PERSPECTIVES: Economic Conditions and Supplemental Security Income Application

by
Social Security Bulletin, Vol. 77 No. 4, 2017

Supplemental Security Income (SSI) is one of the most important means-tested transfer programs in the United States. This article examines whether economic conditions affect the likelihood that jobless adults with disabilities apply for SSI payments. Using data for 1996–2010 from the Survey of Income and Program Participation linked to Social Security administrative records, we examine jobless individuals and observe state unemployment rates at both the time their unemployment spell began and the time they applied for SSI. Hazard model estimates suggest that SSI application is positively associated with an increase in the unemployment rate during an individual's jobless spell but is less likely for an individual whose jobless spell begins when the unemployment rate is comparatively high. Omitting the baseline unemployment rate from the analysis distorts the estimate of the relationship between SSI application and the contemporaneous economic conditions. Our findings suggest long-term fiscal implications for SSI of sustained high unemployment.


Austin Nichols is a researcher with Abt Associates. Lucie Schmidt is associate professor of economics at Williams College. Purvi Sevak is a researcher at Mathematica Policy Research and a professor of economics at Hunter College.

Acknowledgments: A previous version of this article was published as Michigan Retirement Research Center Working Paper No. 2014-318 (http://www.mrrc.isr.umich.edu/publications/papers/pdf/wp318.pdf). We are grateful for funding from the Michigan Retirement Research Center, assistance in preparing the data from Stephan Lindner, and assistance in accessing the data from Thuy Ho, Françoise Becker, Tom Solomon, and Richard Chard at the Social Security Administration. Michael Leonesio and three anonymous referees provided extremely helpful comments. Daniel Jordan Alvarez provided excellent research assistance.

This research was supported by a grant (no. 5-RRC08098401-04-00) from the Social Security Administration to the Michigan Retirement Research Center as part of the Retirement Research Consortium.

The findings and conclusions presented in the Bulletin are those of the authors and do not necessarily represent the views of the Social Security Administration.

Introduction

Selected Abbreviations
DI Disability Insurance
SIPP Survey of Income and Program Participation
SSI Supplemental Security Income
TANF Temporary Assistance for Needy Families
UI unemployment insurance

Over the last 30 years, the Supplemental Security Income (SSI) program, which provides federally funded income support for individuals with disabilities, has become one of the most important means-tested cash aid programs in the United States. In 2015, SSI provided payments to 4.9 million low-income adults aged 18–64 who met its disability criteria (Social Security Administration [SSA] 2017a, Table 7.A1). That figure represents a doubling in the adult SSI caseload since 1990 (Chart 1). The federal government spent $46.9 billion on payments to SSI recipients with disabilities in 2015 (SSA 2017a, Table 7.A4), representing a 155 percent increase in real dollars since 1990.1

Chart 1.
Number of SSI recipients aged 18–64, 1975–2015
Line chart with tabular version below.
Show as table
Table equivalent for Chart 1. Number of SSI recipients aged 18–64, 1975–2015
Year Number
1975 1,699,394
1976 1,713,594
1977 1,736,879
1978 1,747,126
1979 1,726,553
1980 1,730,847
1981 1,702,895
1982 1,655,279
1983 1,699,774
1984 1,780,459
1985 1,879,168
1986 2,010,458
1987 2,118,710
1988 2,202,714
1989 2,301,926
1990 2,449,897
1991 2,641,524
1992 2,910,016
1993 3,148,413
1994 3,335,255
1995 3,482,256
1996 3,568,393
1997 3,561,625
1998 3,646,020
1999 3,690,970
2000 3,744,022
2001 3,811,494
2002 3,877,752
2003 3,953,248
2004 4,017,108
2005 4,082,870
2006 4,152,130
2007 4,221,920
2008 4,333,096
2009 4,451,288
2010 4,631,507
2011 4,777,010
2012 4,869,484
2013 4,934,272
2014 4,913,072
2015 4,888,555
 
SOURCE: SSA (2017a and prior editions).

