Differential Moderation by Aerobic and Muscle‐Strengthening Physical Activity of the Violence Victimization‐Suicidal Ideation Association in Korean Adolescents

ABSTRACT

Purpose

Violence victimization is a known risk factor for adolescent suicidal ideation, yet moderating factors remain poorly understood. This study examined whether aerobic and muscle‐strengthening physical activity (PA) differentially moderated the violence victimization–suicidal ideation association.

Methods

Four waves of the Korea Youth Risk Behavior Survey (KYRBS 2021–2024; n = 214,218) were analyzed using complex sample logistic regression. Aerobic PA (0–7 days/week) and muscle‐strengthening PA (0–5+ days/week) were modeled continuously, with linear and quadratic terms to test interactions with violence victimization, adjusting for psychosocial and sociodemographic covariates. Sex‐ and SES‐stratified analyses and formal three‐way interaction tests were conducted.

Results

Overall, 13.3% reported suicidal ideation, versus 38.9% among violence‐victimized adolescents (OR = 3.43). Aerobic PA showed a non‐monotonic interaction with violence: odds of suicidal ideation were elevated at 1–3 days/week (ORs 1.4–1.8) but attenuated at ≥ 4 days/week, a pattern nominally strongest among males. Muscle‐strengthening PA showed elevated odds at higher frequencies (≥ 3 days/week; ORs 1.4–1.9) rather than low frequencies, though this association was not statistically significant in continuous models.

Discussion

Both PA types showed frequency‐dependent, nonlinear associations inconsistent with single‐threshold framing. Findings are associational and hypothesis‐generating, requiring prospective confirmation.

1. Introduction

Suicide is the number one cause of death in adolescents in South Korea (hereafter Korea), which has had the highest suicide rate among OECD countries in most years since 2003 (Kwak and Ickovics 2019). National surveillance data from the Korea Youth Risk Behavior Survey (KYRBS) indicate that approximately one in seven Korean adolescents seriously considered suicide in the past year, with the prevalence rising from 10.7% in 2020 to 14.0% in 2022 (Park et al. 2023). Therefore, determining modifiable risk and protective factors for suicide is of paramount importance to public health.

Among well‐established risk factors for suicidal ideation, violence victimization, which this study defines per KYRBS as having received hospital treatment due to violence (physical assault, threats, bullying, etc.) inflicted by peers or adults, has some of the largest associated risks. Youth who experience violence victimization are at 2‐ to 4‐times increased risk for suicidal ideation compared to their peers who do not experience violence victimization (Castellví et al. 2017; Turner et al. 2012). While this risk has been well documented, less work has examined whether health behaviors mitigate this pathway. One health behavior with some hypothesized inverse association with suicidal ideation is physical activity (PA); prior meta‐analytic work has supported an inverse association between PA and suicidal ideation (Vancampfort et al. 2018), and exploratory work among US samples has shown that exercise may help to mitigate mental health sequelae of bullying victimization (Iyanda et al. 2022; Sibold et al. 2015). However, these studies have examined PA as one homogeneous construct, and in these samples, violence exposure was operationalized almost entirely as peer bullying rather than more severe forms of victimization. It remains unknown whether the association between violence victimization and suicidal ideation is modified by the frequency and domain (e.g., aerobic, muscle‐strengthening) of PA, particularly among those who experience more severe, hospital‐treated victimization.

This distinction is both theoretically and practically important. Aerobic PA is postulated to decrease stress reactivity and depression via monoaminergic and neuroplastic mechanisms (Schuch and Stubbs 2019). Both pathways are believed to play a role in the violence–suicidal ideation association. On the other hand, there is increasing support for the mental health benefits of muscle‐strengthening PA among adults and adolescents (Zhang et al. 2022). However, these benefits may be mediated via alternate mechanisms, such as self‐efficacy and social identity, that may not be as relevant to violence‐exposed youth. Indeed, there is some preliminary evidence that high‐intensity/resistance‐type training may relate to different (and even paradoxical) mental health outcomes among Korean adolescents (Lee et al. 2013).

Sex and socioeconomic status (SES) were selected as stratification variables for theoretical and empirical rationale in the Korean adolescent sample. Regarding sex, female adolescents have consistently reported higher rates of suicidal ideation than males in Korea and are generally more likely to respond to stress with internalizing symptoms such as depression and suicidal ideation (Donnelly et al. 2021; Jeon et al. 2017). Conversely, PA rates among Korean adolescents have historically been much lower in females than in males for both aerobic and muscle‐strengthening activities; aerobic PA surveillance data suggest that male adolescents are over three times more likely to meet PA guidelines than female adolescents (8.7% vs. 2.9%) (Seo et al. 2022). Because female adolescents are at greater risk for poor mental health outcomes following violence exposure and are less likely to engage in PA overall, this study theorized that sex would be an important consideration for whether PA has a moderating effect. Regarding adolescent SES, violence exposure has been found to be more common among low‐SES adolescents in Korea (Park et al. 2017), and low‐SES adolescents also have fewer resources that allow for PA participation, such as proximity to facilities, affordability of programs, and neighborhood safety (Lee et al. 2023). As both violence victimization (primary exposure) and PA (potential moderating factor) are patterned by SES in Korea, this study assumed it was important to consider whether SES moderated the protective association between PA and suicidal ideation. Covariates, including school level, urbanicity, and family structure, were included in adjusted models but were not chosen as stratification variables as they do not directly bear on the theorized pathway.

