Strict nonselective quarantine and the prediction of suicidal thinking, hopelessness, and substance use in undergraduates: A longitudinal natural experiment 1000 km from a COVID-19 outbreak Authors: Richard J. Linscott Published: 2025 ABSTRACT BackgroundThe costs of quarantine on mental health are difficult to quantify and disentangle from primary and secondary effects of infectious disease. Selective and nonselective quarantines were widely used public health measures during the COVID-19 pandemic. Opportunity arose to test by longitudinal natural experiment the hypothesis that nonselective quarantine predicts elevated suicidal thinking, hopelessness, and drug use in a sample that was almost certainly not concurrently exposed to the COVID-19 virus.MethodsUndergraduates (18- to 22-years) studying psychology in Dunedin, New Zealand, completed a prospective online survey (10 assessment waves, at 3-day intervals) of passive and active suicidal ideation (past 3 days), current hopelessness, and same-day recreational drug use. During the study, detection of one case of COVID-19 in a community that was 1060 km from the study site triggered a strict nationwide 14-day quarantine, implemented with fewer than 6-hours’ notice. Bayesian growth modelling was used to quantify direct and mediated (indirect) effects of the quarantine period.ResultsOf 283 completing baseline assessment, 38.5% (95% CI, 32.8, 44.2) reported past-month ideation and 35.0% (95% CI, 29.4, 40.5) past-fortnight drug use. Retention during follow-up was 89% (n = 253) to 95% (n = 269). At each follow-up, 12.3% to 19.9% reported interval ideation and 1.9% to 4.7% reported drug use. Quarantine predicted a 0.13 standard deviation increase in hopelessness. Quarantine did not directly predict ideation or drug use. However, quarantine had an indirect effect on ideation that was mediated by hopelessness. Among those with average hopelessness, predicted rates of ideation in and out of quarantine were 17.9% and 16.7%, respectively.ConclusionsA nationwide nonselective quarantine predicted increased hopelessness and, in turn, hopelessness predicted small increases in suicidal thinking among emerging adults. The observed effects are unlikely to be attributable to active COVID-19 infection. Public health interventions that ignore developmental, resource, and mental health differences within populations may harm those who are most vulnerable within communities.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40359-025-03890-8. Introduction Many countries have employed quarantine as a public health intervention to reduce COVID-19 infection. Whereas quarantine is typically applied to those exposed to risk of infection [1], nonselective quarantines (e.g., lockdown, stay-at-home orders) restrict whole communities regardless of exposure. During the COVID-19 pandemic, the health benefits of quarantine were evident from changes in infection and mortality rates. Conversely, the health costs of nonselective quarantine—including untoward psychological effects such as suicidal thinking and behaviour (STB)—are difficult to quantify and distinguish from effects of infection, the pandemic, and other interventions. I report a prospective natural experiment on outcomes associated with a strict, nonselective whole-of-country quarantine on suicidal thinking in a sample almost certainly not concurrently exposed to the COVID-19 virus. Early in the COVID-19 pandemic, concern was raised that quarantine may lead to increased STB. Evidence from other epidemics suggests that quarantine increases risk for suicide [1–6], that risk changes remain after quarantine [1], and that increased risk may be irreversible [2]. The effects of public health measures, if any, occur against the background COVID-19 pandemic STB rates, evidence on which is mixed [5, 7–28]. Several years on, the effects of public health responses on background rates remains unclear [1, 29–31]. Across-study comparisons of data from before and during the pandemic suggest declining rates of self-harm and STB in younger people [31]. In contrast, case-control and uncontrolled studies suggest quarantine and isolation have adverse effects on anger, anxiety, depression, stress-related disorders, and substance use, with increased rates of stress-related disorders particularly among younger people [29, 30]. Thus, effects of the pandemic on STB may derive from quarantine itself [29, 32]. In 2020 and 2021, periods of strict nonselective quarantine were key to New Zealand’s COVID-19 public health response. The government closed borders to