Because SSI is a means-tested program, one might expect application trends to be countercyclical—decreasing when the economy is expanding and increasing during recessions. However, the cyclicality of application has varied over time. Chart 2 graphs SSI applications for adults aged 18–64 (left axis) against the unemployment rate (right axis) for 1990–2015. For most of the period—from 1990 through about 2002 and from 2008 to 2015—the trend in SSI application followed the trend in the national unemployment rate fairly closely. For example, the steady decline in SSI application in the 1990s began about 1 year after the unemployment rate began to decline; SSI application increased as unemployment rates rose during the Great Recession of 2008–2010 and application declined during the subsequent recovery. However, the 2003–2007 period presents an anomaly: Although the unemployment rate fell, SSI application continued to rise. Rutledge and Wu (2014) offer a number of explanations for the continuing rise in applications during that period, including the lagged effects of prior welfare reforms that induced Temporary Assistance for Needy Families (TANF) program participants to switch to SSI, persistently high poverty rates, and increases in the share of the population in fair or poor health.

Chart 2.
Number of SSI applications filed by adults aged 18–64, and U.S. unemployment rates, 1990–2015
Line chart with tabular version below.
Show as table
Table equivalent for Chart 2. Number of SSI applications filed by adults aged 18–64, and U.S. unemployment rates, 1990–2015
Year Number of applications Unemployment rate (%)
1990 695,281 5.6
1991 835,058 6.8
1992 955,060 7.5
1993 1,030,912 6.9
1994 982,275 6.1
1995 860,162 5.6
1996 780,058 5.4
1997 644,982 4.9
1998 640,585 4.5
1999 641,973 4.2
2000 680,397 4
2001 753,650 4.7
2002 862,899 5.8
2003 938,352 6
2004 1,038,077 5.5
2005 1,079,252 5.1
2006 1,107,658 4.6
2007 1,100,954 4.6
2008 1,123,307 5.8
2009 1,247,597 9.3
2010 1,254,710 9.6
2011 1,194,291 8.9
2012 1,117,483 8.1
2013 966,650 7.4
2014 852,445 6.2
2015 808,669 5.3
 
SOURCES: SSA (2017b, Table 62); Bureau of Labor Statistics (2017, Table 1).

A number of previous studies looked at the effects of economic conditions on growth in disability program caseloads. However, much of that work focused on Social Security Disability Insurance (DI), which is limited to individuals who meet that program's earnings-history thresholds and who therefore may be more responsive to economic conditions. Most research focusing specifically on SSI dates from the 1990s. Those studies found that higher unemployment was associated with increases in SSI application and caseloads (Rupp and Stapleton 1995; Stapleton and others 1998; Stapleton and others 1999). The relationship between economic conditions and SSI application may have evolved significantly since then. Given rapid growth in the SSI rolls and the slow pace of recovery from the Great Recession, understanding the role that business cycles play in determining SSI participation has become increasingly important.

In this article, we examine the relationship between economic conditions and working-age adult SSI application from 1996 through 2010 using data from the Survey of Income and Program Participation (SIPP) linked to SSA's 831 data file. These restricted-access data allow us to link detailed SIPP information on demographic conditions and unemployment spells with precise SSA records on the month of first application for SSI and DI benefits. Using hazard models, we estimate SSI and DI application risk among individuals who were working when first observed in the SIPP but were unemployed during follow-up surveys in their respective SIPP panels, and examine the effect of the unemployment rate both at the time of job loss (the baseline rate) and at the time of—that is, contemporaneous with—SSI application. Whereas the contemporaneous measure reflects local labor market conditions at the time of application, the baseline rate may reflect differential characteristics of the pool of unemployed workers related to the business cycle. Our results suggest that application risk increases significantly with higher contemporaneous state unemployment rates. The magnitude of this effect is large—suggesting that a 1 percentage point increase in the state unemployment rate would lead to a 20 percent increase in the risk of applying for SSI or DI, raising the probability from 0.30 percent to 0.36 percent. Conversely, workers who began their unemployment spell in a time of high unemployment were less likely to apply for SSI, consistent with the hypothesis that the characteristics of the pool of newly jobless workers varies systematically with the business cycle, generally displaying a lower intrinsic propensity to apply for disability benefits in a period when more workers are being laid off. We also find that omitting the baseline unemployment rate from the analysis would lead to a substantial underestimation of the relationship between SSI application and contemporaneous economic conditions.

Once enrolled in SSI, very few recipients leave the program. Our findings suggest that short-term fluctuations in economic conditions may have substantial long-term effects on program participation and expenditures, and that countercyclical stimulus spending could have greater impacts over time by deterring disability-program application.

