Figure 1 maps the data collection periods for the two surveys and three indicators used in this study onto the phases of the COVID-19 pandemic in Germany. At the start of the COVID-19 pandemic, the third survey wave of the European Health Interview Survey as part of the study “German Health Update” (GEDA 2019/2020-EHIS) for Germany (78) had been in the survey phase conducting telephone interviews since April 2019. This survey included the screening questionnaire PHQ-8 (85), which comprises its abbreviated version PHQ-2 (86), as a measure of depressive symptoms.

The survey was originally not designed for monthly reporting; however, slight adjustments of the sample weighting permitted first analyses of the development of various health indicators in the months preceding the pandemic as well as the first months of the pandemic (87). Given the new informational needs arising from the pandemic, the survey was continued until the beginning of January 2021. After the end of GEDA 2019/2020, short inventories assessing mental health indicators, including the PHQ-2 (86), the GAD-2 (88), and a self-rated mental health (SRMH) item (89) were integrated into a running population-based telephone survey from mid-March to mid-July 2021 with a one-month data collection gap from mid-May to mid-June. This survey, the “COVID-19 vaccination rate monitoring in Germany (COVIMO),” was designed to be sampled on a monthly basis (83). From July 2021 until December 2021 and February until June 2022 (with a data gap in January 2022), continuous monthly interviews were carried out within the frameworks of GEDA 2021 and GEDA 2022 (90), respectively (see Figure 1).

The GEDA surveys and COVIMO were conducted on behalf of the Federal Ministry of Health of Germany. Data was collected by an external market and social research institute (USUMA GmbH). Study design, data collection, and sampling were largely the same across these different surveys, and based on a dual frame approach of mobile and landline numbers. The study population differs between the GEDA surveys and COVIMO, as GEDA targets people aged 15 or older living in private households whose main residence is in Germany (78), whereas COVIMO includes only people aged 18 or older (83). In addition, there is a slight difference in the general content focus of the studies (general health survey in GEDA versus vaccination monitoring in COVIMO). Details of the data pipeline, which cover semi-automated data preparation, data merging, and output creation as well as a description of sample weighting can be found elsewhere (91). We examined potential study-related differences between the GEDA surveys and COVIMO in the indicators of interest: Distributions of outcome variable scores were compared between studies in their one overlapping month using boxplots and violin plots (results not shown). No pronounced differences were detected.

Across the entire survey period, 45,102 participants aged 18 or older were included in the analyses. The distributions by gender, age, and level of education in the different surveys are shown in Table 1, number of monthly cases are shown in Supplementary Figure A1. The GEDA-EHIS 2019, GEDA 2020, 2021, and COVIMO studies surveyed ~1,000 participants per month. GEDA 2022 provided data for 3,000 participants per month. To reduce data gaps, the second half of each month was combined with the first half of the following month across the observation period.

Depressive symptoms were observed prior to and during the pandemic using monthly data from beginning of April 2019 until mid-June 2022 (see Figure 1). There were four short data gaps (the largest between January and mid-March 2021). The indicator was measured with the established ultra-brief screening instrument “Patient Health Questionnaire-2” (PHQ-2) (86), which has been found to perform well as a screening tool for depressive disorders in the German general population (92). The PHQ-2 captures the frequencies of two core symptoms of depressive disorders, asking, “Over the last 2 weeks, how often have you been bothered by the following problems?”: (1) “little interest and pleasure in doing things” (2) “feeling down, depressed or hopeless” (possible responses: 0 = “not at all,” 1 = “several days,” 2 = “more than half the days,” 3 = “nearly every day”). The total score of the PHQ-2 ranges from 0 to 6 (“no symptoms” to “severe symptoms”). According to scoring recommendations (92), scores ≥ 3 represent a positive screen for possible depressive disorder and indicate a potential need for further diagnostic assessment. In our analytical sample, the internal consistency of the PHQ-2 is α = 0.73 [standardized alpha coefficient as recommended for two items (93), unstandardized α = 0.72], slightly higher than in a comparable German sample (45). Two measures are reported in the current study: (1) the mean depressive symptom score, which tracks changes in the mean severity of symptoms in the population (73); (2) the proportion of the adult population screening positive for possible depressive disorder.