non-residents, restricted return to New Zealand by residents and citizens, and implemented widespread screening. As more contagious variants emerged, quarantines were implemented more rapidly. In public statements, leaders and experts emphasised the threat of COVID-19. People were asked to inform Police of others’ noncompliance with quarantine. Together, these strategies led to the elimination of COVID-19 in the community by mid 2020. By coincidence, a strict nation-wide 14-day quarantine occurred during a prospective study of suicidal thinking among undergraduates in 2021. Implemented with fewer than 6 hours’ notice, the quarantine was imposed when COVID-19 was almost certainly not circulating among study participants. The coincidence gave opportunity to examine in a natural experiment whether quarantine, independently of COVID-19 infection, was associated with suicidal thinking, hopelessness, and recreational drug use over 10 waves of assessment. At 18.6 deaths per 100 000, the suicide rate among emerging adults in New Zealand [33] is among the highest in the world [34] and among economically similar countries [35]. In planned analyses, I tested the hypothesis that quarantine directly predicts increased concurrent suicidal thinking, hopelessness, and drug use. In post hoc analyses, I tested the hypothesis that quarantine indirectly predicts increased concurrent suicidal thinking via increases in hopelessness. Materials and methods Participants and procedure Participants were undergraduates in introductory psychology courses at the University of Otago who elected to volunteer for research participation. There were no other inclusion or exclusion criteria. Informed consent was recorded electronically. Participants received a small amount of course credit in recognition of their participation. Following baseline assessment, participants completed nine follow-up assessments at three-day intervals. Participants completed assessments in their own time via the Qualtrics survey platform. Participants were e-mailed URLs for follow-up assessments at 4:00 pm on assessment days and, if required, a reminder 24 h later. The times and dates of assessment completions were recorded. Study period and quarantine Participant recruitment began on 21 July 2021, when there was no evidence of COVID-19 circulating in the country despite high levels of screening [36], and continued until 16 September 2021. Volunteers could begin participation on any date in that period. The last group of participants completed their tenth assessments on 21 October 2021. At 23:59 on 17 August 2021, a whole-of-country quarantine was implemented after the detection, the day before, of a single case of COVID-19 in a community 1068 km from the study site. The case, a 58-year-old male, was presumed to have been infectious in that community for about six days. The government announced the quarantine via news media at 18:15 on 17 August, and via an emergency alert to all cellular devices in the country at 19:16. Within minutes of the emergency alert, the university informed students by e-mail that classes on 18 August were cancelled and that from 19 August, all classes would be taught online. Until that day, classes had been in-person. Quarantine continued until 23:59 on 31 August. During quarantine, all but essential services were closed; the public were to remain in homes or other places of residence. On 1 September, quarantine restrictions were eased but not removed completely; learning remained online. Measures Suicidal ideation was assessed using the Columbia Suicide Severity Rating Scale (C-SSRS) [37]. Using C-SSRS content addressing recent passive and active ideation, four self-report items were composed (“In the last [time-frame], have you wished you were dead?”, “… you could go to sleep and not wake up?”, and “… you were not alive anymore?” for passive ideation, and “In the last [time frame], have you actually had any thoughts of killing yourself?”). Response options were Yes-No, as in the C-SSRS interview [37]. At baseline, the reference time-frame was the past month; at follow-ups, the past three days. For each assessment wave, the dependent measure was a single binary score reflecting the presence or absence of any suicidal ideation. An item sum score, derived from binary scores from each of the four constituent items at the baseline, gave an internal consistency estimate (Kuder-Richardson-20) of rKR = 0.87. Evidence from psychometric studies of the C-SSRS interview with adolescent and adult clinical groups shows strong inter-rater reliabilities for C-SSRS