Depression and anxiety symptoms among Kenyan adolescents: psychometric validation, prevalence, network analysis, and psychosocial determinants Authors: Tom L. Osborn, Rosine Baseke, Faith Kamau, Maureen Ngesa, Rachael Kilonzo, Priscillah Agunda, Kane Vundi, Ray Kongong’a, Christine M. Wasanga, Victoria Mutiso, David Ndetei Published: 2026 ABSTRACT BackgroundAdolescent mental health in Sub-Saharan Africa remains critically understudied despite high burden estimates. We assessed the psychometric properties of Western-derived instruments, prevalence of depression and anxiety symptoms, associations with psychosocial and sociodemographic factors, and symptom network structures in a large multi-year sample of Kenyan adolescents.MethodsWe administered depression (PHQ-8), anxiety (GAD-7), social support, perceived control, and wellbeing measures to 7865 Kenyan secondary school students aged 12–20 across three studies (2021–2023).ResultsBoth PHQ-8 (α = 0.70) and GAD-7 (α = 0.78) demonstrated adequate reliability, and confirmatory factor analyses supported unidimensional structures with acceptable fit (PHQ-8: RMSEA = 0.052, CFI = 0.945; GAD-7: RMSEA = 0.056, CFI = 0.970). Overall, 30.3% (95% CI 29.3–31.3%) and 25.1% (95% CI 24.2–26.1%) of participants met the ≥ 10 cut-off for clinically elevated depression and anxiety symptoms, respectively, with descriptive variation across study years (2021: 42.1% [39.6, 44.6%], 37.5% [35.1, 40.0%]; 2022: 25.3% [23.8, 26.8%], 21.7% [20.3, 23.2%]; 2023: 29.7% [28.1, 31.3%], 22.5% [21.1–24.0%]); direct year-on-year comparison is limited by differences in wave composition. Mixed-effects modelling identified social support (B = − 0.08) and perceived control (B = − 0.29 for depression, B = − 0.24 for anxiety) as robust protective factors. Final-year students showed elevated symptoms (B = 0.93, B = 0.77), as did students in girls’ schools relative to mixed schools (B = 1.35, B = 1.27). Perceived academic ability showed a strong inverse gradient with both outcomes (excellent ability: B = − 2.51 for depression, B = − 1.76 for anxiety). Network analysis identified worry-related symptoms as most central, with hopelessness and nervousness serving as the primary bridges between depression and anxiety domains.DiscussionWestern-derived measures demonstrated adequate psychometric properties with Kenyan adolescents. Depression and anxiety were highly prevalent and varied by sex, school form, and school type. Social support and perceived control were the strongest protective factors; academic pressure and girls’ school attendance were associated with elevated symptoms. Network analyses identify worry, hopelessness, and negative self-evaluation as possible candidate intervention targets warranting empirical evaluation. These findings underscore the urgent need for contextually grounded, school-based mental health provision in Kenya and comparable Sub-Saharan African settings.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13034-026-01117-1. Introduction Adolescent depression and anxiety are leading causes of disability worldwide [1–3], ranking second among contributors to years lived with disability in young people [4]. Their consequences compound across the life course: half of all lifetime mental disorders emerge by age 14 [5–7], onset during adolescence predicts persistence in adulthood [5], and early, untreated episodes carry long-term costs for functioning, socioeconomic prospects, and long-term health [8, 9]. The burden falls hardest where resources are most limited. Adolescents in low-resource settings are disproportionately affected [10, 11] yet have the fewest pathways to care [12]. Sub-Saharan Africa (SSA)—the world’s youngest region [13]—exemplifies this gap: it carries a high youth mental health burden with limited treatment resources [10, 14] while accounting for less than 3% of peer-reviewed adolescent mental health research [15, 16]. The limited evidence that exists is constrained by measurement problem. Most studies have relied on Western-derived instruments to estimate prevalence [10, 17], describe disorder experience [18, 19], identify risk and protective factors [20, 21], and inform intervention design [22, 23]. Yet the validity of these tools in SSA, and other non-Western contexts, cannot be assumed. Developed on culturally homogenous samples, they may fail to capture how distress is experienced and expressed in different cultural contexts [24–28]. In Kenya, for instance, distress commonly manifests through idioms linking rumination to social and economic stressors – a presentation that cuts across what Western nosology separates into discrete depressive and anxiety categories [29–31]. A study in Western Kenya found that while the Western-derived Patient Health Questionnaire (PHQ; [32]) captured core depressive features, a locally co-developed instrument identified dimensions entirely absent from the Western tool: somatic, heart-centered idioms (e.g., “my heart is not happy”), supernatural causal attributions, and a multidimensional structure spanning cognitive, affective, and physical domains [33]. Without contextual validation, these instruments risk producing inaccurate prevalence estimates and misguided intervention strategies [25, 34]. Depression prevalence estimates in Kenyan adolescents range from 26.4 to 45.9%, and anxiety estimates from 12.9 to 49.3% [10, 17, 20, 21, 34]—a range too wide to reflect genuine variation in burden. It more likely reflects differences in instrument validation, sampling methods, and contextual appropriateness [17, 20, 34]. Recent studies suggest that tools like the PHQ and GAD-7 meet basic psychometric standards in this context [17, 21], but the evidence remains limited and further validation work is needed [10, 20]. Even well-validated instruments, however, provide only a partial picture. For instance, summing item scores in an instrument obscures substantial symptom heterogeneity [35, 36] and may misrepresent the true experience of depression and anxiety when applied across diverse contexts [37]. That is why researchers are increasingly supplementing the traditional sum-score approach with additional analytical approaches like the network analysis [38–41]. Rather than treating disorders as latent entities indexed by sum-scores, network analysis treats them as systems of mutually reinforcing symptoms: nodes represent individual symptoms, and edges represent partial correlations between symptom pairs controlling for all other relationship. [38, 40]. This approach identifies which symptoms are most central to maintaining disorder and which serve as bridges between conditions—information with potential direct relevance to intervention targeting [38–40, 42]. Network methods have rarely been applied to Kenyan or SSA samples [18]; one of the few existing studies identified “self-blame” and “depressed mood” as the most central depression nodes, and “too much worry” and “uncontrollable worry” as most central in anxiety [18]. Understanding symptom structure must also be contextualized by the factors that place adolescents at risk or protect them. Prior research in Kenyan samples has identified several consistent patterns: older age and final-year status are associated with higher symptoms, likely reflecting the intense academic pressure of Kenya’s high-stakes end-of-secondary examinations [17, 21]; gender differences are most pronounced for anxiety, with female adolescents reporting higher levels [17, 21, 43]; and psychosocial factors—particularly social support, gratitude, and perceived control—show robust negative associations with both depression and anxiety [17, 21, 43]. Present study We draw on one of the largest adolescent mental health datasets assembled in SSA—7865 participants from 27 schools across four counties, collected over three years (2021–2023)—to address these gaps. Specifically, we: (1) assess the psychometric properties of two commonly used measures for adolescent depression and anxiety (the Patient Health Questionnaire – PHQ-8 [44] and the General Anxiety Disorder – Screener – GAD-7 [18]), including internal consistency, convergent validity, and factor structure; (2) estimate clinically-elevated depression and anxiety prevalence across 2021–2023; (3) apply network analysis to map symptom-level relationships, identify central symptoms, and compare network structures across subgroups; and (4) examine associations between symptom levels and key psychosocial protective factors and demographic characteristics. This study advances mental health research in Sub-Saharan Africa beyond the smaller, narrower studies that have defined the field to date. Its findings have direct implications for culturally grounded intervention design in Kenya and provide methodological frameworks applicable across comparable SSA contexts. Methods Study setting Data were drawn from three cross-sectional studies conducted in 27 secondary schools across four Kenyan counties: Study 1 (2021) in four schools in Nairobi and Kiambu; Study 2 (2022) in ten schools across Nairobi, Kiambu, and Makueni; and Study 3 (2023) in thirteen schools across Nairobi, Kiambu, Makueni, and Machakos. Schools were recruited in partnership with the Kenyan Ministry of Education and local county education offices, prioritizing geographic and school-type diversity across urban and peri-urban settings. The number of participating schools increased across waves as COVID-19 restrictions eased: in 2021, most schools remained reluctant to permit external research activities; by 2022 and 2023, willingness to