Appendix 1
Overall consideration and preliminary analyses.
This study carefully adjusted for clustering effects and sampling weights. Preliminary analyses, including examination of clustering effect, identification of and adjustments for missing data, outliers, and non-normality, were completed. During the analyses, non- linearity, statistical power, and moderation effects were carefully examined. We provided detailed strategies considered as below.
Clustering. Since Add Health used a school-based design for sample recruitment, students were clustered within each school. Intraclass correlations (ICC) were examined to evaluate the size of cluster effects. The cutoff points used for this study were consistent with previous suggestions for low (ICC value lower than 0.059), medium (ICC between 0.059 and 0.138), and high (ICC above 0.138) values (Cohen, 1988). The “CLUSTER” algorithm in Mplus version 8 (Muthén & Muthén, 1998-2017) or “PROC SURVERYLOGISTIC” process in SAS, were used in this study to account for clustering effects.
Weighting scheme. To accommodate non-normality and heteroscedasticity, analyses were adjusted for sampling weights and consistent standard errors (Brownstein et al., 2010;
Brownstein et al., 2011; Harris, Halpern, Biemer, Liao, & Dean, 2019a). Weighted estimation was also used in multivariate analyses with outcome-based sampling to adjust for bias and specification error (Asparouhov & Muthén, 2009; Feinberg, 1989). Given the specific features of the Add Health design previously described (i.e., clustered observations, unequal probability of selection, stratified sampling), when conducting analyses for this study, the survey packages, SAS (“PROC SURVERYLOGISTIC”) and Mplus (Muthén & Muthén, 1998-2017) were used for design-based analysis, which is recommended to adjust estimates for clustering and the unequal probability of selection (e.g., sampling weights) when
computing point estimates and standard errors (Chen & Chantala, 2014). Since the outcome
variables were measured at each of the four waves (Wave I-IV), the weight for Wave IV (i.e.,
“GSWGT4”) was used as the study sampling weight, in keeping with research indicating that the choice of sampling weight for longitudinal analysis should be determined by the data collected at the most recent time point (Chen & Chantala, 2014).
Missing data. Missing data were expected to be minimal for most variables, given the rigorous data collection procedures of Add Health (Harris et al., 2019b). To explore possible missing data bias (i.e., possible differences between missing and non-missing values), a dummy variable representing the presence or absence of missing values was created for each variable in the model. Correlations between this dummy variable and all other variables in the models, including sociodemographic variables (e.g., race/ethnicity, sex, sexual orientation, SES), were used to examine whether the data were missing at random (MAR). A similar approach, through creating a dummy variable indicating loss to follow up (coded as 1 = respondent not followed up, 0 = respondent followed up), was used to examine missing data due to attrition and correlations with a range of sociodemographic variables and
psychological variables (e.g., depressive symptoms) at Wave I to check for attrition bias.
Most variables had less than 5% missingness, except for public assistance receipt (11.52%
missingness). The assumption of MAR was confirmed, and attrition bias was minimum. To manage missingness, multiple imputations were performed using a procedure that accounted for the hierarchical structure of the Add Health data, creating 50 imputed datasets with 20 auxiliary variables (e.g., age, sex, family income, racial/ethnic groups) (Bodner, 2008; I. R.
White, Royston, & Wood, 2011). These analyses were conducted using Mplus version 8 (Asparouhov & Muthén, 2010; Muthén & Muthén, 1998-2017). In addition, full information maximum likelihood (FIML) methods (Asparouhov & Muthén, 2010; Muthén & Muthén, 1998-2017) were implemented in Mplus (using robust maximum likelihood estimator [“MLR”] estimators) and SAS (“PROC SURVERYLOGISTIC” process, specifying a
missing option for MAR) to pursue parameter estimates and model tests (details in Appendix 2).
Outlier analysis. During the preliminary analysis, data were checked for multivariate outliers by examining leverage scores for each case based on the transformed Mahalanobis distance scores (Mahalanobis, 1936). This procedure used the “MAHALANOBIS” command and robust Mahalanobis estimation (Rousseeuw & Driessen, 1999) using the “MCD”
command in SAS, which are the recommended approaches for outlier detection in the multivariate context . The threshold for a case being considered an outlier was a leverage score four times greater than the mean leverage, but no outlying cases were detected.