total scores as well as for ideation items [38], and of high internal consistency [39]. C-SSRS scores are strongly correlated with concurrent measures of ideation [39]. State hopelessness was measured at baseline and follow-ups using the 2-item Brief-H-Neg [40]. The dependent measure was the sum of ratings (range 0 = minimum hopelessness, 8 = maximum). In large adult medical samples, the Brief-H-Neg items are correlated r =.53 [41], have high internal consistency (α = 0.80) [40], and the rating sum correlates r =.93 with the Beck Hopelessness Scale [40]. Alcohol, nicotine, and recreational drug use were assessed using three items from the DSM-5 Self-Rated Level 1 Cross-Cutting Symptom Measure—Adult [42]. For follow-up assessments, the reference time-frame was “today” with Yes-No response options. The dependent measure was a binary score for the presence of drug use. Participants completed a range of other measures among which were distributed six inattention items (e.g., “to show you are paying attention, select the response that means agree”). Inattentive responders were defined as those giving one or more incorrect responses on these items. Measures are described more fully in the supplementary material. Statistical analysis Data from inattentive responders and participants over 22 years were removed. Bayesian growth models, without priors, were tested using Mplus 8.8. Bayesian modelling proceeded in steps. First, exploratory intercept-only, linear, and non-linear (up to quartic) growth models were fitted to outcomes obtained at the nine follow-up waves. (Baseline data had longer reference time-frames.) Second, using the best fitting growth model, quarantine (present = 1, not present = 0) was entered as a time-varying covariate, predicting observed outcomes. Third, to obtain estimates adjusted for sex and minority ethnicity (Māori or Pacific Peoples), these were entered as dichotomous time-invariant covariates (male = 1, minority = 1) predicting intercept and linear slope growth factors. No adjustment was made for attributes that likely have strong antecedent roles or that are concurrent outcomes or corollaries of suicidal ideation (e.g., mental health) because such attributes are not theoretically independent [43]. Supplementary materials include technical details on the Bayesian growth modelling. Time weights were transformed to range from − 1 = Wave 1 to 1 = Wave 9, in steps of 0.25, to allow easier testing of nonlinear growth. As implemented in Mplus, individually-varying times of outcome observation cannot be used in Bayesian growth modelling. Therefore, time of observation was normalised to wave time, however, quarantine status for each outcome was determined from actual participation dates. To retain participants who completed fewer than nine assessment waves, missing participation dates (from which quarantine, the time-varying exogenous predictor, was determined) were prorated by adding the median lag (days) for the follow-up wave to participant’s Wave 0 date. For outcomes and other covariates, Bayes estimation was used to handle missing data. At asymptote, Bayes estimation and maximum likelihood estimation (in frequentist analyses) yield the same results [44]. For dichotomous outcomes, model fit was evaluated principally using the posterior predictive p-value (PPP), which is derived using continuous latent response variables [44]. If the PPP < 0.05, the model does not fit; PPP = ~ 0.5 is optimal. The counts of significant (p <.05) observed vs. estimated discrepancies in bivariate outcome probabilities, univariate probabilities, and the 20 most frequent response patterns were also considered. For the continuous outcome, hopelessness, model fit was evaluated using the PPP, change in the deviance information criterion (DIC; lower is better), the comparative fit (CFI ≥ 0.95) and Tucker-Lewis indices (TLI ≥ 0.95), and root mean square error of approximation (RMSEA < 0.06). Bayesian posterior parameter trace and autocorrelation plots were evaluated visually, and potential scale reduction (convergence) was recorded. Model-based estimates of ideation and drug use probabilities across the nine assessment waves, and change associated with quarantine, were approximated [45]. Results Those completing baseline assessment (n = 331) were aged 18- to 38-years (M = 20.5 years, SD = 2.8) and 13.6% assigned male at birth. After removing data from 30 inattentive responders and 18 participants aged over 22-years, the analysis sample comprised 283 participants (age M = 19.8 years, SD = 1.0; 10.6% male sex). These were predominantly