We assessed 8 demographic and clinical variables: academic course, year of academic education, gender, age, ethnicity, parental educational level, psychotherapeutic treatment, and psychopharmacology intake (ie, prescribed drugs).

We assessed the stress level during the last month using a 4-item short form of the validated Perceived Stress Scale (PSS) [19]. The self-report items assess topics such as confidence in handling problems and overcoming difficulties [19]. The short form of the PSS has been reported to have a Cronbach α of .72 [19]. In our sample, the PSS had an acceptable reliability (Cronbach α=.75; coefficient Ω=0.75). Moreover, we assessed the global stress severity during the last month on a zero-to-hundred slider.

We assessed general mental health using the 12-item version of the General Health Questionnaire (GHQ-12) [20]. The GHQ-12 provides a broad indication of mental health and well-being across the spectrum from good-to-poor mental health but does not act as a measure of diagnosis for mental illness. The self-report items assess topics such as concentration, sleep, or happiness (measured on a 4-point scale). In a previous study, the GHQ was found to have a Cronbach α of between .78 and .95 [21]. The GHQ-12 previously had a mean area under the Receiver Operating Characteristic curves of 0.88 [22]. In our sample, the GHQ-12, in the remainder referred to as mental distress, had a good reliability (Cronbach α=.88; coefficient Ω=0.89).

We assessed 7 individual-level RFs that were empirically supported in our systematic review [3]: high distress tolerance, low ruminative reflection, low ruminative brooding, high self-esteem, high cognitive reappraisal, low expressive suppression, and low aggression potential. The content and psychometric details are shown in Table 2. In Table 2, we report Cronbach α values for previous studies, given that this was the reported internal consistency metric. However, for RESIST we report both Cronbach α and coefficient Ω [23], for completeness.

We assessed 5 family-level RFs that were empirically supported in our systematic review [3]: high immediate family support, high extended family support, high family cohesion, high positive parenting, and high parental involvement. The content and psychometric details are shown in Table 2. As items of the family-related scales may be hard to answer for participants if they have spent a large amount of their childhood in care homes or frequently changed foster families, we added some specific instructions to those survey parts (Supplement I in Multimedia Appendix 1).

We assessed 1 community-level resilience factor, which was empirically supported in our systematic review [3]: high social support, here specifically friendship support. The content and psychometric details are shown in Table 2.

Environmental childhood and youth adversity was assessed using an updated version of the 12-item Youth Trauma Scale (YTS) [40]. The 12 self-report items assess topics such as sexual abuse or severe mental or physical illnesses within the family. A complete list of items and assessment details can be found in Supplement II in Multimedia Appendix 1. The original scale was found to have a low internal consistency (Cronbach α=.63-.67). We adapted the questionnaire [40] so that the words parent and sibling were supplemented with significant other, as our participants were at an age at which some might have invested in significant interpersonal relationships outside the family, such as long-term romantic partners. Besides the presence versus absence of adversity, the questionnaire also assessed the severity of questions for which the presence of the respective adversity was positively confirmed. Moreover, for positively confirmed questions, we further assessed the frequency or duration of the adversity as well as during which age bins the adversity had taken place. We did not assess the frequency or duration for two of the adversity items: (1) “Were you separated from one of your parents for more than 1 year?” and (2) “Was either of your parents unemployed for more than 1 year when they wanted to be working?” as those have an inherent time requirement of at least one year. To also assess the potential criminality of parents, siblings, or significant others, we added such an item to the original scale (“Did parents, siblings, or significant others engage in criminal activities severe enough to cause significant stress or worry?”). In summary, we assessed the presence versus absence of the 13 adversities, and if present, the severity, frequency, and the age bin in which the respective adversity experience occurred. Owing to the adaptations we made, we will thoroughly evaluate the psychometrics of the amended scale in a separate manuscript [40].

Furthermore, we assessed the psychological maltreatment and neglect subscales of the Comprehensive Child Maltreatment Scale (CCMS) [41]. The psychological maltreatment subscale consists of 3 items that assess topics such as how frequently the individual was yelled at, ridiculed, or provoked [41]. The neglect subscale consists of 3 items that assess topics such as whether the individual was provided with sufficient warmth from family members, sufficient nutrition, and protection [41]. In a previous study, the CCMS psychological maltreatment subscale was found to have a Cronbach α of .78, and the CCMS neglect subscale had a Cronbach α of .84 [41]. The only adaptation we made to this scale was that we did not assess the questions up to the age of 13 years [41] but up to the age of 18 years. In our sample, the CCMS psychological maltreatment (Cronbach α=.81; coefficient Ω=0.81) and the CCMS neglect subscale both had a good reliability (Cronbach α=.84; coefficient Ω=0.81).

