This study is a cohort with participants recruited in a cross-sectional survey conducting annual data collections during the pandemic. The first cohort time point (C1) occurred between September 9th and October 16th, 2020. Data for the second time point (C2) was collected between April 29th and May 22th, 2021. The third and final cohort time point (C3) was conducted between September 9th and November 21th, 2022, six months after the revocation of the Brazilian National Public Health Emergency, following the increase in vaccination. For further recruitment details, see the original cross-sectional study (13). We chose these specific timepoints taking local government-issued measures and the rise in vaccinated people into consideration. To be eligible, all participants had to be ≥ 18 years old and residing in Brazil, and each questionnaire had to be completed to the last question. The study was submitted in the National Ethics Committee Platform (“Plataforma Brasil” registry number: 5.376.167) and approved by the Ethics Committee of Research of the Federal University of Health Sciences of Porto Alegre (#4241378).

All follow-ups were collected using the Research Electronic Data Capture (REDCap^®, Vanderbilt University, Nashville, TN, USA) online platform. Participants who explicitly agreed to continue their participation for the next phases of the study were recruited to the subsequent stages through an individualized link that was sent to the e-mail address provided on C1. E-mail reminders were sent periodically according to overall response rate. The first page of the survey contained the informed consent with an accept/decline button, followed by the questionnaire on the next pages. It was composed of 56 questions and was divided into four sections (fully disclosed in Supplementary Material ): 1) sociodemographic data; 2) social distancing and vaccination during the pandemic (questions originally created by authors); 3) the Brazilian-validated version of the ASSIST instrument (18, 21), chosen for its reliability and validity in measuring the use risks of psychoactive substances; 4) the Depression, Anxiety and Stress Scale (DASS-21), also validated in Portuguese (22). At the end of the survey, a counseling message based on the ASSIST calculated score was presented, with their risk of dependence being classified as low, moderate, or high. For moderate and high risk, participants were recommended to seek help from a health care practitioner.

Sociodemographic data were collected in categories, while age was a continuous variable. As determined in the first study published about these data, some sociodemographic strata were collapsed when subgroup samples were too small to be analyzed on their own. Education was collected into eight categories and collapsed into three, “Incomplete Secondary Education”, “Complete Secondary Education” and “Higher Education”. Income was collected in the local currency (Brazilian Real) (USD$ 1.00 ≅ R$ 5.00), and the minimum wage in Brazil during the cohort period ranged from R$ 1,045.00 to R$ 1,212.00. Social distancing perception regarded the level of restriction in social distancing assessed as three categories (Low, Medium and High). All sociodemographic factors were collected in the first survey filled by the subjects.

The Brazilian version of the ASSIST microstructured questionnaire included self-reported frequency of drug use, lifetime use, and use in the three months before the first pandemic wave (C1-September 9th to October 16th, 2020), the second wave (C2-April 29th and May 22nd of 2021), and around the end of the pandemic (C3-September 9th and November 21st of 2022) based on DSM-IV criteria for drug dependence. The score obtained for each substance falls into a “lower” (“occasional use”), “moderate” (“in risk for abuse”), or “high” risk (“dependence”) category which determines the most appropriate intervention for each categorical level of use. Scores from 0 to 3 are considered occasional use (for alcohol: 0 to 10); ≥ 4 to 26 indicates moderate risk for dependence (for alcohol: 11 to 26); ≥ 27 suggests dependence or high risk of dependence for alcohol and all other drugs (18).

The DASS-21 contemplated three subscales (depression, anxiety, and stress) with a Likert format, varying from 0 (did not apply to me at all) to 3 (applied to me very much or most of the time). Cut-offs for levels of subscales are presented as “Normal” - Depression: 0 to 9; Anxiety: 0 to 7; Stress: 0 to 14. “Mild” - Depression: 10 to 13; Anxiety: 8 to 9; Stress: 15 to 18. “Moderate” - Depression: 14 to 20; Anxiety: 10 to 14; Stress: 19 to 25; Severe- Depression: 21 to 27; Anxiety: 15 to 19; Stress: 26 to 33; “Extremely severe” - Depression: >28; Anxiety: >20; Stress: >34 (22).