Crucially, however, these factors were not to be examined as separate dimensions. Rather, informed by intersectionality theory (Crenshaw 2013), or the idea that social identities, including sex and SES, combine on multiple axes to jointly influence people's experiences, this study attempted to shed light on whether the intersecting identities of sex and SES create unique configurations of risk and resilience resource availability for Korean adolescents. For example, one recent study using KYRBS data that applied the intersectional framework found that low‐SES girls were more likely to experience loneliness and depressive symptoms than their higher‐income or male counterparts, results that could not be explained by sex or SES separately (Lee and Lim 2025). Regarding the current study, this intersection might manifest in particularly negative ways for low‐SES girls. Not only are they at heightened risk for internalizing after violence exposure, but they also likely have the fewest opportunities for PA given increased exposure to all the above barriers. For boys, on the other hand, the associations between violence exposure, PA, and suicidal ideation may differ by SES level due to opportunities for participating in competitive sports or social norms supporting physical aggression. Thus, the stratified analyses presented aimed to truly understand nuanced differences at this intersection.

With a nationally representative four‐year pooled sample of Korean adolescents (KYRBS 2021–2024; n = 214,218), the current study had two goals: (1) examine whether aerobic PA and muscle‐strengthening PA moderated the association between violence victimization and suicidal ideation independently; and (2) examine whether interaction effects vary by sex and SES. This study expected aerobic and muscle‐strengthening PA frequency to be associated with the magnitude of the violence–suicidal ideation association, though this study considered it plausible that this relationship could be non‐linear given established U‐shaped or threshold effects of PA frequency on other psychological outcomes.

2. Methods

2.1. Data Source and Participants

This study analyzed data from the KYRBS, 2021–2024. KYRBS is an anonymous web‐based cross‐sectional annual survey sponsored by the Korea Disease Control and Prevention Agency (KDCA) and the Ministry of Education to generate representative national estimates of students' health behaviors in grades 7 through 12. From 2021 to 2024, the sample sizes were 54,848 (2021), 51,850 (2022), 52,880 (2023), and 54,653 (2024), totaling 214,231 participants. Of these records, 13 had missing values on the outcome variable, leaving 214,218 adolescents for this study's analytic sample. Missing data on the exposure (violence victimization), both PA moderators, and covariates other than socioeconomic status (13 missing, 0.01%) and academic performance (10 missing, < 0.01%) were negligible (0 missing). Listwise deletion was used across Models 1–3, which share the same covariate set; the resulting analytic n was 214,218 for all three models, unchanged from the sample size after outcome exclusions.

2.2. Outcome: Suicidal Ideation

Suicidal ideation was assessed by the question: “During the past 12 months, did you seriously consider attempting suicide?” Responses were dichotomized as yes (1) or no (0).

2.3. Primary Exposure: Violence Victimization

Violence victimization was measured using the following item: ‘During the past 12 months, have you been treated at a hospital due to violence (physical assault, threats, bullying, etc.) from friends, older students, or adults?’ Response options captured the number of episodes (0 to ‘6 or more’ times); responses were dichotomized as exposed (≥ 1 episode, coded 1) versus unexposed (0 episodes, coded 0).

2.4. Moderators: Physical Activity

Aerobic PA was measured as the number of days per week during which the student accumulated ≥ 60 min of heart rate‐raising PA (range: 0–7 days) and modeled continuously. Muscle‐strengthening PA was measured as the number of days per week the student engaged in muscle‐strengthening activities (e.g., push‐ups, sit‐ups, weightlifting). Since the 20/21 KYRBS wave top‐coded frequency of muscle‐strengthening PA at “5 or more days/week” while subsequent KYRBS waves (20/22–20/24) distinguished between 5 days, 6 days, and 7 days of muscle‐strengthening PA per week, this study harmonized PA frequency across waves by top‐coding frequency of muscle‐strengthening PA at 5 days/week for all survey years (range: 0–5+ days) to allow pooled analyses. Preliminary analyses found that dichotomizing PA frequency using conventional guideline thresholds (i.e., aerobic PA: ≥ 5 days/week; muscle‐strengthening PA: ≥ 3 days/week) according to South Korean and WHO adolescent PA guidelines (Bull et al. 2020) produced interaction effects that were highly sensitive to the choice of cut‐point and inconsistent with effects observed when PA frequency was modeled as continuous variables (see Supporting Information sensitivity analyses). For this reason, this study modeled aerobic PA frequency and muscle‐strengthening PA frequency as continuous variables in this study's primary analyses, including both linear and quadratic terms to allow for non‐monotonic interaction effects with violence victimization.

2.5. Covariates

Depression (sad or hopeless feeling ≥ 2 weeks in the past year yes vs. no), perceived stress (high vs. low/moderate), insufficient sleep (shorter vs. enough duration of sleep), loneliness (yes vs. no), sex (female vs. male), school level (high school vs. middle school), SES (low/lower‐middle vs. high/upper‐middle; middle vs. high/upper‐middle), academic performance (low or middle vs. high), and survey year (2022–2024 vs. 2021 as reference) were covariates in this study. Because depression, perceived stress, and loneliness may lie on the causal pathway between violence victimization and suicidal ideation rather than acting solely as confounders, sensitivity analyses re‐estimated Models 2 and 3 with these three variables excluded from the covariate set (Table S2). Because covariate missingness was minimal (≤ 0.01% for socioeconomic status and academic performance, and none for the remaining variables), listwise deletion excluded no additional cases beyond those excluded for outcome missingness (analytic n = 214,218; 99.994% of the pooled sample), and multiple imputation was not pursued given this negligible missingness.