Symptoms of anxiety were observed monthly from mid-March 2021 to mid-June 2022 (see Figure 1) with two short data gaps (mid-May 2021 to mid-June 2021 and January 2022). The indicator was measured with the established ultra-brief screening instrument “Generalized Anxiety Disorder-2” (GAD-2), which has been found to perform well as a screening tool for anxiety disorders in the German general population (88). The GAD-2 captures the frequency of two core symptoms of anxiety disorders, asking, “Over the last 2 weeks, how often have you been bothered by the following problems?”: (1) “feeling nervous, anxious or on edge” (2) “not being able to stop or control worrying” (possible responses: 0 = “not at all,” 1 = “several days,” 2 = “more than half the days,” 3 = “nearly every day”). The total score of the GAD-2 ranges from 0 to 6 (no symptoms to severe symptoms). Scores ≥ 3 represent a positive screen for possible anxiety disorder, including generalized anxiety disorder, panic disorder, social anxiety disorder, and posttraumatic stress disorder (88). In our analytical sample, the internal consistency of the GAD-2 is α = 0.67 (standardized alpha unstandardized α = 0.66), almost the same value as in a comparable German sample (45). Just as with depressive symptoms, two measures are reported: (1) the mean anxiety symptom score and (2) the proportion of the adult population screening positive for possible anxiety disorder.

SRMH was observed monthly from mid-March 2021 to mid-June 2022 (see Figure 1) with two short data gaps (mid-May 2021 to mid-June 2021 and January 2022). It was measured using the question: “How would you describe your overall mental health?” (possible responses: 5 = “excellent,” 4 = “very good,” 3 = “good,” 2 = “fair,” 1 = “poor”). The single item is an established way to measure SRMH in population based surveys (89). SRMH has been found to represent a dimension of mental health that is qualitatively distinct from psychopathology (94). Here and elsewhere (58) it is employed as a measure of positive mental health. Two measures are reported: (1) population mean SRMH score; (2) the proportion of the adult population rating their mental health as “very good” or “excellent,” following previous categorization to identify the presence of positive mental health (58).

In order to assess mental health developments over time in the general population and by subgroup, we calculated time series of estimates along with smoothing curves to be represented graphically [for details, see (91)]. Our aim was to achieve high temporal resolution whilst working with sample size restrictions and also to smooth random fluctuations. The estimation procedure described below also ensures that possible fluctuations in distributions of sex, age, and level of education in the sample over time are corrected for and that stratified results are standardized for the other main sociodemographic characteristics.

For each of the three mental health indicators, linear and logistic regressions were used to predict a time series of means and proportions for the adult population in Germany. To handle low cell counts and reduce volatility over time, we estimated centered moving averages rather than monthly averages (97) using weighted data from three-month windows. Some three-month windows only included data from 2 months due to data gaps. The three-month windows move in steps of 1 month. The models for each three-month window regress the mental health indicators on sex, age group, and level of education, and interactions between them. While the linear models include all possible interaction terms, only all possible two-way interactions of the covariates but not the three-way interactions were included in the logistic regression models to avoid problems resulting from empty cells. Nonetheless, there were some empty cells around data gaps, where estimates are based on data from two rather than 3 months, resulting in estimation gaps in the time series of categorical PHQ-2 and GAD-2 estimates.

These regression models are the foundation for standardization for sex, age, and level of education between the three-month windows, which ensures that different distributions of these characteristics between them do not influence the results. For standardization, we calculated averaged predictions in a two-step process. First, we used the models to perform predictions on a standard population. To calculate arithmetic means of mental health indicator scores, we used the model estimates from the linear regressions and predicted the expected values of the indicator in question. To calculate proportions for categorical indicator outcomes, we predicted the expected probabilities. In a second step we averaged over all of the predictions. The standard population was calculated using data from the Microcensus 2018 (95), which approximates Germany's population in 2018.

The calculation of estimates for time series stratified by sex, age, and level of education was similar to the procedure described above. However, in order to exclude different distributions of the respective other two characteristics in different time periods as explanatory factors for temporal developments, stratified results by age group, sex, or level of education were standardized by the remaining two characteristics in the prediction step. For example, the results stratified by age group were standardized for sex and education. This was achieved by making predictions for every subgroup as if all observations in the standard population belonged to this subgroup. The standardization between subgroups means that the subgroup-specific estimates are not representative for the population subgroup. The mathematical and methodological foundations for model-based predictions and standardization can be found elsewhere (98–100).