The inferential analysis investigated the effect of time and sociodemographic factors in the ASSIST scores and DASS-21 scores, as well as the interaction between time and factors through Generalized Linear Models (GLM) or Two Way RM-ANOVA. Post-hoc comparisons were tested using the Bonferroni test or Tukey, respectively. To carry out the GLM, the scores for drug use (ASSIST) were considered the dependent variable. The independent variables were the sample profile (sociodemographic data) and the severity of social distancing. For identifying risks within the social distancing categories, the reference category was “very rigorous”. The covariates defined as controls in the models were listed based on the significant results observed in the bivariate analysis (comparisons of social distancing level with sociodemographic variables and DASS scale).

For GLM analysis, it was considered all the subjects who filled at least one of the surveys (C1 and C2 and/or C3) since this statistical approach allows the use of datasets with missing data, summing n = 2498. For the other statistical approaches, such as ANOVA, Chi-Square and Pearson Correlation, it was considered only the n = 755 who filled all surveys (C1, C2 and C3). Normality and equal variance tests were performed to verify if non-parametric tests were necessary.

The Chi-Square test was performed for dichotomized variables (DASS-21 severity rating prevalence, divided as “Normal” and “Mild or more” and ASSIST Risk Level, divided as Low Risk and Moderate/High Risk). Spearman’s correlation test was used to verify associations between quantitative variables. Descriptive statistics were carried out using Excel^®, and all other inferential analyses were run with IBM^® SPSS Statistics software (v20) or SigmaStat^® 3.1. Differences and associations were considered statistically significant when P ≤.05. F values, t values, z values, as well as the n, confidence interval (CI) are presented next to P values. To assess the potential of false negatives, the model performs a power test according to the hypothesis test used. A β ≤ 0.2 was considered as an adequate value for tests in which the results presented a P value higher than.05. The sample size was calculated before the first study using the most prevalent drugs in the population as targets (alcohol, tobacco and cannabis), comparing the pre-pandemic versus pandemic alcohol use prevalence (main outcome).

The ASSIST mean scores are presented by time points, and the threshold between “Low Risk Level” and “Moderate Risk Level” is also shown in Figure 2 . The “High Risk Level” is not shown in this figure because none of the mean scores were near its threshold. The highest scores besides alcohol, as we can see in Figure 2 , are tobacco and cannabis, followed by hypnotics/sedatives and cocaine/crack. Overall, differences were observed between time points for some drugs, like alcohol, cannabis, hallucinogens and cocaine, when observing mean scores from the ASSIST scale. Differences were found between C1 and C2 and C1 and C3 for alcohol (P < 0.001), cannabis (P = 0.005), hallucinogens (P < 0.001) and cocaine (P < 0.05).

When assessed separately, the effect of time on mean scores for each drug class shows some interesting results. Alcohol mean scores demonstrated a progressive decrease over time, with C2 and C3 being lower than C1 (P < 0.005). Tobacco did not show statistically significant differences between time points, remaining at approximately the same level. Cannabis also presents differences and, although C3 has a slightly higher score than C2, both are lower than C1 (P < 0.05). Even though hypnotics/sedatives showed some greater scores in C1 and C2, it did not show statistically significant differences with C3, which almost fell to low risk level due to the high variability of the use during periods. Hallucinogens and Cocaine/crack presented only one difference between C1 and C2 (P = 0.035), as C3 showed a substantial decrease of score ( Figure 2 ).

