Resilience as a predictor of mental health among medical students: a cross-sectional study using PLS-SEM | BMC Medical Education

Study design and setting

This quantitative study employed a descriptive-analytical, cross-sectional design. Data collection and analysis were carried out during the actual study period, from March to May 2025, after which the manuscript was prepared and submitted. The target population included all medical students enrolled at Kerman University of Medical Sciences (KUMS), located in southern Iran. The study aimed to examine the predictive role of resilience on mental health among this student population using structural equation modeling (PLS-SEM).

Kerman University of Medical Sciences (KUMS) is one of the oldest and most prominent academic health institutions in southeastern Iran. KUMS plays a vital role in training healthcare professionals, conducting medical research, and delivering health services to a broad region across the south and southeast of the Iran.

Participants and sampling

The study sample consisted of 385 medical students selected using a proportionate stratified random sampling method. The required sample size was calculated using Cochran’s formula [29], which is suitable for estimating sample size in large populations with a known confidence level and margin of error. The formula is as follows:

Where:

• Z = 1.96 (corresponding to a 95% confidence level).

• p = 0.5 (assumed proportion for maximum variability).

• d = 0.05 (margin of error).

Substituting these values:

Thus, a sample size of 385 participants was determined to ensure sufficient statistical power and precision. To further justify this sample size, a post hoc statistical power analysis was conducted using G*Power 3.1 software. Based on the model’s effect size (f² = 0.10), an alpha level of 0.05, and 6 predictors in the main regression model, the achieved power (1–β) was 0.95. This indicates that the sample size of 385 was sufficient to detect small to moderate effects with high statistical power.

To obtain a representative sample, proportionate stratified random sampling was used based on the three major phases of medical education at Kerman University of Medical Sciences (KUMS):

1. I.

   Basic sciences phase

2. II.

   Physiopathology phase

3. III.

   Clinical clerkship and internship phase

The sample size required from each academic phase was calculated according to the proportion of students enrolled in each phase. Within each stratum, participants were selected through simple random sampling using student ID numbers and a random number table.

Inclusion and exclusion criteria

To ensure the validity and generalizability of the findings, the following inclusion and exclusion criteria were applied:

• Inclusion Criteria: Participants were eligible to participate in the study if they met all of the following conditions:

  1. 1.

     Being a medical student enrolled in one of the academic phases (basic sciences, physiopathology, clinical clerkship, or internship) at Kerman University of Medical Sciences during the first semester of the 2024–2025 academic year.

  2. 2.

     Having completed at least one full semester of medical training to ensure sufficient exposure to academic stressors.

  3. 3.

     Being within the typical age range for university students (18 years or older).

  4. 4.

     Willingness to participate in the study and provide informed consent.

  5. 5.

     Ability to understand and respond to the study questionnaire (in Persian).

• Exclusion Criteria: Students were excluded from the study if they met any of the following criteria:

  1. 1.

     Refusal to participate or withdrawal of consent at any stage of the research.

  2. 2.

     Current diagnosis of a psychiatric disorder (e.g., major depressive disorder, generalized anxiety disorder, bipolar disorder) as recorded in university health records or self-reported during screening.

  3. 3.

     History of significant neurological conditions, such as epilepsy, multiple sclerosis, traumatic brain injury, or any condition that could interfere with cognitive or emotional functioning.

  4. 4.

     Presence of chronic physical illnesses (e.g., diabetes, cancer, autoimmune diseases) that may independently affect psychological well-being.

  5. 5.

     Completion of the questionnaire with missing or inconsistent responses that would impair data analysis.

These exclusion criteria were purposefully applied to control for clinical confounders that might independently influence mental health outcomes, such as chronic illness or psychiatric diagnoses. This approach allowed us to better isolate the role of resilience within a psychologically and physically stable student population. However, we acknowledge that these criteria may introduce a degree of selection bias and limit the generalizability of the findings to broader or more clinically diverse student populations. Therefore, caution is advised when extrapolating these results to students with complex medical or psychological conditions.

Instruments

Data collection was carried out using a structured questionnaire comprising three sections; The first section included a researcher-designed demographic questionnaire that collected basic participant information such as age, gender, marital status (single/married), place of residence (dormitory, with family, or independent), and employment status (employed/unemployed). These variables were used both for descriptive analysis and as potential moderating factors in the structural model.