In order to improve results interpretation by making trends more visible, we additionally estimated smoothed curves using a general additive model (101) with a smoothing spline (102, 103) and curve by factor interaction (104). Values were predicted on the same standard population. The spline was fitted on weekly observations to maximize temporal resolution given sample size. To avoid over- or underfitting, the smoothing parameter was estimated using restricted maximum likelihood. However, we found that for our shorter time series, the curves based on weekly estimates were less smooth than the three-monthly predictions. Therefore, we only used this procedure for the longer time series.

Missing values in the dependent variables were excluded on a case-by-case basis. Observations without information on sex or age were not included in the survey and were treated as non-responses. Missings in education were imputed in accordance with the weighting procedure (66) by assigning the most frequent value, a medium level of education.

The initial interpretation of the times series was descriptive by visual inspection. Conservative criteria such as confidence interval comparisons were not used to evaluate developments over time at this stage because the first aim was to explore the overall trajectory. In addition to visual inspection, we carried out statistical comparisons between different time periods (see section 2.4.2).

For the longer depressive symptoms time series, we conducted statistical comparisons between the three survey years for two periods of months: (1) mid-March to mid-September [calendar week (CW) 11–37] 2019, 2020, 2021, mid-March to mid-June (CW 11–24) 2022, and (2) mid-September to end of December 2019, 2020, 2021 (CW 38–52). For the shorter time series of SRMH and anxiety symptoms, mid-March to mid-September 2021 (CW 11–37) and mid-March to mid-June (CW 11–24) 2022 were compared. These time periods were chosen based on (1) the declaration of a pandemic on March 11 from WHO (1), (2) data gaps for the initial months of each year, and (3) the turning point in the development of depressive symptoms at the end of the summer of 2020, as shown in Figure 3. Small gaps in some of these time periods could not be avoided; for example, there was no data from mid-March to the beginning of April 2019. We tested for two kinds of trends. First, we tested for differences in mean depressive or anxiety symptom score and proportions at or above PHQ-2/ GAD-2 cutoff between corresponding time periods across years for the overall population as well as within the different subgroups. For SRMH we tested for differences in mean and proportions with very good or excellent SRMH. Second, we tested for possible changes in differences in means or proportions between subgroups over the specified time periods.

To conduct these comparisons, we again used linear and logistic regression models to produce averaged predictions as described above. However, here we calculated estimates for the defined time periods by including a set of dummy variables indicating these periods in the survey year. Furthermore, all possible interactions of these dummies with age group, sex, and level of education were included. The specification of the linear and the logistic regression models again differed with regard to level of interaction. In the logistic regression only, three-way interactions were included, whereas the linear model also included four-way interactions. After model estimation, the standard population was used for prediction of the means of the specified time periods. Contrasts between the time periods and the differences between the subgroups between the time periods were estimated. We used Stata's “margins, contrast” command (105) for estimation and statistical testing using Wald tests, applying a significance level of (p < 0.05).

Before running these contrasts, we conducted joint tests or omnibus tests in order to control for multiple comparisons and reduce the likelihood of false significant results by using protected tests (106). We only performed pairwise comparisons between time periods if the hypothesis that all possible differences were zero could be rejected. To assess the permissibility of pairwise comparisons within subgroups, we jointly tested if the differences within all subgroups defined by sex, age group or level of education, respectively, were zero. To address the question of whether differences between subgroups changed over time, we conducted joint tests including all possible differences, i.e., a test for interaction of time and sociodemographic characteristic.

