Does retirement trigger depressive symptoms? A systematic review and meta-analysis

Aims Retirement is a major life transition that may improve or worsen mental health, including depression. Existing studies provide contradictory results. We conducted a systematic review with meta-analysis to quantitatively pool available evidence on the association of retirement and depressive symptoms. Methods We applied PRISMA guidelines to conduct a systematic review and meta-analysis to retrieve, quantitatively pool and critically evaluate the association between retirement and both incident and prevalent depression and to understand better the potential role of individual and contextual-level determinants. Relevant original studies were identified by searching PubMed, Embase, PsycINFO and the Cochrane Library, through 4 March 2021. Subgroup and sensitivity meta-analyses were conducted by gender, study design (longitudinal v. cross-sectional studies), study quality score (QS) and considering studies using validated scales to diagnose depression. Heterogeneity between studies was evaluated with I2 statistics. Results Forty-one original studies met our a priori defined inclusion criteria. Meta-analysis on more than half a million subjects (n = 557 111) from 60 datasets suggested a protective effect of retirement on the risk of depression [effect size (ES) = 0.83, 95% confidence interval (CI) = 0.74–0.93], although with high statistical heterogeneity between risk estimates (χ2 = 895.19, df = 59, I2 = 93.41%, p-value < 0.0001). Funnel plot asymmetry and trim and fill method suggested a minor potential publication bias. Results were consistent, confirm their robustness and suggest stronger protective effects when progressively restricting the included studies based on quality criteria: (i) studies with the highest QS [55 datasets, 407 086 subjects, ES = 0.81, 95% CI = 0.71–0.91], (ii) studies with a high QS and using validated assessment tools to diagnose depression (44 datasets, 239 453 subjects, ES = 0.76, 95% CI = 0.65–0.88) and (iii) studies of high quality, using a validated tool and with a longitudinal design (24 datasets, 162 004 subjects, ES = 0.76, 95% CI = 0.64–0.90). We observed a progressive reduction in funnel plot asymmetry. About gender, no statistically significant difference was found (females ES = 0.79, 95% CI = 0.61–1.02 v. men ES = 0.87, 95% CI = 0.68–1.11). Conclusions Pooled data suggested that retirement reduces by nearly 20% the risk of depression; such estimates got stronger when limiting the analysis to longitudinal and high-quality studies, even if results are affected by high heterogeneity. As retirement seems to have an independent and protective effect on mental health and depressive symptoms, greater flexibility in retirement timing should be granted to older workers to reduce their mental burden and avoid the development of severe depression. Retirement may also be identified as a target moment for preventive interventions, particularly primary and secondary prevention, to promote health and wellbeing in older ages, boosting the observed impact.


Introduction
Globally, the proportion of older adults (>60 years) is estimated to almost double between 2015 and 2050, from about 12% to 22% (United Nations, 2015). As the world population ages, it is critical to promote and support healthy ageing processes to improve societal wellbeing and limit its clinical and economic burden (Dietz et al., 1987). The prevalence of late-life depression is 7% among the general older population (Mccall and Kintziger, 2013) and accounts for 5.7% of years lived with disability in those over 60 years old (Killinger, 2012). Depressive symptoms are often overlooked and untreated in older populations, are associated with psychosocial and cognitive decline (Nelson, 2001), and result from a complex interaction between psychological, biological and social factors (Alexopoulos, 2019). One significant determinant that could play a role is transitioning into retirement, whose timing, decision and consequences could be influenced by depressive symptoms, such as loneliness and hopelessness, acting as moderators (Gum et al., 2017;Segel-Karpas et al., 2018). Retirement is a major life transition that results in social and psychological transformations (Bosse et al., 1991), which pose both threats and opportunities for mental health. On the one hand, as a potentially stressful life event, retirement can have adverse repercussions on individual physical and psychological wellbeing (Portnoi, 1981). People lose access to social networks, lifestyles and daily routines, as well as potential stimulation, activity and purposes. Conversely, retirement may reduce work-related exposures and improve physical and mental health through complex mechanisms. These could include an increase in social support and in the time available for leisure and healthy activities, and disconnection from work-related stressors (Van Der Heide et al., 2013;Eibich, 2015). These positive health effects were particularly observed among retirees from strenuous jobs (Belloni et al., 2016;Blake and Garrouste, 2019;Ardito et al., 2020;Carrino et al., 2020;Fleischmann et al., 2020). Therefore, as