2.6. Statistical Analysis

Because KYRBS draws an independent nationally representative sample each year, stratum and primary sampling unit (cluster) identifiers are not comparable across survey years even when they share the same numeric code. To pool the four independent annual samples in a valid manner, year‐specific stratum and cluster identifiers were constructed (by concatenating the survey year with the original stratum and cluster codes) prior to specifying the complex sample survey design, ensuring that identically numbered strata or clusters from different years were not erroneously treated as the same sampling unit. Three complex sample logistic regression models were fit within this pooled design. Model 1 included aerobic and muscle‐strengthening PA frequency variables as linear main effects only. Model 2 included linear and quadratic aerobic PA × violence interaction terms in addition to those included in Model 1, with muscle‐strengthening PA frequency entered as a covariate. Model 3 included linear and quadratic muscle‐strengthening PA × violence interaction terms, in addition to those in Model 1, and entered aerobic PA frequency as a covariate. Quadratic terms were included to test for non‐monotonic interaction terms that were suggested by exploratory analyses. Because depression, perceived stress, and loneliness may lie on the causal pathway between violence victimization and suicidal ideation rather than acting solely as confounders, sensitivity analyses re‐estimated Models 2 and 3 with these three variables excluded from the covariate set (Table S2). Odds ratios (OR) and 95% confidence intervals (CI) are reported. Sex (male, female) and SES (high/upper‐middle, middle, low/lower‐middle) stratified analyses were conducted by running each model within strata of the stratifying variable and removing the stratifying variable from the covariate set in each stratum. To formally test whether the PA × violence interaction differed across strata, rather than relying on the significance of separately estimated subgroup interactions, three‐way interaction terms (PA × violence × sex; PA × violence × SES) were additionally tested using joint Wald tests in the full pooled sample for each PA type. Because five strata × two PA types × two interaction terms (linear, quadratic) yielded 20 stratified significance tests, Holm and Benjamini‐Hochberg (FDR) corrections were applied jointly across all 20 p‐values to assess robustness to multiple comparisons (Table S3). A two‐tailed alpha level of 0.05 was used. All analyses were conducted in R (version 4.5.1), using the survey package for complex sample survey design and regression, haven for data import, dplyr/tidyr for data management, and officer/flextable for table generation.

3. Results

3.1. Sample Characteristics

Table 1 displays sample characteristics. Overall prevalence of suicidal ideation was 13.3% (95% CI: 13.1%–13.5%). Past‐year violence victimization was endorsed by 2.3% of the sample (n = 4982), 38.9% of whom reported suicidal ideation, over three times the rate among those without violence victimization (12.7%). Regarding aerobic PA, 32.6% of the sample reported no aerobic PA (None; n ≈69,862), 43.6% reported irregular aerobic PA (1–3 days/week; n ≈93,471), and 23.8% reported regular aerobic PA (4–7 days/week; n ≈50,914). For muscle‐strengthening PA, 52.5% reported none (n ≈112,547), 23.2% reported low frequency (1–2 days/week; n ≈49,786), and 24.2% reported high frequency (3–5+ days/week; n ≈51,886). Aerobic PA group distribution and muscle‐strengthening PA group distribution did not differ significantly between those with and without suicidal ideation (aerobic: 32.4%/43.7%/23.9% vs. 34.0%/43.2%/22.9%; muscle‐strengthening: 52.1%/23.4%/24.5% vs. 55.4%/22.2%/22.4%), though both differences reached statistical significance given the large sample size (both p < 0.001). Participants with suicidal ideation were significantly more likely to report depression (70.7% vs. 20.6%), high perceived stress (76.9% vs. 34.3%), loneliness (86.6% vs. 49.6%), and insufficient sleep (86.8% vs. 75.2%). All between‐group differences were statistically significant (all p < 0.001).

TABLE 1.

Sample characteristics by suicidal ideation status.

Characteristic	No	Yes	p *
	N = 2,257,224	N = 345,790
Violence victimization, %	1.6%	6.7%	< 0.001
Aerobic PA group, %			< 0.001
None	32.4%	34.0%
Irregular (1–3 d/wk)	43.7%	43.2%
Regular (4–7 d/wk)	23.9%	22.9%
Muscle‐strengthening PA group, %			< 0.001
None	52.1%	55.4%
Low (1–2 d/wk)	23.4%	22.2%
High (3–5+ d/wk)	24.5%	22.4%
Depression, %	20.6%	70.7%	< 0.001
High perceived stress, %	34.3%	76.9%	< 0.001
Loneliness, %	49.6%	86.6%	< 0.001
Insufficient sleep, %	75.2%	86.8%	< 0.001
Sex, %			< 0.001
Male	53.6%	38.4%
Female	46.4%	61.6%
School level, %			< 0.001
Middle school	50.3%	55.3%
High school	49.7%	44.7%
Socioeconomic status, %			< 0.001
High/upper‐middle	43.2%	38.1%
Middle	47.0%	43.2%
Low/lower‐middle	9.8%	18.7%
Academic performance, %			< 0.001
High	38.5%	34.8%
Middle	30.4%	26.1%
Low	31.0%	39.0%
Survey year, %			< 0.001
2021	25.4%	24.2%
2022	24.6%	26.7%
2023	24.7%	25.1%
2024	25.3%	24.0%

^*Pearson's χ ²: Rao & Scott adjustment. Weighted percentages within each suicidal ideation group. Aerobic PA group: None (0 d/wk), Irregular (1–3 d/wk), Regular (4–7 d/wk). Muscle‐strengthening PA group: None (0 d/wk), Low (1–2 d/wk), High (3–5+ d/wk., top‐coded). p‐values from Rao–Scott design‐based chi‐square tests. KYRBS 2021–2024, N = 214,218.

3.2. Main Effects and Interaction Models

Table 2 displays results from all three regression models. In Model 1, depression had the largest odds of suicidal ideation (OR = 4.83; 95% CI: 4.67–5.00), followed by violence victimization (OR = 3.43; 95% CI: 3.14–3.73), high perceived stress (OR = 3.06; 95% CI: 2.95–3.17), and loneliness (OR = 2.67; 95% CI: 2.55–2.79). When modeled as continuous variables without interaction terms, neither aerobic PA frequency (OR = 0.99 per 1‐day increase; 95% CI: 0.98–1.00) nor muscle‐strengthening PA frequency (OR = 1.00; 95% CI: 0.99–1.02) was independently associated with suicidal ideation.

TABLE 2.

Complex sample logistic regression models of suicidal ideation (KYRBS 2021–2024, N = 214,218).