Regarding the ASSIST score risk level represented in dashed lines in Figure 2 , it is observed that the mean scores for alcohol, tobacco, hypnotics/sedatives, and cocaine all reach the moderate risk level, with the exception of alcohol’s C2 (10.4 ± 0.6; CI95%: 9.7 – 10.9) and C3 (10.0 ± 0.6; CI95%: 9.3 – 10.5) frequencies and cocaine’s C2 frequency (4.8 ± 1.6; CI95%: 3.0 - 6.2), considering these CIs include the lower threshold of the moderate risk level, which is 11 for alcohol and 4 for the other drugs. The drug users’ risk level prevalence is presented in Figure 3 . ASSIST risk levels showed higher prevalences in the low risk category. Moderate risk was mostly around 30%, and high risk was majorly lower than 10%. When comparing the low and moderate or more prevalences using the Chi-Square test, no statistically significant differences were found between groups.

The gender covariable has shown some statistically significant differences for several drug classes. Alcohol (P = 0.004), tobacco (P = 0.047), cannabis (P < 0.001), and hypnotics (P = 0.030) presented a gender effect, where men reported higher drug use than women. Only amphetamines/ecstasy demonstrated an interaction between the effect of the covariable and time (P = 0.011). Interestingly, for amphetamines/ecstasy, there were no statistically significant differences in the pairwise comparisons, only limitrophe non significance (Men’s C1 > Men’s C2; P = 0.053). This must be related to the pattern observed in their mean scores, since men showed a fall in C2 score, returning to a similar baseline score in C3, while women showed a rise in C2, and returned to a similar score to C1 in C3.

Another factor that showed an interaction between time and itself was “education”, again for amphetamines/ecstasy and also for hypnotics (P < 0.001). Although the majority of drug classes presented differences for covariable effects (all P < 0.001, except for hypnotics), alcohol, tobacco and cannabis were the only ones which did not present small sample sizes in C1 for Incomplete Secondary Education (ISE), while in C2 and C3, all drug classes presented small sample sizes for the mentioned category, hindering their analyses. Considering the most prevalent drug classes, Higher Education (HE) usually had lower scores than Secondary Education (SE). Amphetamines/ecstasy presented differences between SE and HE (SE in C1 > HE in all time points; SE in C2 > HE in C2; SE in C3 > HE in all time points). Tobacco, cannabis, hallucinogens, and cocaine users had higher scores for SE when compared to HE (P < 0.005), when considering only the covariable. Opioids only showed differences between ISE > SE, and ISE > HE ( Figure 4 , represented as lower education presenting a higher ASSIST score than higher education).

Income showed some statistical significant differences as well. An interaction effect was seen for alcohol users (P = 0.012) and hypnotics (P = 0.024), but also a time effect was seen for alcohol, hallucinogens (P < 0.001) and cocaine (P = 0.014) and a covariable effect for tobacco, cannabis (P < 0.001), hypnotics (P = 0.009), hallucinogens (P = 0.028). When we look at alcohol use and time points, we see that moderate to low incomes show higher scores than higher incomes (R$ 751.00 to 1,500.00 in C1 and C2, R$ 1,500.00 to 3,000.00 C1 > R$ 9,000.00 or more, C1 to C3; R$ 1,500.00 to 3,000.00, C1 > R$ 6,000.00 to 9,000.00, C1 and C2; R$ 1,500.00 to 3,000.00, all time points > R$ 6,000.00 to 9,000.00, C3). Hypnotics also showed an interaction between the covariable and time, and the comparisons by pairwise method showed differences between R$ 751.00 to 1,500.00 in C1 > R$ 6,000.00 to 9,000.00 in C2; R$ 3,001.00 to 6,000.00, C1 and C2 > R$ 6,000.00 to 9,000.00 C2; R$ 1,500.00 to 3,000.00, C1 and C2 > R$ 6,000.00 to 9,000.00, C2 ( Figure 4 , represented as lower income presenting a higher ASSIST score than higher income).

