Impact of COVID-19 School Closures on German High-School Graduates’ Perceived Stress: A Structural Equation Modeling Study of Personal and Contextual Resources

Abstract

COVID-19 school closures forced German high-school graduates (Abitur 2022 cohort) to prepare for their final examinations with lengthy learning times at home. Guided by transactional stress theory, we tested how personal resources—self-regulated learning (SRL) skills and academic self-efficacy—and contextual resources—perceived teacher support and teacher digital competence—jointly predicted perceived stress during exam preparation. A cross-sectional online survey (June–July 2022) yielded complete data from N = 2379 students (68% female; Mage = 18.3). Six latent constructs were measured with 23 items and showed adequate reliability (0.71 ≤ α/ω ≤ 0.89). A six-factor CFA fit the data acceptably (CFI = 0.909, RMSEA = 0.064). The structural equation model (CFI = 0.935, RMSEA = 0.064) explained 35% of the variance in stress and 23% of the variance in SRL—action. Academic self-efficacy (β = −0.31, p < 0.001), perceived support (β = −0.28, p < 0.001), teacher digital competence (β = −0.08, p < 0.001), COVID-19 learning disruptions (β = +0.13, p < 0.001), and gender (male = 0.32 SD lower stress, p < 0.001) had direct effects on stress. SRL—action’s direct path was small and non-significant (β = −0.02). Teacher digital competence also reduced stress indirectly through greater perceived support (standardized indirect β = −0.11, p < 0.001). The results highlight self-efficacy and perceived instructional support as the most potent buffers of pandemic-related stress, whereas cancelled lessons added strain. Boosting teachers’ digital pedagogical skills has a dual payoff—raising students’ sense of support and lowering their stress.

1. Introduction

School closures and social-distancing rules introduced to curb COVID-19 profoundly disrupted the daily lives of children and adolescents worldwide. Early multi-national surveys recorded sharp increases in depressive mood, loneliness, and anxiety among pupils forced to engage in home-based learning (Bujard et al., 2021; Döpfner et al., 2021; Zhou et al., 2020). In Germany, the longitudinal COPSY study showed that the share of young people with clinically relevant distress doubled from spring 2020 to winter 2020/2021 and, despite experiencing minor relief, remained well above pre-pandemic baselines one year later (Ravens-Sieberer et al., 2021a, 2021b, 2022). These deteriorations were echoed by reports of greater behavioral problems, heavier media use, and poorer sleep and eating habits (Schlack et al., 2020; Wößmann et al., 2020).

1.1. Personal Resources

However, mean-level declines mask substantial inter-individual variability. Following transactional stress theory (Lazarus & Folkman, 1984), young people’s responses are shaped by the balance between perceived demands and available resources. Evidence indicates that social background, gender, and technological access each tilt this balance in characteristic ways. Pupils from low socioeconomic backgrounds, in single-parent households, or living with relatives who have mental illness reported consistently higher stress levels than peers with more favorable circumstances (Pro Juventute, 2021; Ravens-Sieberer et al., 2022). Girls experienced more anxiety and a greater overall burden than boys (Baier & Kamenowski, 2022; Rogge & Seifert, 2023). The findings for age are mixed: some studies show stronger effects among younger cohorts, possibly because older adolescents maintained peer contact online (Langmeyer et al., 2020; Romero et al., 2020), whereas others point to heightened anxiety and depressive symptoms in 14- to 17-year-olds later in the pandemic (Mansfield et al., 2021; Ravens-Sieberer et al., 2022).

Apart from personal factors such as gender, age, social background, and technological access, individual learning habits seem to have an impact on subjective well-being. Empirical reviews during the COVID-19 pandemic consistently show that students with strong self-regulated learning (SRL) skills reported greater life satisfaction and lower stress levels despite distance learning demands (e.g., Holzer et al., 2023; Brauchle et al., 2025). Building on Nett and Götz’s (2019) comprehensive definition, SRL is the capacity to set learning goals, deploy cognitive and metacognitive strategies, monitor progress, and iteratively adjust actions. Central to this conception are (a) goal setting and (b) the informed, strategic use of learning techniques. Learning strategies are understood as “bundles of cognitions and behaviors that can be purposefully deployed to initiate, maintain, and improve learning processes” (Nett & Götz, 2019, p. 69). Nett and Götz distinguish between hierarchical models, which enumerate components such as specific learning tactics and overarching metacognitive routines, and process models, which conceive SRL as an iterative sequence of phases. For studying learning over extended periods of digitally mediated distance education, a process perspective seems advantageous since temporal dynamics in longer periods of self-regulated learning at home can be described in more detail.

Drawing on Zimmerman (2001), Schmitz (2001) distinguishes a pre-actional phase (planning), an actional phase (performance), and a post-actional phase (reflection).

