Hospitalization Risk Before and After Employment Among Adults With Schizophrenia, Bipolar Disorder, or Major Depression

Clinicians, family members, and people with serious mental illness worry that competitive employment is stressful and consequently may precipitate negative outcomes, such as psychiatric hospitalization (1). Conversely, others have noted the salutary impact of working on health and general well-being (2). Although no studies have shown an association between working and increased psychiatric hospitalization, some evidence suggests an association between working and reduced hospitalization (3).

A recent systematic review identified six prospective studies examining the relationship between employment and hospitalization rates (4). Only one used a cross-lag statistical approach to disentangle the cause-effect relationship between employment and hospitalization risk reduction. A secondary analysis of a multisite controlled trial of supported employment found that employment of at least 90 days at the same job reduced the odds of hospitalization by 18% in the final six months of the two-year study (5). Recently, another randomized controlled trial of supported employment for people with serious mental illness found lower levels of inpatient psychiatric treatment for participants receiving supported employment than for participants in a control group at five-year follow-up, but not at the two-year follow-up (6).

Understanding the relationship between employment and psychiatric hospitalization is important for both economic and humanitarian reasons. The costs of psychiatric hospitalization are well documented (7,8). In addition, hospital care may prevent community integration and thus is a barrier to fulfilling the treatment goals of most people with serious mental illness (9–11).

The Mental Health Treatment Study (MHTS) data are well suited to revisit the relationship between employment and hospital service use. MHTS researchers obtained self-reported data on employment outcomes and health service utilization for more than 2,000 working-age unemployed adults enrolled in the study between 2006 and 2008 in 23 locations in the United States (12). In this secondary analysis, we hypothesized that participants who were employed would be less likely than those who remained unemployed to experience psychiatric hospitalization in the subsequent three months and that those who worked in two successive quarters would have greater reductions in psychiatric hospitalization risk than workers who recently gained work, recently lost work, or never worked.

Methods

Sample

This study is a secondary analysis of the MHTS, a multisite, randomized controlled trial testing three services—supported employment, systematic medication management, and other behavioral health services—plus provision of complete health insurance coverage (all out-of-pocket expenses were covered by the study) among working-age, unemployed individuals who were receiving disability benefits (13). Potential enrollees were Social Security Disability Insurance (SSDI) beneficiaries between 18 and 55 years of age inclusive, with a primary psychiatric diagnosis of schizophrenia or mood disorder, residing within postal zip codes typically served by the study sites and able to give consent. Potential participants were excluded if they resided in a custodial setting, had a legal guardian, or had a life-threatening physical illness. Potential participants were also excluded if they had received supported employment from the study sites prior to recruitment or worked at a competitive job within 30 days prior to study enrollment. The final sample included 2,238 SSDI beneficiaries (13). After removal of those who dropped out early, died, or had unusable records, the final analytic sample consisted of 2,055 participants.

Procedures

Research assistants conducted interviews at baseline and at three-month intervals during follow-up. The baseline interviews were face to face, and the quarterly follow-up interviews were mostly by telephone. During follow-up, participants reported whether they had started, ended, or continued employment and reported similar information for hospitalizations. The interview identified the calendar month in which these events began, but not the exact date.

Baseline Measures

Sociodemographic information.

Baseline characteristics included age, gender, race, primary language spoken, educational status, and marital status.

Diagnosis.

Diagnosis of axis I disorder was determined from Social Security Administration (SSA) records indicating the official cause of disability. A formal diagnostic assessment, typically conducted within the first month of study participation, used the Structured Clinical Interview for DSM-IV (SCID) to assess axis I disorders among participants randomly assigned to the treatment group in the parent study. Concordance between the SCID diagnoses and the SSA diagnostic categories was high (88%) (13).

Previous work history.

Participants were asked at the baseline interview if they had ever worked for pay in the past two years. Adjusting for previous work history helped to isolate the effect of returning to work.

Alcohol and drug use.

Alcohol and drug use were measured by the Addiction Severity Index (ASI) (14). A six-item alcohol composite score was calculated based on frequency of alcohol use, frequency of intoxication, and perceived alcohol-related problems over the previous 30 days. A 13-item ASI drug composite score was calculated on the basis of frequency of use of 11 types of drugs and perceived overall drug-related problems over the previous 30 days (15).