participate had increased substantially, enabling broader recruitment. Kenya’s secondary schools are stratified by student performance on the national Class 8 examination [44]. National schools admit top performers from across the country via government quotas; extra-county schools serve several neighboring counties; county and sub-county schools serve local populations [44]. Approximately 70% of secondary schools are public single-sex boarding institutions, while day schools are typically mixed-gender. Kenya’s 100% transition policy ensures that nearly 80% of primary school completers proceed to secondary education [44]. Secondary education is organized into four consecutive year-groups known as Forms (Forms 1–4), broadly equivalent to school years 9–12 in the United States or Years 10–13 in the United Kingdom, with Form 4 being the final year, culminating in the Kenya Certificate of Secondary Education (KCSE) national examinations. Participants Eligible participants were adolescents aged 12–20 enrolled in the participating schools. The total sample comprised 7865 adolescents (Mage = 15.96, SD = 1.35; 41.3% female): Study 1 (n = 1498; Mage = 16.33, SD = 1.09; 41.4% female), Study 2 (n = 3226; Mage = 16.09, SD = 1.39; 30.0% female), and Study 3 (n = 3141; Mage = 15.68, SD = 1.34; 48.8% female). Full sample characteristics are reported in Table 1. Participants were classified as majority or minority tribal status based on national census data [45] and prior research; tribes comprising more than 5% of the national population were classified as majority [17]. Table 1Sample characteristicsCharacteristic n %Age Mean (SD)15.96 (1.35)Gender Female312241.3 Male443758.7School county Nairobi316040.2 Kiambu232229.5 Makueni137317.5 Machakos101012.8School location Urban/Peri-urban (Nairobi, Kiambu)548269.7 Rural (Makueni, Machakos)238330.3School type National570.7 Extra-county248131.5 County282735.9 Sub-county250031.8School demographics Mixed303038.5 Boys321640.9 Girls161920.6School year (form) Form 1208636.4 Form 2202935.4 Form 3137023.9 Form 458910.3Final school year Yes5899.7 No548590.3Tribal status Majority tribe387049.2 Minority tribe399550.8Parents living with Both parents305269.4 Single parent134430.6 No parents1323.0Parental bereavement None387686.5 Father deceased4079.1 Mother deceased1373.1 Both deceased861.9Father’s education University129229.6 Secondary school94821.7 Primary school111725.6 Not aware109225.0Mother’s education University110225.1 Secondary school113525.9 Primary school92221.0 Not aware133130.3For some variables, there are missing/NA values for participants who did not provide response for specific variables Table 1Sample characteristicsCharacteristic n %Age Mean (SD)15.96 (1.35)Gender Female312241.3 Male443758.7School county Nairobi316040.2 Kiambu232229.5 Makueni137317.5 Machakos101012.8School location Urban/Peri-urban (Nairobi, Kiambu)548269.7 Rural (Makueni, Machakos)238330.3School type National570.7 Extra-county248131.5 County282735.9 Sub-county250031.8School demographics Mixed303038.5 Boys321640.9 Girls161920.6School year (form) Form 1208636.4 Form 2202935.4 Form 3137023.9 Form 458910.3Final school year Yes5899.7 No548590.3Tribal status Majority tribe387049.2 Minority tribe399550.8Parents living with Both parents305269.4 Single parent134430.6 No parents1323.0Parental bereavement None387686.5 Father deceased4079.1 Mother deceased1373.1 Both deceased861.9Father’s education University129229.6 Secondary school94821.7 Primary school111725.6 Not aware109225.0Mother’s education University110225.1 Secondary school113525.9 Primary school92221.0 Not aware133130.3For some variables, there are missing/NA values for participants who did not provide response for specific variables Sample characteristics For some variables, there are missing/NA values for participants who did not provide response for specific variables Procedures All three studies received ethical approval from the Kenyatta University Ethics Review Committee (KUERC) and a research permit from the National Commission for Science, Technology, and Innovation (NACOSTI). Study teams met with interested students in school halls to explain study procedures and answer questions. Parental consent for underage participants was obtained through school administration per KUERC protocols. All students provided written informed consent (age ≥ 18) or assent (age < 18) prior to completing the questionnaire battery. Detailed procedures for each data collection wave—including the number of schools, participating counties, sampling approach, and participant numbers by year—are provided in Supplementary Materials. This study was conducted and reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for cross-sectional studies [46]. Measures Depression and anxiety symptoms Depression symptoms were measured using the Patient Health Questionnaire-8 (PHQ-8) [47], an 8-item self-report scale derived from the PHQ-9 with the suicidal ideation item excluded [32]. Items are rated 0–3; total scores of 5, 10, 15, and 20 correspond to mild, moderate, moderately severe, and severe depression, respectively [32]. The PHQ-8 has demonstrated adequate psychometric properties with Kenyan adolescents [21, 48]; Cronbach’s α in the present study was 0.70. Anxiety symptoms were assessed using the Generalized Anxiety Disorder 7-item scale (GAD-7) [49]. Items are rated 0–3 (total range 0–21); cut-offs of 5, 10, and 15 correspond to mild, moderate, and severe anxiety [49]. The GAD-7 has been validated with Kenyan adolescent samples [21]; Cronbach’s α was 0.78. For prevalence estimation, clinically elevated symptoms were defined as scores ≥ 10 on either scale [47, 49]—the threshold for moderate severity established in North American normative samples [32, 47, 49] and applied in prior Kenyan studies [17, 21]. Psychosocial measures Wellbeing was assessed using the Short Warwick-Edinburgh Mental Well-Being Scale (SWEMWBS) [50], a 7-item measure capturing emotional and functional aspects of positive psychological functioning. The SWEMWBS has demonstrated reliable psychometric performance with Kenyan adolescents specifically [23, 51]. Cronbach’s α was 0.73. Perceived social support was measured using the Multidimensional Scale of Perceived Social Support (MSPSS) [52]; which has demonstrated strong psychometric properties across multiple studies with Kenyan adolescents [17, 21]. Cronbach’s α was 0.79. Perceived control was measured using the Perceived Control Scale for Children (PCS) [53], which assesses beliefs about the ability to achieve desired outcomes through personal effort. Cronbach’s α was 0.68. Sociodemographic and contextual variables Participants self-reported age, gender, tribal affiliation, home setting (urban/rural/village), living arrangements (both parents, single parent, no parents), parental bereavement (none, father deceased, mother deceased, both deceased), current school year (Form 1–4), final school year status, sports participation, and self-rated academic ability. School-level variables—school type (national, extra-county, county, sub-county), demographic composition (boys-only, girls-only, mixed-gender), and county—were drawn from administrative records. The sociodemographic questionnaire has been used in prior research with Kenyan adolescents [22, 23]. Data analyses All analyses were conducted on R Studio (Version: 2024.12.1 + 563). Data and accompanying R code files can be found in the Open Science Framework Repository (10.17605/OSF.IO/Q5AHW). Data nesting and missingness Participants were nested within schools (two-level structure). Missing data were handled using multiple imputation via the Fully Conditional Specification (FCS) method, implemented with the multivariate imputation by chained equations (MICE) algorithm in R [54]. The imputation model included all study variables—PHQ-8 and GAD-7 items, psychosocial scale totals (PCS, MSPSS, SWEMWBS), and all sociodemographic predictors—with school identifier as an auxiliary variable to preserve clustering structure. Five imputed datasets were generated; results from mixed-effects models were pooled using Rubin’s rules [55]. Factor and network analyses, which do not support pooling across imputed datasets, used a single complete-case dataset with full-information maximum likelihood estimation where applicable. Missingness was low for the primary psychometric scales (PHQ-8: 4.5%; GAD-7: 4.6%; PCS: 4.8%; MSPSS: 4.7%), but substantially higher for several sociodemographic variables, ranging from 3.9% (gender) to 44.0% (tribal status). Full missingness rates by variable are reported in Supplementary Materials. Psychometric properties Internal consistency and convergent validity Internal consistency was assessed using Cronbach’s α, with values ≥ 0.70 considered acceptable [56]. Convergent validity was assessed by examining Pearson’s correlations between the PHQ-8 and GAD-7 and the SWEMWBS. The SWEMWBS was selected as the convergent validity comparator because it captures emotional and functional aspects of positive psychological functioning [57] theoretically expected to be inversely related to depression and anxiety [58, 59], has been used as a convergent validity comparator for the PHQ and GAD-7 in prior validation studies [21, 60, 61], and has demonstrated reliable psychometric performance with Kenyan adolescents [21]. Construct validity of the PHQ-8 and the GAD-7 Exploratory factor analysis: The total sample was randomly divided into two equal subsamples (n = 3933 each). Using the first subsample, we conducted separate EFAs for the PHQ-8 