we reported in previous research (Vigezzi et al., 2021), health behaviours changes (e.g. changes in smoke habit, alcohol consumption, physical activity, time use, social interactions) appeared to be among the most relevant mediators of retirement consequences on olders' health, affecting life years after the withdrawal from work (Lang et al., 2007;Vahtera et al., 2009;Celidoni and Rebba, 2017). Nonetheless, current findings are inconclusive. As it has been previously conceptualised (Van Solinge, 2007), health consequences of retiring are influenced by the employment history, the job characteristics (Ardito et al., 2020) and the transition to retirement itself, as well as by the availability of socioeconomic resources at the time of retirement and, last but not least, by individuals' characteristics and appraisal of stress-generating life events (Van Solinge, 2007;Augner, 2018). As a result of such a complex conceptual model, no conclusive evidence exists on the harm-benefit health balance of retirement. In particular, both older and more recent studies have shown contradictory results on the impact of retirement on mental health outcomes (Bossé et al., 1987;Salokangas and Joukamaa, 1991;Gall et al., 1997;Drentea, 2002;Mein et al., 2003;Buxton et al., 2005;Gill et al., 2006;Mojon-Azzi et al., 2007;Van Solinge, 2007; Alavinia and Burdorf, 2008;Vahtera et al., 2009;Jokela et al., 2010;Westerlund et al., 2010).
Here we performed a systematic review and meta-analysis to identify the overall association of retirement with depression. As a second aim, we sought to identify potential modifying individual-and contextual-level factors.

Methods
We followed the Prepared Items for Systematic Reviews and Meta-Analysis (PRISMA) (Liberati et al., 2009;Page et al., 2021) and the Meta-analysis Of Observational Studies in Epidemiology (MOOSE) guidelines (Stroup et al., 2000).

Search methods and inclusion criteria
Studies identified searching the electronic databases PubMed/ Medline, Embase, PsycINFO and the Cochrane Library through 4 March 2021 were included. The search strategy was first developed in Medline and then adapted for use in the other databases (online Supplementary Table 1). Briefly, we used a combination of free text and exploded MeSH headings, identifying: (i) the concept of 'retirement/transition to retirement' and (ii) 'depression/ depressive symptoms'. Further studies were retrieved from manual reference listing of relevant articles and consultation with experts in the field. Details on inclusion and exclusion criteria are reported in Table 1, according to the Population, Exposure, Comparison, Outcomes and Study design (PECOS) framework (Brown et al., 2006;Higgins and Green, 2013). Our inclusion criteria were limited to those studies: reporting original data from quantitative analysis, providing effect sizes (ESs) of the association between retirement (exposure of interest) and depression (outcome of interest); natural experiments (Stuckler, 2017;Ronchetti et al., 2020), observational studies with prospective, retrospective and cross-sectional designs; and written in English. An extensive definition of retirement and retirement status was used: depending on study design, both retired status and transition to retirement were included as exposure of interest; we considered all retirement types, apart from retirement only for disability, which was excluded. Depression-related outcomes of interest included: depressive symptoms, Diagnostic and Statistical Manual of Mental Disorders (DSM), or International Classification of Diseases scale (ICD)-based diagnosis as major depressive disorder and persistent depressive disorder. We excluded opinion papers (i.e. editorials, narrative reviews, commentaries and letters to the Editor) not providing original data. Systematic reviews were also excluded but screened to retrieve relevant original studies. The review's protocol was drafted and approved by authors before conduction (not archived on public databases). retrieved for a second screening. At both stages, disagreements among reviewers were resolved by consensus and by consulting a third senior author (A.O.) when disagreement persisted. Data were independently extracted by two authors (V.G. and G.P.V.), supervised by a third author (A.O.), using an ad-hoc developed data extraction spreadsheet. The data extraction spreadsheet was piloted on ten randomly selected papers and modified accordingly. Data extraction included: full reference details, country of study conduction, study design, study setting, study population details, sample size, exposure details, outcomes of interest, including validated assessment tools for depression, and quantitative results, including ESs and corresponding confidence intervals (CIs). Corresponding authors were contacted by e-mail in case of incomplete data. Quality appraisal of included studies was carried out applying the 14-item scoring system developed by Shim et al. for population-based studies on retirement as a risk factor (Shim et al., 2013). As determined by consensus following the review methodology literature, we consider of high quality the studies with at least ⩾75% of the highest score.