Time series plots suggest that throughout most of the observation period, mean depressive symptom scores were higher in women than in men (Figure 3A). Likewise, percentages of positive screens for possible depressive disorder appear to be higher in women than in men for the most part, although the overlap was greater than for mean scores (Figure 3B). The overall shape of the plotted time series stratified by sex roughly matches that of the whole population: initial symptom declines followed by two increases. However, women experienced less relief in the early phases of the pandemic and earlier symptom increases in the later phases:

Declines in both mean symptom scores and proportion of positive screens in the early phases of the pandemic are seen in both sexes, but limited to the summer plateau in women. Statistical comparisons between spring/summer 2019 and 2020 only show significant declines in men (Tables 2, 4). The plotted time series suggest two increases in both measures in both sexes, the first between autumn 2020 and spring 2021, the second at the end of 2021/beginning of 2022. Statistical comparisons reveal significant increases above 2019 levels in spring/summer 2021 for women but not for men (p-values for CW 38–52 cannot be reported due to non-significant joint tests). Men's depressive symptom levels surpass 2019 levels for the first time in 2022, also rising significantly above 2021 levels. In women, this second visible symptom level increase does not result in levels significantly above 2021 levels.

Over the course of the observation period, mean scores and the percentage of positive screens among women standardized by age and level of education increased by 0.33 points (from 1.00) and 5.9 percentage points (from 11%) between spring/summer 2019 and spring/summer 2022; among men, by 0.26 points (from 0.93) and 5.5 percentage points (from 11.5%) (Tables 2, 4).

Time series plots show a tendency for lower mean symptom scores among those aged 65+ years and, less consistently, higher mean symptom scores among 18–29-year-olds compared to the other age groups in the observation period (all age groups standardized by sex and education; Figure 3C). This pattern is less pronounced in the proportions of positive screens time series (Figure 3D). Supplementary Figure A2 shows these time series in a separate plot for each age group.

While no declines in depressive symptom levels are visible among those aged 65+ for the early stages of the pandemic, plotted time series show decreasing means and proportions of positive screens from the beginning of the outbreak among the middle age groups and among the youngest in summer 2020. However, these declines resulted in a significant difference between means in spring/summer 2019 and 2020 only for those aged 45–64 years (Table 2).

At different times between the second wave and the end of the observation period, every age group then showed increases in depressive symptom levels beyond pre-pandemic levels:

Among 18- to 29-years-olds, means and positive screens rose very markedly compared to other groups from autumn 2021 to the end of the year. The standardized proportion of positive screens reached 18.6% in September-December 2021, representing an 8 percentage point-increase from the same period in 2019, and means rose from 1.01 to 1.57. However, statistical uncertainty is fairly high in this group, and this increase did not reach significance compared to pre-pandemic levels (Table 5; p-values for means cannot be reported due to non-significant joint tests). Following the sharp increase at the end of 2021, symptoms returned to lower levels numerically but not statistically significantly above 2019 levels in 2022.

30- to 44-year-olds showed a temporary increase in both PHQ-2 measures in spring/summer 2021. Standardized mean symptom scores increased significantly from 2019 in this time and markedly compared to other groups (from 0.89 in 2019 to 1.21 in 2021, Table 2). The temporary increase of 3.7 percentage points (from 11.5% in 2019) in the standardized proportion of positive screens did not reach statistical significance (Table 4). Mean symptom scores (but not proportions) again significantly surpassed 2019 levels in spring/summer 2022, but not 2021 levels. However, this increase does not stand out in magnitude.

45- to 64-year-olds exhibited an increase in depressive symptom levels above pre-pandemic levels later than the other age groups. The time series graphs show increases from about mid-2021. Both PHQ-2 measures significantly surpassed 2019 levels for the first time in 2022 (Tables 2, 4), also surpassing 2021 levels, resulting in a 5.9 percentage point increase (from 12.7%) in the standardized proportion of positive screens between spring/summer 2019 and the same period in 2022.

Smoothing curves suggest an overall continuous increase in symptoms of depression from autumn 2020 among 65+-year-olds. Statistical time period comparisons also point to a particularly steady trend of increase in this age group, with significant differences compared to 2019 in proportion of positive screens as early as end of 2020 and again end of 2021 (Table 5; p-values for comparisons of means in CW 38–52 cannot be reported due to non-significant joint tests) and significant differences compared to 2019 levels in means from spring/summer 2021 (Table 2). The marked increase in symptom levels within 2022 (significantly surpassing 2021 levels) also stands out in this group. The standardized proportion of positive screens reached 18.0% in spring/summer 2022–A 9% point increase from the same period in 2019 (Table 4).