and GAD-7 using maximum likelihood estimation with promax oblique rotation, which allows factors to correlate. Factor retention was guided by eigenvalue > 1.0 and scree plot inspection [62]. Data suitability was evaluated using the Kaiser–Meyer–Olkin (KMO) measure of sampling adequacy; values between 0.8 and 1.0 indicate adequate sampling [63, 64]. Bartlett’s Test of Sphericity determine whether the correlation matrix was suitable for data reduction; significant results (p < 0.05) suggest appropriateness for factor analysis [62, 65]. Confirmatory factor analysis: Using the second subsample, we tested the factor structure identified by EFA using confirmatory factor analysis (CFA) implemented in the lavaan package with maximum likelihood estimation [66]. Latent factors were standardized to allow free estimation of all factor loadings. Model fit was evaluated using the root mean square error of approximation (RMSEA; ≤0.05 good, 0.05–0.08 acceptable, > 0.08 poor), the comparative fit index (CFI; ≥0.90 acceptable, ≥ 0.95 excellent), and the Tucker-Lewis Index (TLI; ≥0.90 acceptable, ≥ 0.95 excellent) [67, 68]. Following single-group CFA, we tested measurement invariance across sex and age groups by comparing constrained and unconstrained models using chi-square difference tests. We also tested for invariance across majority and minority tribal groups. See Supplementary Materials. Factor analyses were conducted at the individual item level without school-level clustering, which is standard practice for item-level psychometric work. Joint factor structure of depression and anxiety measures: To examine whether PHQ-8 and GAD-7 items reflect a single dimension or distinct constructs, we conducted an EFA on all 15 items simultaneously (EFA subsample) and tested the resulting structure with CFA (CFA subsample), comparing one-factor and two-factor models on fit indices. Prevalence rates of depression and anxiety symptoms Prevalence of clinically-elevated symptoms was estimated as the proportion of participants scoring ≥ 10 on the PHQ-8 (depression) or GAD-7 (anxiety) [47, 49], calculated overall and stratified by study year and county. Network analyses Network analyses had three aims: (1) estimate the PHQ-8/GAD-7 symptom network in the full sample; (2) identify central symptoms; and (3) compare networks between elevated-symptom and low-symptom subgroups. Network estimation We estimated a Gaussian graphical model (GGM) using EBICglasso regularization [69–71], which penalizes spurious edges to reduce false positives and address multiple-comparison concerns [72]. Network estimation utilized the estimateNetwork function from the bootnet package with default EBICglasso parameters and a tuning parameter of 0.5 [73]. To evaluate network precision, we conducted nonparametric bootstrap procedures with 1000 iterations to generate confidence intervals around edge weights. Wider confidence intervals indicate greater uncertainty in edge weight estimation, while narrower intervals suggest more reliable connections between symptoms [74]. Centrality and stability: Four centrality indices were calculated [75]: strength (sum of absolute edge weights) [76], expected influence (accounts for both positive and negative connections) [77], betweenness (frequency on shortest paths between node pairs) [78], and closeness (proximity to all other nodes) [78]. We report all four indices but focus on strength centrality in the discussion, as it most directly identifies symptoms suitable as intervention targets. Network stability was assessed via case-dropping bootstrap: the correlation stability (CS) coefficient indicates the maximum proportion of cases removable while maintaining centrality rank-order correlations ≥ 0.70 with the full sample [74]. CS values ≥ 0.25 indicate adequate stability; ≥ 0.50 indicate good stability [74]. Bridge symptoms: Bridge centrality indices were calculated to identify symptoms that most strongly connect the depression and anxiety domains [79]. Bridge strength quantifies the sum of a node’s edge weights to nodes in the other domain [80]. Network comparison: Networks were compared between an elevated-symptom group (PHQ-8 ≥ 10 or GAD-7 ≥ 10) and a low-symptom group (PHQ-8 < 10 and GAD-7 < 10) using the Network Comparison Test (NCT) [81], a permutation-based procedure that tests for differences in global network structure, overall connectivity (global strength), and individual edge weights while controlling for multiple testing [21, 47, 49]. Multidomain network: To examine how symptom domains interact with protective factors, we constructed 36-node network spanning five domains: depression symptoms (PHQ-8; 8 items), anxiety symptoms (GAD-7; 7 items), perceived control (PCS; 6 items), social support (MSPSS; 8 items), and wellbeing (SWEMWBS; 7 items). The same EBICglasso approach and stability procedures were applied. Bridge centrality was calculated across the five domains [79], with community boundaries defined by the theoretical domain structure [82]. This multidomain approach allows examination of how psychopathology symptoms interact with protective factors and positive mental health indicators within a single network structure [38]. Association of depression and anxiety symptoms with psychological factors and socio-demographic factors Two parallel linear mixed-effects models were estimated using the lmer function in the lme4 package in R [83]—one with PHQ-8 total score as the outcome, one with GAD-7 total score. Both models included an identical fixed-effects structure to enable direct comparison across outcomes. Fixed effects included: psychosocial variables (perceived social support, perceived control); demographic variables (age, gender, tribal status); school and family context (school type, school demographic composition, home setting, parental bereavement, living arrangements, perceived academic ability, final school year status); and sports participation. Random intercepts were specified for schools to account for unmeasured school-level factors and clustering within schools. As a sensitivity analysis, we additionally estimated models including study year (2021, 2022, 2023) as a fixed-effect covariate, with 2022 as the reference category, to examine whether temporal variation in symptoms persisted after covariate adjustment; results of these models are reported in Supplementary Materials, alongside models additionally including school county as a fixed-effect covariate to examine regional differences. Results Prevalence of depression and anxiety symptoms Across the full sample, participants reported moderate levels of depression and anxiety symptoms. For depression (PHQ-8), the overall mean was 7.43 (SD = 4.89); anxiety (GAD-7) symptoms averaged 6.58 (SD = 4.89). When examining clinically elevated symptoms (scores ≥ 10 indicating moderate or above severity), 30.3% (95% CI 29.3–31.3%) of participants met criteria for elevated depression and 25.1% (95% CI 24.2–26.1%) for elevated anxiety symptoms. Figure 1; Table 2 present prevalence estimates with breakdown by gender, school form, school type, county, and year. Fig. 1Panel 1 shows mean symptom scores by study year, compared with mean symptoms scores from 2018 [17] and 2019 [21]. Error bars represent 95% confidence intervals. Panel 2 shows percentage of youth with clinically-elevated symptoms by study year with values from 2018 [17] and 2019 [21] Panel 1 shows mean symptom scores by study year, compared with mean symptoms scores from 2018 [17] and 2019 [21]. Error bars represent 95% confidence intervals. Panel 2 shows percentage of youth with clinically-elevated symptoms by study year with values from 2018 [17] and 2019 [21] Table 2Prevalence of clinically-elevated depression and anxiety symptoms by demographic subgroupGroupSubgroupN*PHQ-8GAD-7M (SD)*% Elevated** [95% CI]M (SD)% Elevated [95% CI]Full sample78657.43 (4.89)30.3 [29.3, 31.3]6.58 (4.89)25.1 [24.2, 26.1]Study year202114989.12 (5.39)41.9 [39.4, 44.4]8.25 (5.54)37.5 [35.1, 40.0]202232266.63 (4.73)25.3 [23.8, 26.8]5.94 (4.78)21.7 [20.3, 23.2]202331417.45 (4.57)29.7 [28.1, 31.3]6.43 (4.46)22.5 [21.1, 24.0]CountyKiambu23228.02 (4.98)35.1 [33.2, 37.1]7.19 (5.03)30.3 [28.5, 32.2]Nairobi31607.99 (4.94)33.6 [32.0, 35.3]7.09 (4.97)27.9 [26.4, 29.5]Machakos10106.97 (4.37)26.7 [24.1, 29.5]5.93 (4.33)18.6 [16.3, 21.1]Makueni13735.47 (4.37)16.9 [15.0, 19.0]4.81 (4.32)14.3 [12.5, 16.3]GenderFemale31227.77 (5.03)33.3 [31.7, 35.0]6.98 (5.01)28.3 [26.7, 29.9]Male44377.20 (4.79)28.2 [26.9, 29.5]6.31 (4.80)23.1 [21.9, 24.4]School formForm 120866.25 (4.41)22.2 [20.5, 24.0]5.61 (4.48)18.2 [16.6, 19.9]Form 220297.24 (4.53)28.0 [26.1, 30.0]6.33 (4.47)22.7 [20.9, 24.6]Form 313707.24 (4.93)29.9 [27.5, 32.4]6.22 (4.80)23.4 [21.2, 25.7]Form 45898.66 (4.87)38.9 [35.0, 42.9]7.64 (5.01)31.7 [28.1, 35.6]School typeExtra-County24818.23 (5.07)35.9 [34.0, 37.8]7.24 (5.08)29.3 [27.5, 31.1]National577.52 (4.51)31.6 [21.0, 44.5]6.66 (4.29)28.1 [18.1, 40.9]County28277.15 (4.63)28.2 [26.6, 29.9]6.34 (4.73)23.5 [22.0, 25.1]Sub-County25006.95 (4.89)26.9 [25.2, 28.7]6.17 (4.81)22.4 [20.8, 24.1]School demographicsGirls16198.71 (5.13)40.0 [37.6, 42.4]7.78 (5.15)34.4 [32.1, 36.7]Boys32167.47 (4.93)30.5 [28.9, 32.1]6.59 (4.98)25.4 [23.9, 26.9]Mixed30306.70 (4.54)24.7 [23.2, 26.3]5.91 (4.50)19.7 [18.3, 21.2]Tribal statusMajority38708.01 (4.88)34.1 [32.6, 35.6]7.04 (4.88)27.3 [25.9, 28.7]Minority39956.87 (4.82)26.5 [25.2, 27.9]6.12 (4.85)22.9 [21.6, 24.2]Final year (KCSE)No (Forms 1–3)54856.86 (4.61)26.3 [25.2, 27.5]6.03 (4.57)21.2 [20.1, 22.3]Yes (Form 4)5898.66 (4.87)38.9 [35.0, 42.9]7.64 (5.01)31.7 [28.1, 35.6]“% Elevated” refers to the proportion of participants scoring ≥ 10 on the PHQ-8 or GAD-7, indicating clinically-elevated (moderate or above) symptom severity. Year, County, Gender, School Form, School Type, School Demographics, and Tribal Status subgroup N values do not always sum to the total sample due to missing data on those variables. Tribal Status N is based on non-missing responses (total missing = 44.0%). Form 4 and Final Year (KCSE) subgroups are identical populations. KCSE = Kenya Certificate of Secondary Education Table 2Prevalence of clinically-elevated depression and anxiety symptoms by demographic subgroupGroupSubgroupN*PHQ-8GAD-7M (SD)*% Elevated** [95% CI]M (SD)% Elevated [95% CI]Full sample78657.43 (4.89)30.3 [29.3, 31.3]6.58 (4.89)25.1 [24.2, 26.1]Study year202114989.12 (5.39)41.9 [39.4, 44.4]8.25 (5.54)37.5 [35.1, 40.0]202232266.63 (4.73)25.3 [23.8, 26.8]5.94 (4.78)21.7 [20.3, 23.2]202331417.45 (4.57)29.7 [28.1, 31.3]6.43 (4.46)22.5 [21.1, 24.0]CountyKiambu23228.02 (4.98)35.1 [33.2, 37.1]7.19 (5.03)30.3 [28.5, 32.2]Nairobi31607.99 (4.94)33.6 [32.0, 35.3]7.09 (4.97)27.9 [26.4, 29.5]Machakos10106.97 (4.37)26.7 [24.1, 29.5]5.93 (4.33)18.6 [16.3, 21.1]Makueni13735.47 (4.37)16.9 [15.0, 19.0]4.81 (4.32)14.3 [12.5, 16.3]GenderFemale31227.77 (5.03)33.3 [31.7, 35.0]6.98 (5.01)28.3 [26.7, 29.9]Male44377.20 (4.79)28.2 [26.9, 29.5]6.31 (4.80)23.1 [21.9, 24.4]School formForm 120866.25 (4.41)22.2 [20.5, 24.0]5.61 (4.48)18.2 [16.6, 19.9]Form 220297.24 (4.53)28.0 [26.1, 30.0]6.33 (4.47)22.7 [20.9, 24.6]Form 313707.24 (4.93)29.9 [27.5, 32.4]6.22 (4.80)23.4 [21.2, 25.7]Form 45898.66 (4.87)38.9 [35.0, 42.9]7.64 (5.01)31.7 [28.1, 35.6]School typeExtra-County24818.23 (5.07)35.9 [34.0, 37.8]7.24 (5.08)29.3 [27.5, 31.1]National577.52 (4.51)31.6 [21.0, 44.5]6.66 (4.29)28.1 [18.1, 40.9]County28277.15 (4.63)28.2 [26.6, 29.9]6.34 (4.73)23.5 [22.0, 25.1]Sub-County25006.95 (4.89)26.9 [25.2, 28.7]6.17 (4.81)22.4 [20.8, 24.1]School demographicsGirls16198.71 (5.13)40.0 [37.6, 42.4]7.78 (5.15)34.4 [32.1, 36.7]Boys32167.47 (4.93)30.5 [28.9, 32.1]6.59 (4.98)25.4 [23.9, 26.9]Mixed30306.70 (4.54)24.7 [23.2, 26.3]5.91 (4.50)19.7 [18.3, 21.2]Tribal statusMajority38708.01 (4.88)34.1 [32.6, 35.6]7.04 (4.88)27.3 [25.9, 28.7]Minority39956.87 (4.82)26.5 [25.2, 27.9]6.12 (4.85)22.9 [21.6, 24.2]Final year (KCSE)No (Forms 1–3)54856.86 (4.61)26.3 [25.2, 27.5]6.03 (4.57)21.2 [20.1, 22.3]Yes (Form 4)5898.66 (4.87)38.9 [35.0, 42.9]7.64 (5.01)31.7 [28.1, 35.6]“% Elevated” refers to the proportion of participants scoring ≥ 10 on the PHQ-8 or GAD-7, indicating clinically-elevated (moderate or above) symptom severity. Year, County, Gender, School Form, School Type, School Demographics, and Tribal Status subgroup N values do not always sum to the total sample due to missing data on those variables. Tribal Status N is based on non-missing responses (total missing = 44.0%). Form 4 and Final Year (KCSE) subgroups are identical populations. KCSE = Kenya Certificate of Secondary Education Prevalence of clinically-elevated depression and anxiety symptoms by demographic subgroup “% Elevated” refers to the proportion of participants scoring ≥ 10 on the PHQ-8 or GAD-7, indicating clinically-elevated (moderate or above) symptom severity. Year, County, Gender, School Form, School Type, School Demographics, and Tribal Status subgroup N values do not always sum to the total sample due to missing data on those variables. Tribal Status N is based on non-missing responses (total missing = 44.0%). Form 4 and Final Year (KCSE) subgroups are identical populations. KCSE = Kenya Certificate of Secondary Education PHQ-8 and GAD-7 psychometrics Internal consistency and convergent validity Cronbach’s alpha was 0.70 for the PHQ-8 and 0.78 for the GAD-7; both exceeded the 0.70 cutoff for acceptable internal consistency. For the psychosocial measures, Cronbach’s alpha was 0.79 for the MSPSS, 0.73 for the SWEMWBS, and 0.68 for the PCS. Convergent validity was assessed using Pearson’s correlations. The PHQ-8 was negatively correlated with the SWEMWBS (r = − 0.26), PCS (r = − 0.27), and MSPSS (r = − 0.24), and positively correlated with the GAD-7 (r = 0.69). Similarly, the GAD-7 was negatively correlated with the SWEMWBS (r = − 0.25), PCS (r = − 0.22), and MSPSS (r = − 0.24). Factor structure of the PHQ-8 and GAD-7 Using subsample 1 (N = 3933), EFA yielded a 1-factor structure for the PHQ-8 explaining 23.77% of the variance (KMO = 0.82; Bartlett’s test of sphericity = 3899.35, p < 0.001; Eigen-value = 2.64). For the GAD-7, EFA also yielded a 1-factor structure explaining 34.72% of the variance (KMO = 0.86; Bartlett’s test = 5976.43, p < 0.001; eigenvalue = 3.06). See Table 3. Table 3Summary of factor analyses for PHQ-8 and GAD-7MeasureModelEFAχ²pRMSEA (90% CI)CFITLIFactor loadings rangeInter-factor correlationEigenvalueVariance Explained (%)PHQ-81-Factor2.6423.77223.36< 0.0010.05 (0.04–0.06)0.9460.9250.32–0.60–2-Factor2.64, 0.9229.8932.110.0020.02 (0.01–0.03)0.9890.9890.26–0.950.69GAD-71-Factor3.0634.72191.63< 0.0010.06 (0.05–0.06)0.9690.9540.50–0.71–2-Factor3.06, 0.8639.0817.10< 0.0010.02 (0.01–0.03)0.9960.9960.36–0.850.76EFA conducted on subsample 1 (N = 3933), CFA conducted on subsample 2 (N = 3933). RMSEA = Root Mean Square Error of Approximation; CFI = Comparative Fit Index; TLI = Tucker-Lewis Index. High inter-factor correlations in 2-factor solutions supported selection of more parsimonious 1-factor models Table 3Summary of factor analyses for PHQ-8 and GAD-7MeasureModelEFAχ²pRMSEA (90% CI)CFITLIFactor loadings rangeInter-factor correlationEigenvalueVariance Explained (%)PHQ-81-Factor2.6423.77223.36< 0.0010.05 (0.04–0.06)0.9460.9250.32–0.60–2-Factor2.64, 0.9229.8932.110.0020.02 (0.01–0.03)0.9890.9890.26–0.950.69GAD-71-Factor3.0634.72191.63< 0.0010.06 (0.05–0.06)0.9690.9540.50–0.71–2-Factor3.06, 0.8639.0817.10< 0.0010.02 (0.01–0.03)0.9960.9960.36–0.850.76EFA conducted on subsample 1 (N = 3933), CFA conducted on subsample 2 (N = 3933). RMSEA = Root Mean Square Error of Approximation; CFI = Comparative Fit Index; TLI = Tucker-Lewis Index. High inter-factor correlations in 2-factor solutions supported selection of more parsimonious 1-factor models Summary of factor analyses for PHQ-8 and GAD-7 EFA conducted on subsample 1 (N = 3933), CFA conducted on subsample 2 (N = 3933). RMSEA = Root Mean Square Error of Approximation; CFI = Comparative Fit Index; TLI = Tucker-Lewis Index. High inter-factor correlations in 2-factor solutions supported selection of more parsimonious 1-factor models Using subsample 2 (N = 3933), CFA confirmed the 1-factor structure for both measures. The model demonstrated good fit for the PHQ-8 (χ² = 223.36; p < 0.001; RMSEA = 0.051; CFI = 0.946; TLI = 0.925) and for the GAD-7 (χ² = 191.63; p < 0.001; RMSEA = 0.057; CFI = 0.969; TLI = 0.954). Factor loadings for the PHQ-8 ranged from 0.321 to 0.595; GAD-7 loadings ranged from 0.459 to 0.708, indicating adequate to strong item-factor relationships. See Table 3. Although the 1-factor models demonstrated acceptable fit, 2-factor solutions were also examined during EFA. For the PHQ-8, the 2-factor model showed improved fit indices (TLI = 0.989, RMSEA = 0.020) but factors were highly correlated (r = 0.69), suggesting the additional factor did not represent a meaningful separate dimension. For the GAD-7, the 2-factor model demonstrated excellent fit (TLI = 0.996, RMSEA = 0.017) but factors were again highly correlated (r = 0.76), indicating that a single-factor solution was most parsimonious and theoretically appropriate for both measures. See Table 3. Joint factor structure of depression and anxiety measures We investigated whether the PHQ-8 and GAD-7 items reflect a single underlying factor or distinct constructs by conducting EFA on all 15 items simultaneously (EFA subsample, N = 3933), testing one-, two-, and three-factor solutions. Exploratory factor analysis: The one-factor solution explained 28% of the variance with adequate fit (TLI = 0.947, RMSEA = 0.040). All items loaded positively on this general factor, with GAD worry-related items showing the highest loadings (GAD_2 = 0.69, GAD_3 = 0.66). The two-factor solution provided improved fit (TLI = 0.970, RMSEA = 0.031) and explained 30% of the variance; however, the two factors were highly correlated (r = 0.80), with some cross-loadings observed. The three-factor solution showed the best statistical fit (TLI = 0.978, RMSEA = 0.026) but exhibited extremely high inter-factor correlations (r = 0.72–0.77), suggesting over-extraction. Confirmatory factor analysis: Using the CFA subsample (N = 3933), the two-factor model demonstrated good fit (χ² = 727.871, df = 89, p < 0.001; CFI = 0.949; TLI = 0.939; RMSEA = 0.043 [90% CI 0.040–0.046]; SRMR = 0.030). All factor loadings were significant and substantial, ranging from 0.329 to 0.612 for PHQ items and 0.469 to 0.686 for GAD items. Most critically, the correlation between the PHQ and GAD latent factors was extremely high (r = 0.921, p < 0.001), indicating that the two constructs share approximately 85% of their variance. Network structure of depression and anxiety symptoms The symptom network contained 15 nodes and 103 non-zero edges out of 105 possible connections (98.1% density), with a mean edge weight of 0.059. Centrality analysis and network stability The highest strength centrality values were observed for GAD_2 (feeling unable to stop or control worrying; strength = 1.085), GAD_3 (worrying too much about different things; strength = 1.028), and GAD_1 (feeling