New Zealand or other European (93%) with smaller numbers of Māori (12%), Pacific Peoples (3%), or other ethnicities (18%). Among those aged 18- to 22-years, those excluded for inattention did not differ in age (t < 1) from those retained in analyses. Comparison of all those excluded (n = 48) with those retained in the analyses indicated no differences in assigned sex at birth or ethnic identity (all p >.15). Of those completing the baseline assessment, follow-up response rates fluctuated from 89% to 95% (Table 1). Results of Little’s MCAR test, used to determine whether nonresponse was random across waves, indicated response patterns were consistent with data missing completely at random (MCAR). Figure 1 shows outcomes on ideation, hopelessness, and drug use at each wave. At baseline, 38.5% reported past month suicidal ideation. Three participants completed no follow-up assessments, leaving n = 280 for growth modelling. Table 2 shows key bivariate correlations. Table 1Follow-up assessment retention and latencies relative to baseline (n = 283 at baseline)Follow-up assessment wave123456789 n 269260262269260262253264265Response rate (cf. baseline)0.950.920.930.950.920.930.890.930.94Latency (M days)3.96.79.812.615.918.621.624.727.9Latency (SD days)1.11.01.00.81.20.90.91.11.7Latency range (days)3, 96, 149, 1612, 1615, 2318, 2321, 2724, 3327, 39 Table 1Follow-up assessment retention and latencies relative to baseline (n = 283 at baseline)Follow-up assessment wave123456789 n 269260262269260262253264265Response rate (cf. baseline)0.950.920.930.950.920.930.890.930.94Latency (M days)3.96.79.812.615.918.621.624.727.9Latency (SD days)1.11.01.00.81.20.90.91.11.7Latency range (days)3, 96, 149, 1612, 1615, 2318, 2321, 2724, 3327, 39 Follow-up assessment retention and latencies relative to baseline (n = 283 at baseline) Fig. 1Observed suicidal ideation, hopelessness, and substance use. Legend: (A) Proportions reporting passive or active ideation and rated intensity of ideation experiences. Baseline (Wave 0) = past month; follow-up (Waves 1 to 9) = past three days. Intensity scale range = 2 (fleeting, once) to 10 (persistent or continuous, many times each day). (B) Mean Brief H Neg total score. Total ranges from 0 (least hopelessness) to 8 (most hopelessness). (C) Proportions reporting alcohol, nicotine, and drug use. Baseline = past two weeks; follow-up = “so far today”. Alcohol threshold = ≥ 4 drinks; nicotine = any; drugs = any. For all plots, error bars are 95% confidence intervals. Baseline n = 283; follow-up n = 252 to 267 Observed suicidal ideation, hopelessness, and substance use. Legend: (A) Proportions reporting passive or active ideation and rated intensity of ideation experiences. Baseline (Wave 0) = past month; follow-up (Waves 1 to 9) = past three days. Intensity scale range = 2 (fleeting, once) to 10 (persistent or continuous, many times each day). (B) Mean Brief H Neg total score. Total ranges from 0 (least hopelessness) to 8 (most hopelessness). (C) Proportions reporting alcohol, nicotine, and drug use. Baseline = past two weeks; follow-up = “so far today”. Alcohol threshold = ≥ 4 drinks; nicotine = any; drugs = any. For all plots, error bars are 95% confidence intervals. Baseline n = 283; follow-up n = 252 to 267 Table 2Matrix of bivariate correlations among key variables1234561. Suicidal ideation, W02. Suicidal ideation, W1-90.75(0.67, 0.81)0.74(0.61, 0.86)3. Drug use, W00.00− 0.01(−0.21, 0.10)4. Drug use, W1-90.07(−0.45, 0.45)− 0.01(−0.37, 0.41)0.54(0.38, 0.70)0.53(0.08, 0.96)5. Hopelessness, W00.380.43(0.37, 0.50)0.070.02(−0.13, 0.23)6. Hopelessness, W1-90.39(0.34, 0.45)0.43(0.26, 0.60)0.03(−0.07, 0.10)0.04(−0.22, 0.40)0.63(0.59, 0.69)0.77(0.61, 0.77)7. Male− 0.16− 0.20(−0.33, − 0.08)0.090.12(−0.07, 0.30)− 0.17− 0.12(−0.23, 0.00)8. Minority0.020.03(−0.08, 0.12)0.060.18(−0.12, 0.37)− 0.05− 0.01(−0.10, 0.11)Coefficients are observed tetrachoric, polyserial, biserial, and Pearson correlation coefficients; cells for W1-9 show average (minimum, maximum) correlation coefficients across the 9 waves; n = 234 to 283. W0 – Wave 0 (baseline), W1-9 – Wave 1 to Wave 9 Table 2Matrix of bivariate correlations among key variables1234561. Suicidal ideation, W02. Suicidal ideation, W1-90.75(0.67, 0.81)0.74(0.61, 0.86)3. Drug use, W00.00− 0.01(−0.21, 0.10)4. Drug use, W1-90.07(−0.45, 0.45)− 0.01(−0.37, 0.41)0.54(0.38, 0.70)0.53(0.08, 0.96)5. Hopelessness, W00.380.43(0.37, 0.50)0.070.02(−0.13, 0.23)6. Hopelessness, W1-90.39(0.34, 0.45)0.43(0.26, 0.60)0.03(−0.07, 0.10)0.04(−0.22, 0.40)0.63(0.59, 0.69)0.77(0.61, 0.77)7. Male− 