Data pooling and meta-analysis
We performed descriptive analysis to report and pool the characteristics of included studies using ranges and average values. With regard to the pre-specified outcomes of interest, we would expect variability between studies, e.g. by study design and population. We, therefore, applied random-effects meta-analyses to acquire estimates of the association between retiring and risk of depression/depressive symptoms, rather than to assume a single true value in a fixed-effects approach (Higgins and Green, 2013). Moreover, a random effect model is highly recommended when high heterogeneity is expected or detected. Pooled ESs were calculated as odd ratios (ORs) (Ter Hoeve et al., 2020). When the included studies reported ESs as regression beta coefficients with corresponding standard errors (S.E.s), we mathematically converted them into ORs with corresponding CIs (Bland and Altman, 2000;Hailpern and Visintainer, 2003). We also included studies that reported ESs as χ 2 or ρ correlation coefficients with corresponding total sample sizes or as mean differences with sample sizes and corresponding correlations. Heterogeneity was assessed using the I 2 statistic (see online Supplementary Table 3 for details) and visual inspection of funnel plots. We performed sensitivity analyses progressively limiting meta-analysis to: (i) high-quality studies; (ii) high-quality studies using validated scales to diagnose depression; (iii) high-quality longitudinal studies using validated scales to diagnose depression. Moreover, we conducted a subgroup meta-analysis by gender strata and study design.
We assessed publication bias with funnel plot visual inspection (Higgins et al., 2011) and the Begg and Mazumdar (1994) and Egger et al. (1997) tests. A 'trim and fill' method was used if publication bias was detected Gianfredi et al., 2020) to estimate potential missing studies which contribute to the funnel plot's asymmetry (Sutton et al., 2000). This method assumes that the most extreme ES studies have not been reported, biasing the overall ES estimates (Shi and Lin, 2019). Meta-analyses were conducted using ProMeta3 ® (Internovi, Milan, Italy) software.
Nineteen (46.3%) studies were longitudinal studies; their follow-up time ranged from 2 to 25 years, with most of them (n = 12, 60.0%) having less than 10 years of follow-up. Most of the longitudinal analyses were derived from the Survey on Health and Ageing and Retirement in Europe (SHARE, n = 8), followed by the Health and Retirement Study (HRS, n = 4) and the Gaz et Electricité cohort study (GAZEL, n = 3). Twenty-one  (Schwingel et al., 2009). Overall, sample sizes of included studies ranged from 30 (Farakhan et al., 1984) to 245 082 subjects (Olesen et al., 2015) (mean: 14 423 subjects, median: 4189 subjects); longitudinal studies sample sizes ranged between 458 and 245 082 subjects (mean: 21 884 subjects, median: 7134 subjects). The majority of included study populations' age ranged between 45 and 80 years (n = 38, 92.7%). One study included only males (Tuohy et al., 2005) and one only females (Sheppard and Wallace, 2018). Details on study populations are reported in Table 2, which also reports information on the type of retirement, available for 76% of studies, and outcomes assessment. More than ninety per cent of included studies used validated tools to diagnose depression-related outcomes, including the Center for Epidemiologic Studies Depression scale (CES-D) in 17 studies (41.5%), the Euro Depression-scale (EURO-D) in eight studies (19.5%), the Geriatric Depression Scale (GDS) in three studies (7.1%) and the Zung Self-rating Depression Scale (ZSDS) in two studies (4.9%). The International Classification of Diseases-10 (ICD-10) was used to identify depression-related conditions in three studies (7.1%). The Patient Health Questionnaire-8 (PHQ-8), the Hospital Anxiety and Depression Scale (HADS), the Composite International Diagnostic Interview (CIDI), the Depression Adjective Check List (DACL) and the Clinical Interview Schedule-Revised (CIS-R) were used in only one study each (Table 2). Three studies (7.1%) used non validated tools to identify depression-related outcomes (Mojon-Azzi et al., 2007;Sheppard and Wallace, 2018;Anxo et al., 2019). Included studies' quality score (QS) is also reported in Table 2. The mean QS was 15.5/20. The lowest QS was 6 (Farakhan et al., 1984), whereas the highest was 19 (Schwingel et al., 2009;Olesen et al., 2015;Park and Kang, 2016;Van Den Bogaard and Henkens, 2018). Question 7 [Is retirement a main effect, co-variable, confounder, or interaction in the study?] (n = 15) and Question 13 [Was the loss to follow-up appropriately addressed and/or adequately described in the study?] (n = 11) reported the lowest scores (for details on quality appraisal, see online Supplementary Table 2).