Variable	Model 1: main effects		Model 2: aerobic PA × violence		Model 3: muscle PA × violence
	OR (95% CI)	p	OR (95% CI)	p	OR (95% CI)	p
Violence victimization	3.43 (3.14–3.73)	< 0.001	3.11 (2.66–3.63)	< 0.001	2.82 (2.47–3.22)	< 0.001
Aerobic PA frequency (per 1 d/wk)	0.99 (0.98–1.00)	0.070	0.96 (0.94–0.98)	< 0.001	0.99 (0.98–1.00)	0.091
Aerobic PA frequency2	—		1.01 (1.00–1.01)	0.001	—
Violence × Aerobic PA frequency	—		1.15 (1.02–1.30)	0.021	—
Violence × Aerobic PA frequency2	—		0.98 (0.96–1.00)	0.010	—
Muscle PA frequency (per 1 d/wk)	1.00 (0.99–1.02)	0.431	1.00 (0.99–1.01)	0.683	0.98 (0.95–1.02)	0.341
Muscle PA frequency2	—		—		1.00 (1.00–1.01)	0.407
Violence × Muscle PA frequency	—		—		0.98 (0.83–1.16)	0.837
Violence × Muscle PA frequency2	—		—		1.03 (0.99–1.06)	0.106
Depression	4.83 (4.67–5.00)	< 0.001	4.84 (4.67–5.00)	< 0.001	4.84 (4.67–5.01)	< 0.001
High perceived stress	3.06 (2.95–3.17)	< 0.001	3.06 (2.95–3.17)	< 0.001	3.07 (2.96–3.18)	< 0.001
Loneliness	2.67 (2.55–2.79)	< 0.001	2.67 (2.56–2.79)	< 0.001	2.67 (2.56–2.79)	< 0.001
Insufficient sleep	1.22 (1.17–1.28)	< 0.001	1.22 (1.17–1.28)	< 0.001	1.22 (1.17–1.28)	< 0.001
Female sex	1.34 (1.29–1.39)	< 0.001	1.34 (1.29–1.39)	< 0.001	1.33 (1.28–1.38)	< 0.001
High school	0.73 (0.71–0.76)	< 0.001	0.73 (0.71–0.76)	< 0.001	0.73 (0.70–0.75)	< 0.001
SES: Middle	1.01 (0.97–1.05)	0.659	1.01 (0.97–1.05)	0.666	1.01 (0.97–1.04)	0.699
SES: Low/lower‐middle	1.53 (1.46–1.61)	< 0.001	1.53 (1.46–1.61)	< 0.001	1.53 (1.46–1.60)	< 0.001
Academic: Middle	0.92 (0.88–0.96)	< 0.001	0.92 (0.88–0.96)	< 0.001	0.92 (0.89–0.96)	< 0.001
Academic: Low	1.08 (1.03–1.12)	< 0.001	1.07 (1.03–1.12)	< 0.001	1.08 (1.04–1.12)	< 0.001
Survey year: 2022	1.06 (1.01–1.12)	0.010	1.06 (1.01–1.12)	0.011	1.06 (1.01–1.12)	0.011
Survey year: 2023	1.07 (1.02–1.12)	0.004	1.07 (1.02–1.12)	0.004	1.07 (1.02–1.12)	0.004
Survey year: 2024	0.89 (0.85–0.94)	< 0.001	0.89 (0.85–0.94)	< 0.001	0.89 (0.85–0.94)	< 0.001

Note: Model 1 includes both PA types as linear main effects only (no interaction terms). Model 2 tests the Aerobic PA × violence interaction (linear and quadratic), adjusting for Muscle PA as a covariate. Model 3 tests the Muscle‐Strengthening PA × violence interaction (linear and quadratic), adjusting for Aerobic PA as a covariate. Aerobic PA frequency ranges 0–7 days/week; Muscle‐strengthening PA frequency ranges 0–5+ days/week (top‐coded; see Methods). Reference categories: socioeconomic status (SES), high/upper‐middle; academic performance, high; survey year, 2021. All models additionally adjust for depression, high perceived stress, loneliness, insufficient sleep, sex, and school level. Key interaction terms are bolded.

Abbreviations: CI, confidence interval; OR, odds ratio; PA, physical activity.

Because aerobic PA frequency was entered as a continuous linear and quadratic predictor, the violence main effect in Model 2 (OR = 3.11; 95% CI: 2.66–3.63) represents the violence–suicidal ideation odds ratio among adolescents with 0 days/week of aerobic PA, holding muscle‐strengthening PA and covariates constant. Model 2 tested the aerobic PA × violence interaction using both linear and quadratic terms. The linear term was significant (OR = 1.15; 95% CI: 1.02–1.30; p = 0.021), as was the quadratic term (OR = 0.98; 95% CI: 0.96–1.00; p = 0.010), indicating a statistically significant non‐monotonic (inverted U‐shaped) interaction pattern. As illustrated in Figure 1a (Table S1, Panel B), exploratory day‐by‐day analyses confirmed this pattern: relative to adolescents who reported no aerobic PA (0 days/week), violence‐victimized adolescents who reported irregular aerobic PA (1–3 days/week) showed markedly elevated odds of suicidal ideation (ORs: 1.51–1.78; all p < 0.01), whereas adolescents reporting more frequent aerobic PA (≥ 4 days/week) showed no significant interaction with violence victimization (ORs: 0.95–1.21; all p > 0.20). These findings indicate that the aerobic PA × violence interaction is frequency‐dependent and non‐linear rather than consistent with a monotonic inverse association and was sensitive to dichotomization cut‐point selection (Table S1, Panel A). Because these interaction effects were not evident in unadjusted comparisons but emerged after covariate adjustment (Table S2), we examined which covariates were responsible: excluding depression, perceived stress, and loneliness from the model eliminated the aerobic PA interaction (p = 0.896, 0.992) while revealing a significant muscle‐strengthening PA interaction that was not evident in the fully adjusted model (p = 0.012, 0.004). This indicates that the observed interactions are highly sensitive to adjustment for these plausible mediators, and should be interpreted with corresponding caution.