0.16− 0.20(−0.33, − 0.08)0.090.12(−0.07, 0.30)− 0.17− 0.12(−0.23, 0.00)8. Minority0.020.03(−0.08, 0.12)0.060.18(−0.12, 0.37)− 0.05− 0.01(−0.10, 0.11)Coefficients are observed tetrachoric, polyserial, biserial, and Pearson correlation coefficients; cells for W1-9 show average (minimum, maximum) correlation coefficients across the 9 waves; n = 234 to 283. W0 – Wave 0 (baseline), W1-9 – Wave 1 to Wave 9 Matrix of bivariate correlations among key variables 0.75 (0.67, 0.81) 0.74 (0.61, 0.86) − 0.01 (−0.21, 0.10) 0.07 (−0.45, 0.45) − 0.01 (−0.37, 0.41) 0.54 (0.38, 0.70) 0.53 (0.08, 0.96) 0.43 (0.37, 0.50) 0.02 (−0.13, 0.23) 0.39 (0.34, 0.45) 0.43 (0.26, 0.60) 0.03 (−0.07, 0.10) 0.04 (−0.22, 0.40) 0.63 (0.59, 0.69) 0.77 (0.61, 0.77) − 0.20 (−0.33, − 0.08) 0.12 (−0.07, 0.30) − 0.12 (−0.23, 0.00) 0.03 (−0.08, 0.12) 0.18 (−0.12, 0.37) − 0.01 (−0.10, 0.11) Coefficients are observed tetrachoric, polyserial, biserial, and Pearson correlation coefficients; cells for W1-9 show average (minimum, maximum) correlation coefficients across the 9 waves; n = 234 to 283. W0 – Wave 0 (baseline), W1-9 – Wave 1 to Wave 9 Growth models and direct effects of quarantine The supplementary materials (including Tables S1-S7) show full results for linear and non-linear growth models. For all outcomes, non-linear models were superior to linear models. For ideation and drug use, quadratic models performed the best (Table S1), and for hopelessness, the quadratic model was no worse than the cubic model when adjustment was included for sex and ethnicity (Table S2). Therefore, quadratic models were used to estimate the relationship of quarantine with outcomes. Mean linear growth coefficients showed a declining trajectory for ideation, but did not differ from zero for drug use and hopelessness (Tables S3 & S4). For hopelessness, quarantine predicted a small but significant quarter-point increase in the Brief-H-Neg total (M = 0.27, 95% credibility interval = 0.11, 0.41; effect size d = 0.13; Fig. 2, Table S4). In contrast, the coefficients for the direct influences of quarantine on ideation and drug use were not significant (p =.285 and p =.211, respectively; Table S3). The probit coefficient for ideation (−0.08, 95% credibility interval = −0.36, 0.20) corresponded to a 0.4 to 0.5%-point decrease in the mean predicted probability of ideation, from the background rate of 17.1% (range across assessment waves = 14.6% to 18.9%; Fig. 2), dropping non-significantly to 16.6% under quarantine (range across assessment waves = 14.2% to 18.4%). The probit coefficient for drug use (0.17, 95% credibility interval = −0.26, 0.59) equated to a 0.9 to 1.7%-point increase in the probability of drug use, increasing non-significantly from the background rate of 3.8% (range = 3.0% to 5.2%) to 5.2% (4.1% to 6.9%). Adding quarantine did not affect the fit of ideation or drug use models (Table 3), but improved fit of the hopelessness model (Table 4). Fig. 2Modelled growth in suicidal ideation, hopelessness, and substance use. Legend: Observed outcomes (plot symbol, circle), predicted quadratic growth trajectories (solid line), and model-based estimates of trajectories with and without quarantine (dashed = strict quarantine in place, dotted = no strict quarantine) on (A) suicidal ideation, (B) hopelessness ratings, and (C) drug use across nine follow-up assessments, all with n = 280 Modelled growth in suicidal ideation, hopelessness, and substance use. Legend: Observed outcomes (plot symbol, circle), predicted quadratic growth trajectories (solid line), and model-based estimates of trajectories with and without quarantine (dashed = strict quarantine in place, dotted = no strict quarantine) on (A) suicidal ideation, (B) hopelessness ratings, and (C) drug use across nine follow-up assessments, all with n = 280 Table 3Model convergence and fit indices for suicidal ideation and drug use (n = 280)Model (free parameters)Suicidal ideationDrug usePPPNon-fitting residualsn (%)PPPNon-fitting residualsn (%)Quadratic (9)0.51400.5034 (2.3)Quadratic adjusted (13)0.56900.4895 (2.9)Quadratic quarantine (10)0.43600.4754 (2.3)Quadratic quarantine adjusted (14)0.45000.5324 (2.3)PPP posterior predictive p-value, (%) percentage of 173 standardized residuals in the model, DNC did not converge, adj adjusted for male sex and minority status. See Table S1 for full fit statistics Table 3Model convergence and fit indices for suicidal ideation and drug use (n = 280)Model (free parameters)Suicidal ideationDrug usePPPNon-fitting residualsn (%)PPPNon-fitting residualsn (%)Quadratic (9)0.51400.5034 (2.3)Quadratic