Social distancing showed some effects of the covariable and time. Alcohol use showed a covariable effect (P < 0.001) and a Time effect (P < 0.001). Cannabis use showed a covariable effect (P = 0.001) and a very close to significance time effect (P = 0.052). Amphetamines/ecstasy use also presented a covariable effect (P < 0.001). Hallucinogens presented a covariable and time effect (P < 0.001 and P = 0.001, respectively). Alcohol, cannabis, cocaine and hallucinogens showed differences as follows: lower social distancing greater than median social distancing and Lower social distancing greater than higher social distancing. Amphetamines/ecstasy had different patterns, with amphetamines/ecstasy showing one more difference (lower social distancing > median social distancing; lower > higher; median > higher, P < 0.05). Opioids only showed a difference between Low and Median (P < 0.05) ( Figure 4 , represented as lower social distancing presenting a higher ASSIST score than higher social distancing).

Mean scores for depression, anxiety and stress traits mean scores are displayed in Figure 5 and showed differences in all time points. For depression, mean scores were 14.70 (SEM: ± 0.27), 12.76 (SEM: ± 0.38) and 10.23 (SE: ± 0.36), for C1, C2 and C3, respectively. For anxiety, means were, also respectively, 8.68 (SEM: ± 0.21), 6.99 (SE: ± 0.28), 5.46 (SE: ± 0.24). And for stress, 16.24 (SEM: ± 0.24), 13.82 (SEM: ± 0.33), 11.22 (SEM: ± 0.32), also respectively. Considering the DASS-21 scale cutoffs, it can be observed that the C1 mean score for depression is in the moderate rating, while the others are in the mild rating. C2 mean scores are in the borderline between mild and normal ratings, and lastly, C3 mean scores are in the Normal range, except for the depression mean score.

DASS-21 prevalences from the DASS-21 scale are presented in Figure 6 . In view of the prevalences observed for each time point, a few interesting considerations can be taken. The most extreme categories (“normal” and “extremely severe”) were the ones which changed the most from C1 to C3. The depression scale suffered increases in the “normal” category (44.51% to 61.19%), while the “extremely severe” suffered reductions (17.99% to 10.61%). “mild” and “moderate” reduced slightly (approximately -2% for each time point), and “severe” barely reduced percentages. Anxiety scale also suffered similar changes in percentages, although “normal” was the only one showing increases (66.07% to 74.17%). “mild” category showed an increase in C2, but reduced again in C3. “moderate”, “severe”, and “extremely severe” changed approximately -3% for each time point, suggesting a normalization of symptoms. Stress scale exhibited similar changes to anxiety. The only category with raises in percentages was “normal”, the rest decreased around 3 to 4% (“moderate” to “extremely severe” falling almost in half in C3). A chi-square test was run for DASS-21 ratings prevalences, comparing normal and mild to extremely severe ratings, showing statistically significant differences for all three subscales, indicating C1 had lower frequency than expected and C3 had more than expected.

All three traits in C1 exhibited higher mean scores than C2 and C3, while C2 was higher than C3 (P < 0.001) ( Figure 7 , represented as dichotomous variables). As for the sociodemographic factors, when analyzed through the Two-Way RMAnova, marital status exhibited an interaction with time for all subscales (Two-Way RMAnova; P < 0.05). Single/divorced/widowed presented greater scores for depression, anxiety and stress than married/stable union (P < 0.001). Gender demonstrated a covariable effect for anxiety and stress subscales (P < 0.001) and the scores were greater for women. Education level presented associations with depression, anxiety and stress. HE showed lower scores of depression (P = 0.002) and stress (P = 0.027). For income, stress and anxiety showed higher scores for the R$ 1,501 to 3,000 and R$ 3,001 to 6,000 ranges than higher incomes (more than R$ 6,001), especially in C2 and C3. Also, social distancing presented higher scores of anxiety and stress for lower social distancing when compared to medium and high social distancing (P < 0.05).