Background

SSI provides means-tested cash assistance to the elderly, to children, and to adults who are blind or have disabilities. Enacted in 1972, SSI replaced an uneven range of state programs and thereby standardized income support for those groups (Berkowitz and DeWitt 2013). The SSI disability determination process is complicated; most claims must pass through five stages before the applicant receives payments.2 At the first stage, individuals must meet the income and asset eligibility requirements and show that they are not involved in “substantial, gainful” economic activity. The second and third stages involve medical evaluations. Applicants with impairments that are deemed “nonsevere” or that are not expected to end in death or to last at least 12 months are denied in stage 2; those with impairments deemed “extremely severe” are allowed in stage 3. Stages 4 and 5 assess capacity to work. Applicants who are able to work in jobs that they held in the past are denied in stage 4, and applicants who, given their age, education, and work experience, are judged able to work in any type of employment in the economy are denied in stage 5. As noted by Chen and van der Klaauw (2008), the vocational grid used in stage 5 creates age discontinuities in eligibility determinations beginning with age 45. Less than half of SSI applicants are ultimately approved (Duggan, Kearney, and Rennane 2016). Although the majority of SSI payments are federally funded, many states supplement payments with state funds.3 The maximum monthly individual federal benefit was $733 in 2015. Benefit levels are adjusted annually for increases in the cost of living.

SSI is one of two major U.S. programs targeted to individuals with disabilities. The other, DI, provides benefits to individuals with disabilities who are insured by the contributions they made to the Social Security system while they were working. The disability determination process for DI is the same as that for SSI. However, DI benefits are not means-tested; instead, they depend on an individual's having a sufficient earnings history. DI enrollment is greater than that of SSI and is growing more rapidly. In 2015, 8.9 million workers with disabilities received DI benefits, an increase of 196 percent since 1990 (SSA 2017a, Table 5.D3). Primarily because of the work-history requirements, DI applicants and beneficiaries are less economically disadvantaged than are those who apply for and receive SSI payments. DI applicants are typically older, more highly educated, and wealthier than SSI applicants are. They are also more likely to be male, white, non-Hispanic, and married (Bailey and Hemmeter 2015).

However, many individuals are concurrently eligible for benefits from both SSI and DI. These beneficiaries have work histories sufficient to qualify for DI but their asset, income, and benefit levels are low enough that they still meet the means test to qualify for SSI. Of the 12.9 million working-age (18–64) adults receiving benefits administered by SSA on the basis of a disability in 2015, 8.0 million (62 percent) received DI benefits only, 3.5 million (27 percent) received SSI payments only, and 1.3 million (10 percent) received SSI and DI benefits concurrently (SSA 2017a, Table 3.C6.1).4 Many applicants may not know how the eligibility rules for the two programs differ, and SSA staff may need to direct them toward one program or the other. In fact, the online application (https://www.socialsecurity.gov/applyfordisability) does not mention either program; it only describes how to apply for “disability benefits.”

Macroeconomic Conditions and SSI Participation

For several reasons, changes in macroeconomic conditions may reduce the extent to which the stringency of the disability determination process discourages SSI participation. First, the SSI means test examines family (not individual) income, so if other family members experience income declines related to the business cycle, the lower family-level means may establish SSI eligibility. In addition, as a local labor market declines, an individual's physical or mental impairment may represent a relatively greater impediment to employment, making SSI participation more viable. Finally, an economic downturn that leads to an exogenous job separation may lower the opportunity cost of remaining out of work for an individual with disabilities who applies for SSI. These effects are all consistent with evidence suggesting that the rates at which individuals self-report disabilities respond to the relative costs and benefits of disability program participation (Waidmann, Bound, and Schoenbaum 1995).

A number of studies have looked at the effects of economic conditions on disability program caseloads. Autor and Duggan (2003) found that shifts in state-level labor demand predict changes in DI participation. Black, Daniel, and Sanders (2002) used changes in coal prices as a shock to local earnings growth to examine the effects of earnings on disability program participation. They found that both DI and SSI participation respond to earnings shocks, but that SSI participation is less responsive than that of DI.