A social gradient is apparent throughout the observation period, with higher mean scores and proportions of positive screens for possible depression in those with the lowest levels of education, followed by the middle and high-level groups (all standardized by sex and age; Figures 3E, F). Just like with stratification by sex, the overall shape of the plotted time series stratified by level of education roughly matches that of the whole population: initial symptom declines followed by two increases.

However, the declines in means and positive screens in the first pandemic spring and summer result in statistically significant differences between 2019 and 2020 only in the middle level of education group (Tables 2, 4).

Smoothing curves suggest that the subsequent increases between autumn 2020 and spring 2021 as well as end of 2021 to beginning of 2022 amount to a particularly steady increasing trend in the high level of education group. Indeed, statistical time period comparisons show significant increases beyond 2019 levels in spring/summer 2021 in the high level group but not in the other two groups (p-values for CW 38–52 cannot be reported due to non-significant joint tests). Comparisons between 2022 and corresponding 2019 and 2021 time windows show significant increases compared to both years in both the high and medium level of education groups. In the low level of education group, means also increased significantly beyond 2019 (but not 2021) levels in 2022, and a p-value of 0.057 suggests a possible increase in the proportion of positive screens as well.

Looking at the whole observation period, the standardized proportions of those with a positive screen rose by 5.2 percentage points (from 5.5%) in the high level of education group, 4.6 (from 10.6%) in the middle group, and 4.8 percentage points (from 16.2%) in the low level of education group (not significant) between spring/summer 2019 and 2022 (Table 4).

While most studies on mental health in the COVID-19 pandemic in Germany examine its first months only (43), our results shed light on the development of symptoms of depression in the adult population until June 2022 and reveal two increases. Consistent with our finding of increased mean depressive symptom scores as well as positive screens between the last months of 2020 and spring 2021, i.e., during the second wave of infections, a German study reports lower subjective psychological wellbeing measured using a screening tool for depression in December 2020 compared to May and September 2020 (116). Also in keeping with our findings, a representative survey of the German resident adult population showed that a far larger percentage of the population found the overall situation “depressing” in the second lockdown than in the first (47).

A significantly higher proportion of positive PHQ-2 screens (45) and mean symptoms scores (59) in early 2021 compared to 2019 were also found in the German representative panel study (the “Socio-Economic Panel”). However, in contrast to our finding that symptoms of depression first increased in the second wave following an initial decline in the first pandemic months, scores and percentages were actually found to be lower in January/February 2021 than in April through June 2020 in the SOEP. Despite this discrepancy, the January/February 2021 proportion of positive screens in this other study is 12%, very similar to the September-December 2020 levels (12.1%) in our study (we do not have data from January and February 2021).

While we have no data on symptoms of anxiety and SRMH from before the pandemic or in its early stages, our findings of a potential increase in symptoms of anxiety and a clear decrease in SRMH between March 2021 and the end of 2021 are in keeping with the picture of worsening mental health following the onset of the second wave.

These changes occurred in the context of a second wave of infections much larger than the first, followed very quickly by a third wave and a fourth, very severe wave with only short periods of lower infection rates in between. Although vaccinations began at the end of 2020, measures to slow transmission were in place for much of this time, mortality rates were high, and hospitals were reported to have come dangerously close to their limits (20, 21, 23–25). While, again, our results do not allow for conclusions on causal relationships between infection rates, mortality, NPIs, or other pandemic factors and mental health, associations between mean PHQ-4 scores and “pandemic intensity” have been reported in a meta-analysis (117).

The sheer increased duration of the cumulative pandemic stressors may also explain potential pandemic-related changes in mental health later on (author?) (19). A resilient response is more likely in the face of brief stressors than in the face of more chronic stress (18). In general, longer-term experiences of lack of control and helplessness threaten mental health and may be particularly related to depressive symptoms (16). Reductions in protective factors such as social contact, leisure activities (19), and access to the full spectrum of health services (16) may also grow more harmful with longer durations. Finally, most mental disorders take time to develop and manifest with a prodromal phase (118, 119). While our study does not address the prevalence of any mental disorders, the individual symptoms that comprise these disorders might be subject to the same dynamics.