Retirement and depression: quantitative reporting and meta-analysis
Quantitative pooling of effect estimates was conducted on a total of 557 111 subjects from 60 different databases. Overall, the pooled ES for the risk of depression when retired is 0.83 (95% CI = 0.74-0.93, p-value = 0.001, Fig. 2a), with high statistical heterogeneity (χ 2 = 895.19, df = 59, I 2 = 93.41, p-value < 0.001). The funnel plot resulted slightly asymmetrical at visual inspection, showing a low potential for publication bias, not confirmed by Egger's linear regression test (Intercept 0.53, t = 0.78, p-value = 0.439). Moreover, the ES change after the trim and fill method was minor [0.84 (95% CI = 0.75-0.94)], and two studies were trimmed in the lower right quarter of the funnel plot (Fig. 2b), suggesting few studies of poor quality could be missing.
Results of the sensitivity and subgroup analyses are summarised in Table 3. We performed a sensitivity analysis, progressively increasing the quality of included studies in order to test our overall results' consistency. First, we limited the analysis to studies of the highest quality (QS equal or higher than 15): 47 datasets and 485 092 subjects were included in the meta-analysis, reporting a consistent statistically significant association between retirement and decreased risk of depression (ES = 0.79, 95% CI = 0.68-0.91, p-value = 0.001, online Supplementary Fig. 1a). Then, we limited the analysis to studies with high QS and using validated assessment tools to diagnose depression. In this analysis, 44 datasets were included, for a total of 239 453 subjects, strengthening the significant association between retirement and decreased risk of depression (ES = 0.76, 95% CI = 0.65-0.88, p-value < 0.001, online Supplementary Fig. 1b). Finally, only studies (i) with a QS equal or higher than 15, (ii) using validated assessment tools to diagnose depression and (iii) with a longitudinal study design were included. We report a statistically significant association between retirement and depression (ES = 0.76, 95% CI = 0.64-0.90, p-value = 0.001, Fig. 3a) based on 24 datasets and 162 004 subjects. High statistical heterogeneity and slight visual asymmetry of the funnel plot were observed at each step of the analysis (Table 3), with the exception of the last one restricted to longitudinal studies of the highest quality, when estimated ES did not change with the trim and fill method.
Gender-strata meta-analyses are reported in online Supplementary Fig. 2b and 2c. When only considering women, the analysis included 21 datasets and a total of 219 655 subjects, reporting no statistically significant association between retirement and depression (pooled ES = 0.79, 95% CI = 0.61-1.02, p-value = 0.074) and high heterogeneity (Table 3). About men, the analysis included 20 datasets, for a total of 223 840 participants, reporting a pooled ES of 0.87 (95% CI = 0.68-1.11, p-value = 0.258) and high heterogeneity between studies (Table 3). In both cases, evidence of publication bias was suggested by funnel plot.

Discussion
Pooled data from 41 original studies and more than half a million subjects suggested that retirement or transition to retirement reduce by nearly 20% the risk of depression or depressive symptoms; such estimates remain consistent when limiting the analysis to longitudinal and high-quality studies. Before interpreting our findings further, we must account for the considerable heterogeneity among the included studies, which might limit the generalisability of pooled effect estimates. To overcome this and test the results level of strength, we first applied a random-effect model. Secondly, we conducted sensitivity and stratified meta-analyses by study design and QS. The reasons behind the high level of heterogeneity among the included studies are to be explored in light of, on one side, the wide variety of studies' designs, settings and populations, definitions and methodological quality and, on the other side, of the complex, multi-determinant and multi-mediator relationship between the process of retirement and mental health and wellbeing (Pesaran et al., 1999;Rabe-Hesketh and Skrondal, 2008;Behncke, 2012;Oksanen and Virtanen, 2012;Insler, 2014;Eibich, 2015). We could not retrieve further evidence on the reasons: even excluding one dataset at a time in the meta-analysis to identify potential outliers, heterogeneity persisted (online Supplementary Table 3). However, sensitivity analyses confirmed the results' consistency.