nervous, anxious, or on edge; strength = 1.004). Among depression symptoms, PHQ_6 (feeling bad about yourself or that you are a failure; strength = 0.915) and PHQ_2 (feeling down, depressed, or hopeless; strength = 0.879) showed the highest centrality. The least central symptom was PHQ_1 (little interest or pleasure in activities; strength = 0.441). The correlation stability coefficient for strength centrality was 0.75. See Figs. 2 and 3; Table 4. Fig. 2Regularized partial correlation networks of depression (PHQ-8) and anxiety (GAD-7) symptoms. Nodes represent individual symptoms and edges indicate regularized partial correlations between symptoms, with thicker edges indicating stronger associations Regularized partial correlation networks of depression (PHQ-8) and anxiety (GAD-7) symptoms. Nodes represent individual symptoms and edges indicate regularized partial correlations between symptoms, with thicker edges indicating stronger associations Fig. 3Regularized partial correlation networks of depression (PHQ-8) and anxiety (GAD-7) symptoms in networks of non-elevated symptoms (low-symptom network) and elevated-symptom networks (high-symptom network) groups, showing significantly higher global connectivity in the elevated-symptom network (global strength = 3.58 vs. 3.26, p = 0.010) Regularized partial correlation networks of depression (PHQ-8) and anxiety (GAD-7) symptoms in networks of non-elevated symptoms (low-symptom network) and elevated-symptom networks (high-symptom network) groups, showing significantly higher global connectivity in the elevated-symptom network (global strength = 3.58 vs. 3.26, p = 0.010) Table 4Centrality metrics for depression and anxiety symptomsSymptomContentStrengthClosenessBetweennessExpected InfluenceAnxiety symptoms GAD_2Unable to stop/control worrying1.0850.004461.085 GAD_3Worrying too much about different things1.0280.004221.028 GAD_1Feeling nervous, anxious, or on edge1.0040.0049131.004 GAD_7Feeling afraid, as if something awful might happen0.8860.004340.886 GAD_5Being so restless that it’s hard to sit still0.8210.004020.821 GAD_6Becoming easily annoyed or irritable0.7930.003830.793 GAD_4Trouble relaxing0.7040.003700.704Depression symptoms PHQ_6Feeling bad about yourself or that you are a failure0.9150.004350.915 PHQ_2Feeling down, depressed, or hopeless0.8790.004530.879 PHQ_4Feeling tired or having little energy0.8100.004240.810 PHQ_7Trouble concentrating on things0.7650.004180.765 PHQ_3Trouble falling or staying asleep0.7140.003850.714 PHQ_8Moving or speaking slowly, or being fidgety0.6980.003830.698 PHQ_5Poor appetite or overeating0.6180.003530.618 PHQ_1Little interest or pleasure in doing things0.4410.002900.441Strength = sum of absolute edge weights; Closeness = inverse of average shortest path distances; Betweenness = frequency of lying on shortest paths between other nodes; Expected Influence = sum of edge weights preserving sign. Network stability: CS coefficient = 0.75 (excellent stability) Table 4Centrality metrics for depression and anxiety symptomsSymptomContentStrengthClosenessBetweennessExpected InfluenceAnxiety symptoms GAD_2Unable to stop/control worrying1.0850.004461.085 GAD_3Worrying too much about different things1.0280.004221.028 GAD_1Feeling nervous, anxious, or on edge1.0040.0049131.004 GAD_7Feeling afraid, as if something awful might happen0.8860.004340.886 GAD_5Being so restless that it’s hard to sit still0.8210.004020.821 GAD_6Becoming easily annoyed or irritable0.7930.003830.793 GAD_4Trouble relaxing0.7040.003700.704Depression symptoms PHQ_6Feeling bad about yourself or that you are a failure0.9150.004350.915 PHQ_2Feeling down, depressed, or hopeless0.8790.004530.879 PHQ_4Feeling tired or having little energy0.8100.004240.810 PHQ_7Trouble concentrating on things0.7650.004180.765 PHQ_3Trouble falling or staying asleep0.7140.003850.714 PHQ_8Moving or speaking slowly, or being fidgety0.6980.003830.698 PHQ_5Poor appetite or overeating0.6180.003530.618 PHQ_1Little interest or pleasure in doing things0.4410.002900.441Strength = sum of absolute edge weights; Closeness = inverse of average shortest path distances; Betweenness = frequency of lying on shortest paths between other nodes; Expected Influence = sum of edge weights preserving sign. Network stability: CS coefficient = 0.75 (excellent stability) Centrality metrics for depression and anxiety symptoms Strength = sum of absolute edge weights; Closeness = inverse of average shortest path distances; Betweenness = frequency of lying on shortest paths between other nodes; Expected Influence = sum of edge weights preserving sign. Network stability: CS coefficient = 0.75 (excellent stability) Bridge symptom analysis PHQ_2 (feeling down, depressed, or hopeless) and GAD_1 (feeling nervous, anxious, or on edge) emerged as the strongest bridge symptoms (bridge strength = 0.51 and 0.50, respectively), followed by PHQ_6 (feeling bad about yourself; bridge strength = 0.46) and GAD_6 (becoming easily annoyed or irritable; bridge strength = 0.43). See Figs. 2 and 3; Table 5. Table 5Network comparison test results: high vs. low symptom groupsGlobal network propertiesLow-symptom groupHigh-symptom groupTest STATISTICp-valueGlobal strength3.263.58S = 0.318 0.010 Network structure invarianceM = 0.116 0.010 Significant edge differencesTest statistic (E)p-valuePHQ_3—PHQ_40.101 0.010 PHQ_2—PHQ_60.102 0.010 PHQ_6—PHQ_70.072 0.010 PHQ_7—GAD_20.075 0.020 GAD_2—GAD_30.116 0.010 PHQ_3—GAD_10.053 0.030 Significant centrality differencesLow-symptom strengthHigh-symptom strengthTest statistic (C)p-valueDepression symptoms PHQ_3 (Sleep problems)0.3690.259− 0.1090.010 PHQ_4 (Fatigue)0.5000.334− 0.1660.010 PHQ_5 (Appetite)0.2680.146− 0.1220.010 PHQ_7 (Concentration)0.3780.223− 0.1550.020 PHQ_8 (Psychomotor)0.3910.221− 0.1700.010Anxiety symptoms GAD_1 (Nervous/anxious)0.5660.728+ 0.1620.010 GAD_4 (Trouble relaxing)0.3390.474+ 0.1350.010 GAD_5 (Restlessness)0.4560.588+ 0.1320.010 GAD_6 (Irritability)0.5470.280− 0.2670.010 GAD_6 (Irritability)0.5470.280− 0.2670.010Network Comparison Test with 1000 permutations. Low-symptom group: PHQ-8 < 10 AND GAD-7 < 10; High-symptom group: PHQ-8 ≥ 10 OR GAD-7 ≥ 10. Only significant differences (p < 0.05) are shown. Positive test statistics indicate higher values in high-symptom group; negative values indicate higher values in low-symptom group Table 5Network comparison test results: high vs. low symptom groupsGlobal network propertiesLow-symptom groupHigh-symptom groupTest STATISTICp-valueGlobal strength3.263.58S = 0.318 0.010 Network structure invarianceM = 0.116 0.010 Significant edge differencesTest statistic (E)p-valuePHQ_3—PHQ_40.101 0.010 PHQ_2—PHQ_60.102 0.010 PHQ_6—PHQ_70.072 0.010 PHQ_7—GAD_20.075 0.020 GAD_2—GAD_30.116 0.010 PHQ_3—GAD_10.053 0.030 Significant centrality differencesLow-symptom strengthHigh-symptom strengthTest statistic (C)p-valueDepression symptoms PHQ_3 (Sleep problems)0.3690.259− 0.1090.010 PHQ_4 (Fatigue)0.5000.334− 0.1660.010 PHQ_5 (Appetite)0.2680.146− 0.1220.010 PHQ_7 (Concentration)0.3780.223− 0.1550.020 PHQ_8 (Psychomotor)0.3910.221− 0.1700.010Anxiety symptoms GAD_1 (Nervous/anxious)0.5660.728+ 0.1620.010 GAD_4 (Trouble relaxing)0.3390.474+ 0.1350.010 GAD_5 (Restlessness)0.4560.588+ 0.1320.010 GAD_6 (Irritability)0.5470.280− 0.2670.010 GAD_6 (Irritability)0.5470.280− 0.2670.010Network Comparison Test with 1000 permutations. Low-symptom group: PHQ-8 < 10 AND GAD-7 < 10; High-symptom group: PHQ-8 ≥ 10 OR GAD-7 ≥ 10. Only significant differences (p < 0.05) are shown. Positive test statistics indicate higher values in high-symptom group; negative values indicate higher values in low-symptom group Network comparison test results: high vs. low symptom groups Network Comparison Test with 1000 permutations. Low-symptom group: PHQ-8 < 10 AND GAD-7 < 10; High-symptom group: PHQ-8 ≥ 10 OR GAD-7 ≥ 10. Only significant differences (p < 0.05) are shown. Positive test statistics indicate higher values in high-symptom group; negative values indicate higher values in low-symptom group Network comparison between symptom severity groups Network Comparison Test results revealed significant differences between elevated-symptom and low-symptom groups. The global test for network invariance was significant (M = 0.116, p = 0.010). The elevated-symptom group demonstrated significantly higher overall connectivity than the low-symptom group (global strength = 3.58 vs. 3.26; S = 0.318, p = 0.010). Most depression symptoms (PHQ_3, PHQ_4, PHQ_5, PHQ_7, PHQ_8) had significantly lower centrality in the elevated-symptom group (p = 0.010 for all), while anxiety symptoms GAD_1, GAD_4, GAD_5, and GAD_6 showed higher centrality in the elevated-symptom group (p = 0.010 for all). See Table 5. Bridge connectivity patterns also differed between groups. In the low-symptom group, PHQ_2 (bridge strength = 0.43), GAD_2 (bridge strength = 0.41), and GAD_6 (bridge strength = 0.40) were the strongest bridges. In the elevated-symptom group, PHQ_2 (bridge strength = 0.29), GAD_1 (bridge strength = 0.26), and PHQ_6 (bridge strength = 0.23) were the primary bridges. See Figs. 2 and 3. Both networks demonstrated adequate to excellent stability (CS coefficients: low-symptom group = 0.67, elevated-symptom group = 0.75). Multidomain network analysis The multidomain network contained 36 variables across five domains and 322 non-zero edges out of 630 possible connections (51.1% density), with a mean edge weight of 0.039. Centrality in the multidomain network The highest strength centrality was observed for SWEMWBS_5 (feeling optimistic about the