FIGURE 1.

Violence × physical activity frequency interaction odds ratios by aerobic and muscle‐strengthening PA frequency (KYRBS 2021–2024). Interaction odds ratios (OR) and 95% confidence intervals (CI) from complex sample logistic regression, comparing violence‐victimized versus non‐victimized adolescents at each level of PA frequency relative to the 0 days/week reference group (black square). Panel (a): Aerobic PA frequency (0–7 days/week); Panel (b): Muscle‐strengthening PA frequency (0–5+ days/week; top‐coded to harmonize 2021 wave). Significant interactions (p < 0.05) are indicated by filled orange triangles; non‐significant interactions are indicated by open gray circles, to ensure legibility in grayscale reproduction and for colorblind readers. The y‐axis is displayed on a log scale. All models adjust for depression, high perceived stress, loneliness, insufficient sleep, sex, school level, socioeconomic status, academic performance, and survey year, with the non‐focal PA type included as a continuous covariate. See Table S1, Panels (B) and (C) for numeric estimates.

Similarly, the violence main effect in Model 3 (OR = 2.82; 95% CI: 2.47–3.22) represents the violence–suicidal ideation odds ratio among adolescents with 0 days/week of muscle‐strengthening PA, holding aerobic PA and covariates constant. Model 3 tested the muscle‐strengthening PA × violence interaction. The linear term was not significant (OR = 0.98; 95% CI: 0.83–1.16; p = 0.837), nor was the quadratic term (OR = 1.03; 95% CI: 0.99–1.06; p = 0.106). However, day‐by‐day analyses (Figure 1b; Table S1, Panel C) revealed a distinct frequency‐dependent pattern: no significant interaction was observed at low muscle‐strengthening PA frequencies (1–2 days/week), whereas significantly elevated odds emerged at higher frequencies (3–5+ days/week; ORs: 1.40–1.76; all p ≤ 0.022). Notably, dichotomizing muscle‐strengthening PA at ≥ 3 days/week, the threshold used in the original analysis, reversed the sign of the interaction (dichotomized OR = 0.60; 95% CI: 0.50–0.71), illustrating how cut‐point selection can fundamentally alter the observed direction of an effect (Table S1, Panel D).

3.3. Sex‐ and SES‐Stratified Interaction Effects

Table 3 and Figure 2 display the sex‐ and SES‐stratified results. For the aerobic PA × violence interaction, the linear term was nominally significant among males (OR = 1.21 per 1‐day increase; 95% CI: 1.04–1.41; p = 0.015; Holm‐adjusted p = 0.28; FDR‐adjusted p = 0.10), with a nominally significant complementary quadratic term (OR = 0.97; 95% CI: 0.95–0.99; p = 0.006; Holm‐adjusted p = 0.13; FDR‐adjusted p = 0.08). Among females and all SES strata, neither the linear nor quadratic aerobic PA interaction terms reached statistical significance, though point estimates for the quadratic term were directionally consistent across strata (ORs: 0.97–0.99).

TABLE 3.

Sex‐ and SES‐stratified aerobic and muscle‐strengthening PA × violence interactions on suicidal ideation.

	Interaction term	Aerobic PA × violence		Muscle‐strengthening PA × violence
		OR (95% CI)	p	OR (95% CI)	p
Male	Linear (per 1‐day)	1.21 (1.04–1.41)	0.015	1.17 (0.94–1.46)	0.156
	Quadratic	0.97 (0.95–0.99)	0.006	0.99 (0.95–1.04)	0.747
Female	Linear (per 1‐day)	1.09 (0.90–1.31)	0.366	0.80 (0.61–1.05)	0.110
	Quadratic	0.99 (0.96–1.02)	0.501	1.08 (1.02–1.15)	0.008
High/upper‐middle SES	Linear (per 1‐day)	1.17 (0.98–1.40)	0.083	0.88 (0.69–1.13)	0.318
	Quadratic	0.97 (0.95–1.00)	0.044	1.05 (1.00–1.10)	0.054
Middle SES	Linear (per 1‐day)	1.08 (0.88–1.32)	0.479	1.24 (0.95–1.62)	0.114
	Quadratic	0.99 (0.96–1.02)	0.455	0.98 (0.93–1.03)	0.424
Low/lower‐middle SES	Linear (per 1‐day)	1.21 (0.93–1.59)	0.159	0.81 (0.54–1.22)	0.317
	Quadratic	0.97 (0.93–1.01)	0.101	1.07 (0.99–1.16)	0.097

Note: Aerobic PA frequency (0–7 days/week) and muscle‐strengthening PA frequency (0–5+ days/week, top‐coded) modeled continuously, with linear and quadratic interaction terms estimated within each stratum using a single pooled survey design with year‐specific strata and cluster identifiers (see Methods), subset to each stratum. Each model adjusts for depression, insufficient sleep, high perceived stress, loneliness, school level, academic performance, and survey year, plus sex (when stratifying by SES) or socioeconomic status (when stratifying by sex). KYRBS 2021–2024. Key interaction terms are bolded. *p < 0.05. **p < 0.01.

Abbreviations: CI, confidence interval; OR, odds ratio; PA, physical activity; SES, socioeconomic status.

FIGURE 2.

Sex‐ and SES‐Stratified Linear Interaction Odds Ratios: Aerobic and Muscle‐Strengthening PA × Violence on Suicidal Ideation (KYRBS 2021–2024). Forest plot of linear interaction odds ratios (OR per 1‐day increase in PA frequency) and 95% confidence intervals (CI) from sex‐ and SES‐stratified complex sample logistic regression. Panel (a): Aerobic PA × violence interaction; Panel (b): Muscle‐strengthening PA × violence interaction. Significant interactions (p < 0.05) are indicated by filled orange triangles; non‐significant interactions are indicated by open gray circles, to ensure legibility in grayscale reproduction and for colorblind readers. The dashed reference line at OR = 1.0 indicates no interaction effect. Note that significant non‐linear (quadratic) interaction terms were additionally estimated in these models and are reported in Table 3; the forest plot displays linear terms only to facilitate cross‐stratum visual comparison. All models adjust for depression, high perceived stress, loneliness, insufficient sleep, school level, academic performance, and survey year, with the stratifying variable (sex or SES) excluded from the covariate set within each stratum.