1.1.1. Pre-Actional Phase

Learners first appraise the task and filter it through available resources. Straight-forward tasks may be executed “automatically”, whereas challenging tasks trigger an explicit check of motivational and emotional resources, self-efficacy beliefs, and the energy required. Only after this appraisal do learners set goals—the core of SRL—and devise a process plan (Schmitz & Schmidt, 2007). Unlike Zimmerman’s original model, Schmitz conceptualizes planning broadly, encompassing considerations of prior knowledge, concrete cognitive strategies and metacognitive activities such as self-monitoring.

1.1.2. Actional Phase

Goals are translated into behavior through the application of cognitive, metacognitive, and resource management strategies. Volitional aspects—e.g., sustaining concentration and protecting motivation—become pivotal. Quantitative indicators (learning time) and qualitative indicators (appropriate strategy use) can be used to index learning quality. Continuous self-monitoring allows learners to compare actual with intended behavior and to make in situ adjustments that feed forward into subsequent learning episodes.

1.1.3. Post-Actional Phase

Learners evaluate outcomes with regard to subjective satisfaction, quantitative criteria (e.g., the amount of material covered), and qualitative criteria (e.g., the depth of understanding). These judgements inform future motivational, emotional, and strategic appraisals, thus closing the regulatory loop.

1.2. Contextual Resources

Beyond personal dispositions, the quality of learners’ immediate environment further tilts the demand–resource balance. The JuCo surveys suggested that having a private room and adequate digital devices buffered subjective stress (Andresen et al., 2022), a finding that mirrors modest but reliable associations between technical equipment and lower levels of negative emotions in the first lockdown (Huber & Helm, 2020). Still, material assets alone rarely sufficed. Perceived social support—from parents or, crucially, from teachers—emerged as a stronger predictor of well-being than hardware availability (Cohen et al., 2022; Gayatri & Irawaty, 2022; Grommé et al., 2023; Tang et al., 2021; Wenter et al., 2022). Self-determination theory (SDT; Ryan & Deci, 2000) would interpret this pattern as the fulfilment of basic needs for relatedness and competence, thereby sustaining intrinsic motivation even under adverse conditions.

A special form of support during distance learning is digital pedagogical competence. Eickelmann and Drossel (2020) showed that pupils attending schools already advanced in digitalization before the pandemic were reached more consistently by online instruction, implying they received more guidance of higher quality. Among German high-school graduates (Abiturienten), perceived teacher media competence correlated positively with perceived instructional support and negatively with stress (Rogge & Seifert, 2023). Whether similar links hold for younger students remains unclear: recent evidence suggests that primary-age pupils benefit even more from parental and teacher scaffolding (Grommé et al., 2023), yet other studies found that this age group was initially the most burdened by the disruption of structured school routines (Langmeyer et al., 2020).

Despite the rapidly growing literature, two gaps persist. First, most mental health studies analyze broad cohorts; few drill down into high-stakes subgroups such as Abitur candidates, whose final examinations determine access to tertiary education and thus may amplify appraisal-related stress. Second, existing studies often focus on either personal or contextual resources, making it difficult to see how these layers operate together. Guided by transactional stress theory, we therefore examine personal factors (SRL skills and academic self-efficacy) alongside two key contextual supports (perceived instructional help and teachers’ digital pedagogical competence) to provide a more integrated picture of the forces that buffer students’ stress during prolonged disruptions.

1.3. The Present Study

The present study addresses these gaps, focusing on 2379 German students who graduated in 2022, after almost two years of intermittent distance instruction. Drawing on transactional stress theory (Lazarus & Folkman, 1984) and self-determination theory (Ryan & Deci, 2000), we specify a structural equation model in which self-regulated learning (SRL) and academic self-efficacy are conceptualized as personal resources that counteract subjective stress. While well-being is multidimensional (WHO-5, life satisfaction, and positive affect), pandemic studies consistently report perceived stress as the most proximal and sensitive marker of adolescents’ mental load (Betthäuser et al., 2023; Ravens-Sieberer et al., 2022). We therefore treat stress as the parsimonious affective outcome in this study, noting that higher levels of stress co-occur with lower levels of life satisfaction and positive affect in the German COPSY data set.

Recent evidence supports this choice: higher SRL skills—and the concomitant reduction in procrastination—predicted less stress and greater life satisfaction in German and Swiss secondary students during school closures (Brauchle et al., 2025). SRL also showed an age-graded increase, suggesting that older adolescents, such as Abitur candidates, may derive protective benefits (Brauchle et al., 2025). Parallel work indicates that academic self-efficacy buffered the negative impact of forced distance learning on perceived competence and, indirectly, on psychological well-being (Holzer et al., 2023; Fomina et al., 2020; both summarized in Brauchle et al., 2025).