A series of related studies found that increased unemployment rates associated with the recession of the early 1990s played an important role in the growth of SSI application and awards, with the effect on application being the stronger of the two (Rupp and Stapleton 1995; Stapleton and others 1998; Stapleton and others 1999). More recent studies found positive relationships between unemployment rates and DI application (Soss and Keiser 2006; Guo and Burton 2012; Coe and others 2011). Recent evidence that focuses primarily on SSI is less plentiful. Soss and Keiser (2006) found a positive and significant relationship between the unemployment rate and SSI application when estimated jointly with DI application. Rutledge and Wu (2014) found that SSI enrollment is negatively and significantly correlated with unemployment rates, but that the relationship has grown less negative over time and even turned positive during the Great Recession. They also found a relationship between SSI application flows and local unemployment rates that has weakened in recent years.

In this article, we extend the existing literature in three important ways. First, we examine the relationship between state-level economic conditions and SSI application in the late 1990s and 2000s, focusing on policy-relevant changes in the application flows rather than in the program rolls.5 Second, we demonstrate the importance of examining the unemployment rates at both the time of job loss and the time of application. The first measure provides a baseline that captures differential selection into unemployment. The second measure reflects the labor market conditions contemporaneous with the application decision. Omitting the baseline rate—which can be correlated with the contemporaneous rate—could lead to biased estimates of the relationship between SSI application and contemporaneous economic conditions. Third, we provide evidence suggesting that higher current unemployment rates alone do not predict higher application rates, but that persistently higher unemployment rates do, with important consequences for policies aiming to limit the long-term consequences of a recession, both for program budgets and for individuals with extended periods of unemployment.

Data

We use SIPP data matched to Social Security administrative records. The SIPP is a nationally representative longitudinal survey which collects data on a number of topics including employment, demographics, income, and program participation. Because it focuses on program participation, the SIPP oversamples low-income households. Monthly data are available for sample members for as long as about 3 years. We use data from the four most recent SIPP panels, which began in 1996, 2001, 2004, and 2008, respectively. Each panel lasted 3 to 5 years, and taken together, they covered calendar years 1996–2010. The initial sample size for individuals of all ages in each panel ranges from 95,315 in the first wave of the 1996 panel to 105,663 in the first wave of the 2008 panel. We link the SIPP data to SSA's 831 file, which provides data on the timing of the first application for SSI and DI.

There are several advantages to using the matched SIPP/SSA data. First, SIPP's monthly data allow us to examine dynamics related to employment, unemployment, and benefit receipt. Second, because the 831 file contains records on applications for DI or SSI that cleared the financial screen (stage 1) of the disability determination process—including the decisions on those applications—we are able to avoid standard concerns about survey respondents underreporting program participation (Meyer, Mok, and Sullivan 2009). In addition, the SSA data allow us to observe the exact date of application, whereas data from the SIPP alone would provide only the month of first benefit receipt (subject to error, as respondents do not always report the correct source of income). This is potentially important because applicants must remain out of the labor force until their application is resolved. Applicants may wait months or years before receiving benefits. As a result, the date of allowance or first receipt of benefits is much less likely to be tied to economic conditions than is the date of first application.6 A third benefit of using the matched SIPP/SSA data is that they include information on nonapplicants (and on demographic characteristics of applicants) that are unavailable from administrative-only data. Research has shown that using matched administrative records in this fashion provides more accurate estimates of SSI participation and payment amounts than using SIPP self-reported information does (Huynh, Rupp, and Sears 2002).

The matched SIPP/SSA data have several limitations. First, SSA cannot match records for all SIPP respondents. Survey respondents are matched to the Protected Identification Key (PIK) using a Social Security number, where available; or name, date of birth, and location, where other identifiers are unavailable. Behind a firewall, the survey data are matched to administrative earnings and disability-determination records using the PIK. SIPP panels vary in the proportion of cases that are matched to administrative records: The 2008 panel exceeded a 90 percent match rate, and the 1996 and 2004 SIPP panels achieved close to 90 percent match rates. However, the 2001 SIPP match rate was less than 70 percent. One possible way to deal with the imperfect matching would be to reweight the sample. However, that would require untenable assumptions about the selection process for those cases not matched to administrative data, namely that selection is on the observable characteristics in the survey only. For this reason, we do not reweight; however, we do include dummies for individual characteristics such as education, age, race, time period, and state in our regressions. This means that reweighting on those factors would have little impact on our regression estimates. That is, the factors used to adjust the baseline hazard also effectively adjust for differences in sample characteristics. Second, because an 831 record is not created for applicants whose claims are rejected because of financial ineligibility at stage 1 of the determination process, those applicants are miscoded as nonapplicants in our data. Our estimates should be biased toward zero if those applicants are particularly sensitive to the unemployment rate.7,8