future; strength = 1.139), followed by GAD_2 (feeling unable to stop or control worrying; strength = 1.132), GAD_3 (worrying too much about different things; strength = 1.093), and GAD_1 (feeling nervous, anxious, or on edge; strength = 1.082). Among depression symptoms, PHQ_6 (feeling bad about yourself; strength = 1.088) and PHQ_2 (feeling down, depressed, or hopeless; strength = 1.017) maintained high centrality. The correlation stability coefficient for strength centrality was 0.75. See Fig. 4. Fig. 4Centrality differences between the full-sample (full), the elevated-symptom sub-sample (high), and non-elevated symptom sub-sample (low) Centrality differences between the full-sample (full), the elevated-symptom sub-sample (high), and non-elevated symptom sub-sample (low) Cross-domain bridge analysis Depression symptoms, particularly PHQ_2 (feeling hopeless; bridge strength = 0.699) and PHQ_6 (feeling bad about self; bridge strength = 0.695), emerged as the strongest cross-domain bridges, followed by GAD_1 (nervous/anxious; bridge strength = 0.590). Among protective factors, several wellbeing items showed substantial bridge connectivity, including SWEMWBS_5 (optimistic about future; bridge strength = 0.371) and SWEMWBS_6 (dealing with problems well; bridge strength = 0.355). See Fig. 5. Fig. 5Multidomain network incorporating 36 variables across five psychological domains: depression symptoms (PHQ-8; 8 items), anxiety symptoms (GAD-7; 7 items), perceived control (PCS; 6 items), social support (MSPSS; 8 items), and wellbeing (SWEMBS; 7 items) Multidomain network incorporating 36 variables across five psychological domains: depression symptoms (PHQ-8; 8 items), anxiety symptoms (GAD-7; 7 items), perceived control (PCS; 6 items), social support (MSPSS; 8 items), and wellbeing (SWEMBS; 7 items) Association of depression and anxiety symptoms with psychological and socio-demographic factors Table 6 presents the results of the linear mixed-effects models. Social support was significantly associated with lower depression (B = − 0.08, 95% CI [− 0.09, − 0.07], p < 0.001) and anxiety (B = − 0.08, 95% CI [− 0.09, − 0.07], p < 0.001) symptoms. Perceived control was similarly associated with lower depression (B = − 0.29, 95% CI [− 0.32, − 0.26], p < 0.001) and anxiety (B = − 0.24, 95% CI [− 0.27, − 0.20], p < 0.001) symptoms. Table 6Linear mixed-effects model results for depression and anxiety symptomsPredictorPHQ-8GAD-7B95% CI p B95% CI p Intercept13.50[11.68, 15.31] < 0.001 12.29[10.18, 14.41] < 0.001 Social support (MSPSS)−0.08[− 0.09, − 0.07] < 0.001 −0.08[− 0.09, − 0.07] < 0.001 Perceived control (PCS)−0.29[− 0.32, − 0.26] < 0.001 −0.24[− 0.27, − 0.20] < 0.001 Age0.15[0.06, 0.25] 0.002 0.13[0.02, 0.25] 0.023 Gender [male]−0.06[− 0.41, 0.29]0.752−0.37[− 0.73, − 0.02] 0.037 Sports [yes]−0.02[− 0.29, 0.25]0.868−0.09[− 0.41, 0.23]0.546Final school year [yes]0.93[0.46, 1.40] < 0.001 0.77[0.31, 1.23] 0.001 Tribal status [minority]−0.35[− 0.70, 0.00] 0.047 −0.39[− 0.73, − 0.06] 0.022 School [boys]0.48[− 0.46, 1.42]0.3190.58[− 0.26, 1.43]0.175School [girls]1.35[0.28, 2.43] 0.014 1.27[0.30, 2.23] 0.010 Home [small town]0.20[− 0.12, 0.52]0.208−0.13[− 0.41, 0.16]0.385Home [big town]0.21[− 0.11, 0.54]0.202−0.11[− 0.51, 0.28]0.564Home [city]0.24[− 0.11, 0.58]0.181−0.13[− 0.55, 0.30]0.547Parents deceased [father]−0.06[− 0.45, 0.34]0.774−0.29[− 0.67, 0.09]0.133Parents deceased [mother]0.27[− 0.55, 1.09]0.4990.30[− 0.48, 1.09]0.433Parents deceased [both]−0.64[− 1.30, 0.02]0.057−0.74[− 1.51, 0.03]0.060Perceived academic ability [satisfactory]−1.46[− 1.84, − 1.09] < 0.001 −1.29[− 1.69, − 0.88] < 0.001 Perceived academic ability [good]−2.18[− 2.52, − 1.84] < 0.001 −1.41[− 1.75, − 1.07] < 0.001 Academic ability [very good]−2.35[− 3.02, − 1.67] < 0.001 −1.60[− 2.09, − 1.11] < 0.001 Perceived academic ability [excellent]−2.51[− 2.93, − 2.08] < 0.001 −1.76[− 2.20, − 1.32] < 0.001 Living with parents [no parents]0.33[− 0.34, 1.00]0.3261.04[0.34, 1.75] 0.005 Living with parents [single parent]0.30[0.04, 0.56] 0.026 0.64[0.36, 0.92] < 0.001 Reference categories: Year = 2022; Gender = Female; Tribal Status = Majority; Sports = No; Final Year = No; Home = Rural/Village; Parents Deceased = None; Academic Ability = Not Satisfactory; Living Arrangements = Both Parents. Models included random intercepts for schools. Significant effects (p < 0.05) are shown in bold. No correction for multiple comparisons was applied Table 6Linear mixed-effects model results for depression and anxiety symptomsPredictorPHQ-8GAD-7B95% CI p B95% CI p Intercept13.50[11.68, 15.31] < 0.001 12.29[10.18, 14.41] < 0.001 Social support (MSPSS)−0.08[− 0.09, − 0.07] < 0.001 −0.08[− 0.09, − 0.07] < 0.001 Perceived control (PCS)−0.29[− 0.32, − 0.26] < 0.001 −0.24[− 0.27, − 0.20] < 0.001 Age0.15[0.06, 0.25] 0.002 0.13[0.02, 0.25] 0.023 Gender [male]−0.06[− 0.41, 0.29]0.752−0.37[− 0.73, − 0.02] 0.037 Sports [yes]−0.02[− 0.29, 0.25]0.868−0.09[− 0.41, 0.23]0.546Final school year [yes]0.93[0.46, 1.40] < 0.001 0.77[0.31, 1.23] 0.001 Tribal status [minority]−0.35[− 0.70, 0.00] 0.047 −0.39[− 0.73, − 0.06] 0.022 School [boys]0.48[− 0.46, 1.42]0.3190.58[− 0.26, 1.43]0.175School [girls]1.35[0.28, 2.43] 0.014 1.27[0.30, 2.23] 0.010 Home [small town]0.20[− 0.12, 0.52]0.208−0.13[− 0.41, 0.16]0.385Home [big town]0.21[− 0.11, 0.54]0.202−0.11[− 0.51, 0.28]0.564Home [city]0.24[− 0.11, 0.58]0.181−0.13[− 0.55, 0.30]0.547Parents deceased [father]−0.06[− 0.45, 0.34]0.774−0.29[− 0.67, 0.09]0.133Parents deceased [mother]0.27[− 0.55, 1.09]0.4990.30[− 0.48, 1.09]0.433Parents deceased [both]−0.64[− 1.30, 0.02]0.057−0.74[− 1.51, 0.03]0.060Perceived academic ability [satisfactory]−1.46[− 1.84, − 1.09] < 0.001 −1.29[− 1.69, − 0.88] < 0.001 Perceived academic ability [good]−2.18[− 2.52, − 1.84] < 0.001 −1.41[− 1.75, − 1.07] < 0.001 Academic ability [very good]−2.35[− 3.02, − 1.67] < 0.001 −1.60[− 2.09, − 1.11] < 0.001 Perceived academic ability [excellent]−2.51[− 2.93, − 2.08] < 0.001 −1.76[− 2.20, − 1.32] < 0.001 Living with parents [no parents]0.33[− 0.34, 1.00]0.3261.04[0.34, 1.75] 0.005 Living with parents [single parent]0.30[0.04, 0.56] 0.026 0.64[0.36, 0.92] < 0.001 Reference categories: Year = 2022; Gender = Female; Tribal Status = Majority; Sports = No; Final Year = No; Home = Rural/Village; Parents Deceased = None; Academic Ability = Not Satisfactory; Living Arrangements = Both Parents. Models included random intercepts for schools. Significant effects (p < 0.05) are shown in bold. No correction for multiple comparisons was applied Linear mixed-effects model results for depression and anxiety symptoms Reference categories: Year = 2022; Gender = Female; Tribal Status = Majority; Sports = No; Final Year = No; Home = Rural/Village; Parents Deceased = None; Academic Ability = Not Satisfactory; Living Arrangements = Both Parents. Models included random intercepts for schools. Significant effects (p < 0.05) are shown in bold. No correction for multiple comparisons was applied Age was positively associated with both depression (B = 0.15, 95% CI [0.06, 0.25], p = 0.002) and anxiety (B = 0.13, 95% CI [0.02, 0.25], p = 0.023) symptoms. Male gender was associated with lower anxiety symptoms (B = − 0.37, 95% CI [− 0.73, − 0.02], p = 0.037) but not depression symptoms (B = − 0.06, 95% CI [− 0.41, 0.29], p = 0.752). Minority tribal status was associated with lower depression (B = − 0.35, 95% CI [− 0.70, 0.00], p = 0.047) and anxiety (B = − 0.39, 95% CI [− 0.73, − 0.06], p = 0.022) symptoms. Sports participation showed no significant associations with either outcome. Being in the final school year was associated with elevated depression (B = 0.93, 95% CI [0.46, 1.40], p < 0.001) and anxiety (B = 0.77, 95% CI [0.31, 1.23], p = 0.001) symptoms. Girls’ school attendance was significantly associated with higher depression (B = 1.35, 95% CI [0.28, 2.43], p = 0.014) and anxiety (B = 1.27, 95% CI [0.30, 2.23], p = 0.010) symptoms relative to mixed schools; boys’ schools showed no significant association with either outcome. Home environment showed no significant associations with either outcome across all categories. Parental bereavement was not significantly associated with depression or anxiety symptoms for loss of father only, mother only, or both parents, though the coefficient for both parents deceased approached significance for depression (B = − 0.64, 95% CI [− 1.30, 0.02], p = 0.057) and anxiety (B = − 0.74, 95% CI [− 1.51, 0.03], p = 0.060). Perceived academic ability showed a consistent gradient relative to the poor performance reference category. For depression: satisfactory (B = − 1.46, 95% CI [− 1.84, − 1.09], p < 0.001), good (B = − 2.18, 95% CI [− 2.52, − 1.84], p < 0.001), very good (B = − 2.35, 95% CI [− 3.02, − 1.67], p < 0.001), and excellent (B = − 2.51, 95% CI [− 2.93, − 2.08], p < 0.001) ability levels were all associated with significantly lower symptoms. The same pattern held for anxiety: satisfactory (B = − 1.29, 95% CI [− 1.69, − 0.88], p < 0.001), good (B = − 1.41, 95% CI [− 1.75, − 1.07], p < 0.001), very good (B = − 1.60, 95% CI [− 2.09, − 1.11], p < 0.001), and excellent (B = − 1.76, 95% CI [− 2.20, − 1.32], p < 0.001). Living with a single parent was associated with elevated depression (B = 0.30, 95% CI [0.04, 0.56], p = 0.026) and anxiety (B = 0.64, 95% CI [0.36, 0.92], p < 0.001) symptoms. Not living with any parent was associated with elevated anxiety symptoms (B = 1.04, 95% CI [0.34, 1.75], p = 0.005) but not depression symptoms (B = 0.33, 95% CI [− 0.34, 1.00], p = 0.326). Discussion Across a sample of 7865 