In addition to these personal resources, we incorporate contextual supports—perceived learning support, teacher digital competence, and the infrastructure of the home-learning environment—because prior German studies emphasized their role in shaping pandemic learning experiences (Andresen et al., 2022; Rogge & Seifert, 2023; Grommé et al., 2023).

This study aims to address the following research questions:

RQ1. 

To what extent are SRL skills and academic self-efficacy associated with students’ psychological and affective well-being (operationalized as perceived stress)?

RQ2. 

How do perceived instructional support and teacher digital competence affect stress, directly and indirectly, via SRL/self-efficacy?

RQ3. 

What is the unique contribution of demographic (gender) and contextual (COVID-19 learning disruption) factors to stress after personal and instructional resources are taken into account?

Related to these research questions, the following was hypothesized:

H1a. 

Higher SRL skills predict lower stress.

H1b. 

Higher academic self-efficacy predicts lower stress.

H2a-1. 

Greater perceived instructional support predicts lower stress.

H2a-2. 

Greater perceived instructional support predicts higher SRL skills.

H2b-1. 

Higher teacher digital competence predicts greater perceived instructional support.

H2b-2. 

Teacher digital competence reduces stress indirectly via perceived instructional support.

H3. 

Greater COVID-19 learning disruption predicts higher stress, whereas male gender predicts lower stress, controlling for all other predictors.

Figure 1 depicts the a priori hypothesized relations that guided both the CFA and the subsequent SEM analyses. The solid arrows represent hypothesized direct effects; the dashed arrow indicates an indirect path (teacher digital competence → stress via support). The plus and minus signs denote expected positive and negative associations, respectively.

By modeling personal (SRL and self-efficacy) and contextual (perceived support and teacher digital competence) resources simultaneously, the study moves beyond descriptive prevalence estimates toward an explanatory account of how Abitur candidates navigated pandemic learning demands. The findings aim to inform targeted interventions capable of buffering mental health risks and sustaining academic engagement during future disruptions—whether triggered by pandemics, teacher shortages, or climate-related closures (Clarke et al., 2022; Marani et al., 2021).

2. Materials and Methods

2.1. Design and Setting

The study employed a cross-sectional online survey design. Data collection took place between 6 June and 27 July 2022, depending on the federal state, either during the final phase of or immediately following the Abitur examinations. A questionnaire was administered using the online survey platform Typeform. A total of 96,000 high-school graduates from the 2022 cohort across Germany were invited via email to participate. Participation was voluntary and anonymous. The survey asked participants to retrospectively assess their experiences with school-guided home learning during Abitur preparation in upper secondary education between 2020 and 2022, as well as their experiences with their final examinations, a limitation discussed below. In total, N = 3896 students completed the survey (opened: 7475; started: 5556; drop-off: 1660 [28.8%]; completion rate: 70.2%). To attenuate self-selection bias, respondents could enter a lottery for three EUR 100 vouchers.

2.2. Participants

Eligibility was restricted to students enrolled in Grade 12 or 13 of a German secondary institution (gymnasiums and comprehensive or vocational schools) during the 2021/2022 academic year who intended to sit the final high-school examination in spring 2022. A total of 27 non-binary respondents were excluded owing to insufficient cell size. After exclusions, 2 379 students remained (68.1% female). Participants attended gymnasiums (80%), comprehensive schools (10%), vocational upper-secondary schools (9%), or private schools (1%). Mean age was 18.30 years (SD = 0.84). Compared with national graduation statistics, gymnasium graduates and female students are over-represented (see

Table 1. Sample characteristics.

School Type	N	%	Sex				Age
			Female		Male
			n	%	n	%	M	SD
Gymnasium	1909	80.2	1298	68	611	32	18.1	0.78
Comprehensive	237	10	160	67.5	77	32.5	18.8	0.66
Vocational	211	8.9	147	69.7	64	30.3	19	0.85
Private	22	0.9	16	72.7	6	27.3	18.8	0.53
Total	2379	100	1621	68.1	758	31.9	18.3	0.84

Notes. N = Number of participants, M = Mean, SD = Standard deviation, % within school type, Percentages might not have a sum of 100 due to rounding.

).

2.3. Measures

All scales were administered in German. Item wording for the SRL sub-scales was adapted from the Lernstrategien-Inventar (Baumert, 1993) and the LIST questionnaire (Wild & Schiefele, 1994). Items were mapped to the three-phase process model (pre-actional, actional, and post-actional phases). Cognitive interviews with eight Abitur graduates ensured comprehensibility. SRL and learning environment items had five-point Likert formats (from 1 = “trifft gar nicht zu” to 5 = “trifft vollkommen zu”). Perceived stress and academic self-efficacy had seven-point Likert formats (from 1 = “stark belastet” to 7 = “gar nicht belastet”). Cronbach’s α values refer to the current sample.