Non-normality. Normality was examined using the “PROC UNIVARIATE” process in SAS (O'Rourke & Hatcher, 2013). Because the sample size in this study was larger than 2,000, the Kolmogorov-Smirnov test was used, and variables with p-values less than .05 were considered to have a skewed distribution (Lilliefors, 1967; Massey Jr, 1951). Robust
maximum likelihood (MLR) estimator using the Huber-White (1980) sandwich standard error estimator was employed to handle skewed variables in latent variable modeling using Mplus (Muthén & Muthén, 1998-2017). This estimation strategy works well for sample sizes over 200 (Bentler & Yuan, 1999). As logistic regression does not have strict assumptions for normality, data obtained from Mplus can be used directly for further analyses by the “PROC SURVEYLOGISTIC” process in SAS (O'Rourke & Hatcher, 2013).
Nonlinear trajectories and nonlinear relational forms. Based on previous studies of suicidal trajectories among other populations (Rueter & Kwon, 2005), it was possible that there would be multiple subgroups of suicidal trajectories, including increasing, decreasing, U-shape, and inverted U-shaped trajectories. Hence, both linear and nonlinear relationship forms were estimated. This study followed a sequential protocol to examine linear models, quadratic or cubic models: 1) generate linear models specifying an initial score reflecting the mean initial
adolescent suicidal ideation score, as well as a linear trend factor showing the mean linear change in suicidal ideation over time (linear slope), 2) estimate quadratic models by adding a third factor (a quadratic term of suicidal ideation) representing the mean quadratic change in suicidal ideation (quadratic slope) over time, and 3) when the quadratic model did not fit significantly better than the linear model, the cubic model (generated by adding a cubic term of suicidal ideation) was not pursued. Model comparisons were examined using 1) chi‐square statistics (χ2) of the linear, quadratic, and cubic models, and 2) whether the quadratic and cubic terms were significant. A statistically significant reduction in χ2 indicated that the latter model better represented the trends in actual data relative to the prior model. Our initial results suggested the linear model provided a better fit and explanations based on the data.
Thus, latent class growth analysis (LCGA) was used, and corresponding results were presented.
Appendix 2
Strategies of multiple imputations
Missing data in Add Health and study sample
Add Health is advantageous in collecting longitudinal, social, and behavioral data beginning in early adolescence and extending to adulthood. Meanwhile, there are three types of missing data biases embedded in the Add Health longitudinal design (Brownstein et al., 2010; Brownstein et al., 2011; Harris et al., 2019a). We carefully considered and evaluated the missingness pattern, impact of missing data, and decided technologies to handle potential bias. First, the sample included planned missing values (i.e., not eligible as defined by the study design). For instance, adolescents in 12th grade in Wave I was not included in Wave II, and individuals younger than 18 years old were not included in Wave III (Kathleen M Harris, 2013). These individuals were not included in this study, as there was no theoretical
estimation of their value importance at one of the time points, and no meaningful ways to retrieve their responses.
Second, attrition would occur at each Wave in the longitudinal design: Nonrespondents in Waves I-IV for several reasons, including eligible individuals with whom interviewers (e.g., filed contractors) were not able to establish contact, groups who were located but unavailable to participate or unable to participate for physical, linguistic, or mental
incapability, and groups who refused to participate (Brownstein et al., 2010; Brownstein et al., 2011). Previous studies found differential attrition rates in Add Health by sex, age, race/ethnicity, socioeconomic status (SES), urban residence, and immigrant status across time, with higher attrition rates for males, older-age individuals, racial/ethnic minorities, lower SES, rural and immigrant respondents at Waves III and IV (Brownstein et al., 2010;
Brownstein et al., 2011; Harris et al., 2019a). In general, these attrition patterns are consistent with most national longitudinal studies (e.g., National Survey of Families and Households;
Midlife in the United States; Harris et al., 2019b). Specifically, at each Wave, in order to examine whether patterns of attrition contribute to biased estimation of survey outcomes, nonresponse analyses were conducted by comparing respondents and nonrespondents on a range of sociodemographic, health, behavioral and attitudinal indicators at baseline, and by evaluating the extent to which significant differences between respondents and
nonrespondents may introduce bias in study results (Brownstein et al., 2010; Brownstein et al., 2011; Harris et al., 2019a). Results indicated that total and relative biases were minimum after adjusting final sampling weights when conducting analyses.
