Vocational rehabilitation service receipt,service expenditures,and ruralness

2.1 Data sources

For these analyses, we used 2015 Rehabilitation Services Administration Case Service Report (RSA-911) data. The 2015 RSA-911 dataset includes case level demographic, disability, VR services, and employment status information for consumers whose cases closed during the 2015 fiscal year. The 2015 RSA-911 dataset also includes county of residence, which we used to match with Federal Information Processing System (FIPS) codes. FIPS codes are used across multiple data sets and can be used to match case level data to a variety of county-level indicators.

2.2 Data cleaning

Data cleaning included several steps. First, we removed all cases where we were unable to link FIPS codes. This resulted in the removal of all observations from Alaska and U.S. territories plus additional individual cases that did not collect accurate county indicators. Second, we removed individual observations with missing data on key demographic variables including gender, race, and age. Finally, we removed any observations that fell outside the 18–65 age range. Table 1 reports on our data cleaning process and the final number of observations included in the study. The remaining data represented approximately 86% of all cases.

We used 2015 RSA-911 data for sociodemographic controls and received services. The advantages of the RSA-911 national data are that it covers all states in one data set and it is readily available to researchers. Alternatively, we could have collected individual state VR agency service data. With state agency data, one can get RSA-911 data elements along with some extra variables (see Dean et al., 2015, 2017, 2018, 2019; Schmidt et al., 2019). However, securing data at the state level requires much administrative, data cleaning, and data-merging effort.

Gender. Coded as dummy variable for female.

Race/ethnicity. Coded as dummy variables for White, Black or African American, American Indian or Alaska Native, Asian, Native Hawaiian or other Pacific Islander, and Hispanic or Latino.

Age. Calculated as 2015 minus birth year.

Education level at application. Classified into four groups: less than high school graduation, special education certificate of completion, high school graduate/GED, and post-secondary education.

Disability type. Counselors report disability in terms of impairment codes and cause or source codes. These data were used to collapse disability types into 10 categories including intellectual disability, learning disability, attention deficit hyperactivity disorder (ADHD), traumatic brain injury (TBI), Autism, mental illness, substance abuse, physical impairment, vision impairment, and hearing impairment.

Disability severity. VR counselors classify severity of disability as 0 = no significant disability, 1 = significant disability, and 2 = most significant disability. Those with the most significant disabilities have the highest barriers to employment and require multiple VR services over time. Also, when agencies are required to go into order of selection (where VR state budgets do not allow for service delivery to all eligible applicants), those with the most significant disabilities are served first.

VR services. The 2015 RSA 911 collected data on 27 service categories. For each service, there is an indicator variable: service was not received (= 0), was provided by the VR agency (= 1), was purchased by the VR agency (= 2), was provided by comparable services and benefits providers (= 3), or was provided by a combination of agency and/or purchased and/or comparable services (= 4). There are follow-up questions about purchased services in terms of provider type and costs of purchased services from Title 1 and Title 4, Part B funds.

For our analyses, we collapsed service categories into nine groups based on past return on investment research (Dean et al., 2015, 2017, 2018, 2019; Schmidt et al., 2019; Stern et al., 2019). Table 2 provides this categorization.

For each service category, we created three dependent variables:

Agency services = 1 if the consumer received only agency-provided services in the category (as opposed to any purchased services from an outside vendor).

Distance. The distance variable captures how far a rural (micropolitan or noncore) county is from the nearest metropolitan county, based on Office of Management and Budget (OMB) county designations. 1 The distance measure was constructed using county-level population centroids from the 2010 U.S. Census and the Haversine formula, which is commonly used to calculate the shortest distance between two points. 2 Within each state, we calculated the county’s distance from every other county, and retained the minimum distance of a rural county from an urban county. For metro counties, the minimum distance is naturally zero. We added 1 to each distance measure for mathematical reasons. Using these data, we classified VR cases into three groups: (1) live under 20 miles from a metro office, (2) live 20 to 50 miles from a metro office, and (3) live greater than 50 miles from a metro office.

Employment ratio. The employment ratio was constructed using employment data from the U.S. Census, Bureau of County Business Patterns and county-level population estimates from the U.S. Census, Population Division. The employment ratio for each county is defined as (county employment of nearest metro area/county population of nearest metro area) * (1/distance) (Dean et al., 2015, 2017, 2018, 2019; Schmidt et al., 2019). This measure assesses economic opportunity based on the employment rate and distance from the nearest metro area. Note that this variable declines as distance increases.