• Perceived stress (5 items, α = 0.85);
• Self-regulated learning—planning (4 items, α = 0.77);
• Self-regulated learning—actional phase (5 items, α = 0.89);
• Self-regulated learning—reflection (3 items, α = 0.79);
• Academic self-efficacy (3 items, α = 0.71);
• Learning environment (3 items, α = 0.74).

In addition, four single-item predicators/covariates were recorded: (a) perceived support, (b) teacher digital competence, (c) COVID-19 learning disruption, and (d) gender (male = 1, female = 0). Of these, perceived support enters the SEM as a predictor of stress and SRL—action.

Confirmatory Factor Analysis

A confirmatory factor analysis (CFA) was conducted to verify that the six theoretically posited constructs—perceived stress (Belastung), three phases of self-regulated learning (SRL—planning, SRL—action, and SRL—reflection), academic self-efficacy, and learning environment—were empirically distinguishable.

Because all indicators were collected according to five- or seven-point ordered categories and showed moderate skew, we first evaluated a robust DWLS solution. Fit patterns and factor loadings were virtually identical to maximum likelihood (ML); therefore, to maintain full comparability with the subsequent structural models, we report the ML results while noting that the robust DWLS method yielded the same substantive conclusions. The results are as follows: χ²(215) = 2283.03, CFI = 0.909, TLI = 0.892, RMSEA = 0.064 [0.061, 0.066], and SRMR = 0.059. Although TLI falls slightly below the conventional 0.90 guideline, RMSEA and SRMR lie well within the “acceptable” region (<0.08). Given the large sample (N = 2379) and the high indicator-to-parameter ratio, we interpret the overall pattern as borderline-good and sufficient for further modeling.

Standardized loadings (|λ|) ranged from 0.54 to 0.84 (median = 0.73) and were all highly significant (z = 26–49, p < 0.001), indicating each item made a substantive contribution to its intended factor (see

Table 2. Confirmatory factor analysis.

Factor	CR	AVE	Indicator	Loading (std.)	SE	R2
Stress	0.86	0.61	Exhausted (Exam Prep)	0.78	0.03	0.61
			Burnt Out (School)	0.74	0.03	0.55
			Workload (Exam Prep)	0.75	0.02	0.56
			Drained (School)	0.71	0.03	0.51
			School Demands	0.73	0.02	0.53
SRL—planning	0.80	0.55	Goal Formulation	0.69	0.03	0.48
			Thoughtful Preparation	0.73	0.03	0.53
			Adaption Learn. Plan I	0.66	0.04	0.43
			Adaption Learn. Plan II	0.63	0.03	0.40
SRL—reflection	0.81	0.63	Review Misunderstood II	0.65	0.02	0.43
			Review Learned III	0.83	0.03	0.69
			Review Learned II	0.78	0.03	0.60
SRL—action	0.89	0.66	Focus (Home)	0.79	0.03	0.63
			Learning Success	0.84	0.03	0.71
			Motivation (Home)	0.79	0.03	0.63
			Organization (Home)	0.74	0.03	0.55
			Overview (Home)	0.73	0.03	0.54
Self-efficacy	0.74	0.52	Good Student	0.82	0.03	0.67
			Effective (School)	0.69	0.04	0.48
			Effective (Exam Prep)	0.55	0.03	0.30
Environment	0.76	0.56	Technical Equipment	0.79	0.03	0.62
			Internet Quality	0.70	0.03	0.49
			Quiet Place to Study	0.62	0.03	0.39

Note. λ = standardized factor loading; CR = composite reliability; AVE = average variance extracted. Loadings ≥ 0.50 bold-faced. Fit statistics (ML): χ²(215) = 2283.03, CFI = 0.909, TLI = 0.892, RMSEA = 0.064 [0.061, 0.066], SRMR = 0.059.

). Item-specific R² values averaged 0.54, exceeding the 0.40 benchmark for adequate reliability. Inter-factor correlations were modest (|r| ≤ 0.47; mean = 0.26), with the strongest associations—expectedly—among the three SRL dimensions. The stress factor correlated negatively with SRL—action (r = −0.24) and self-efficacy (r = −0.23) and most strongly with learning environment support (r = −0.37), supporting the constructs’ discriminant validity.

Despite marginally sub-optimal TLI, all other diagnostics, combined with theoretically plausible loadings and moderate latent correlations, confirm that the six-factor measurement model provides an adequate representation of the data. These latent variables therefore constitute the measurement block in subsequent structural equation modeling reported in the manuscript.

2.4. Procedure

After following a study information screen and providing digital consent, students completed the 15 min questionnaire. A debriefing page supplied mental health resources and instructions for entering the voucher lottery. The survey did not store IP addresses or indirect identifiers.

2.5. Data Preparation

Item-level missingness was <2%. Listwise deletion produced N = 2379 for SEMs. All interval-scaled variables were mean-centered.