We limit our sample to individuals who were newly unemployed during the SIPP panel so that we observe periods out of work for which the onsets are not censored. This restriction also helps select a sample whose members are subject to the strict income limits for SSI eligibility. SSI application is a relatively low-probability event, and those who are either continuously working or already in a long unemployment spell in our sample are much less likely to apply. Consistent with Mueller, Rothstein, and von Wachter (2016), who examine the risk of disability application at the time of unemployment insurance (UI) benefit exhaustion, our sample restriction allows us to focus on the population and the time period in which individuals are most at risk of applying for SSI.

To implement this restriction, we select all SIPP respondents aged 20–59 who were employed at the time of the first wave of their SIPP panel and were newly out of work for at least 1 month during the remainder of the period in which they were observed. Hereafter, we use “unemployed” to describe individuals with this specific experience (including those who separate from a job and do not report looking for another). We limit our analysis sample to those with an unemployment spell that began during the survey observation period for several reasons. First, we rely on SIPP data to identify state of residence so we can accurately measure local labor market conditions. If a respondent's unemployment spell begins before the initial SIPP interview, we cannot be confident that we are matching the correct unemployment rate for their location. Second, we would not be able to identify when an unemployment spell began for individuals who were unemployed at the start of their SIPP spell. To include such individuals in our sample would left-censor our data, meaning that we would not know the duration of their unemployment spell nor the labor market conditions when the spell began. Restricting the sample to individuals whose unemployment spell began during the SIPP panel may exclude a disproportionate share of long unemployment spells. Therefore, our results should be interpreted as identifying the relationship between unemployment rates and SSI application over the first few years of an unemployment spell. Individuals who have long unemployment spells are less likely to apply for SSI; again, we are interested in examining the population most at risk for application.

We identify spells of nonwork, which we define as months during which individuals are out of work after an observed job separation, whether or not they are actively looking for work. We observe a job separation for about one-quarter of the sample in each SIPP panel. Respondents enter our analysis sample in the month of job separation, and we follow them until they apply for disability benefits, become reemployed, or leave the SIPP sample. Our sets of matched survey and administrative data thus represent the U.S. noninstitutionalized population who had earnings during the time covered by the first wave's survey (that is, the year for which the panel is named) but stopped working in a subsequent month. We measure the duration of the unemployment spell in months.

As discussed earlier, we are primarily interested in SSI. However, many potential beneficiaries may not fully understand eligibility rules and may not be sure which programs they should apply for. Staff at Social Security field offices may steer individuals toward one program versus the other. People may apply for both, and then find out which program they are eligible for (or whether they are eligible for both). As noted earlier, 10 percent of working-age beneficiaries with disabilities receive concurrent benefits from both programs. Therefore, in addition to examining SSI applications (which comprise SSI-only and concurrent SSI/DI applications), we examine applications to any federal disability program (that is, the sum of applications for DI only, for SSI only, and for DI and SSI concurrently).

We merge state-level measures (including unemployment rate and a number of policy variables) to the matched SIPP/SSA data by state and month. Although SSI payments are determined at the federal level, a number of states supplement payments, so we include the dollar amount of state-level SSI supplements. Because research has documented a link between welfare reform and SSI participation (for example, Schmidt and Sevak 2004), we also control for a number of TANF-related variables, including the maximum TANF benefit for a family of three9 and indicators of whether state TANF programs have strict sanctions, strict time limits, and few exemptions from work requirements. Finally, we include a state fiscal distress measure. Kubik (2003) shows that states undergoing unexpected fiscal distress in the 1990s were likely to have SSI caseloads increase more sharply than did participation in Aid to Families with Dependent Children, the program that TANF replaced. Further information on the policy variables is provided in Appendix A.