We were interested in learning how receipt of services was influenced by distance to a metro area while controlling for sociodemographic characteristics. We hypothesized that some services, particularly those that rely on purchased services, would be sensitive to distance due to the unavailability of vendors or community rehabilitation programs in more rural locations (Ipsen et al., 2019). For each service category, we explored this question with two dependent variables including receipt of services and log expenditure of purchased services, conditional on service receipt.

Methodology for services receipt. We used probit to produce estimates for each combination of service and service source (agency services and purchased services); for a discussion of probit, see Maddala (1993). We felt comfortable using probit (with normally distributed errors) instead of some other binary choice error assumption because of results in Stern (1996). For each estimation exercise associated with agency services, the choice was either receipt of agency services or no receipt of agency services. There was a similar specification for purchased services. This resulted in 18 different estimation exercises based on the 9 different service categories for both agency services and purchased services. We used the same set of explanatory variables in each of the estimation procedures.

Methodology for conditional log expenditure of purchased services. RSA 911 collected expenditure data only on purchased services. Services provided by the agency or as a comparable benefit were not assigned a monetary value. For this reason, our analyses of expenditure data were confined to services that were purchased. We used log expenditures to reduce the influence of large outliers and because log expenditure data appear to have a normal distribution. The way to interpret coefficient estimates associated with the log expenditure equations is the percentage increase in expenditure as the variable of interest increases by one. For each explanatory variable that increases log expenditure, it also increases expenditure, and for each explanatory variable that decreases log expenditure, it also decreases expenditure.

We explored both Ordinary Least Squares (OLS) and semi-parametric regression models to estimate log expenditures for each service category. For many services, including assessment, job training, job search and placement, supported employment, and other supports, the linear specification imposed by OLS resulted in similar estimates to the semi-parametric estimates. For other categories including disability treatment, education, disability accommodations, and benefits, the OLS estimates were not proportional to the semi-parametric estimates, and therefore OLS was not a good specification of the model.

Estimation results. Although all probit, OLS, and non-parametric model results are available in supplementary tables (see https://sites.google.com/site/stevensterneconomics/research-interests/vocational-rehabilitation/ipsen-jain-stern-estimates), we focused our discussion on receipt of job search and placement services. Past research indicated that receipt of VR job search and placement services strongly predicts employment outcomes across disability types (Dutta et al., 2008; Leahy et al., 2018; Nord & Hepperlen, 2016). Our results included three models with identical explanatory variables listed in Table 3. Model 1 used probit to estimate the effects of explanatory variables on receiving agency-based job search and placement services. Model 2 used probit to estimate the effects of explanatory variables on receiving purchased job search and placement services. Model 3 used OLS to estimate the effects of explanatory variables on purchased job search and placement services, conditional on receipt of the service. Because OLS and semi-parametric models were proportional or comparable for purchased job search and placement services, we reported only OLS estimates to streamline discussing results.

The sample was predominantly male (55%), White (70%), and with a high school diploma or less (70%). The most common disability types were mental illness (38%), followed by physical impairment (17%). Table 3 shows descriptive statistics for the study sample. The reference groups were male, not-White or Black, not-Hispanic, and education less than a high school diploma.

2.6 Descriptive statistics for receipt of services

Table 4 shows the percent of cases that received services for each service category by source. The most common services received include assessment services (55%), benefits counseling (49%), other supports (34%), and job search and placement services (33%). While assessment, benefits, and other supports help individuals get linked with additional resources and address barriers to employment, job search and placement services are related to finding employment.

Table 5 reports on the ratio of cases that received agency services vs purchased services by distance from metro (i.e., < 20 miles, 20–50 miles, or  > 50 miles). As distance to a metro area increased, ratios of agency service receipt to purchased service receipt increased substantially for assessment, job training, and job search and placement service categories.

Table 6 shows the mean value of purchased services for each category. The mean expenditures presented in Table 6 are likely to be a misleading estimate of costs due to indications of kurtosis. This implies that the data were not normally distributed, and a lognormal distribution was more appropriate for analyses of purchased services conditional on purchased service receipt. We compared the empirical distribution function to that of an appropriately chosen log normal distribution and found excellent fits. These are reported in the supplementary tables.

4.1 Receipt of services

These data highlight differences in service provision based on demographic, disability, and environmental characteristics. While some rationale might be provided for service delivery differences based on age, education, disability type, and distance to metro, it is more difficult to describe why gender and race were associated with differences. Each probit parameter estimate should be interpreted as the effect of the associated explanatory variable on the value of the service choice (agency or purchased) relative to any alternative choice and, therefore, the probability of choosing that choice. Models for receipt of agency services and purchased services showed that certain demographic groups consistently received different rates of job search and placement services regardless of delivery type (agency and/or purchased). For instance, (1) relative to males, females received significantly lower rates of both agency and purchased services, (2) relative to other race, Whites received significantly higher rates of agency services, and (3) relative to non-Hispanics, Hispanics received significantly lower rates of both agency and purchased services. Additionally, there were groups that received services in different ways. For instance, relative to other race, Blacks received significantly more agency services and significantly fewer purchased services. After controlling for other sociodemographic and geographic characteristics, differences in service provision based on race and gender suggest that the VR agencies may not have provided services equitably.