2.6. Analytic Strategy

Analyses were conducted in R 4.4.2 with lavaan 0.6.19 (Rosseel, 2012). CFA and SEM were estimated with robust maximum likelihood (MLR); the DWLS method was inspected for the CFA and yielded virtually identical results. Standardized estimates and 95% percentile bootstrap CIs (2000 draws) are reported. Model fit was evaluated against the joint cut-offs of CFI ≥ 0.90 (0.95 desirable), RMSEA ≤ 0.06–0.08, and SRMR ≤ 0.08 (Hu & Bentler, 1999; Schermelleh-Engel et al., 2003).

2.6.1. Measurement Model

See Section 2.3 for CFA details.

2.6.2. Structural Equation Modelling (SEM)

The structural model was built on the six-factor measurement model described in Section 2.3. All latent variables were specified in lavaan with factors fixed to unit variance (std.lv = TRUE), so structural paths can be interpreted as fully standardized β-coefficients.

• Model specification

• Perceived stress was regressed on;
• SRL—action;
• Academic self-efficacy;
• Perceived support;
• Teacher digital competence.

SRL—action and academic self-efficacy were, in turn, regressed on contextual supports to capture indirect effects. Indirect paths (e.g., support → SRL → stress) were defined with lavaan’s = operator and evaluated with bias-corrected bootstrap confidence intervals (2000 resamples). Model comparisons indicated that a more parsimonious specification (details in Section 4.1) provided an equally substantive account of the data.

• Estimation and fit criteria

Models were estimated with robust maximum likelihood (MLR) and full-information maximum likelihood for missing data. Goodness of fit was judged by the joint criteria of χ²/df < 3, CFI ≥ 0.90 (≥0.95 preferred), TLI ≥ 0.90, RMSEA ≤ 0.06–0.08, and SRMR ≤ 0.08. Nested models (e.g., with vs. without contextual supports) were compared using scaled Δχ² tests and ΔCFI ≥ 0.010.

2.7. Power Considerations

A Monte Carlo simulation (500 replications, simsem) indicated 0.88 power for paths of |β| ≈ 0.08 (e.g., teacher digital competence → stress) and ~0.10 power for very small paths (|β| ≈ 0.02). Analytic RMSEA power for the close-fit test (df = 284) was 1.00.

3. Results

3.1. Preliminary Analyses

• Normality

The absolute skew ranged from −0.44 (stress) to +0.10 (teacher digital competence), and the excess kurtosis ranged from −0.82 (perceived support) to −0.25 (self-efficacy). With N = 2379, deviations of this magnitude are unlikely to bias the robust maximum-likelihood estimation.

• Scale reliabilities

The Cronbach’s α ranged from 0.71 (self-efficacy) to 0.89 (SRL—action), confirming adequate internal consistency (see Section 2.3).

3.1.1. Descriptive Statistics

The analytic sample comprised 2379 cases (68.1% female, n = 1621; 31.9% male, n = 758; Mage = 18.30, SD = 0.84). Perceived stress (M = 4.39, SD = 1.16) correlated modestly with SRL—action (r = −0.21, p < 0.001).

3.1.2. Assumption Checks

Multicollinearity among the predictors was assessed using the variance inflation factor (VIF) and corresponding tolerance values (

Table 3. Test for multicollinearity within analytic sample.

	VIF	Tolerance
SRL—action	1.38	0.72
Self-efficacy	1.36	0.74
Gender	1.04	0.96
Teacher digital competence	1.11	0.90
COVID-19 learning disruption	1.05	0.95

Note. Dependent variable: academic stress; VIF < 10 and tolerance > 0.10 indicate no problematic multicollinearity.

). VIFs ranged from 1.04 for gender to 1.38 for SRL—action, all well below the conventional cutoff of 10. Likewise, tolerance values spanned from 0.72 (SRL—action) to 0.96 (gender), exceeding the recommended minimum of 0.10. These diagnostics indicate that none of the independent variables exhibit problematic multicollinearity, justifying their joint inclusion in subsequent regression models predicting academic stress.

3.2. Measurement Model

A six-factor confirmatory factor analysis (stress, SRL—planning, SRL—reflection, SRL—action, academic self-efficacy, and learning environment) was estimated with the maximum likelihood (ML). The fit indices indicated an acceptable representation of the data—particularly given the large indicator-to-parameter ratio (23 items and 61 free parameters): χ²(215) = 2283.03, p < 0.001; CFI = 0.909; TLI = 0.892; RMSEA = 0.064, 90% CI [0.061, 0.066]; and SRMR = 0.059. A supplementary CFA with a robust DWLS method (WLSMV) treating all 23 Likert indicators as ordered yielded virtually identical standardized loadings (0.66–0.88). The model fit was slightly different: χ²(215) = 2964.21, CFI = 0.974 (ΔCFI = +0.065), TLI = 0.969, RMSEA = 0.073 [0.071, 0.076], and SRMR = 0.059. Using the scaled (robust) test statistic lowered the CFI to 0.885 (RMSEA = 0.079), but the substantive pattern of results remained unchanged. Thus, the conclusions do not hinge on the estimator choice.