The second section consisted of the Connor-Davidson Resilience Scale (CD-RISC), a validated 25-item self-report instrument designed to measure resilience across five dimensions: Perception of competence (8 items), Trust in individual instincts (7 items), Positive acceptance of change and secure relationships (5 items), Control (3 items), and Spiritual effects (2 items). Each item is rated on a 5-point Likert scale ranging from 1 (not true at all) to 5 (true nearly all the time), yielding a total score between 25 and 125. Based on the total score, resilience levels are classified as: very low [25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45], low (46–66), moderate (67–87), high (88–108), and very high (109–125). The CD-RISC has been previously validated in Iranian populations and cultural settings. For example, Kalbasi and Zahabion (2019) assessed its psychometric properties among university students in Iran and reported acceptable internal consistency (Cronbach’s alpha = 0.84) [30]. The Persian version was adapted through forward-backward translation, reviewed by bilingual experts, and pilot-tested to ensure cultural and linguistic appropriateness. In the present study, a pilot test was also conducted on a subsample of 30 students prior to the main data collection. The internal consistency of the Persian version of CD-RISC was re-evaluated, yielding a Cronbach’s alpha of 0.87, indicating good reliability in the current sample.

The third part consisted of the standardized mental health questionnaire (Symptom Checklist-90-Revised, SCL-90-R). This instrument includes 90 items that assess nine major dimensions of psychological symptoms, including: Depression (13 items), Anxiety (10 items), Somatization (11 items), Obsessive-Compulsive symptoms (10 items), Interpersonal Sensitivity (9 items), Hostility (6 items), Phobic Anxiety (7 items), Paranoid Ideation (6 items), and Psychoticism (10 items), along with 7 additional items. Responses are rated on a 5-point Likert scale (none, a little, moderately, quite a bit, extremely), scored from 0 to 4 to reflect symptom severity. The frequency and percentage for each dimension were calculated based on the categorization of items. To compute the Global Severity Index (GSI), the total score from all 90 items was summed and divided by the number of items (90). Based on this index, an average score ≤ 0.5 indicates no psychological symptoms (psychologically healthy), 0.51 to 1.5 reflects mild symptoms, 1.51 to 2.5 indicates moderate symptoms, and scores above 2.5 suggest severe psychological distress. The Persian version of the SCL-90-R has been extensively validated in numerous Iranian studies, confirming its reliability and validity in this cultural context. For example, Akhavan Abiri and Shairi (2020) reported a test-retest reliability coefficient of 0.97, and sensitivity, specificity, and overall accuracy of 0.94, 0.98, and 0.96, respectively [31]. In the present study, a pilot test was also conducted on a subsample of 30 students before the main data collection, and the internal consistency of the SCL-90-R was assessed, yielding a Cronbach’s alpha of 0.94, indicating excellent reliability in the current sample.

Procedures

After obtaining ethical approval from the Research Ethics Committee of Kerman University of Medical Sciences and securing informed consent from participants, data collection was carried out in a structured and ethical manner. All eligible students were informed in advance about the study and invited to participate voluntarily. The objectives of the study, the nature of the questions, and the confidentiality of the data were clearly explained to them in a transparent and comprehensible manner. Participants were also explicitly informed about any potential risks, including possible emotional discomfort from responding to sensitive questions related to mental health and stress. They were advised that they could skip any question they felt uncomfortable answering or withdraw from the study at any time without penalty. On the designated data collection day, printed questionnaires were distributed and collected in person during scheduled class sessions. A trained member of the research team, who had received instruction on ethical data collection procedures, supervised the process to ensure consistency and accuracy. The researcher provided necessary guidance while ensuring that participants could complete the forms privately and without pressure. Informed written consent was obtained from all participants prior to questionnaire distribution. They were explicitly assured that participation was voluntary, their responses would remain anonymous, and the collected data would be used solely for academic research purposes. All questionnaires were completed and returned on the same day to minimize response bias and maximize response rates.