Kenyan secondary school students—one of the largest adolescent mental health datasets assembled in Sub-Saharan Africa—depression and anxiety were highly prevalent, symptom networks were tightly organized around worry and negative self-evaluation, and social support and perceived control emerged as the most consistent protective factors. Western-derived instruments demonstrated adequate psychometric properties, lending confidence to both the prevalence estimates and the analytic findings. Together, these results extend the evidence base on adolescent mental health in SSA and carry direct implications for measurement practice, intervention design, and policy. Psychometric validation of western-derived measures Both the PHQ-8 and GAD-7 demonstrated adequate to good psychometric properties among Kenyan adolescents, with unidimensional factor structures and acceptable model fit. These findings replicate prior work [21, 48] and support the continued use of these brief, freely-available instruments in low-resource settings where validated assessment tools are scarce [22, 51]. Two-factor solutions showed improved fit while maintaining very high inter-factor correlations, suggesting that locally adapted versions warrant exploration to better capture nuances in symptom presentation among Kenyan youth. The joint factor analysis of PHQ-8 and GAD-7 items bears on the conceptual distinction between depression and anxiety in this population. Although a two-factor model fit the data better than a one-factor solution, the correlation between latent factors was exceptionally high (r = 0.92), indicating that the two constructs share approximately 84% of their variance and are nearly indistinguishable in this sample. This is consistent with hierarchical models of psychopathology positing a general distress factor underlying specific internalizing disorders [84–86] and with the high comorbidity rates observed across study years and in comparable research [17, 21]. Measurement invariance testing indicated that both instruments perform comparably across key subgroups. Invariance across sex and age groups was evaluated by comparing constrained and unconstrained models using chi-square difference tests; results supported adequate invariance across both groupings and are reported in full in the Supplementary Materials. For tribal group comparisons (majority versus minority), the GAD-7 demonstrated full scalar invariance, indicating that anxiety scores are directly comparable across tribal groups and that observed mean-level differences reflect genuine variation. The PHQ-8 supported metric confirming equivalent factor loadings across groups, but scalar invariance was not fully established, suggesting some item intercepts differ slightly between tribal groups. Prevalence estimates based on observed sum scores remain interpretable across groups, but comparisons of latent depression means across tribal groups should be made with caution. Correlations between the SWEMWBS and the PHQ-8 and GAD-7 were in the expected negative direction (r = − 0.26 and r = − 0.25, respectively) but weaker than typical clinical benchmarks (r ≈ − 0.60 to − 0.65; [57, 60]). Three factors likely account for this. We used a universal community-based sample which possibly attenuates correlations relative to clinical populations. The dual continuum model underlying SWEMWBS development anticipates only partial overlap between positive wellbeing and symptom severity [57]. Finally, SWEMWBS items reflecting individualistic wellbeing constructs may not fully capture how Kenyan adolescents experience positive mental health [18]. Future validation work should incorporate locally co-developed wellbeing measures alongside Western comparators. Prevalence and temporal patterns Depression and anxiety were highly prevalent: 30.3% of participants met the clinical threshold for depression and 25.1% for anxiety. These rates exceed those reported in many Western contexts [1, 87] but are consistent with recent Kenyan studies [21] and with SSA data from Ghana and Nigeria [88, 89], reinforcing that elevated internalizing symptom burden is a regional pattern rather than a Kenya-specific phenomenon. Symptom levels varied across study years, with 2021 showing the highest burden—41.9% for depression and 37.7% for anxiety—followed by a decrease in 2022 and a partial increase in 2023. This pattern aligns with documented rises in adolescent mental health difficulties during the COVID-19 pandemic and partial recovery as restrictions eased [43, 90–92]. Kenya’s school disruption was severe: schools closed nationwide in March 2020 and remained closed for approximately ten months before phased reopening in January 2021 [93, 94], meaning the 2021 cohort was assessed at the immediate return to in-person schooling after prolonged disruption. Year-on-year comparisons nonetheless require caution. The three waves differed not only temporally but in school composition—the 2021 sample comprised exclusively boarding schools in Nairobi and Kiambu, while 2022 and 2023 introduced day schools and more rural counties—meaning that observed changes in prevalence may partly reflect compositional differences rather than true temporal trends. A sensitivity analysis including study year as a fixed-effect covariate is reported in Supplementary Materials; the cross-sectional design and absence of a pre-pandemic baseline from the same schools limit causal inference about pandemic-related effects. Network structure of depression and anxiety symptoms Near-complete network density indicates that depression and anxiety symptoms form a tightly interconnected system in this population, contributing to ongoing debate about the relationship between the two conditions. Globally, depression and anxiety are strongly correlated and highly comorbid [95–97], and the field has yet to resolve whether they represent distinct disorders [98] or manifestations of a shared underlying syndrome [99]. The high symptom interconnectivity observed here adds SSA-contextual evidence to that discussion. Anxiety symptoms occupied the most central network positions, with uncontrollable worry, excessive worry, and nervousness showing the highest centrality indices. This is consistent with theoretical models positioning worry as the core maintaining mechanism in anxiety disorders [100, 101] and parallels findings from adult clinical samples, suggesting that worry-focused interventions—including cognitive-behavioral and acceptance-based approaches—may be particularly effective for Kenyan adolescents [102]. Among depression symptoms, negative self-evaluation and depressed mood showed the highest centrality. The prominence of negative self-evaluation aligns with Beck’s cognitive theory of depression and with network studies identifying self-criticism as a central node [103, 104], suggesting that cognitive interventions targeting negative self-referential thinking merit particular attention in this context. Bridge analysis identified hopelessness and nervousness as the strongest cross-domain bridges, indicating that these symptoms serve as primary pathways through which depression and anxiety mutually reinforce one another. Targeting bridge symptoms therapeutically could disrupt the cascading processes that generate comorbid presentations [79, 80]. Comparing elevated- and low-symptom networks yielded findings that partially support and partially complicate existing network theory. The elevated-symptom group showed higher overall connectivity—consistent with predictions that symptoms become more tightly coupled as severity increases [38]—but depression symptoms became less central while anxiety symptoms became more central, a reorganization that warrants further investigation. Reduced bridge connectivity in the elevated-symptom group was unexpected and may reflect a tendency for severe symptom states to organize more rigidly around specific clusters rather than maintaining flexible cross-domain connections. An alternative interpretation is that symptom severity moderates the optimal intervention target, with anxiety-focused approaches becoming relatively more relevant at higher overall symptom levels. In the multidomain network, optimism about the future emerged as the most central variable across all five domains. This suggests that strengthening positive future orientation may produce benefits that propagate through the broader psychological system, supporting the growing evidence base for positive psychology approaches in African contexts [22, 23, 51, 105]. Hopelessness and negative self-evaluation showed strong bridge connectivity between psychopathology and broader functioning domains, reinforcing the potential value of transdiagnostic approaches targeting shared maintaining mechanisms [106]. The relatively peripheral positioning of social support and perceived control in the network suggests these factors may function more effectively as adjunctive than as standalone intervention components. Association of symptoms with psychological and socio-demographic factors Social support and perceived control were robustly and negatively associated with both depression and anxiety [107, 108], consistent with theoretical expectations and prior Kenyan research. Perceived control showed particularly strong associations with both outcomes, suggesting that interventions targeting adolescents’ sense of agency over their circumstances could yield meaningful clinical benefits. Older age was associated with higher symptom levels for both outcomes, consistent with SSA research identifying