The standardized loadings ranged from 0.55 to 0.84 (median = 0.73) and were all highly significant (z = 25–49, p < 0.001). The composite reliabilities (CRs) met or exceeded the 0.70 benchmark for all constructs (CR = 0.73–0.89). The AVE ranged from 0.46 to 0.61. Although the AVE for SRL—planning fell slightly below 0.50, the combination of the strong loadings and satisfactory CR supports convergent validity. Discriminant validity was likewise supported: latent correlations were modest (|r| ≤ 0.47, M = 0.28), with the largest associations—expectedly—among the three SRL dimensions. The stress factor correlated negatively with SRL—action (r = −0.24) and self-efficacy (r = −0.23) and most strongly with learning environment (r = −0.37). Given these diagnostics, the six latent variables were kept unchanged as the measurement block for all subsequent structural modeling.

3.3. Structural Equation Modeling

The structural model (Figure 2) combined the six latent factors established in the CFA with four single-item covariates (gender, teacher digital competence, COVID-19 learning disruption, and perceived support) and was estimated with the robust ML (MLR). The global fit met conventional cut-offs (

Table 4. Model fit indices (CFA vs. full SEM).

	χ2	df	CFI	TLI	RMSEA	RMSEA_CI	SRMR
CFA	2283.03	215	0.909	0.892	0.064	[0.061, 0.066]	0.059
SEM	1122.90	106	0.94	0.92	0.06	[0.06, 0.067]	0.07

Note: VIF < 10 and tolerance > 0.10 indicate no problematic multicollinearity.

).

The results were as follows: χ²(106) = 1122.90, p < 0.001; CFI = 0.935; TLI = 0.918; RMSEA = 0.064, 90% CI [0.060, 0.067]; and SRMR = 0.075. Although the χ² test was significant—as expected, with N = 2379—the incremental and residual indices indicate an adequate overall fit for a model of this complexity.

Table 5. Descriptive statistics and intercorrelations within analytic sample.

	M	SD	Skew	Kurtosis	1.	2.	3.	4.	5.
1. Stress	4.39	1.16	−0.44	−0.44	-
2. SRL—action	3.46	1.26	0.09	−0.72	−0.21 ***	-
3. Self-efficacy	4.62	1.05	−0.23	−0.25	−0.26 ***	0.41 ***	-
4. Perceived support	3.54	1.40	0.01	−0.82	−0.38 ***	0.33 ***	0.29 ***	-
5. Teacher digital competence	3.21	1.20	0.10	−0.55	−0.25 ***	0.22 ***	0.22 ***	0.40 ***	-

Note: M = mean, SD = standard deviation, skew = skewness, kurtosis = excess kurtosis. Values below the diagonal represent bivariate Pearson correlations. Significance levels: *** = p < 0.001 (bold-faced).

shows that perceived stress was moderately high (M = 4.39 on a seven-point scale), whereas SRL—action sat around the scale midpoint (M = 3.46). Both self-efficacy (M = 4.62) and perceived support (M = 3.54) were slightly above average, while teacher digital competence received the lowest mean rating (M = 3.21). Bivariate correlations followed the expected pattern: all personal and contextual resource variables correlated negatively with stress (|r| = 0.21–0.38, p < 0.001) and positively with one another (up to r = 0.41). These zero-order relations justified entering the full set of predictors into the SEM and provided an initial benchmark for evaluating unique versus shared contributions.

The final SEM accounted for 35% of the variance in perceived stress (R² = 0.35) and 23% of the variance in SRL—action (R² = 0.226), matching the values returned by lavInspect (sem_fit, “r2”).

Direct Effects

As shown in

Table 6. Structural path coefficients (standardized with 95% Wald CI).

Path	β	SE	CI	p
Support ← teacher digital competence	0.40	0.02	[0.36, 0.45]	<0.001 ***
SRL—action ← support	0.27	0.02	[0.24, 0.30]	<0.001 ***
SRL—action ← self-efficacy	0.39	0.03	[0.38, 0.51]	<0.001 ***
Stress ← support	−0.28	0.02	[−0.29, −0.2]	<0.001 ***
Stress ← teacher digital competence	−0.08	0.02	[−0.13, −0.04]	<0.001 ***
Stress ← SRL—action	−0.02	0.03	[−0.08, 0.03]	0.435
Stress ← self-efficacy	−0.31	0.04	[−0.46, −0.31]	<0.001 ***
Stress ← COVID-19 learning disruption	0.13	0.02	[0.1, 0.18]	<0.001 ***
Stress ← gender (male)	−0.32	0.06	[−0.44, −0.20]	<0.001 ***
Indirect_teacher digital competence	−0.11	0.01	[−0.14, −0.09]	<0.001 ***