Table 1 provides summary statistics for the individual variables included in the SIPP for our analysis sample. Each of the four SIPP panels contributes 20 to 30 percent of the full sample.10 Most individual variables are measured in the month that the individual enters the sample (is first jobless); we report the means of those values. As described earlier, individuals who enter the sample are considered to be at risk for SSI application beginning with the month in which they are newly reported in the SIPP as not employed. Fifty-six percent of sample members are female. Half of the sample members were married when they entered. Thirteen percent of the sample is foreign-born. Respondents range in age from 20 to 59, but are disproportionately found at the younger end of that range. Roughly three-quarters are non-Hispanic whites and 12 percent are non-Hispanic blacks. Ninety-one percent have graduated from high school and 62 percent have attended college. On entry, approximately one-third of the sample had family income of less than twice the federal poverty level. The sample statistics differ from those for a nationally representative sample because our sample is restricted to individuals who are first observed as employed and then lose employment during the SIPP panel.

Table 1. Selected characteristics of sample members in the first month of their unemployment spell
Variable Mean Standard deviation
  Demographic characteristics (%)
SIPP panel
1996 27.4 0.450
2001 20.9 0.406
2004 30.4 0.460
2008 21.3 0.410
Foreign-born 13.1 0.337
Married 50.5 0.500
Female 56.4 0.496
Age
20–24 20.5 0.404
25–29 14.8 0.355
30–34 13.8 0.345
35–39 12.6 0.331
40–44 12.1 0.327
45–49 10.9 0.312
50–54 9.4 0.292
55–59 5.9 0.236
Race
White non-Hispanic 76.3 0.426
Black non-Hispanic 12.4 0.330
Educational attainment
High school graduate 91.4 0.281
Attended college 62.5 0.484
Family income < twice the federal poverty level 35.1 0.477
  Monthly application rates (per 1,000)
DI only, SSI only, or both 3 0.053
SSI only or both 1 0.039
  Monthly state-level economic indicators
Unemployment rate (%) 5.267 1.893
Maximum TANF benefit, family of three ($) 385.60 149.50
State SSI supplement ($) 28.49 61.48
Per capita unexpected deficit shock 0.0 0.3
Percentage of states with strict TANF
Time limits 32.0 0.467
Sanctions 32.3 0.468
Work exemptions 86.0 0.347
SOURCE: Authors' calculations using SIPP 1996, 2001, 2004, and 2008 panels matched to Social Security administrative records.
NOTES: Sample size = 26,077; application person-months = 199,870; indicator state-months = 9,180.

We report summary statistics for selected time-varying variables across person-month records. In a given month out of work after a job separation, about 3 out of 1,000 sample members apply for either SSI or DI (or both) and about 1 in 1,000 apply for SSI (alone or concurrently with DI). This means that more than twice as many individuals apply for DI as apply for SSI, a finding that is consistent with the relative caseloads of the two programs. The mean value of our key variable of interest, the monthly state unemployment rate, is 5.3 percent. We also report summary statistics for state policy variables at the state-month level.

Model Specification

We examine the relationship between application for federal disability benefits and prevailing economic conditions by estimating with a series of discrete-time hazard models the risk of application for (1) any such benefits (DI alone, SSI alone, or both concurrently) and (2) SSI (alone or concurrently with DI). In other words, the “SSI application” category differs from the “any-program application” category in that it excludes DI-only applications. To address the fact that both current and lagged labor-market conditions should be related to one's current employment status and risk of program application, we use two unemployment rate measures: The baseline rate, which is current in the month when the individual's unemployment spell begins; and the contemporaneous rate, which is current in the month of application.

The contemporaneous measure captures an individual's perception of his or her chance of gaining employment. The baseline measure provides an indicator of two factors. For the first, a persistent economic downturn might be indicated if the baseline rate is higher than the level of the unemployment rate in a subsequent observation month. For the second, differential characteristics of the jobless population during periods of high versus low unemployment might be captured. For example, in recessions, the pool of unemployed individuals shifts toward those with higher skill or employability (Mueller 2012). Those individuals should be less likely to apply for SSI. The baseline and contemporaneous unemployment rates are highly correlated; it is therefore important to include both, even if the research objective focuses on the relationship between SSI application and the contemporaneous rate.