Before starting the content part of the web-based survey, participants were asked to read the information sheet, which contained the major study aims, and to complete a consent form. Before completion of the survey, participants were enabled to download their consent form and the information sheet. Moreover, we provided details on how to get help and support, in case the study would bring up difficult feelings or in case a participant would want to report childhood maltreatment or a crime, in a mental health services information sheet, which could be downloaded from the web-based survey. Further details regarding participant safety considerations are provided in Supplement III in Multimedia Appendix 1.

RESIST was approved by the Cambridge Psychology Research Ethics Committee (PRE.2017.096). RESIST was funded by JF’s Medical Research Council Doctoral Training Grant and by POW’s personal research account.

To include both participants with incomplete and complete data, we used a full information maximum likelihood (FIML) estimator. The use of this estimator has been shown to function well in longitudinal structural equation models. For example, Kievit et al [42] report that “FIML usually performs as well or better than alternative methods.” We decided to treat missing data, instead of performing a complete case analysis, to increase statistical power, reduce standard errors, decrease the probability of biased parameter estimates, and improve generalizability. We tested several potential variables as predictors for missing data patterns, including (1) perceived stress, global stress severity (ie, stress slider), mental distress, psychotherapeutic treatment, and psychopharmacological treatment on all 3 occasions; (2) gender, ethnicity, academic year, age, and childhood adversity on occasion 1; and (3) retrospective subjective stress levels for occasion 2 assessed on occasion 3. This was primarily done to enhance the understanding of missingness in the RESIST sample, and because variables that qualify as predictors for missing data patterns can, in conjunction with an FIML estimator, be used as auxiliary variables and thereby potentially enhance the estimation precision.

We conducted a series of latent growth models (LGMs) to explore the mean change trajectory of perceived stress and mental distress over the 3 occasions. We fixed the slope loading of occasion 2 to 1, expecting this occasion to have the highest level of perceived stress (and mental distress), and the slope loading of occasion 3 to 0, expecting this occasion to have a lower level of perceived stress (and mental distress) than occasion 2. Hence, the slope loading of occasion 1 was freely estimated and provides an indication of where the (scaled) mean level lies in comparison to occasion 2 (fixed to 1) and occasion 3 (fixed to 0). We conducted the LGMs with invariant residual variances for the 3 occasions (M1). To test whether our latent growth model is significantly different from a no-change trajectory (ie, no change in overall mean levels), we estimated an additional LGM (M2) with the latent slope mean set to 0, modeling no overall change. In sum, we mainly used the models to identify the change trajectories of perceived stress and mental distress and to test whether these trajectories fit better than a no-change trajectory. For completeness, we refitted the LGMs with freely estimated residual variances (Supplement IV in Multimedia Appendix 1).

We conducted a series of bivariate latent change score models (BLCSMs; as described by Kievit et al [42]) to investigate the change of perceived stress and mental distress in conjunction. More specifically, we conducted 3 BLCSMs, 1 for each pairwise combination of the 3 occasions, to allow for direct comparisons without estimating overly complex models. To enable the computation of the models, we used the standard BLCSM estimation (additional details are provided by Kievit et al [42]). We then investigated the relationship between perceived stress and mental distress on the earlier occasion as well as the relationship between the change scores of perceived stress and mental distress on the later occasion. Moreover, we investigated the autoregressive paths of perceived stress and mental distress with their respective change scores as well as the potentially mutualistic relationship between perceived stress and mental distress, that is, perceived stress predicting change in mental distress and mental distress predicting change in perceived stress.

All analyses were conducted in R version 3.5.1 (The R Foundation) [43], mainly using the packages lavaan [44] and semTools [45]. The analysis script can be found on the Open Science Framework [46] and the anonymized data used for the analyses in this manuscript have been uploaded to the Cambridge Data Repository [47].

Students were approximately uniformly distributed over all 6 academic years, with percentages ranging from 12.4% (56/451) to 20.6% (93/451) per year. A total of 57.4% (259/451) of the students were female (1.3% [6/451] preferred not to answer) and 58.3% (263/451) were White. Most students were between 18 and 23 years of age and had parents with higher education after secondary school. About 13.5% (61/451) of the students received psychotherapeutic treatment and 10.9% (49/451) received psychopharmaceutic treatment in the 6 months before occasion 1. Table 3 contains the descriptive statistics for all students who took part on occasion 1. Supplement V in Multimedia Appendix 1 contains the same table, with the inclusion of all students who provided data for at least two occasions (n=324).

The mean levels suggest that perceived stress and mental distress increased from the time before the exams to the exam period and decreased after the exams to a lower level than before the exams. Table 4 depicts perceived stress and mental distress levels on the 3 occasions for participants who took part in the respective occasion. The corresponding table with only those participants who provided data for at least two occasions can be found in Supplement V in Multimedia Appendix 1.