Health status.

The 12-item Short Form Health Survey (SF-12) is a brief self-reported health assessment with two subscales measuring physical and mental components of functioning (16). The physical component of the SF-12 comprises the following domains: physical functioning, physical role limitations, bodily pain, and general health. The mental component of the SF-12 includes the following domains: vitality, social functioning, social role limitation, and mental health.

Previous psychiatric hospitalizations.

Participants reported psychiatric hospitalizations for the year prior to baseline, which served as a potential covariate to adjust for individual differences in hospital service utilization.

Measures During Follow-Up

Quarterly work status during follow-up.

The primary covariate of interest was quarterly work status defined as at least one day of paid employment during a three-month period, as reported during the quarterly interviews. Work status is nearly universally used as a measure of employment outcome in vocational studies (17) because of its ease of measurement and its robust correlation with other, more complex measures of employment outcome (18). The validity of self-reported employment status, especially for short-term recall periods, has been shown in prior research (19). Job stability referred to participants’ employment status in two consecutive three-month periods, defined as either nonworker (not working during both three-month periods), lost job (working during first but not second period), gained job (not working during first and working during second period), or maintained employment (working in both three-month periods, nearly always the same job) to determine the impact of developing a work routine.

Psychiatric hospitalizations.

Psychiatric hospitalization, determined by self-report during quarterly interviews, served as the primary outcome of interest. Participants reported hospitalization events, each event’s main cause (mental health; physical health; and alcohol, drug, or other problem) and the number of nights in the hospital. The outcome measure was limited to psychiatric hospitalizations (that is, for mental health problems).

Validity of patient self-reported service use has been mixed (20). Several studies have found that respondents with serious mental illness underreport hospital days (21,22). We assumed that respondents would be more accurate reporting hospital episodes than specific number of days hospitalized. Moreover, shorter time frames (three months or less) greatly reduce recall bias (21,22), especially for “self-reported inpatient hospitalizations, which tend to be rare and highly memorable” (20).

Statistical Analyses

A total of 1,884 (92%) participants completed the final (two-year) follow-up interview: 477 (93%) of 514 participants who were coded as working in the preceding quarter and 1,407 (91%) of 1,541 participants who were coded as not working in the preceding quarter, a nonsignificant difference. When participants missed a quarterly interview, they were asked to report any hospitalizations and employment that occurred since the previous interview. Thus subsequent interviews filled in gaps in both hospitalization and employment information during the missing periods for the first 21 months of follow-up. The missing data for the final quarter were imputed by using procedures that combined traditional methods for hot-deck imputation with modern model-dependent chained parametric procedures (23). Sensitivity analyses of the main study outcomes suggested that the imputation procedures did not affect any conclusions regarding statistical significance (13).

After preliminary data examination through basic descriptive statistics and graphs to examine linearity, normality, and outliers, we performed several longitudinal analyses of the cohort of working-age unemployed adults with severe psychiatric disorders. We used multilevel modeling approaches to capitalize on the hierarchical structure of the repeated-measures data (24,25). The multilevel models control for within-person correlation over time by determining a random intercept, or baseline likelihood of hospitalization, for every person separate from the between-person fixed effects of the covariates (25). Multilevel modeling allows estimation of both group-level and individual-level effects (both between-person and within-person effects in our case); accounts for within-person correlation due to repeated measures; and partitions total variance in outcomes into components: those attributable to between-person difference, within-person observation difference, and residual error term. Because of a skewed distribution with many zeros and because of more accurate recall of episodes than days of hospitalizations, we dichotomized psychiatric hospitalization within each three-month period. We used SAS proc glimmix to model these dichotomous outcomes with random effects logistic regression that included a random intercept for each subject (26).

