The Impact of Job Accommodations on Employment Outcomes Among Individuals With Psychiatric Disabilities

The data in this investigation came from the federally funded Employment Intervention Demonstration Project (EIDP) that evaluated the impact of an evidence-based approach to delivering employment services to individuals with psychiatric disabilities between 1996 and 2000 (18). Individuals were recruited from existing clinical populations in eight states by provider referral, self-referral, family referral, and word of mouth and, in one state, by newspaper advertisement. Inclusion criteria were being age 18 or older and meeting the requirement of psychiatric disability (severe and persistent mental illness) as defined by the Center for Mental Health Services of the Substance Abuse and Mental Health Services Administration. Eligible participants were unemployed but willing to work, and they completed a written informed consent approved by each site’s institutional review board. The total eligibility pool was 10,653, from which 2,883 were invited to join the study. Of those, 1,655 completed the first interview, 1,648 of whom underwent random assignment; details are described in Cook and colleagues’ 2005 article (18). Participants were compensated between $10 and $20 per interview, and 1,159 (70%) completed all five interviews as of month 24, indicating that attrition was not a major problem. With a lack of outcome information, work hours could not be estimated for 110 individuals, leaving an analytical sample of 1,538 for the hours analysis. For the duration analysis, the 1,040 individuals who obtained jobs during the 48-month period were included in the sample.

Interview protocols consisting of structured assessments that have been psychometrically validated (19) were administered by trained interviewers. The protocol included clinical assessments of symptoms and diagnosis (including the Positive and Negative Syndrome Scale) and numerous self-rated measures of functioning, service and job satisfaction, employment history, demographic information, financial condition, and quality of life. Questions were also asked about frequency and age of first psychiatric hospitalization, substance use, and work motivation and self-esteem. Hours worked and amount earned were recorded on a weekly basis for the initial 24 months covered by the interviews and 24 months after, resulting in 48 months of employment outcome data. Additional employment information was recorded each time a participant started, changed, or terminated employment and included an open-ended detailed description of any job accommodations, which was subsequently summarized as a count of number of accommodations reported for each job. All data collection protocols for this study are available from the authors on request.

Quality assurance of data was conducted on two levels. At the program level, the research team took elaborate steps to ensure that the interview protocol met the strongest psychometric requirements and independently verified all employment outcomes data through payroll documentation. At the study level, the research team purged the data set of input and calculation errors and generated descriptive statistics on all study variables to detect and correct for illegal values.

The generalized linear model (GLM) with log link was selected to test the hypothesis on hours worked because the cumulative distribution of average hours worked exhibited a mixed distribution. The GLM is a weighted nonlinear regression model that allows for the estimation of mean hours worked in which the assumption of independent and identically distributed disturbance with constant variance may be relaxed (20–22). Goodness-of-fit tests, including the modified Park test, log likelihood, and Akaike and Bayesian information criteria, were conducted to confirm that the log-link function was the correct assumption made about the distribution of hours worked (22,23). Marginal effects were calculated for all coefficients to facilitate interpretation (24). A positive relationship between accommodations and the marginal effect would indicate that the receipt of job accommodations led to increased work hours, with other factors held constant.

We used duration model analysis, also known as survival analysis, to test the hypothesis that individuals who report receiving job accommodations experienced longer employment tenure across multiple spells of employment. The model evaluates duration in a given job by estimating how long it takes before transition to another job occurs (25,26). By focusing on transition rather than average job duration, we accounted for the fact that some individuals did not experience job termination during the study period.

Duration analysis examines the probability that a transition over a given period has occurred, modeled through a hazard function and defined as the instantaneous probability of transitioning out of a spell of employment (25). Whereas ordinary regression describes the conditional mean and variance of a distribution (27), the hazard function describes the conditional rate of job separation. In duration models, covariates affect the expected duration of employment by changing the position of the baseline hazard, or the rate at which a given person experiences the transition event—job loss in this case. Covariates shift the hazard function up or down. An important consideration is whether covariates change over the observation period, because time-varying covariates may also inhibit our ability to determine the impact of covariates on the duration of employment. This concern was addressed by averaging all time-varying covariates (25).

Another important consideration in our analysis was how to account for multiple jobs or changes in employment for the same individual during the time frame studied. Failure to account for the correlation among multiple job durations for the same person would lead to biased estimates of the effect of accommodations on the risk of job termination. The estimator that best addresses this challenge is a parametric form known as Weibull. Results of testing six competing parametric models confirmed that Weibull was the best fit for the employment duration data.

Table 2 summarizes results from the GLM estimation on average hours worked per month and demonstrates that the pattern of effects generally followed the direction hypothesized [see online data supplement], providing empirical support for the conceptual framework. Some limitations to activities, such as frequency of seeing non–family members, and environmental barriers predicted fewer hours worked, as did some measures of health status, such as reported schizophrenia (p≤.01) and blunted affect (p≤.01). More hours worked were predicted by the environmental facilitator supported employment and by labor market predictors, such as number of months employed in the past five years and frequency in phoning non–family members (p≤.01). Human capital variables, such as being married (p≤.01), predicted more work hours. Concerning the first research question, the marginal effects column in Table 2 shows that individuals who reported receiving job accommodations on average worked seven hours more per month (p<.01), after analyses controlled for all other covariates.

Duration of employment was measured in days across multiple spells of employment. The length of a completed employment spell was 135.81±194.03 days, or 27.16 weeks. With data broken down by number of jobs and converted into weeks, the first job averaged 37.39 weeks (186.97±243.28 days), the second job 24.88 weeks (124.43±161.61 days), the third job 19.93 weeks (99.66±136.17 days), the fourth 16.66 weeks (83.30±121.41 days), and the fifth and subsequent jobs averaged roughly the same duration, 12.33 weeks (61.63±111.74 days). Approximately half of all first jobs lasted more than 13 weeks. More than 75% of these lasted between 36 and 265 weeks. The more jobs one worked, the shorter the spell of employment with each successive job.

The direction of effect shown in Table 3 indicates that numerous covariates affected the risk of job termination in the expected direction, again providing empirical support for the conceptual framework. [The online data supplement provides further details.] In regard to the second research question, the accommodations hazard ratio of .873 (p<.01) indicates that workers receiving accommodations tended to remain employed longer, with each additional job accommodation decreasing the risk of job termination by 12.7%. For those with job accommodations, the predicted mean period of employment was 206.96 days, 31% higher than those without job accommodations (p≤.05). Figure 1 shows the impact of job accommodations on the cumulative risk of job loss and demonstrates that with each additional accommodation, the risk of job loss diminished over time.