The LGM showed that, on average, students experience most perceived stress during exams (occasion 2: slope loading fixed to 1) and least perceived stress after the exams (occasion 3: slope loading fixed to 0); before the exams, they experienced more perceived stress than after the exams, but less than during the exams (occasion 1: estimated slope loading=0.29; Table 5). The mean level trajectory is shown in Figure 2 (left panel). We further found that the model estimating the change trajectory of perceived stress fits significantly better than the no-change model (χ²₁=72.4; P<.001; fit indices are presented in Table 6), indicating that perceived stress changed significantly over the 3 occasions. Additional post-hoc analyses confirmed that the mean levels differed significantly between all three occasions.

The LGM showed that, on average, students experience most mental distress during exams (occasion 2, slope loading fixed to 1) and least mental distress after exams (occasion 3, slope loading fixed to 0); before the exams, they experienced more mental distress than after the exams, but less than during the exams (occasion 1, estimated slope loading=0.60; Table 5). The mean level trajectory is shown in Figure 2 (right panel). We further found that the model estimating the change trajectory of mental distress fits significantly better than the no-change model (χ²₁=58.8; P<.001; fit indices are presented in Table 6), indicating that mental distress changed significantly over the 3 occasions. Additional post-hoc analyses confirmed that the mean levels differed significantly between all three occasions.

The BLCSM showed that perceived stress and mental distress on occasion 1 are significantly positively associated, which indicates that students with higher perceived stress report on average higher mental distress (Figure 3, upper-left panel; for exact coefficients, see Supplement VIII in Multimedia Appendix 1). Similarly, changes in perceived stress and mental distress from occasion 1 to 2 are significantly positively associated, which indicates that students with more increase in perceived stress report on average also a greater increase in mental distress. The BLCSM revealed a significant and negative autoregressive effect for perceived stress and its change score, which indicates that higher perceived stress on occasion 1 results, on average, in less increase in perceived stress from occasion 1 to 2. Equally, mental distress had a significant and negative autoregressive effect with its change score, indicating that higher mental distress on occasion 1 results, on average, in less increase in mental distress from occasion 1 to 2. Perceived stress was not significantly associated with the change in mental distress from occasion 1 to 2. However, mental distress was significantly positively associated with the change in perceived stress from occasion 1 to 2, which indicates that higher mental distress on occasion 1 results on average a greater increase in perceived stress from occasion 1 to 2.

The BLCSM showed that changes in perceived stress and mental distress from occasion 1 to 3 are significantly positively associated, which indicates that individuals with a greater decrease in perceived stress report, on average, also a greater decrease in mental distress (Figure 3, upper-right panel; for exact coefficients, see Supplement VIII in Multimedia Appendix 1). Moreover, the BLCSM revealed a significant and negative autoregressive effect for perceived stress and its change score, which indicates that higher perceived stress on occasion 1 results, on average, in a greater decrease in perceived stress from occasion 1 to 3. Equally, mental distress had a significant and negative autoregressive effect with its change score, indicating that higher mental distress on occasion 1 results, on average, in a greater decrease in mental distress from occasion 1 to 3. Mental distress was not significantly associated with change in perceived stress from occasion 1 to 3. Perceived stress was not significantly associated with change in mental distress from occasion 1 to 3; however, the P value was marginal (β=.15; P=.06). If one would opt to interpret the directionality of the effect, it would suggest that, on average, higher perceived stress is marginally associated with less decrease in mental distress from occasion 1 to 3. However, given that the effect was not significant, we suggest to err on the side of caution and shall not interpret it.

The BLCSM showed that perceived stress and mental distress on occasion 2 are significantly positively associated, which indicates that individuals with higher perceived stress report on average higher mental distress (Figure 3, lower-left panel; for exact coefficients, see Supplement VIII in Multimedia Appendix 1). Similarly, changes in perceived stress and mental distress from occasion 2 to 3 are significantly positively associated, which indicates that individuals with a greater decrease in perceived stress report, on average, also a greater decrease in mental distress. Moreover, the BLCSM revealed a significant and negative autoregressive effect for perceived stress and its change score, which indicates that higher perceived stress on occasion 2 results on average in a greater decrease in perceived stress from occasion 2 to 3. Equally, mental distress had a significant and negative autoregressive effect with its change score, indicating that higher mental distress on occasion 2 results on average in a greater decrease in mental distress from occasion 2 to 3. Mental distress was not significantly associated with change in perceived stress from occasion 2 to 3. However, perceived stress was significantly positively associated with change in mental distress from occasion 2 to 3, which indicates that higher perceived stress results on average in less decrease in mental distress from occasion 2 to 3.

Future analyses on the RESIST data are primarily set out to shed light on which RFs lend themselves best as prevention targets (before the stressor) and which as treatment targets (at times of stress) for the mitigation of mental health problems that are triggered or accelerated by natural exam stress. Therefore, the RESIST study may lay the foundations necessary to inform student support services, as well as mental health services, as well as resilience and transdiagnostic mental health theory.

We aim to publish all articles that are based on RESIST data in peer-reviewed journals, ideally under an open access agreement. Alongside the manuscripts, we aim to release the related and anonymized data on the Cambridge Data Repository.