Note: CI: Confidence interval. Significance levels: *** = p < 0.001 (bold faced).

and controlling for gender and all other predictors, five of the seven structural paths to the stress latent factor reached statistical significance. Academic self-efficacy showed the largest unique association (standardized β = −0.31, SE = 0.04, 95% CI [−0.46, −0.31], p < 0.001), followed closely by perceived support (β = −0.28, SE = 0.02, 95% CI [−0.29, −0.20], p < 0.001). Smaller but reliable effects were observed for teacher digital competence (β = −0.08, SE = 0.02, 95% CI [−0.13, −0.04], p < 0.001) and for COVID-19 learning disruption, which related positively to stress (β = +0.13, SE = 0.02, 95% CI [+0.10, +0.18], p < 0.001). The gender contrast indicated lower stress among male students (β = −0.32, SE = 0.06, 95% CI [−0.37, −0.27], p < 0.001). By contrast, the direct path from SRL—action to stress was small and non-significant (β = −0.02, SE = 0.03, 95% CI [−0.08, +0.03], p = 0.435). Together, the full predictor set accounted for 35% of the variance in the stress construct (R² = 0.35).

3.4. Indirect Pathways

The only sizeable mediation concerned digital pedagogical competence: higher teacher digital competence → greater perceived support → lower stress (β = −0.11, p < 0.001). No significant indirect path emerged for SRL—action.

The latent factors are shown in ovals; single-item observed covariates are shown in rectangles. Stress is regressed on SRL—action, academic self-efficacy, perceived support, teacher digital competence, pandemic-related COVID-19 learning disruption, and gender (male = 1, female = 0). SRL—action is, in turn, predicted by self-efficacy and perceived support, while perceived support is predicted by teacher digital competence. The standardized path coefficients are shown by the arrows; the solid lines indicate p < 0.001; and the dashed line denotes the non-significant direct path from SRL—action to stress (p = 0.44). The double-headed arrows represent residual covariances added for parsimony (see Section 2.3). All latent variables were specified with the unit variance (std.lv = TRUE). The overall model fit was as follows: χ²(106) = 1122.90, CFI = 0.935, TLI = 0.918, RMSEA = 0.064 [0.060, 0.067], and SRMR = 0.075. The model explains 35% of the variance in stress (R² = 0.35) and 23% of the variance in SRL—action (R² = 0.23), supporting R² = 0.16.

4. Discussion

Drawing on transnational stress theory and self-determination theory, we modeled how personal (SRL skills and academic self-efficacy) and contextual (perceived support and teacher digital competence) resources predicted perceived stress in a cohort of 2379 German Abitur candidates who had studied under pandemic conditions for almost two years.

4.1. Model Parsimony and Alternative Specifications

The three-item “learning environment” factor (hardware, internet quality, and a quiet study space) loaded adequately in the CFA, but its inclusion in the SEM reduced the overall fit (CFI ↓ 0.038; SRMR ↑ 0.027; and RMSEA ↑ 0.009) while adding virtually no explanatory power (R²_stress 0.35 → 0.34). Its direct path to stress was small and non-significant (β = −0.07, p = 0.06). For parsimony, we therefore excluded “learning environment” from the structural model and retained the two contextual variables that provided clear incremental predictions—perceived support and teacher digital competence.

Initial models that also contained SRL—planning and SRL—reflection produced high collinearity with SRL—action (r > 0.70) and non-significant direct paths to stress. Consistent with Zimmerman’s emphasis on the execution phase as the most proximal to affective outcomes, we retained only the actional component.

The final structural model explained more than one-third of the variance in stress—substantially more than most single-predictor studies of adolescent well-being during the COVID-19 pandemic (Betthäuser et al., 2023). Below, we discuss three major patterns and their implications.

4.2. Personal Resources: Self-Efficacy Matters More than Enacted SRL

Consistent with H1, students who believed they were effective learners reported markedly lower stress (β = −0.31). This aligns with work showing that self-efficacy buffers the depleting effects of distance learning by sustaining competence appraisals (Holzer et al., 2023). Surprisingly, the execution phase of SRL did not add unique predictive value once self-efficacy and support were controlled (β = −0.02, ns). One interpretation is that there is a conceptual overlap: a large share of the variance in SRL—action was already captured by self-efficacy (r = 0.41) and perceived support (r = 0.33). When the more proximal appraisals are in the model, behavioral tactics such as scheduling or monitoring confer little incremental relief. This pattern resonates with Zimmerman’s hierarchy: strategy use operates through efficacy beliefs rather than direct effects. Nevertheless, SRL—action retained substantive variance (R² = 0.23), suggesting that it may still predict academic outcomes (e.g., grades) that we did not model here.