Additionally, the sample at each follow-up wave (Waves II-IV) accurately represented the same national population as the Wave I sample: 7-12th graders in the US in 1994-1995 (Brownstein et al., 2010; Brownstein et al., 2011; Harris et al., 2019a). In particular, analysis of attrition bias at Wave IV (by calculating total, component, and relative bias between respondents and nonrespondents using known answers of Wave I and comparing with Wave IV) indicated that total and relative bias of variables that have been repetitively measured (e.g., demographic characteristics, school experiences, health attitudes and physical activities, substance abuse, violence, and delinquency) rarely exceeded 1%, which is much smaller than the 20-50% prevalence rates of most indicators at Wave I (Brownstein et al., 2011).1 Besides, for variables measuring Health (e.g., overall health) and health risk behaviors (e.g., fight at school), none of the bias due to nonresponse was statistically significant. In sum, nonresponse bias due to attrition is negligible, and differences in measurements between respondents and nonrespondents were due to random variation. Therefore, the results of this study are not
1
Total bias is defined as the bias due to any form of non-response. Component bias is bias due to an individual category of non-response. The four components are “No Contact,”
“Unable,” “Refusal,” and “Other.” Components are additive in that the sum of the four component biases equals the total bias. Relative bias is defined as the total bias for a
particular measure (e.g., smoking) divided by the prevalence of that measure (Brownstein et al., 2011).
statistically affected by attrition across waves, and the attrition bias (or nonresponse bias) and selective survival bias were minimal.
Third, missing data at each Wave may occur when participants were included in the interviews but failed to answer some or all the measures. In this study, the study sample included in each Wave had various amounts of missing data in terms of different variables. In Study 1, the magnitude of missing ranged from no missing (e.g., sex, race/ethnicity), to a low amount of missing (sexual orientation, 0.11%), and to a rather high amount of missing
(federal welfare assistance, 11.52%). In Study 2, the magnitude of missing ranged from no missing (e.g., sex, race/ethnicity), to a low amount of missing (family size, 0.06%), and to a rather high amount of missing (peer network density, 28.02%). In Study 3, the magnitude of missing ranged from no missing (e.g., sex, race/ethnicity), to a low amount of missing (sexual orientation, 0.11%), and to a rather high amount of missing (closeness to father at Wave IV, 17.54%). In order to improve the accuracy of parameter estimation and reduce bias due to missing, multiple imputations were conducted at the item level of each indicator before the preliminary descriptive analysis, and imputed variables were used to calculate a full score whenever applied (Schafer & Olsen, 1998).
Multiple imputations and application in the current study
Multiple imputations (MI) create several different plausible imputed datasets by
replacing missing values repeatedly with plausible values and appropriately combines results from analyzing each completed imputed dataset (Graham, 2009; Graham, Hofer, & Piccinin, 1994; Schafer & Graham, 2002; Sterne et al., 2009). Compared to traditional procedures to handle missing data (e.g., simple listwise deletion, pairwise deletion, mean substitution, or regression substitution), MI is particularly powerful to minimize bias among longitudinal data with complex survey design (Allison, 2002; Graham et al., 1994; Little & Rubin, 1989;
Schafer & Graham, 2002; Schafer & Olsen, 1998). MI improves the trustworthiness of
parameter estimates, ensures that the standard errors associated with the parameter estimates are not artificially decreased, and enhance the confidence of inferences (Graham et al., 1994;
Schafer & Graham, 2002; Schafer & Olsen, 1998).
The main assumption of MI is that the missing data are missing at random (MAR;
Schafer & Graham, 2002). Based on prior analysis of missing data patterns in Add Health, results indicated that the missingness in Add Health is due to random variation (Brownstein et al., 2010; Brownstein et al., 2011; Harris et al., 2019a). Thus, this study applies MI to estimate missing data and improve further estimation. To maximize the likelihood of meeting the assumption that the data are MAR, this study also included relevant auxiliary variables (see details below) in MI (Collins, Schafer, & Kam, 2001).
Detailed procedures of multiple imputations
Four stages of MI were conducted for each of the study studies (Sterne et al., 2009).
Stage 1: Determine auxiliary variables and imputed variables
The number of missing time points for suicidal ideation and suicide attempts was used as a grouping variable, and a series of one-way analyses of variance were conducted to determine what auxiliary variables were significantly related to missingness in the grouping variable. Together with variables without missing in each Wave (i.e., sex, age at each Wave), several auxiliary variables, including stratification, weights (Waves I-IV), race/ethnicity, age (Waves I-IV), sex were incorporated into the MI models. The original measurement scales (e.g., suicidal ideation as categorical variables; depressive symptoms treated as continuous variables) were used in the imputation process in Mplus. Including auxiliary variables helps to ensure that the imputed data are unbiased (Davis-Kean et al., 2008).