Despite half of the retrieved studies being cross-sectional, which did not allow us to explore causality, they accounted for less than one-third of included subjects. Another limitation to consider is that duration of retirement was not reported in most studies, so we could not differentiate among the potential risk of depression for short-or long-term exposure to retirement. A subgroup analysis considering the work before retiring was not possible since only two included studies stratified results for this variable (Belloni et al., 2016;Kolodziej and García-Gómez, 2019). Lastly, even if most of the analysed data came from administrative databases or surveys designed for other purposes, some studies had small sample sizes with poor precision in effect estimates.
To the best of our knowledge, this is the first systematic review and meta-analysis pooling all original studies investigating the association of retirement with prevalent and incident depression. We used a comprehensive range of databases and search terms to maximise the number of studies retrieved and minimise the chance of publication bias. Besides, further studies were retrieved from the reference listing of relevant articles. Such a comprehensive and rigorous summary of the available evidence offers several meaningful insights, valuable to plan, implement and evaluate public health and preventive strategies, public policies, as well as future avenues of research.
Despite the well-known assumption that considers retirement as a potentially stressful life event (Kremer, 1985;Ekerdt, 1987;Salokangas and Joukamaa, 1991), one of our review's critical findings is that retiring does not necessarily harm an individual's mental health but possibly decrease the risk of depression, as a balance of contextual and individual-level variables impact on such association. In conceptual frameworks proposed in the ageing research literature (Van Solinge, 2007), these variables were categorised into: (i) characteristics of the retirement transition, (ii) characteristics of the job, (iii) access to resources, (iv) individual appraisal and (v) gender.
Characteristics of the transition refer to the type and conditions of retirement, which were available in 76% of included studies. For instance, we report different impacts on depression between voluntary and involuntary retirement, with the more considerable impact of the latter (Mosca and Barrett, 2016), suggesting elements of desirability and degree of control might play a role in the association (Van Solinge, 2007).
There is extensive literature on how employment characteristics influence health after retirement (Hernberg, 2001;Robroek et al., 2013;De Wind et al., 2014Soh et al., 2016;Ardito et al., 2020). As emerges from original data, among job characteristics, employment history, time pressure, workload and physical demand may impact the risk of mental health disorders' onset after retirement (Thoits, 1983;Shultz et al., 1998).
With reference to resources, access to social and financial resources around retirement might compensate and mitigate the impact of lifestyle changes and the psychological consequences of retiring. We reviewed data where the risk of depression at retirement is differentially distributed by household socioeconomic status (Arias-De La Torre et al., 2018), marital and family relations (Park and Kang, 2016), social engagement (Sabbath et al., 2015;Shiba et al., 2017): as the studies suggest, reliable financial resources, social networks and marriage can mitigate negative health repercussions of retirement (Deeg and Bath, 2003).
Concerning individual appraisal, personality characteristics influence the meaning assigned to retirement and the ability to cope with this change. Negative expectations and fears about retirement are more likely related to adverse repercussions on individuals' wellbeing ( Barnes-Farrell, 2003). Moreover, having confidence in coping with changes determines fewer difficulties in adjusting to retirement (Van Solinge and Henkens, 2005).
Regarding gender, differences in primary role between women and men, at home and work, respectively, could explain differences in adapting to the event and in health outcomes by gender (Moen, 1996), but need to be further explored.
Overall and sensitivity analyses results are consistent with other reviews on the topic. Van Der Heide et al. (2013) focused on mental health and antidepressant use in longitudinal studies. They registered an improvement in mental health shortly after retirement, possibly linked to work pressure reduction, even if with gender differences. Schaap et al. (2018) analysed the health effects of an exit from work across different socioeconomic groups. They found out that, despite significant heterogeneity, withdrawal from work had more positive effects among employees with a higher socioeconomic status than with a lower position. On the contrary, a systematic review was previously conducted on the effects of working or volunteering beyond statutory retirement ages on mental health by Maimaris et al. (2010); they suggested that, through the mechanism of maintaining a productive societal role with a continued income and social support, working beyond retirement age might be beneficial for mental health. Nevertheless, the benefits were not universal, but they varied greatly by lifestyles, self-esteem and socioeconomic status.