Statistical analysis

Data analysis was conducted using the Partial Least Squares Structural Equation Modeling (PLS-SEM) approach with SmartPLS (a software for PLS-SEM) version 4.0.9.6. This method was chosen due to its suitability for predictive modeling, its ability to manage complex models involving multiple latent variables and observed indicators, and its robustness with smaller sample sizes. Furthermore, PLS-SEM was preferred over covariance-based SEM (CB-SEM) because the primary objective of this study was to predict endogenous constructs and explore complex relationships, including moderation effects, rather than to confirm a well-established theoretical model.

The analysis followed a two-step procedure. In the first step, the measurement model (outer model) was evaluated to assess the reliability and validity of the constructs [32]. Indicator reliability was verified through factor loadings, with values ≥ 0.70 considered acceptable [33]. Internal consistency was assessed using both Cronbach’s alpha and composite reliability (CR), with thresholds of 0.70 or higher [32]. Convergent validity was confirmed if the Average Variance Extracted (AVE) for each construct was ≥ 0.50 [34]. Discriminant validity was assessed using both the Fornell-Larcker criterion and the Heterotrait-Monotrait ratio (HTMT) [35], ensuring that each latent variable was sufficiently distinct from the others.

In the second step, the structural model (inner model) was analyzed to test the study hypotheses. Prior to interpretation, multicollinearity was checked using the Variance Inflation Factor (VIF), with acceptable values being below 5 [36]. The significance of hypothesized paths was determined through a bootstrapping procedure to estimate standard errors, t-values, and confidence intervals [37]. The model’s explanatory power was evaluated using the coefficient of determination (R²), where values of 0.25, 0.50, and 0.75 were interpreted as weak, moderate, and substantial, respectively [38]. Effect sizes (f²) were interpreted according to Cohen’s guidelines for structural equation models, where 0.02 indicates a small effect, 0.15 medium, and 0.35 large [39]. These values may differ from correlation coefficients due to the inclusion of multiple predictors and model complexity. Additionally, the model’s predictive relevance was assessed using the Stone-Geisser Q² statistic obtained via the blindfolding procedure, where Q² values greater than zero indicated acceptable predictive relevance [40].

Although PLS-SEM does not traditionally rely on global model fit indices like the Root Mean Square Error of Approximation (RMSEA) or the Comparative Fit Index (CFI), recent developments have introduced approximate model fit criteria for PLS-SEM [41]. In this study, model fit was assessed primarily using the Standardized Root Mean Square Residual (SRMR), with additional reference to the Normed Fit Index (NFI) and the Chi-square statistic (χ²), as generated by SmartPLS 4. An SRMR value below 0.08 was considered indicative of good model fit. While RMSEA, CFI, and the Non-Normed Fit Index/Tucker-Lewis Index (NNFI/TLI) are more commonly used in covariance-based SEM, their reference values (e.g., RMSEA < 0.08, CFI and NNFI > 0.90, and χ²/df < 3) are reported here for comparison purposes, in line with established thresholds [42]. However, interpretations of these indices in the context of PLS-SEM should be made with caution.

In the correlation analysis section, effect sizes based on Pearson’s r were interpreted using Cohen’s criteria, where r values around 0.50 represent large effects, 0.30 medium, and 0.10 small [39]. However, in line with recommendations for psychological research, correlations in the range of 0.40–0.45 and above were also considered strong and practically meaningful.

Moderation hypotheses (involving age, gender, marital status, residence, and employment status) were tested by creating interaction terms between resilience and each moderator. These moderators were selected based on theoretical and empirical evidence suggesting their potential influence on mental health outcomes and resilience. Marital status was included because it can affect social support systems, which may buffer stress and thus modulate the impact of resilience on mental health. Employment status influences financial stability and daily stress levels, which can interact with an individual’s resilience capacity. Age and gender have been widely reported in the literature as important demographic factors that shape coping mechanisms and psychological resilience. Additionally, residence type, whether living in a dormitory, with family, or independently, affects environmental stressors and social networks that can moderate the relationship between resilience and mental health. Therefore, these variables were chosen to explore how different social and demographic contexts might influence the predictive role of resilience. The moderation effects were evaluated using the bootstrapping method to determine their statistical significance at the 0.05 level.

All statistical procedures adhered to a 95% confidence level, and data analyses were interpreted accordingly.