adjusted (13)0.56900.4895 (2.9)Quadratic quarantine (10)0.43600.4754 (2.3)Quadratic quarantine adjusted (14)0.45000.5324 (2.3)PPP posterior predictive p-value, (%) percentage of 173 standardized residuals in the model, DNC did not converge, adj adjusted for male sex and minority status. See Table S1 for full fit statistics Model convergence and fit indices for suicidal ideation and drug use (n = 280) PPP posterior predictive p-value, (%) percentage of 173 standardized residuals in the model, DNC did not converge, adj adjusted for male sex and minority status. See Table S1 for full fit statistics Table 4Model convergence and fit indices for hopelessness (n = 280)Model (free parameters)PPPDICRMSEACFITLIQuadratic (Q) (10)0.0358220.040.0360.9920.994Q adjusted (14)0.1538223.770.0220.9960.996Q quarantine (11)0.2678209.040.0070.9990.999Q quarantine adjusted (15)0.4068212.390.0001.001.00PPP posterior predictive p-value, DIC deviance information criterion, RMSEA root mean square error of approximation, CFI comparative fit index, TLI Tucker-Lewis index, quad quadratic, adj adjusted for male sex and minority status. See Table S2 for full fit statistics Table 4Model convergence and fit indices for hopelessness (n = 280)Model (free parameters)PPPDICRMSEACFITLIQuadratic (Q) (10)0.0358220.040.0360.9920.994Q adjusted (14)0.1538223.770.0220.9960.996Q quarantine (11)0.2678209.040.0070.9990.999Q quarantine adjusted (15)0.4068212.390.0001.001.00PPP posterior predictive p-value, DIC deviance information criterion, RMSEA root mean square error of approximation, CFI comparative fit index, TLI Tucker-Lewis index, quad quadratic, adj adjusted for male sex and minority status. See Table S2 for full fit statistics Model convergence and fit indices for hopelessness (n = 280) PPP posterior predictive p-value, DIC deviance information criterion, RMSEA root mean square error of approximation, CFI comparative fit index, TLI Tucker-Lewis index, quad quadratic, adj adjusted for male sex and minority status. See Table S2 for full fit statistics Mediation model for the indirect effect of quarantine on suicidal ideation The absence of direct effects does not preclude indirect effects [46]. Therefore, I tested whether quarantine predicted ideation indirectly through hopelessness. Modelling ran incrementally (Supplementary Figure): Model 1 comprised simultaneous quadratic models for ideation and hopelessness, with linear slopes of one regressed onto intercepts of the other. In Model 2, observed ideation and hopelessness were regressed onto quarantine. In Model 3, observed ideation was regressed onto observed hopelessness, as a time-varying predictor. In Model 4, direct paths from quarantine to ideation were removed. Models 3 and 4 were tested without and with adjustment for sex and ethnicity. Tables S5 to S7 show model convergence and fit, coefficients, and correlations among growth factors, respectively. Model 3 with adjustment and Model 4 without adjustment were the best fitting. As with planned analyses, quarantine directly predicted hopelessness (standardized b = 0.12, 95% credibility interval = 0.05, 0.19) but not ideation (probit = −0.23, 95% credibility interval = −0.59, 0.11). Hopelessness predicted suicidal ideation (probit = 0.42, p <.001, 95% credibility interval = 0.32, 0.54). At the mean level of hopelessness, the approximate mean rate of suicidal ideation was 16.7% (range across waves = 11.6% to 22.2%), at 1 SD below average hopelessness, the mean rate was 9.4% (6.7% to 14.2% across waves), and at 1 SD above average hopelessness, the mean rate was 27.0% (17.9% to 32.7%). Quarantine had a significant indirect effect on suicidal ideation (effect = 0.11, p <.001, 95% credibility interval = 0.04, 0.19; Table S6); at the mean level of hopelessness, the mean rate of suicidal ideation was 1.2% greater during quarantine periods (M = 17.9%, with 12.3% to 23.4% across waves), and at 1 SD above average hopelessness, the mean rate of ideation was 1.5% greater during quarantine (M = 28.5%, with 18.8% to 34.3% across waves). Removal of the direct path from quarantine to ideation had no meaningful effect on results (Tables S5 & S6). Discussion A strict, nationwide quarantine indirectly predicted greater odds of suicidal ideation in an unselected sample of 18- to 22-year-old undergraduates. Hopelessness played an integral mediating role: quarantine directly predicted increased hopelessness that, in turn, predicted increased ideation. Contrary to expectation, there was no evidence of direct effects of quarantine on ideation or drug use. Others have reported