For the muscle‐strengthening PA × violence interaction, no linear or quadratic terms reached significance in any sex or SES stratum after adjustment for multiple comparisons. The exception was the quadratic term among females (OR = 1.08; 95% CI: 1.02–1.15; p = 0.008; Holm‐adjusted p = 0.15; FDR‐adjusted p = 0.08), which likewise did not survive correction. All other stratified estimates were non‐significant.

Formal three‐way interaction tests (PA × violence × sex; PA × violence × SES), conducted in the full pooled sample using joint Wald tests, were non‐significant for both aerobic PA (sex: p = 0.419; SES: p = 0.916) and muscle‐strengthening PA (sex: p = 0.064; SES: p = 0.301), providing no statistical evidence that the magnitude of either PA × violence interaction differed significantly by sex or SES. Taken together with the multiplicity‐corrected subgroup estimates, these stratified patterns should be interpreted as exploratory and hypothesis‐generating rather than confirmatory evidence of differential effect modification.

4. Discussion

This study, drawing on a nationally representative four‐year pooled sample of 214,218 Korean adolescents (KYRBS 2021–2024), examined whether aerobic and muscle‐strengthening PA differentially moderate the association between violence victimization and suicidal ideation. Three key findings emerge.

First, victimization due to violence emerged as a strong positive predictor of suicidal ideation (OR = 3.42). Nearly 39% of violence‐victimized youth endorsed suicidal ideation, speaking to the public health significance of this exposure. This effect size is comparable to previous systematic data findings, 2‐ to 4‐fold increased odds of suicide after interpersonal violence exposure (Castellví et al. 2017; Turner et al. 2012), and clarifies violence‐exposed adolescents as a targeted subpopulation for mental health treatment.

Second, aerobic PA × violence interaction had a frequency‐dependent, non‐monotonic trend. Compared with never engaging in aerobic PA (0 days/week), violence‐exposed adolescents who engaged in low or irregular levels of aerobic PA (1–3 days/week) had significantly increased odds of suicidal ideation across all models (ORs: 1.51–1.78), while those engaging in higher frequencies of aerobic PA did not meaningfully interact with violence victimization on suicidal ideation (≥ 4 days/week: ORs: 0.95–1.21). The inclusion of a quadratic term for the aerobic PA × violence interaction in Model 2 revealed a significant inverted U‐shape trend (OR = 0.98, p = 0.010), indicating that the association between aerobic PA frequency and violence‐suicidal ideation is not consistent with a linear inverse association. Non‐monotonic dose–response curves between PA frequency and health outcomes have been observed in other adolescent samples, challenging the single‐threshold PA cut points used in categorical definitions of PA (Zhang et al. 2025). We speculate that biological processes such as upregulation of BDNF and serotonergic signaling, and normalization of hypothalamic–pituitary–adrenal (HPA) axis responsivity to stressors, could plausibly contribute to exercise‐related psychological benefits (Cotman et al. 2007; Li et al. 2025; Schuch and Stubbs 2019), though these mechanisms were not directly tested in this study and are drawn from general exercise‐depression literature rather than violence‐specific research. We note that the aerobic PA interaction motivating this speculation was itself modest in magnitude and sensitive to adjustment for depression, perceived stress, and loneliness (Table S2); these mechanistic possibilities should therefore be read as speculative hypotheses for future research rather than an explanation of a well‐established effect. Irregular exercise performed at low frequencies (1–3 days/week) may not reach a sufficient “dose” to activate these neurobiological mechanisms of stress attenuation, and it is also possible that incident exercise among violence‐exposed youth at these low frequencies is reflective of stress‐induced rather than routine healthy behavior; a causal pathway this cross‐sectional study cannot distinguish. Among participants in the stratified analyses, the non‐monotonic relationship with aerobic PA reached nominal significance among males (both linear and quadratic interaction terms), but formal three‐way interaction tests found no statistical evidence that this pattern differed significantly from females or from other strata (PA × violence × sex: p = 0.419; see Results), a reminder that a significant estimate in one subgroup and a non‐significant estimate in another do not themselves constitute evidence of a difference between subgroups. This finding also differs from our a priori hypothesis that female adolescents would experience the largest moderating association of aerobic PA given their increased risk of internalizing symptoms after violence exposure (Nolen‐Hoeksema and Girgus 1994). Because none of the stratified estimates, including the nominally significant male pattern, survived correction for multiple comparisons (Table S3) or the formal interaction tests, these subgroup findings should be interpreted as exploratory rather than confirmatory. Point estimates across SES strata were likewise broadly similar in magnitude (aerobic linear OR: 1.08–1.21) despite differing statistical significance, suggesting that this variation more likely reflects differences in subgroup precision and sample size than genuine effect modification by SES. Returning to the intersectionality framework introduced above, the non‐significant three‐way interaction tests (PA × violence × sex; PA × violence × SES) do not necessarily indicate that sex and SES are irrelevant to how PA relates to the violence–suicidal ideation association (Gelman and Stern 2006). Violence victimization was rare in this sample (2.3%), which substantially limited power to detect differences across subgroups, particularly at the intersection of sex and SES. It is also possible that the intersectional patterns observed in prior KYRBS research on loneliness and depressive symptoms (Lee and Lim 2025) operate through mechanisms not fully captured by PA frequency alone, or that the specific population examined here (adolescents with hospital‐treated violence victimization) differs in relevant ways from the broader samples used in that prior work. Future research with larger violence‐exposed subsamples is needed to more directly test whether sex‐ and SES‐based vulnerabilities jointly shape the PA–violence–suicidal ideation association. More broadly, the interaction odds ratios observed in the full sample (ORs: 1.51–1.78 at 1–3 days/week; attenuation to 0.95–1.21 at ≥ 4 days/week) are modest in absolute magnitude, and the very large sample size (n = 214,218) substantially increases statistical power such that even small effects will achieve conventional significance thresholds. The wide confidence intervals observed across frequency levels (e.g., aerobic PA at 2 days/week: OR = 1.78, 95% CI: 1.34–2.37; muscle‐strengthening PA at 5+ days/week: OR = 1.76, 95% CI: 1.39–2.23) further underscore the imprecision of these estimates and the need for replication in independent samples. These findings should therefore be considered exploratory and hypothesis‐generating rather than definitive evidence of a clinically meaningful interaction.