4.3. Contextual Resources: Instructional Support as a Dual Buffer

H2 received strong support. Perceived learning support emerged as the second-strongest direct predictor of lower stress (β = −0.28) and simultaneously promoted SRL—action (β = 0.27). This dovetails with SDT’s claim that responsive scaffolding fulfils the needs for relatedness and competence, thereby fostering both well-being and autonomous regulation. Teacher digital competence exerted a small direct effect on stress (β = −0.08) but, more importantly, a medium-size indirect effect via support (β ind = −0.11). Put differently, graduates felt less strained when teachers used technology in a way that was experienced as helpful. The mediation underscores that hardware skills alone are insufficient; they must translate into pedagogical practices perceived as supportive.

4.4. Adverse Conditions: Pandemic Disruption and Gender Differences

Consistent with H3, more frequent COVID-related lesson cancellations were associated with higher stress (β = +0.13). This finding dovetails with UK data collected during the 2020 school closures, when uncertainty about syllabus coverage and exam grading was a salient source of anxiety for 14- to 17-year-olds (Mansfield et al., 2021). The male students in our cohort again reported markedly lower stress levels (β = −0.32), echoing the robust gender gap observed across Europe. A systematic review and meta-analysis of nine European countries found a small-to-medium overall rise in depressive symptoms during the pandemic (SMD ≈ 0.31), with girls showing a steeper increase in clinically relevant cases, while boys showed a larger jump in mean symptom scores (Ludwig-Walz et al., 2022). Longitudinal Austrian data likewise document elevated internalizing problems one and two years after the first lockdown (Wenter et al., 2022), situating the average stress level of our German graduates toward the upper end of comparable European samples.

Across contexts, instructional scaffolding appears to buffer distress more reliably than sheer hardware availability. Daily diary data from the United States showed that emotional well-being during virtual schooling was tied more to perceived adult support than to access to devices or quiet workspaces (Cohen et al., 2022), and a Chinese survey study during prolonged closures reached a similar conclusion (Tang et al., 2021). Together with the indirect path from teacher digital competence → perceived support → lower stress in our SEM, these converging findings underscore that technology only protects students when it is translated into responsive pedagogical support.

Because the gender path remained substantial after controlling for self-efficacy and support, additional mechanisms—such as rumination tendencies or differential emotion socialization—warrant exploration in future work.

4.5. Integration with Theory

The findings outline a coherent demand–resource chain: (1) higher teacher digital competence translates into (2) stronger perceived instructional support, which, in turn, enhances (3) students’ self-efficacy and, indirectly, their self-regulated learning behavior—ultimately lowering stress. This sequence aligns with transactional stress theory, which posits that coping hinges on the balance between personal and contextual resources, and it specifies how digital-age pedagogy slots into that balance. Notably, the data indicate that learners’ belief in their own academic competence—rather than the sheer frequency of their strategy use—constitutes a decisive personal buffer against stress.

4.6. Practical Implications

• Invest in pedagogically meaningful digital training. Workshops that show teachers how to translate media competence into a clear structure, timely feedback, and emotional check-ins are likely to yield the greatest well-being dividends;
• Nurture students’ self-efficacy alongside their skills. Brief mastery/experience reflections or verbal persuasion techniques could accompany study skills courses to convert know-how into “can-do” beliefs;
• Prepare contingency plans that minimize uncertainty. Transparent criteria for grade-relevant content and flexible exam schedules may curb disruption-driven stress spikes.

4.7. Limitations and Future Directions

First, the cross-sectional, retrospective design precludes causal inference and may suffer recall bias. Second, self-report measures raise common method bias concerns; however, the large effects of objectively verifiable covariates (e.g., lesson cancellations) mitigate this risk. Third, the measurement model, while adequate, showed a borderline CFI/TLI; future studies should refine the item wording or explore bi-factor solutions. Finally, well-being was operationalized solely as distress. Integrating positive indicators (life satisfaction and vitality) would enable a dual-continuum test of flourishing versus languishing during crises.

Longitudinal designs that track how gains in teacher competence ripple through support perceptions, efficacy beliefs, and stress over time would sharpen the temporal logic implied by our study. Experimental micro-interventions targeting each link (e.g., a two-week teacher feedback protocol) could establish causal leverage and practical effect sizes.

4.8. Conclusions

Among German high-stakes graduates, perceived instructional support and academic self-efficacy were the pivotal levers for tempering pandemic-era stress, outweighing enacted study tactics. Crucially, these levers are malleable: when teachers translate digital skills into supportive interactions and when learners internalize mastery beliefs, the strain of disrupted schooling can be substantially alleviated. Policies that couple digital infrastructure funding with pedagogical and motivational interventions are therefore vital for building resilience against future disruptions—whether viral, climatic, or organizational.