Our findings of a further increase in symptoms of depression and symptoms of anxiety between late 2021 and early 2022 resulting in by far the highest levels in our over three-year observation period lend further support to the assumption of a potential build-up of pressure on mental health. The pandemic context at this time was a fifth wave of infections driven by the omicron variant immediately following the fourth wave and reaching the largest ever peak in new infections in April 2022 (21, 22), and yet a suspension of most NPIs from the beginning of April (35). Another major and acute population-wide stressor in the final 4 months of our observation period was the war in the Ukraine beginning on February 24^th, 2022. It is of note that symptoms of depression and anxiety increased rather markedly in the January/February-centered estimate of 2022, which includes data until mid-March 2022, suggesting potential mental health impacts of the war (120) and emerging economic developments (11, 18). The fact that subsequent estimates remained elevated raises the possibility of developments beyond a short-lived reaction to a discrete event.

The absence of a decline in SRMH between late 2021 and early 2022 following its decline within 2021 suggests that increases in depressive and anxiety symptoms did not translate to people reporting worse overall mental health within this specific time window. Importantly, the temporal reference frame of these measures is very different: while the PHQ-2 and GAD-2 ask about the previous 2 weeks, SRMH has no reference time. Perhaps SRMH shows different dynamics in this particular time window as a more global and less acute measure. Also, mental health problems are known not to necessarily translate into poor SRMH in general (121). The present observation period is too short for conclusions about differences in dynamics between these indicators, but this would be interesting to analyze in longer time series.

Turning to subgroups, stratification by sex, age, and level of education in our uniquely long and continuous time series revealed increases in depressive symptom levels at different times after the onset of the second wave of infections in all groups. Increases in symptom levels were more pronounced in women than in men until the end of 2021. This finding is as expected based on previous literature [e.g., 41, 52, 53, 109] and considering factors such as a greater burden from increased care work among women (53) and increases in domestic violence (14, 16, 53). In 2022, however, men also showed a significant increase above 2019 levels, as well as above 2020 and 2021 levels (and, just like women, also increases in anxiety and declines in SRMH). This later increase may be due to new stressors or simply a delay in negative mental health developments.

While the sexes and the level of education groups all showed relatively similar overall trajectories, age groups differ in the shape of their time series after the onset of the second wave, suggesting that stressors and protective factors may differ by age in particular. In keeping with previous findings of mental health vulnerabilities among young adults in the pandemic (41, 49, 54–57), the youngest age group stands out in our study for its steep increase in depressive symptoms (and also symptoms of anxiety) at the end of 2021. Vulnerabilities in this group could be related to the transitional nature of young adulthood, the particularly great importance of social contact with peers when leaving the parental home (19, 51), and an overall greater disruption of life in this group (58) during the pandemic. On the other end of the age spectrum, those aged 65+ years stand out for early significant increases beyond pre-pandemic levels and a particularly constant trend of increase throughout the observation period from 9% positive screens in spring/summer 2019 to 18% in 2022 (standardized estimates). While most studies highlight risks among younger adults, a German study using primary care data found early increases of mental health diagnoses among those aged 80 and over (122), consistent with our results. A greater risk of severe disease and death from COVID-19 (16) may have resulted in greater stress and isolation throughout the pandemic (19) in this age group, with less relief from temporary suspensions of NPIs. While 45-64-year-olds do show worsening mental health, but only late 2021/early 2022, 30-44-years-olds stand out for somewhat steadier levels across indicators (except for one temporary increase in depressive symptoms), suggesting they may have been more resilient in the observation period.

The feared widening of disparities in mental health (16, 19, 50) by SES in the pandemic was not evident in our study, which looks at educational differences as one of the dimensions of SES. We found a similar overall trend in all education level groups, with an increase of about five percentage points in each group between spring/summer 2019 and spring/summer 2022. The high level of education group stands out for the greatest relative increase given baseline levels and in terms of how early and continuous increases are across the observation period. Previous international studies from other countries in the Organization for Economic Co-operation and Development (OECD) have also found greater increases in psychological distress in higher SES groups (52, 56, 69, 70). In Germany, for example, greater declines in life satisfaction during the pandemic have been reported for higher income individuals (57, 123). Discussed reasons include more working from home in this group (124), which has been shown to be linked to mental health declines in the pandemic (46). Moreover, this group may have experienced a more substantial change in lifestyle more generally (56), perhaps with concomitant greater expectations for the constant availability of resources (70). However, a complex set of risk factors is likely to be at play in all education groups. Occupational and financial difficulties were identified as particularly crucial for an increase in depressive and anxiety symptoms in Germany (46). Importantly, the established social gradient in the risk of depressive symptoms remains unchanged until the end of the observation period in our study, with twice the percentage of positive screens in the low as in the high level of education group in spring/summer 2022.