Third, the interaction between muscle‐strengthening PA and violence followed a frequency‐dependent pattern, inconsistent with a main inverse association of muscle‐strengthening PA. Although continuous linear and quadratic terms of the PA × violence interaction were not statistically significant in the full sample analysis, individual day‐contrast analyses demonstrated a clear dose–response pattern of increasingly positive interaction odds at higher levels of muscle‐strengthening frequency: no significant interaction was detected at 1–2 days/week, while significantly positive interaction odds appeared ≥ 3 days/week (ORs: 1.40–1.76). Importantly, had muscle‐strengthening PA been dichotomized according to the standard ≥ 3 days/week cutoff, the interaction would have appeared inversely associated (OR = 0.60) (Royston et al. 2006), highlighting how dichotomizing a PA variable at a single cut‐point can flip the apparent direction of an effect if the true relationship is non‐monotonic/frequency‐dependent. Testing of continuous and nonlinear interaction specifications should be prioritized prior to applying binary operationalizations of PA frequency. Given the coding of this interaction term, this finding indicates that the violence–suicidal ideation association was stronger among adolescents meeting muscle‐strengthening PA guidelines (≥ 3 days/week) than among those who did not, a pattern consistent with previous findings among Korean adolescents implying paradoxical mental health signals from vigorous/high‐intensity exercise (Lee et al. 2013; Roh et al. 2015, 2024). We propose two untested hypotheses for this pattern. First, adolescents who participate regularly in formalized muscle‐strengthening disciplines (e.g., martial arts, weightlifting, organized team sports) may spend more time engaged in competitive peer groups or physically aggressive environments, where exposure to violence may co‐occur for some youth (Lee et al. 2013; Roh et al. 2015, 2024). Second, this association could reflect residual confounding by sport or team participation itself (e.g., contact‐sport environments that independently elevate both muscle‐strengthening PA frequency and violence exposure) rather than a direct effect of muscle‐strengthening PA. Neither mechanism was directly tested in this study, and both should be treated as hypotheses for future research rather than established explanations.

An important alternative interpretation is reverse causation: because violence victimization, PA, and suicidal ideation were all assessed concurrently over the same 12‐month period, adolescents experiencing suicidal ideation or depressive symptoms may have reduced their PA levels, rather than PA frequency causally moderating the psychological response to violence exposure. This possibility is supported by longitudinal evidence outside the violence‐exposure context: a cross‐lagged panel analysis of Japanese adolescents found that worsening depressive symptoms significantly predicted subsequent reductions in PA duration, with no corresponding effect of PA on later depressive symptoms (Kawai and Yamagata 2024), and a 6‐year prospective study of adolescent girls similarly found that depressive symptoms and depression onset significantly reduced future PA levels (Jerstad et al. 2010). If a similar depression‐ or ideation‐driven reduction in activity occurred among violence‐exposed adolescents in our sample, the observed interaction between PA and violence victimization could partly reflect adolescents' behavioral withdrawal from activity in response to psychological distress, rather than PA frequency buffering or amplifying the impact of violence exposure. Our cross‐sectional design cannot distinguish this reverse‐causal pathway from the moderation interpretation we have emphasized, and this ambiguity should be considered when interpreting the direction of the observed associations.

The contrasting patterns observed across PA types highlight an important methodological point; considering PA as one‐dimensional construct (especially when dichotomizing frequency) could obscure heterogeneous frequency‐dependent effects that not only vary in magnitude, but may also vary in direction (Ashdown‐Franks et al. 2020; Royston et al. 2006; Zhang et al. 2022). Previous studies reporting no or only small association between PA and suicidal ideation in Korean youth (Lee et al. 2013) may have missed this opposing signal. More importantly, aerobic and muscle‐strengthening PA do not appear to simply have different magnitudes of association with violence‐related suicidal ideation; they may have different frequency thresholds at which risk is increased or decreased. PA frequency should be modeled as continuous, and each PA type should be examined separately before conclusions are drawn about the PA–mental health relationship in violence‐exposed youth.