We described temporal relationships between work activity and psychiatric hospitalization in three-month periods. Temporal separation is one requirement for causal inference, so we examined the relationship in each three-month period between employment status in the previous three-month period and the likelihood of hospitalization in the current period. To reduce the likelihood that observed effects were due to enhanced services in the intervention group, the analyses controlled for experimental-arm allocation. Other statistical controls included nonlinear time trends; psychiatric hospitalization status in the previous three-month period, because prior psychiatric hospitalization may be a marker of condition severity; and variables associated with worker status during the study (age, illicit drug use, primary language, physical health, and work history). Consistent assessment intervals made these data appropriate for this type of longitudinal analysis, in which we controlled for the effect of a participant’s employment status on subsequent hospitalization as a within-person effect.

We also compared the separate effects for losing, gaining, and keeping work compared with never having worked (that is, the interaction of having worked in the previous quarter [“lagged work”] with working in the current quarter). In further exploratory analyses, we tested whether the work-hospitalization relationship was sensitive to previous psychiatric hospitalization status (that is, the interaction of having been hospitalized and having worked in the previous quarter); whether the work-hospitalization relationship was sensitive to treatment arm (that is, the interaction of lagged work and treatment arm); and, to replicate previous cross-sectional research, whether inferences from the lagged work–hospitalization relationship presented in the main models were similar to the relationship between worker status and hospital service use measured in the same quarter. In all models, baseline levels of the outcome were allowed to vary across participants. Significance levels were set at .05 (two-tailed) for hypothesis testing.

Results

Table 1 summarizes baseline characteristics of the 2,055 adults with serious mental illness in the total sample and of those who worked (N=1,038, 51%) or did not work (N=1,017, 49%) during the study. Because of the exclusion criterion, no participants were working at baseline. Among participants who ever worked during the study, the mean±SD hours worked per week in the “main job” (the job in which the participant worked the most hours) was 9.51±7.84 and the rate of pay was $9.71±$4.23 per hour. Over the course of the study, 433 (21%) of the 2,055 participants were hospitalized at least once for a psychiatric reason.

Table 2 summarizes longitudinal patterns for psychiatric hospitalization in relation to lagged employment status. For the total sample during the follow-up period, there were 675 three-month periods with at least one psychiatric hospitalization. During the second year of the follow-up period, employment in the previous three-month period predicted participants’ psychiatric hospitalization status.

Table 3 shows the main results of statistical modeling. The odds ratio for lagged work status was .65, indicating that employment predicted significantly fewer psychiatric hospitalizations in the subsequent three-month period, even after the analysis adjusted for many potential confounders.

Table 4 presents a comparison of the effect of losing, gaining, or maintaining employment over two consecutive three-month periods with remaining unemployed during those periods. Participants who maintained employment (those who were working during two consecutive three-month periods) were significantly less likely to experience psychiatric hospitalization than participants who remained unemployed. The likelihood of psychiatric hospitalization did not significantly change among those who lost or gained work compared with those who did not work in the previous three months.

The observed association between employment and psychiatric hospitalization risk was further explored through a series of supplementary analyses. These tested alternative explanatory models for predicting hospitalization by examining hospitalization and employment status during the same quarter and lagged work, which was further controlled for the effect of past hospitalization by double-lagging previous hospitalization. These alternative statistical models did not yield stronger results than the statistical model presented here (results not shown).

Discussion

Overall, adults with serious mental illness in this prospective analysis were less likely to experience a psychiatric hospitalization in the three-month period after gaining or sustaining employment, even after the analysis adjusted for many baseline characteristics. Specifically, the subgroup of participants who worked in two successive quarters was less likely than nonworkers to experience a psychiatric hospitalization in the next quarter. By contrast, those who started or ended a job in the preceding quarter were neither more nor less likely than nonworkers to be hospitalized. Differences between participants by worker status were statistically significant in only the second half of the two-year follow-up period, possibly suggesting a delayed or cumulative impact of working. Together these findings suggest a cumulative protective effect of employment on risk of rehospitalization, consistent with two prior studies (5,6).

Also noteworthy were factors that did not predict hospitalization. The association between employment and psychiatric hospitalization was unaffected by previous psychiatric hospitalization, enrollment in an intervention condition with enhanced outpatient services, or other covariates, such as self-reported physical health and drug use.