Because the duration of an unemployment spell may be related to SSI application in ways that could also be related to the unemployment rate, it is important to control for it. First, the likelihood of financial eligibility for SSI may increase with spell duration, as individuals deplete their assets; this circumstance can also lead individuals who applied for DI at the outset of the spell to subsequently apply for SSI as well, placing their applications in the “concurrent” category. For this reason alone, we would want to compare application for SSI to application for either DI or SSI; but it is also of substantive interest to compare SSI application rates to the any-program application rates.

Second, many unemployed workers will receive UI benefits during the early months of their spell, and some of them will apply for disability benefits when the UI benefits expire (Lindner 2016). On the other hand, a long unemployment spell may indicate ineligibility for SSI because disability has not been determined; or, it may reflect increasing selection into the pool of applicants who have not applied. In this circumstance, duration could negatively affect application. In our preferred model specification, we control for the duration of unemployment with a measure of the natural log of months of unemployment. We also present results from an alternative specification that controls for duration nonparametrically with a series of variables indicating unemployment spells ranging from 3–5 months to 36–38 months (with 0–2 months being the reference category). Duration variables measure the baseline hazard (that is, the chance of application in each month, conditional on not yet having applied) for a case with all covariates set to zero. The covariates are then used to adjust hazards proportionally at all points in time.

All models control for age (in 5-year bands), sex, race, and educational attainment, as well as for indicators of marital status and immigrant status. We also control for whether an individual had low income at the start of the unemployment spell (indicator is equal to one if the respondent's family income was less than twice the federal poverty level) and for the state TANF policy parameters discussed earlier.

All specifications include state fixed effects and controls for secular shifts over time. In our preferred specification, we control for secular shifts with an indicator for the SIPP panel (1996, 2001, 2004, or 2008). The inclusion of dummy variables for each panel adjusts not only for different initial labor market conditions in the panel, but also for differential inclusion probabilities by panel because of SIPP/SSA data matches. We also test the robustness of our results by flexibly controlling for secular trends with year-fixed effects.

Although DI and SSI application trends are likely to be strongly influenced by additional factors such as disability status, spousal and other income, and asset levels over the duration of the panel, we do not control for them because they are likely to be endogenous. Waidmann, Bound, and Schoenbaum (1995), Benítez-Silva and others (2004), and Benítez-Silva, Disney, and Jiménez-Martín (2010) found that survey measures of work-related disability are sensitive to labor market conditions and the availability of disability benefits. There is substantial disagreement about the quality of survey measures of disability or of related health conditions or limitations. Self-reported disability is both a subjective judgment of inability to work and a perception about one's ability to find work given both health and labor market conditions. Including self-reported measures of disability in regression analysis could be problematic, given that changes in labor market conditions after job separation could affect both a decision to apply for benefits and a self-report of disability. That is, the variable of self-reported disability is an outcome as well as a potential control. Assets and individual and spousal income are also outcomes of processes that may evolve in response to both time out of work and choices about spending and work in response to health and labor market conditions. Because these factors may therefore reflect an individual's plan to apply for benefits, we do not adjust for them in the regressions.

As a result, we omit health and asset variables as controls because of potential endogeneity. Although the SIPP data sets include information on health and assets that would support formal endogeneity tests, we do not have reliable instruments. We have therefore traded a potential endogeneity problem for a potential omitted-variable bias problem because people with longer unemployment spells and greater exposure to weaker labor markets may also be in marginally better health or have greater financial assets than do those who have already applied for disability benefits or returned to work. The omission has no consequence for the empirical work if the correlation between the omitted variables and the regressors of interest is small or zero, but it could bias our findings on current unemployment rates downward if individuals who remain exposed to higher unemployment rates without applying are those who are better able to weather the downturn. That is, we expect that a positive measured effect of contemporaneous unemployment rates on application could represent a lower-bound estimate if those in the risk pool for longer periods have better health and asset levels because those characteristics are negatively correlated with application but are positively correlated with ongoing high unemployment rates in that situation.

We examine the heterogeneity of our results by estimating the hazard of SSI application by age group (44 or younger, 45 or older) for a number of reasons. First, the vocational grid used in stage 5 of the disability determination process introduces discontinuities by age beginning at age 45. Second, the hazard could also differ because of variations in human capital and health that are correlated with age.