Since the attrition bias (or nonresponse bias) and selective survival bias were found to be negligible (Brownstein et al., 2011), there was no statistically significant difference between respondent and nonrespondent. Therefore, missing variables due to attrition are not necessary
to consider in the MI procedure (Brownstein et al., 2010; Brownstein et al., 2011; Harris et al., 2019a).
Stage 2: Generate imputed datasets
Mplus version 8 (Muthén & Muthén, 1998-2017) was used to conduct MI in this study.
Following the recommended number of imputation datasets in previous simulation study (Graham, Olchowski, & Gilreath, 2007), we created m=50 imputed datasets in Mplus version 8 using the TYPE=COMPLEX command, which accounted for clustering, stratification, and sample weights (Asparouhov & Muthén, 2010; Muthén & Muthén, 1998-2017). Continuous variables (e.g., age) and categorical variables (e.g., suicidal ideation) were separately
specified according to the original measurement of each item in the Mplus commands
(Asparouhov & Muthén, 2010; Grund, Lüdtke, & Robitzsch, 2018). For categorical variables with multiple categories (e.g., race/ethnicity, maternal education), dummy variables were created prior to enter them in the imputation models.
Stage 3: Analyze imputed dataset and generate pooled results
In this study, LCGA was conducted using the 50 imputed datasets, with results averaged across the 50 analyses of imputed datasets using the Rubin method (Muthén & Muthén, 1998-2017; Rubin, 1987). All analyses were conducted in Mplus v.8 using the
TYPE=IMPUTATION command (Muthén & Muthén, 1998-2017).
Stage 4: Output dataset for further analysis
Since Mplus v.8 does not support a series of combined analysis of LCGA and logistic regression in the imputed dataset through one step, either does it support conducting further analyses with averaged class designation from the imputed dataset while adjusting the complex sampling design, after LCGA, this study randomly selected one of the 50 imputed datasets for further analyses (e.g., bivariate analyses, multivariate multinomial logistic regressions). In order to keep a consistency of the parameter estimates and class separation of
LCGA, final estimates in the output of LCGA were saved and used to give starting values when outputting the one selected imputed dataset (Muthén & Muthén, 1998-2017). Full information maximum likelihood (FIML) methods were implemented in Mplus (using robust maximum likelihood estimator [“MLR”] estimators) and SAS (“PROC
SURVERYLOGISTIC” process, specifying a missing option for MAR) to pursue parameter estimates and model tests (Asparouhov & Muthén, 2010; Muthén & Muthén, 1998-2017). To check the robustness and certainty of latent class and parameter estimation, we further
conducted sensitivity analyses using different imputed datasets. Results stayed the same interpretation as the ones presented.
Appendix 3
Sample characteristics and group comparisons by suicidal behaviorsa
Total Suicidal Ideation at W1
P b Suicide Attempts at W1 P b (n = 9,421) (n = 1,276, 13.82%) (n = 366, 4.13%)
Age at Baseline (M±SD) 14.99 (1.61) 15.20 (1.53) <.001 15.03 (1.52) .96
Sex, n(%) <.001 <.001
Male 4279 (46.85) 33.95 22.94
Female 5142 (53.15) 66.05 77.06
Sexual Orientation, n (%) <.001 <.001
Heterosexual 8101 (85.70) 75.77 68.61
Sexual minority 1310 (14.30) 24.23 31.39
Race/ethnicity, n (%) .30 .55
White 5319 (70.48) 72.24 73.11
Black 1931 (13.84) 11.89 13.79
Latinx 1448 (10.99) 10.76 10.22
Asian 597 (3.49) 3.87 2.47
American Indian 73 (0.71) 0.95 0.31
Other races 35 (0.43) 0.22 /
Multiracial 13 (0.07) 0.06 0.10
Maternal Education, n (%) .21 .46
No school/Less than high school 1514 (16.21) 17.75 17.49
High school or equivalent 3627 (44.28) 39.73 43.09
Some college 1162 (13.01) 13.92 16.49
Graduate school and higher 2599 (26.50) 28.61 22.93
Public Assistance, n (%) .04 .07
None 6089 (73.68) 70.41 67.88
At least one 2247 (26.32) 29.59 32.12
Depression (M±SD) 0.64 (0.47) 1.04 (0.57) <.001 1.12 (0.57) <.001
Mental Health Treatment (W1) 1089 (12.57) 354 (29.60) <.001 147 (41.40) <.001
Note. The left side presents the sociodemographic and depressive symptoms of the total sample based on unimputed data. The right side shows the bivariate analysis between suicidal behaviors and sociodemographic characteristics as well as depression. W1=Wave I.
a Participant numbers based on unweighted data and percentages based on weighted data.
b Comparing participants with suicidal ideation and suicide attempts to those without suicide attempts. χ² analyses were used for categorical variables, whereas the t-test was used for continuous variables.