What accounts for the reduced likelihood of psychiatric hospitalization after an individual gains employment? Past research has indicated that employment may reduce social isolation (27), provide structure (28–30), encourage self-esteem (31–33), reduce stigma (34), reduce positive and negative symptomatic burden (35–38), and decrease risk of symptomatic relapse (39,40). Employment could reduce hospitalization rates by improving one or more of these proximal outcomes to prevent the personal crises that precede psychiatric hospitalization (41). Alternatively, improved financial security may be the dominant function leading to positive outcomes. Because employment is a nonrandomizable experience, we also cannot rule out the possibility that participants who achieved employment were healthier than those who did not in unobserved ways for which we did not adjust.

The finding that employment predicted reductions in psychiatric hospitalization risk but that loss of employment did not increase hospitalization risk may surprise some, because job instability—and, in particular, job loss—are hazardous in the general population (42). When a stable job is lost, the would-be worker loses the associated social network, financial security, and self-esteem. But an adult with a serious mental illness faces many disparities, not only in employment experience but also in other social, economic, and health resources.

Longitudinal qualitative studies indicate that people with schizophrenia, bipolar disorder, or major depressive disorder who achieve worker status often experience a significant improvement in life circumstance (43–45). Moreover, no published study has found that gaining employment increases rates of hospitalizations, incarcerations, suicide attempts, or symptomatic relapse, even if the job is subsequently lost (46).

The strengths of this study included the use of a very large cohort of people with severe psychiatric disorders, which allowed us to model psychiatric hospitalizations, a relatively rare event; adjustment for a range of potentially associated variables; use of lagged predictors to make stronger inferences about causality than did previous observational studies; and the criterion that excluded employed individuals at baseline, which reduced the likelihood of selection bias. Furthermore, the parent study was conducted in diverse community service settings, which makes the findings more generalizable to routine care. Methodologically, our analytic approach afforded theoretical insights that would not have been observable with ordinary logistic regression approaches that collapse longitudinal data into cross-sectional analytic frameworks. Thus our lagged analyses were consistent with a stronger inference regarding causality: gaining employment preceded reductions in the likelihood of psychiatric hospitalization.

Several limitations of this study also deserve attention. First, the parent study was designed to test for differences in the effectiveness of a package of health services and not the effect of return to work on other life domains. Second, the adults in this study were SSDI recipients, who are generally older and have more work experience than others with psychiatric disabilities, limiting generalizability. Because only 14% of persons in the initial sampling frame wanted to join the study, however, the final sample is not representative of the typical SSDI recipient. Third, most of the measures used in this study were self-report. Fourth, quarterly interviews did not include measures of clinical status, which would have greatly improved our capacity to understand the mediating role of symptoms in the employment-hospitalization relationship. Fifth, the study had a relatively brief follow-up period. Sixth, the analyses included multiple statistical tests, raising concerns about alpha inflation. We estimated several alternative models as sensitivity analyses to bolster the level of confidence in our findings and explore more nuanced findings, capitalizing on the largest sample size of any longitudinal study investigating the relationship between employment and use of health services. We preserved the type I error rate by basing conclusions on our main a priori hypothesis. Seventh, hospitalization was a relatively low-frequency event in the study sample, so results may not generalize to people who are frequently hospitalized. Finally, in any observational study, statistical associations do not unequivocally demonstrate causality.

This study raises important questions about the benefits of employment: Do the findings generalize to other health service use outcomes? Is the employment-hospitalization relationship even stronger for lengthier periods of employment? Are the findings sensitive to work hours or tenure? Would higher wages amplify the effects? Does the quality of coworker relationships modify the effects? Findings from empirical investigation of these and other research questions will inform insurers, policy makers, program managers, clinicians, people with serious mental illness, and their families, all of whom are working toward the provision of rehabilitation-focused, cost-effective services.

Conclusions

This study found that the likelihood of future psychiatric hospitalizations was reduced when unemployed people with serious mental illness gained employment, especially when employment was maintained over six months. The implications are that clinicians should support client aspirations to gain and maintain employment, not only because working is a valued outcome in its own right but also because working is associated with fewer hospitalizations and greater community tenure. These findings also merit special attention from those engaged in policy development: managed care organizations that incorporate supported employment into packages of psychosocial treatment for beneficiaries with schizophrenia, bipolar, or major depressive disorder may observe reduced per capita hospital admissions.