older adolescents as a particularly at-risk group [17, 109, 110] and with meta-analytic evidence that mean age moderates depression prevalence in SSA youth samples [10]. This association likely reflects the cumulative academic, social, and developmental pressures that intensify across secondary school years in Kenya’s high-stakes educational environment. Gender was differentially associated with anxiety but not depression, with male adolescents reporting lower anxiety. This aligns with global literature documenting higher anxiety rates in females [111, 112] and with SSA research showing that gender differences are more pronounced for anxiety than for depression [17, 21, 113]. The association between perceived academic ability and symptom levels was among the strongest in the models. A clear graded bi-directional association emerged: students who rated their academic performance more highly reported progressively lower symptoms. Students in their final school year showed elevated symptoms for both depression and anxiety, consistent with the intensifying pressure of Kenya’s Certificate of Secondary Education (KCSE) examinations [17, 21, 114]. These associations held after controlling for year of data collection, age, and all other covariates, indicating that academic pressure operates as a specific burden independent of broader temporal trends. Interventions targeting academic self-efficacy, study skills, and academic support may therefore carry mental health benefits that extend well beyond their educational value [23, 115]. Girls’ school attendance was associated with elevated depression and anxiety relative to mixed-gender schools, while boys’ school attendance showed no significant associations with either outcome. This effect appeared in primary models but not in sensitivity analyses adjusting for study year, warranting interpretive caution. Two theoretical mechanisms are plausible. Social comparison theory predicts that homogeneous peer environments—with a narrower range of comparison targets—may intensify evaluative processes around academic achievement, social status, and appearance [116–120]. The gender intensification hypothesis similarly proposes that single-sex settings may amplify certain gender-role pressures [116–120]. These dynamics may be compounded in boarding contexts, where peer environments are continuous rather than bounded by school hours. All schools in this study were public institutions operating under Kenya’s near-universal secondary enrolment policy, so the girls’ school finding reflects school environment rather than selection into education. Dedicated longitudinal investigation is needed to disentangle these mechanisms. Minority tribal status was associated with lower depression symptoms—a finding that runs counter to minority stress models, which typically link minority status to elevated internalizing symptoms in Western contexts [121]. In Kenya, strong cultural identity and community belonging may function as resilience resources that buffer against distress [122, 123], and minority tribal students in boarding schools may draw on these resources within the school environment. This finding requires cautious interpretation given 44% missingness in tribal status data and modest effect sizes. Measurement invariance analyses across tribal groups (see Supplementary Materials) demonstrated full scalar invariance for the GAD-7 but only metric (not full scalar) invariance for the PHQ-8, meaning that anxiety-score comparisons across tribal groups can be interpreted at face value while depression-score mean comparisons should be made with caution. Replication with complete tribal data and qualitative investigation of underlying mechanisms is warranted. Living arrangements showed differential associations across outcomes. Living with a single parent was associated with elevated symptoms for both depression and anxiety, while living with no parent was associated with elevated anxiety but not depression. These patterns suggest that family support structures are particularly salient for anxiety, possibly reflecting the role of care-related uncertainty in maintaining anxious symptomatology. County-level differences emerged in sensitivity analyses—with Kiambu and Nairobi showing markedly higher prevalence than Makueni and Machakos—but should be interpreted cautiously given that county and year are partly confounded in the sampling design. They nonetheless point to regional variation in symptom burden that warrants dedicated future investigation. Parental bereavement, sports participation, and home environment setting showed no significant associations with either outcome in the primary models. Limitations Several limitations qualify our conclusions and should guide interpretation of the findings reported above. First, the cross-sectional design precludes causal inference. Network models can suggest potential causal pathways, but longitudinal data are required to establish temporal precedence and track how symptom dynamics unfold over time; experience sampling or ecological momentary assessment methods would provide finer-grained insight into within-person processes. The same constraint applies to the COVID-era interpretation: observed differences across waves are descriptively consistent with pandemic-era disruption and recovery, but with no pre-pandemic baseline from the same schools we cannot adjudicate between pandemic effects, secular trends, and cohort composition. Second, the three data-collection waves differed in school composition. We addressed this analytically through sensitivity analyses including study year and school county as fixed-effect covariates (see Supplementary Materials), but the year-on-year prevalence comparisons are descriptive rather than estimates of pure temporal change. Additionally, participants were drawn from secondary schools and may not represent Kenyan adolescents who have dropped out of school or lack consistent educational access. This limits the generalizability of findings to the full population of Kenyan youth. Fourth, reliance on self-report measures introduces potential biases including social desirability and common method variance. Despite adequate psychometric properties, the Western origin of the PHQ-8 and GAD-7 means they may still fail to capture culturally specific idioms of distress. Convergent validity was assessed against a single comparator (SWEMWBS), which yielded weaker correlations than Western clinical benchmarks. Future validation work should employ a broader range of convergent indicators, including locally co-developed measures of both distress and positive mental health. Fifth, measurement invariance testing across tribal groups supported full scalar invariance for the GAD-7 but only metric (not full scalar) invariance for the PHQ-8. Mean-level depression comparisons across tribal groups should therefore be made with caution; the lower-symptom association observed for minority-tribe adolescents rests more securely for anxiety than for depression. Sixth, no correction for multiple comparisons was applied across regression models, as predictors were entered simultaneously within a single model per outcome rather than tested independently. Findings should be treated as requiring replication in independent samples. Finally, network models represent static snapshots of symptom relationships and cannot capture dynamic processes. Centrality estimates identify symptoms that co-vary with many others in this sample; whether targeting them therapeutically actually disrupts symptom maintenance is an empirical question requiring intervention trials, and our centrality and bridge findings should be read as hypothesis-generating rather than as direct treatment recommendations. Longitudinal network studies are needed to clarify the directionality of associations between symptoms and the psychosocial factors examined here. Conclusions Depression and anxiety are highly prevalent among Kenyan secondary school adolescents, with rates substantially exceeding Western norms and comparable to other SSA contexts. Western-derived instruments demonstrated adequate psychometric properties in this population, supporting their continued use while underscoring the need for locally grounded validation work. Network analyses identify worry, negative self-evaluation, and hopelessness as possible priority candidate symptom-level intervention targets, warranting empirical testing in trials of mechanism-targeted interventions, while optimism about the future emerges as the most systemically central protective variable across the full psychological network. Social support and perceived control emerged as the strongest psychosocial correlates of lower symptoms in this cross-sectional sample, and the academic pressure associated with Kenya’s examination system was a robust correlate of elevated symptoms and a strong candidate target for future intervention research. These findings have implications for mental health service design in Kenya and comparable SSA settings. The scale of unmet need documented here—nearly one in three adolescents meeting the threshold for clinically elevated depression—makes a compelling case for expanded school-based mental health provision. The symptom- and factor-level evidence generated by this study provides a principled basis for intervention targeting that goes beyond broad prevalence surveillance toward a mechanistic understanding of how depression and anxiety develop and persist in this population. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 1.