pandemic-related elevations in ideation [11, 15, 23, 29], attempts [20, 32], completed suicide [19, 24], and STB-related risk factors [21, 29, 47]. These effects on STB appear nonuniform across age, with greater effects in older adolescents than others [29], and across clinical vs. non-clinical populations [1, 31]. Quarantine poses developmentally-specific challenges for emerging adults [31]. Quarantine, school closures, and exposure to COVID-19 are associated with poorer outcomes across a range of other measures, including depression, anxiety, stress, sleep disturbance [29, 30, 48, 49]. This study, which shows quarantine is associated with increased ideation, adds to understanding of the possible effects of quarantine in two important ways: It relied on a strong research design, and it relied on a sample of volunteers for whom quarantine has a homogeneous threat value. In respect of the first of these, most available evidence on the outcomes of quarantine derives from uncontrolled cross-sectional work, comparisons with normative data, between-group comparisons, or simple before vs. during quarantine comparisons. These designs cannot exclude many competing explanations for observed relationships [50, 51]. In contrast, in this study, quarantine was a discrete 14-day epoch that fell 29 days into a 93-day assessment period. Given that design—given quarantine was coded as a discrete 14-day period—alternative explanations that are not constrained within that same epoch can be ruled out (e.g., the pandemic itself, prior infection or long-COVID, psychological sequelae of infection, secondary effects of the pandemic). Additionally, it is unlikely that there was concurrent infection in the sample because the sample was from a community in which COVID-19 was not known to be present despite high screening rates. Equally, the effect of quarantine on ideation was not direct. Whereas hopelessness served a mediating role in a manner consistent with suicide theories [52, 53], other unmeasured individual differences and corollaries of the pandemic, for example, anxiety and encumbered access to health services [54], or other pandemic stressors [16], may have contributed to the observed effect of quarantine. Secondly, as research tools, umbrella reviews and meta-analyses make important inductive contributions to knowledge about the effects or risks associated with quarantine. However, these tools can obscure legitimate or genuine effects and risks because of the (often statistical) assumption that the underlying population is homogeneous when it is not. For example, important sources of heterogeneity go unmeasured or are under-represented in umbrella reviews and meta-analyses [31]. This study adds to understanding of what can occur within distinct communities. Whereas undergraduate samples are often viewed unfavourably, a particular advantage of these is their relative homogeneity of immediate goals and activity. That is, undergraduates commit to investing significantly into intensive learning activity in the hope of achieving desired vocational outcomes or future socioeconomic benefit. The homogeneity of the threat value of quarantine rests on the disruption of these common goals and activities. In future research, investigators could seek to understand these and other nuances in the effects of quarantine associated with populations, development, and the parameters of quarantine implementation. Homogeneity of threat or stressors in non-experimental research, although methodologically challenging, could be achieved by constraining samples. For example, strict quarantine might have a more specific and uniform threat value to owners of small businesses dependent on in-person custom than to owners of businesses providing online services. When these and other sources of heterogeneity are not considered, conclusions from umbrella reviews and meta-analyses may do a disservice to important communities in our societies [31]. Investigators could consider whether the associations of quarantine with ideation are age-specific. Age-related effects associated with the pandemic, for which evidence appears to be quite robust, may arise from developmental factors interacting with quarantine [5]. For example, whereas for older adults quarantine may yield valued outcomes (e.g., more time with family) [26], the same restrictions (i.e., more time with family, less with friends) may impede or threaten independence in emerging adults [21, 55]. Whereas greater access to technology may be valued by older adults [25], for emerging adults the same technology