The Spearman’s correlation demonstrated positive correlations between ASSIST scores and DASS-21 scores for all subscales and drugs ( Supplementary Table S1 ). However, r coefficients hardly surpassed the 0.20 mark, which indicates rather weak associations, taking in account that the P-values were statistically significant. The strongest correlations were found for opioids and hypnotics (around 0.30-0.40), especially in C1 and C2. Amphetamines/ecstasy showed somewhat stronger correlations in C2 (around 0.30) as well, but not in other time points. Only the correlation between Cocaine/crack versus stress was not significant (P = 0.116). This means both scores are somehow linked, although not very strongly, and might be influencing one another, as both scores suffered reductions through time.

One important thing to note is that the sample was obtained from an online survey, disclosed and spread through social media (Facebook^® and Instagram^®), message apps and emails. This might introduce a selection bias, as this constitutes a convenience sample, which may lead to impaired internal validity.

As could be seen from the sample size we obtained from this 3-year cohort, there was an unsurprisingly significant participant dropout rate (41). It is important to note that prospective studies are always prone to loss of follow up from participants (61), as they may not feel as engaged or motivated to participate in the study as they were in the beginning. This is especially true when taking into account how diverse were the COVID-19 waves and prospects during 2022 when compared to 2020. Nonetheless, this shortage in sample size did not hinder statistical analyses completely, as the model used still had enough power to perform several comparisons between various subgroups. There is a loss of follow-up rate for all drug classes when we observe absolute numbers. However, they maintain proportional numbers when percentages are calculated from the total number of each time point. This indicates no biased loss for one or more drug classes or in one group. Although use of inhalants was excluded from our results, we believe this had little impact in our findings due to the remarkably low prevalence of abuse for this particular class of substances in Brazil.

An aspect of the GLM is that it makes analyses of data in a manner that enables the comparison of time points. Some participants have answered all 3 segments of the cohort, while some only answered the first and the second sequels, or the first and the third ones. The model is designed to handle varying types of error distributions and varying sample sizes along remitted measures.

Drug-specific questions from the ASSIST questionnaire required information about their drug use in the last three months. This introduces a memory bias (61) because, even if this period of time is not considered to be that extensive, some people might have more difficulty in remembering their past drug use pattern than others. Importantly, C2 and C3 participants were asked to make this assessment years later. However, they were most likely motivated to pay attention to their own habits due to media’s constant disclosure of mental health issues emerging during the pandemic, as well as to the fact that they were being “observed” by researchers from this study, making them less prone to forget their drug use and mood state symptoms. Nevertheless, this study highlights the dynamic nature of drug use and emotional well-being during prolonged periods of social disruption, emphasizing the need for targeted public health strategies to support at-risk populations during such crises.

Through this cohort, it was possible to address a great mental health issue very much enlivened in the pandemic, which was the extension of the impact of the pandemic on drug use and emotional state, in the first year of the pandemic when very strict physical distancing measures took place and how it might have changed in subsequent years (more or less flexible physical distancing measures).

In conclusion, drug use for several drugs presented a reduction in scores when comparing the time points soon after the end of the pandemic with its beginning, and the same happened to the emotional state symptoms, implying a development of resilience or an accommodation to the context. An important aspect to note is that sociodemographic strata showed different drug use patterns and emotional states. Men, single/divorced/widowed, lower education, income and social distancing respondents perceived more drug use score, which comes in line with the literature. Finally, women, younger age, single/divorced/widowed, lower education, income and social distancing showed higher scores for the emotional/mood state.

Altogether, these insights are important for comprehending what steps can be taken for future research in drug use and abuse/mental health of similar populations. Even though the general Brazilian population might be much more diverse than our sample, our findings still provide a valid source of insights for a great part of the population. By acknowledging the dynamic relationship between drug use and emotional well-being, policymakers and healthcare providers can develop targeted interventions that address the unique challenges faced by at-risk populations, ultimately promoting resilience and improving overall well-being during times of crisis.