Appendix 4
Sociodemographic and psychological differences across suicidal trajectoriesa
Suicidal Ideation Trajectory Suicide Attempt Trajectory
Trajectory 1 [Low- stable]
Trajectory 2 [High- decreasing]
Trajectory 3 [Moderate- decreasing-
increasing] P χ² / F
Trajectory 1 [Low-stable]
Trajectory 2 [Moderate-
decreasing] P χ² / F
(n = 8,608,
91.37%) (n = 322,
3.42%) (n = 491,
5.21%) (n = 9,213,
97.79%) (n = 208, 2.21%)
Age (M±SD) 14.98 (0.12) 15.26 (0.14) 14.97 (0.17) .15 F(2, 127) = 1.92 15.04 (0.11) 14.91 (0.18) .40 F(1, 128) = 0.73
Sex, n (%) < .001 χ²(2) = 55.10 .62 χ²(1) = 0.25
Male 4008
(46.56) 96 (29.81) 175 (35.64) 4181 (45.38) 98 (47.12)
Female 4600
(53.44) 226 (70.19) 316 (64.36) 5032 (54.62) 110 (52.88)
Sexual Orientation, n (%) < .001 χ²(2) = 179.29 < .001 χ²(1) = 27.84
Heterosexual 7523
(87.50) 246 (76.40) 332 (67.62) 7948 (86.36) 153 (73.56)
Sexual minority 1075
(12.50) 76 (23.60) 159 (32.38) 1255 (13.64) 55 (26.44)
Race/ethnicity, n (%) .05 χ²(12) = 15.74 .17 χ²(6) = 9.00
White 4834
(56.19) 190 (59.01) 295 (60.08) 5205 (56.53) 114 (54.81)
Black 1803
(20.96) 47 (14.60) 81 (16.50) 1897 (20.60) 34 (16.35)
Latinx 1318
(15.32) 57 (17.70) 73 (14.87) 1405 (15.26) 43 (20.67)
Asian 539 (6.27) 25 (7.76) 33 (6.72) 585 (6.35) 12 (5.77)
American Indian 65 (0.76) 2 (0.62) 6 (1.22) 70 (0.76) 3 (1.44)
Other races 33 (0.38) 1 (0.31) 1 (0.20) 34 (0.37) 1 (0.48)
Multiracial 11 (0.13) 0 (0.00) 2 (0.41) 12 (0.13) 1 (0.48)
Maternal Education, n (%) .05 χ²(6) = 12.42 .37 χ²(3) = 3.12
No school/Less than high 1362 74 (24.03) 78 (17.14) 1473 (16.91) 41 (21.24)
school (16.73) High school or equivalent 3329
(40.90) 111 (36.04) 187 (41.10) 3552 (40.79) 75 (38.86)
Some college 1069
(13.13) 33 (10.71) 60 (13.19) 1135 (13.03) 27 (13.99)
Graduate school and higher 2379
(29.23) 90 (29.22) 130 (28.57) 2549 (29.27) 50 (25.91)
Public Assistance, n (%) .36 χ²(2) = 2.05 .03 χ²(1) = 4.62
None 5572
(73.13) 208 (74.82) 309 (70.39) 5966 (73.20) 123 (66.13)
At least one 2047
(26.87) 70 (25.18) 130 (29.61) 2184 (26.80) 63 (33.87)
Depression (M±SD) 0.60 (0.44) 1.17 (0.58) 0.92 (0.60) < .001 F(2, 127) =
359.77 0.63 (0.01) 1.13 (0.06) < .001 F(1, 128) = 73.62 Mental Health Treatment
(W1) 875 (10.95) 78 (28.68) 136 (29.35) < .001 χ²(2) = 191.52 1004 (11.80) 85 (44.08) < .001 χ²(1) = 179.90 Note. a Chi-square (χ²) tests for categorical variables, Analysis of variance (ANOVA) analyses for continuous variables by suicidal ideation trajectories, t-tests for continuous variables by suicide attempt trajectories. W1=Wave I.