Implications for public health policies and practice
Regarding public health and preventive strategies, we demonstrated that, besides other factors influencing the risk of late-life depression, transition to retirement, as a life event that almost the entire population experience at some point (Clark and Oswald, 1994), has an independent effect in itself. The transition is differentially distributed by contextual and individual-level characteristics and, as such, could be identified as a target point for mental health prevention, including both primary and secondary interventions. We claim that primary prevention interventions, aimed at promoting healthy lifestyles and supporting social roles, could be effectively directed towards subjects who do not benefit from retirement flexibility and its protective effect on short-and long-term risk of late-life depression (Smit et al., 2006;Barnett et al., 2012;Lindwall et al., 2017). As life-course transitions tend to bring along lifestyle changes, synchronising them with public health interventions might be a successful approach (Ben-Shlomo and Kuh, 2002;Werkman et al., 2010;Heaven et al., 2013Heaven et al., , 2016. Along the same lines, secondary prevention, including early depressive symptoms detection, could be effectively targeted to older workers still employed, with particular reference to interventions implemented at the primary care level (Okereke et al., 2013;Costantini et al., 2021).
About public policies, our data complement the accumulating evidence on the impact of pension reforms on health and mental health (Eibich, 2015;Carrino et al., 2020), suggesting that older workers should be granted greater flexibility in the timing of retirement in order to reduce their mental burden and avoid the development of severe depression. As many countries are implementing budget reductions to social welfare (Hall and Soskice, 2001), it is crucial to retrieve solid evidence on how different retirement policies might impact healthy ageing to balance money saved from cuts to pension systems with direct and indirect costs passed onto healthcare and social support systems. Although our review only focuses on mental health, the burden of mental health and, in particular, of depression is known to be associated with the burden of chronic physical conditions that significantly affect people's quality of later life, their demands for healthcare and other publicly funded services, generating significant societal consequences (Bech et al., 2011;Hughes et al., 2011;Rechel et al., 2013).

Recommendations for future research
Concerning research, it clearly emerges from our analysis that, in order to reduce heterogeneity and accumulate solid evidence, shared methodological standards and definitions should be followed in the future. More extended longitudinal studies should be preferred so as to reduce inverse causality issues and might help disentangle and quantify the different components that mediate the effects of retirement on the risk of depression and its determinants and monitor such association's temporal evolution. It would also be necessary to further differentiate between contextual and individual characteristics to adapt coping strategies at the public health and clinical levels. Special attention should be paid to health inequalities to investigate better socioeconomic status indicators role in the relationship between retirement and health (Adler et al., 1994) and address the impact of specific policies focusing on health promotion for disadvantaged groups (Rechel et al., 2013). Stratifying results by job and retirement type and by socioeconomic status might be helpful to fill the gaps in current literature.

Conclusions
As a matter of fact, despite current trends in extending working lives, life expectancy after regular retirement is projected to grow faster than increases in the pension age, reaching 20.3 years for men and 24.6 years for women in 2050 (OECD, 2011). Therefore, from a societal, welfare and public health perspective, it is essential to invest in 'third age' health and wellbeing (Crimmins, 2015). In a progressively ageing society, strengthened efforts are needed to make health interests count in welfare and pension policies and promote health protection after retirement (Moen, 1996). We call for a coordinated advocacy action to identify retirement as a gateway for healthy lifestyles and an entry point for mental health prevention. Multidisciplinary collaborations between social sciences, public and community health, preventive medicine and psychiatry could be fruitfully put in place to generate much-needed evidence on the determinants, mediators and effect modifiers of the association between retirement and depression, as well as to design preventive interventions targeting older workers.
Supplementary material. The supplementary material for this article can be found at https://doi.org/10.1017/S2045796021000627 Data. The datasets supporting the conclusions of this study are available from the corresponding author upon request.