Three features of the study should be highlighted as strengths: (1) Continuous, representative data spanning 1 year before the outbreak of COVID-19 and over 2 years of the pandemic: While most of the existing literature on mental health developments during the COVID-19 pandemic in Germany covers limited time periods and focuses on the early phases of the pandemic only, we present results on the whole course of the pandemic until June 2022, including pre-pandemic data for depressive symptoms. (2) Development of a method for assessing trends at higher temporal resolution: A method for deriving robust month-by-month results from relatively small samples was developed. Using graphical representations of monthly moving estimates, multiple adjustments of the sample, and smoothing spline curves, we were able to produce graphic time series for the visual identification of trends which were nearly all verified by statistical time period comparisons. This demonstrates the feasibility of this approach to high-frequency mental health surveillance. (3) Examining developments over time both in mean scores and using scale cutoffs: The relevance of population means for public mental health in connection to Geoffrey Rose's ideas about prevention and health promotion at the population level has been previously discussed (84). Changes in the population symptom level are of interest irrespective of whether they result in more positive screens. The additional examination of positive screens permits conclusions on whether changes manifest in increases or reductions in cases of potential immediate clinical significance.

Limitations in the interpretation and evaluation of our findings include: (1) Time periods of observation and comparison: Because the time series on anxiety symptoms and SRMH span only about 16 calendar months during the pandemic and include no pre-pandemic data, observed developments cannot be contextualized temporarily and are more difficult to interpret than the longer time series for depressive symptoms (11 months pre-pandemic, 27 months during the pandemic). However, even for depressive symptoms, 11 months of pre-pandemic data are not sufficient to control for seasonal trends and long-term secular trends. By providing more context, longer time series would also facilitate our understanding of how meaningful the observed magnitudes of change are. (2) Gaps in data collection: Data collection was interrupted four times for depressive symptoms and twice for anxiety symptoms and SRMH. Also, for some months the number of observations was low. These monthly periods with fewer than 1,000 observations were minimized by using months ranging from the middle of one calendar month to the following calendar month. Additionally, predictions on a three-month window were still made when only 2 months were included. Thus, results assigned to the central months might be biased toward the first or the third months in the window or just averages of the first and the third month. Also, one gap in the GEDA study was filled with data from the COVIMO study, which had a comparable design but a different overall framing and focus. However, we checked for and did not find systematic, study-related differences. (3) Representativity of the sample for the general population and statistical power for subgroups: The response to population-based telephone surveys typically varies systematically by sociodemographic factors (125). In particular, younger individuals and those with lower levels of education are underrepresented in our study. We used weighting factors to account for population structure, but the small number of cases in these groups may mean that possible changes over time within a group and differences from other groups might not be detected. In order to reliably achieve statistical significance in the subgroup analyses, larger sample sizes within certain subgroups would be required. Mentally ill and especially severely mentally ill individuals may also be less likely to participate (126, 127), a bias for which we cannot correct. Similarly, we cannot rule out the possibility that willingness to participate in a survey conducted by a governmental public health institute during the pandemic was related to subjective pandemic-related psychological distress. (4) Measurement and scaling: Using short versions of screeners to measure depressive and anxiety symptoms results in a restricted range of scores compared to the full questionnaires. This might have decreased the likelihood of detecting changes compared to the respective long versions with more items. Additionally, the PHQ-2 and the GAD-2 as well as its long versions PHQ-8, PHQ-9, and GAD-7 measure the severity of depressive and anxiety symptoms on an ordinal scale. However, validation supports the interpretation as a metric scale (86, 128). According to this assumption, distributions of PHQ-based measures are commonly described by means of sum scores [e.g., 46, 57, 73, 114]. Furthermore, information based on self-report can be subject to recall-bias and social desirability (129). While telephone surveys have the advantage of not limiting the sample to those who are able to complete a survey online or via an app, social desirability may represent more of a confound with this survey mode (130).