Appendix 5
Results of the multinomial logistic regression with interactions
Suicidal Ideation Trajectoriesa. Suicide Attempt Trajectoriesa.
Trajectory 2 Trajectory 3 Trajectory 2
[High-decreasing] [Moderate-decreasing-
increasing] [Moderate-decreasing]
OR (95% CI) P OR (95% CI) P OR (95% CI) P
Sex Male
Female 1.53 (0.95-2.45) .08 0.74 (0.47-1.17) .20 0.52 (0.29-0.92) .02
Sexual Orientation Heterosexual
Sexual minority 1.80 (1.08-3.01) .02 2.93 (1.85-4.62) < .001 2.33 (1.28-4.24) .006
Age (Wave I) 1.05 (0.96-1.16) .30 0.95 (0.88-1.03) .23 0.95 (0.83-1.08) .44
Race/Ethnicity White
Black 0.29 (0.07-1.17) .08 0.20 (0.07-0.53) < .001 0.55 (0.15-2.02) .37
Latinx 2.15 (0.77-6.00) .14 1.34 (0.49-3.69) .56 1.83 (0.61-5.47) .28
Others 2.19 (0.34-14.2) .41 0.59 (0.22-1.56) .28 1.19 (0.34-4.14) .78
Maternal Education College and higher
Less than college 0.93 (0.60-1.44) .73 0.99 (0.73-1.34) .92 1.08 (0.70-1.68) .72
Public Assistance None
At least one 0.71 (0.46-1.12) .14 0.84 (0.56-1.25) .38 1.46 (0.80-2.67) .22
Mental Health Treatment (MHT, W1) No
Yes 1.82 (1.14-2.91) .01 2.43 (1.73-3.42) < .001 3.80 (2.07-7.00) < .001 Race/Ethnicity × Sex
White × Sex
Black × Female 1.26 (0.42-3.79) .68 1.75 (0.73-4.17) .21 0.72 (0.20-2.64) .62
Latinx × Female 0.93 (0.34-2.53) .89 1.41 (0.43-4.62) .57 0.99 (0.28-3.46) .99
Other race × Female 0.23 (0.029-1.90) .17 1.70 (0.52-5.60) .38 0.43 (0.079-2.32) .32 Race/Ethnicity × Sexual Orientation
White × Heterosexual
Black × Sexual minority 2.03 (0.70-5.90) .19 0.98 (0.42-2.28) .97 1.61 (0.45-5.72) .46 Latinx × Sexual minority 0.59 (0.21-1.64) .31 0.91 (0.30-2.72) .86 1.34 (0.38-4.76) .65 Other race × Sexual minority 1.08 (0.17-6.84) .93 0.60 (0.21-1.66) .32 0.047 (0.0047-0.48) .01 Race/Ethnicity × Maternal Education
White × College and higher
Black × Less than college 0.93 (0.32-2.74) .89 1.49 (0.66-3.36) .33 1.32 (0.39-4.47) .65 Latinx × Less than college 0.32 (0.11-0.94) .04 0.28 (0.13-0.59) .00 0.24 (0.07-0.75) .02 Other race × Less than college 0.22 (0.07-0.69) .01 1.02 (0.24-4.35) .98 0.20 (0.016-2.54) .21 Race/Ethnicity × Public Assistance
White × None
Black × At least one 1.20 (0.44-3.27) .72 1.49 (0.60-3.68) .39 1.03 (0.28-3.70) .97 Latinx × At least one 2.15 (0.70-6.61) .18 2.11 (0.86-5.20) .10 1.75 (0.53-5.76) .35 Other race × At least one 1.36 (0.28-6.65) .71 2.05 (0.40-10.5) .39 3.12 (0.23-42.2) .39 Race/Ethnicity × MHT (W1)
White × MHT (W1)
Black × MHT (W1) 2.22 (0.87-5.69) .10 3.41 (1.50-7.80) .004 0.70 (0.25-1.92) .48
Latinx × MHT (W1) 0.78 (0.28-2.23) .65 0.51 (0.18-1.46) .21 0.70 (0.20-2.41) .57
Other race × MHT (W1) 2.83 (0.67-12.00) .16 0.28 (0.04-2.11) .22 1.30 (0.25-6.76) .75 Depression 7.02 (5.34-9.21) < .001 3.43 (2.72-4.34) < .001 5.83 (4.11-8.28) < .001
Note. OR = odds ratio; 95% CI = 95% confidence interval. W1=Wave I. MHT = mental health treatment.
a. Reference group = Trajectory 1.