Differences in service provision also occurred because of differences in individual characteristics such as education level, disability type, and disability severity. At an intuitive level, these differences make logical sense. For instance, maybe those with significant and most significant disabilities received significantly higher rates of both agency services and purchased services, relative to those with no significant disabilities, because they required more services to achieve an employment outcome. Differences may also be explained by the degree or type of services needed to achieve an employment outcome. For instance, those with TBI, intellectual disability, or Autism received significantly fewer agency services but significantly more purchased services. These differences might show the niche that purchased services fill in terms of more intensive employment services that agency personnel are unable or unequipped to provide. 3

Geography also played a role in the type or mix of services received. Specifically, after controlling for sociodemographic characteristics and disability type, consumers living at a distance from the VR office received significantly higher rates of agency services and significantly lower rates of purchased services. Ipsen, Goe, and Bliss (2019) reported gaps in provider availability in rural communities due to increased economic, transportation, and delivery barriers. Specifically, fewer economic opportunities in rural communities make placements harder to secure; drive time to reach rural consumers adds to the costs of providing services; and it is difficult to support full-time staff in locations with limited numbers of consumers to serve. They posited that providers are less willing to serve rural consumers because it is more expensive to do so and placement success is not as favorable. They suggested that VR payment models do not adequately address these barriers or the added costs of doing business in rural locations (Ipsen et al., 2019). When vendors are unavailable, VR agency personnel are responsible for providing services.

4.2 Purchased services

For the subset of consumers who received purchased services, we conducted a linear regression of log expenditures of purchased job search and placement services on the same set of explanatory variables in the probit models to see which characteristics caused the differences in expenditures. We know that females and Blacks received fewer agency services and purchased services overall (as shown in the probit models). In addition, females and Blacks received significantly lower expenditures for purchased services (see Table 8: OLS log expenditure model beta estimates). For example, the t-statistic associated with the differences in estimates for Blacks and Whites was 3.058, indicating a statistically significant difference. Though there are other possible explanations for this difference, a possible cause of the result is racial discrimination.

In terms of disability, those with a learning disability, TBI, Autism, a hearing impairment, and significant disability received significantly higher levels of purchased services, while those with substance abuse or physical impairment received significantly lower levels of purchased services. As mentioned previously, this may relate to the intensity of services required for certain disability types.

The two “distance to metro” variables provided conflicting evidence. Those living between 20 and 50 miles from a metro office received significantly higher expenditures of purchased services relative to those living at most 20 miles from the VR agency, while those at a distance greater than 50 miles received significantly lower expenditures. In part, this may relate to differences in economic opportunity. Very rural locations may have such limited economic opportunity that costs for job search and placement services are limited, while places that are closer to an economic center may have more job opportunities to explore. This is similar to the employment ratio variable, which indicated employment opportunities had a positive impact on the amount of purchased services received.

Ipsen and Stern (2020) found that people with disabilities from rural counties were less likely to apply for VR services based on population estimates of the population with disabilities. One hypothesis for this lower-than-expected application rate was focused on expectations for positive employment outcomes. In part, this analysis builds evidence that this might be the case insomuch as rural consumers received fewer purchased services overall, and cases that were most rural received less expenditure.

4.3 Limitations

These analyses were limited in many ways. First, we used RSA 911 data from 2015, which provided a dated picture of VR service delivery. This choice was made for two reasons. First, these data reflect a timespan where the RSA 911 included geographic indicators and were collected annually on case closures versus quarterly as cases were progressing through the VR system. Second, we were interested in using data that allowed us to make time-relevant comparisons with other papers focused on rural VR service delivery (see Ipsen & Stern, 2020; Ipsen et al., 2019).

Another limitation relates to RSA 911 data collection protocols related to services received. While we know whether a consumer received agency services, we do not know how much agency services were received. Presently, many state VR agencies do not collect information on time or monetary expenditure for agency services. Information on agency expenditure would provide each agency a better understanding of how staff resources are used. This lack of agency expenditure data is quite different for purchased services. Having a common indicator of both agency services and purchased services received, such as dollars or time, would aid in (1) understanding parity across consumers, (2) controlling for regional differences in VR-related costs, and (3) providing a measure of service intensity required to achieve employment outcomes for specific groups.