducted among freshmen in one Belgian college. The findings might not generalize to college students from other universities in different countries or cultures. Finally, although we included a large set of known risk indicators for MDD onset, some important risk indicators were not assessed, such as subsyndromal depression, chronic somatic conditions, personality traits/disorders, psychotic experiences/disorders, poor self‐perceived health, low emotion regulation skills, low self‐esteem, low resilience, and neuroticism (Berking, Wirtz, Svaldi, & Hofmann, 2014; Cole & Dendukuri, 2003; Ebert, Hopfinger, & Berking, 2017; Korten, Comijs, Lamers, & Penninx, 2012; Pelkonen, Marttunen, Kaprio, Huurre, & Aro, 2008; Wild et al., 2016). As a result, the strength of the composite risk index found here should be considered a lower bound estimate compared to the estimate that might be obtained in future research that includes additional predictors. A related limitation is that we used conventional research analysis methods to develop the risk model. It is likely that we will be able to improve on this performance in planned cross‐national analyses using machine learning methods (Kessler et al., 2016, 2017). Finally, we did not assess serious life events during follow‐up. Therefore, we do not know which MDD incidence cases are due to baseline vulnerability and which due to exposure to random traumas that could not be predicted at baseline (e.g., sexual assault, death of a parent). Such information would not only be important to inform about strategies to improve prediction accuracy of the algorithm, but also relevant for the development of appropriate prevention strategies. This should be explored in future studies.

Implications for clinical practice and future research

Our study has relevant implications for clinical practice and future research. First, to the best of our knowledge, this study is among the first that prospectively estimated the 1‐year incidence proportion of MDD in students during their first year of college. The reported incidence proportion is somewhat higher than the estimated incidence of MDD among college students based on the World Health Organization (WHO) World Mental Health Surveys (Auerbach et al., 2016). Differences may be explained by geographical or methodological differences (i.e., adjustment for college attriters or the use of retrospective designs in the WHO surveys). Our data suggest that the first year in college constitutes a risk period for the onset of MDD. In fact, the vast majority of observed MDD cases were incidence cases (94.8%), thus this period in life seems to be an opportune point in time to intervene preventively.

Second, our study further adds to the cumulating evidence that the development of risk‐prediction for psychiatric disorders is feasible (Bernardini et al., 2017) and provides evidence that a multivariate prediction model can be a useful tool to accurately predict the onset of MDD during college. Prediction accuracy (AUC = 0.73) was comparable to the few prediction algorithms that have been evaluated for depression within a general population (AUC = 0.71) (Nigatu, Liu, & Wang, 2016) and primary care samples (AUC = 0.82) (Bellon et al., 2011) and are also comparable to other fields of medicine (Karnes et al., 2017; ten Haaf et al., 2017). However, to achieve optimal performance, recalibration of models is needed prior to applying the models to a new population. As predictors included in the model contribute to a model's calibration capacity, it is important to develop target group‐specific prediction algorithms because predictors for the risk of MDD onset and their predicted values may differ among different population segments (i.e., college students). The risk prediction algorithm could be used to predict future MDD among incoming freshmen. More research on the validation of such specific risk prediction models is warranted; nevertheless, it is a promising methodology and enables interesting opportunities for the development of individualized approaches for MDD in emerging adults. Data on self‐reported risk factors could easily be collected by means of regular student surveys. The assigned predicted probabilities could then be used as a way to delineate those at highest risk for onset of MDD in the following year. However, although risk‐prediction algorithms might be of high value for detecting students at risk, for developing mental health problems, it should be noted that relying only on procedures based on students self‐reports might be not sufficient to detect students at risk, and other measures such as staff training and awareness campaigns should not be neglected. Students considered to be at high risk could be offered preventive interventions, for example delivered through the internet (Buntrock et al., 2016, 2017; Ebert et al., 2018; Harrer et al., 2018). Based on our model, over one third of MDD cases will occur in the 10% of students at highest predicted risk. However, this does not imply that students at lower risk do not warrant preventive interventions. More research is needed to obtain information on the needs of students who are associated with different risk levels and which interventions work best at varying levels. Due to high comorbidity rates between emotional disorders (Beekman et al., 2000) and overlapping risk factors (de Graaf, Bijl, Smit, Vollebergh, & Spijker, 2002), such studies should also explore relative advantages of disorder versus transdiagnostic and individual tailored preventive interventions (Weisel et al., 2018). In addition, clinical outcome and cost‐effectiveness research based on varying risk thresholds should be conducted so that intervention decisions derived from the prediction model are evidence‐based.

Third, the population‐level estimates offered relevant insights into the design of future interventions. Based on individual‐level effect sizes, one could argue preventive interventions should focus on students who have been either sexually abused or who had suicide plans or attempts (OR > 8). However, the impact of these factors on a population level overall was very low (PARP < 5%) due to low prevalence. In contrast, targeting students who experience any childhood–adolescent trauma, such as emotional abuse, could have a beneficiary effect for about one third of subsequent depression onsets. Also, the incidence of depression among these students designated to be at high risk (24.7%) is sufficiently high that the cost‐effectiveness of a preventive intervention has a reasonable chance of being within an actionable range. Likewise, targeting students whose parents have a mental disorder could potentially reduce one fifth of depression cases (19.7%). Targeting students at college entry who broke recently up with a romantic partner, could have preventive effects for one fifth of subsequent MDD cases. In general, offering such specific interventions, subsequent to a screening at college student entrance, might result in a more developmental approach to the prevention of depression during adolescence and emerging adulthood which may ultimately help decrease the large burden associated with this disorder in young people.

Finally, prevalence estimates of STBs were, potentially due to the exclusion of MDD baseline cases in the present study, somewhat lower than recent estimates of STB cross‐national prevalence rates (Mortier et al., 2018). These low prevalence rates lead, as stated above, to a comparable low proportion of MDD cases in the population attributable to STBs. However, due to the disabling nature of STBs and their adverse consequences, there nevertheless is a clear need for interventions that are specifically designed to reach this underreached population (Mortier et al., 2018) and help affected students to cope effectively.

Supporting Information

Click here for additional data file.