may contribute to increased distress and STB [56]. If effects of quarantine on mental health are age-specific, greater caution is required when considering indiscriminate quarantine as a public health response to infectious diseases. Finally, investigators could examine the influence of variations in the form and manner of quarantine procedures and requirements across jurisdictions. In the present study, quarantine was nationwide and nonselective, and the population was required to isolate within their usual residential “bubbles.” Fear and anxiety had previously been employed—intentionally or not—to build public adherence to quarantine orders. Government announcements, advertising, and statements from the Prime Minister, ministers of government, medical officers, epidemiologists, and other scientists emphasised the threat and risks of COVID-19. Other provisions, such as the enabling of reporting to Police of others’ violations of quarantine, likely added distrust to the environment of fear and anxiety. In other jurisdictions, quarantine was selective and centralised [57]. Understanding the influence of variations in the form and manner of quarantine on mental health would enable public health officials to mitigate the costs of quarantine. Limitations The design does not afford causal inferences about effects of quarantine. Although described as natural experiment, there was no randomization or experimental control to neutralize unmeasured confounding [50, 51]. Undergraduate or college samples are often considered inferior to community and clinical samples. However, undergraduate samples do confer a particular advantage in this context, namely homogeneity in the repercussions (and therefore threat value) of nonselective quarantine. The generalisability of the findings to other non-help-seeking and clinical populations is unclear. In clinical samples, chronic and intermittent ideation may be differently responsive to those aspects of quarantine that increased hopelessness in undergraduates [31, 58]. Also, infection following the August outbreak, prior COVID-19 infection, recent international travel, COVID-related bereavement, or participants’ physical location during quarantine were not assessed in this study. Finally, the post hoc nature of the tests for mediation increases the need for replication. The observed effects were small. The effect size shift in hopelessness associated with quarantine (d = 0.13) is somewhat smaller than Cohen’s [59] small effect size (d = 0.2). Construing the clinical significance (albeit in a nonclinical sample) of that effect is difficult because the measure of hopelessness was a two-item instrument designed for efficiency in epidemiological and repeated assessment studies [40]. However, the clinical significance of that shift in hopelessness can be inferred from the associated shift in suicidal ideation within the mediation model. Specifically, the 0.13 standard deviation shift in hopelessness transmuted into an increase in the rate of ideation from 16.7% to 17.9% at average hopelessness. At one standard deviation above average hopelessness, the rate shifted from 27.0% to 28.5%. While these effects were small in size, they were observed in a country with one of the highest rates of suicide among emerging adults in the world [34]. Conclusions and public health significance In a natural experiment, a nationwide quarantine predicted increased hopelessness and, in turn, hopelessness predicted increased suicidal thinking among emerging adults in a community that was almost certainly free of COVID-19 infection. That is, for emerging adults, increased hopelessness and worsening suicidal thinking are among the health costs associated with strict nonselective quarantine. The findings imply that decision-makers must take seriously the immediate psychological costs of public health interventions, weighing these alongside the anticipated gains. Moreover, interventions such as quarantine, applied in ways that ignore developmental, mental health, and other aspects of difference within populations (e.g., deprivation), may harm those who are most vulnerable within communities. Similarly, research efforts could be designed to build understanding of heterogeneity in the effects of quarantine within populations and communities to enable decision-makers to optimise the use of quarantine. Supplementary Information Additional file 1. pdf: Supplementary methods, analysis information, results, and model figures. Additional file 1. pdf: Supplementary methods, analysis information, results, and model figures.