Several limitations should be considered. First, because this was a cross‐sectional study, causality cannot be determined. Relatedly, violence victimization, PA, and suicidal ideation were all assessed with reference to the same 12‐month recall window, so temporal ordering among these variables cannot be established; the observed interactions should therefore be interpreted as concurrent associations rather than evidence that PA precedes or moderates the psychological sequelae of violence exposure. Second, compared to other behaviors or chronic diseases, the prevalence of violence victimization is low (2.3%), providing less power to detect differences in stratified analyses. On the other hand, the very large overall sample size (n = 214,218) affords much statistical power, such that even relatively small interaction effects can be statistically significant. Therefore, the interaction ORs presented here (ORs: 1.15–1.78 for aerobic PA; 1.40–1.76 for muscle‐strengthening PA, at specified frequency categories) are intended as exploratory results of small magnitude and should not be considered practically or clinically meaningful based on statistical significance alone. Interpretation of findings should prioritize effect size magnitude and the width of confidence intervals alongside p‐values, as the latter alone may overstate the reliability of small interaction effects in large survey samples. Third, all measurements were self‐reported, which introduces potential recall and social desirability bias. Although the primary analyses modeled PA frequency continuously, muscle‐strengthening PA was top‐coded at 5+ days/week in the 2021 wave due to survey coding changes across waves, limiting the precision of estimates at the highest frequency levels. Additionally, the KYRBS does not assess exercise intensity, the type of formalized activity, or the social context of exercise, thereby precluding mechanistic interpretation of the observed patterns. Fourth, the KYRBS does not inquire about the type, context, or duration of students' exercise sessions; therefore, this study cannot fully interpret the mechanisms behind findings. Fifth, as this study was conducted using a population based in East Asia, these findings may not be generalizable to other populations. Sixth, although missingness was minimal (< 0.01% for socioeconomic status and academic performance, and none for the exposure, PA moderators, or remaining covariates), listwise deletion was used, and the analytic sample (n = 214,218) may not perfectly represent the full pooled sample if missingness was not completely random. Seventh, the sex‐ and SES‐stratified analyses involved 20 significance tests (5 strata × 2 PA types × 2 interaction terms); although we applied Holm and Benjamini‐Hochberg corrections and additionally conducted formal three‐way interaction tests (PA × violence × sex; PA × violence × SES), none of which were significant, the stratified subgroup patterns discussed above should still be interpreted cautiously as exploratory rather than confirmatory given the number of comparisons involved. Lastly, the significance of the aerobic and muscle‐strengthening PA × violence interactions was sensitive to the inclusion of depression, perceived stress, and loneliness in the covariate set: excluding these variables rendered the aerobic PA interaction non‐significant, whereas the muscle‐strengthening PA interaction remained significant (Table S2). This pattern underscores that these findings should be interpreted as associational rather than causal, and that the choice to adjust for plausible mediators materially shapes which interaction appears more robust. Because “buffering” is typically a public health claim concerned with absolute rather than relative risk, this study also examined these interactions on the additive scale. Absolute risk differences between violence‐exposed and non‐exposed adolescents varied substantially across PA frequency levels rather than remaining constant, ranging from 0.13 to 0.16 across aerobic PA frequencies and from 0.11 to 0.22 across muscle‐strengthening PA frequencies, and the relative excess risk due to interaction (RERI) was positive for both aerobic PA (RERI = 0.22, comparing 0 vs. ≥ 5 days/week) and muscle‐strengthening PA (RERI = 0.38, comparing 0 vs. ≥ 3 days/week). These results indicate that the observed interactions reflect a real difference on the absolute risk scale, rather than an artifact confined to the multiplicative odds ratio scale.

Despite these limitations, this study has notable strengths: a large, nationally representative four‐year pooled sample, rigorous accounting for complex sampling design, and the explicit, simultaneous testing of two PA types using continuous modeling with nonlinearity checks. The finding that aerobic and muscle‐strengthening PA show distinct, frequency‐dependent interaction patterns with violence victimization, rather than uniform inverse associations, has methodological implications for how PA is operationalized in future mental health research. The current findings do not provide sufficient evidence to recommend specific PA interventions for violence‐exposed youth, as the underlying patterns were non‐monotonic, sensitive to model specification, and cross‐sectional in nature. However, they do suggest that future prospective and experimental work should examine whether sustained, high‐frequency aerobic PA, rather than sporadic engagement, is associated with lower mental health risk in violence‐exposed adolescents, and whether the high‐frequency muscle‐strengthening signal represents a contextual risk (e.g., aggressive peer environments) amenable to targeted intervention.

5. Conclusions

This study examined whether aerobic and muscle‐strengthening PA frequency moderated the association between violence victimization and suicidal ideation in a large, nationally representative sample of Korean adolescents. Both PA types showed frequency‐dependent, non‐linear interaction patterns with violence victimization that diverged from simple inverse associations, and both were sensitive to the choice of analytical operationalization. Aerobic PA showed an inverted U‐shaped interaction, with elevated risk concentrated at low‐to‐moderate frequencies (1–3 days/week) and attenuation at higher frequencies (≥ 4 days/week), a pattern most pronounced among males. Muscle‐strengthening PA showed an approximately monotonic pattern of elevated interaction odds at higher frequencies (≥ 3 days/week), which reversed when the variable was dichotomized at the conventional ≥ 3 days/week threshold, illustrating a methodological artifact of cut‐point selection. Together, these findings highlight that PA type, PA frequency, and analytical modeling decisions all shape the observed relationship between PA and violence‐related suicidal ideation. Simple categorical PA classifications may obscure meaningful frequency‐dependent patterns. Future longitudinal research that models PA frequency continuously, separates PA types, and accounts for exercise context is needed before firm conclusions about the moderating role of PA for violence‐exposed adolescents can be drawn.

Author Contributions

Youngjun Lee: conceptualization, data curation, formal analysis, methodology, investigation, supervision, validation, software, writing – original draft, writing – review and editing.

Funding

The author has nothing to report.

Disclosure

IRB Statement: This study used publicly available, fully anonymized secondary data from the Korea Youth Risk Behavior Survey (KYRBS). As this study involved no identifiable human subjects, it did not meet the definition of human subjects research under 45 CFR 46 and was therefore not subject to IRB review.

Conflicts of Interest

The author declares no conflicts of interest.

Supporting information

Table S1: Sensitivity analyses: physical activity × violence victimization interaction effects on suicidal ideation across alternative model specifications (KYRBS 2021–2024, N = 214,218).

Table S2: Missing data summary, analytic sample size, and sensitivity analyses excluding potential mediator covariates.

Table S3: Multiplicity‐adjusted stratified interaction tests and formal three‐way interaction tests.

Data Availability Statement

The data that support the findings of this study are publicly available from the Korea Disease Control and Prevention Agency (KDCA) at https://www.kdca.go.kr (Korea Youth Risk Behavior Survey, KYRBS). Access requires registration and approval through the KDCA's data request system. Restrictions apply to the availability of these data, which were used under permission for this study.