We also estimate the hazard of application by sex. Evidence suggests that SSI has become a more important part of the safety net since welfare reforms were enacted in the 1990s, and this development is likely to affect women disproportionately (Wittenburg and others 2015).

Results

Table 2 presents coefficient estimates and z-statistics from four hazard models of application for disability benefits, estimated on the full sample at risk. The table presents results under a specification that controls only for the contemporaneous unemployment rate and under a specification that also controls for the baseline unemployment rate (respectively labeled “without” and “with” the baseline unemployment rate). Results are given for any federal disability benefit application (DI only, SSI only, or both) and for SSI application (alone or concurrent with DI). The coefficient on the contemporaneous state unemployment rate is positive in all four specifications, suggesting that as the unemployment rate increases, the hazard of application increases. In addition, using the model that includes the baseline unemployment rate roughly doubles the magnitude of the coefficient on the contemporaneous unemployment rate and increases its statistical significance. This result is consistent with a conclusion that the baseline unemployment rate reflects unobserved variation in the composition of newly unemployed individuals. When the baseline rate is excluded, the estimated coefficient on the contemporaneous unemployment rate is biased downward. This finding has important implications for other research examining the effects of contemporaneous unemployment rates on disability program participation. As a result, in the rest of the article, we focus on results from specifications that control for the baseline unemployment rate.

Table 2. Logistic regression results: Hazards of filing an application for any disability program and for SSI
Variable Without baseline unemployment rate With baseline unemployment rate
Any application a SSI application b Any application a SSI application b
Coefficient z-statistic Coefficient z-statistic Coefficient z-statistic Coefficient z-statistic
State unemployment rate at the time of—
Unemployment onset (baseline) . . . . . . . . . . . . -0.117** -2.49 -0.127 -1.58
Application (contemporaneous) 0.100* 1.65 0.108 1.14 0.186** 2.73 0.202* 1.81
Log duration of unemployment -0.309** -5.99 -0.343** -4.98 -0.312** -5.92 -0.348** -4.80
SIPP panel
1996 (reference category) . . . . . . . . . . . . . . . . . . . . . . . .
2001 -0.064 -0.32 -0.091 -0.35 -0.085 -0.42 -0.112 -0.44
2004 0.033 0.18 -0.183 -0.68 0.035 0.19 -0.183 -0.69
2008 -0.830* -1.93 -1.112 -1.57 -0.783* -1.85 -1.064 -1.53
Foreign-born -1.012** -3.06 -1.053** -3.94 -1.010** -3.07 -1.054** -3.96
Married -0.285* -1.88 -0.769** -3.84 -0.291* -1.90 -0.773** -3.83
Female -0.193* -1.87 0.148 0.90 -0.189* -1.83 0.151 0.92
Age
20–24 -2.516** -11.26 -1.419** -5.59 -2.529** -11.23 -1.433** -5.56
25–29 -1.299** -3.19 -0.352 -0.75 -1.310** -3.23 -0.363 -0.78
30–34 -1.308** -6.54 -0.508* -1.66 -1.319** -6.59 -0.517* -1.68
35–39 -0.827** -4.96 -0.162 -0.45 -0.835** -5.01 -0.171 -0.47
40–44 -0.451** -2.59 0.264 0.84 -0.458** -2.62 0.260 0.83
45–49 -0.088 -0.54 0.542* 1.72 -0.098 -0.60 0.533* 1.69
50–54 0.191 1.47 0.797** 3.89 0.186 1.42 0.794** 3.88
55–59 (reference category) . . . . . . . . . . . . . . . . . . . . . . . .
Race
White non-Hispanic -0.194 -0.99 0.176 0.54 -0.193 -1.00 0.177 0.55
Black non-Hispanic -0.076 -0.31 0.195 0.50 -0.077 -0.32 0.194 0.50
Educational attainment
High school graduate -0.247* -1.71 -0.244* -1.75 -0.243* -1.66 -0.239* -1.71
Attended college -0.220* -1.90 -0.477** -2.39 -0.220* -1.89 -0.476** -2.37
Family income < twice the federal poverty level 0.483** 4.02 1.056** 6.27 0.480** 4.04 1.057** 6.32
SOURCE: Authors' calculations using SIPP 1996, 2001, 2004, and 2008 panels matched to Social Security administrative records.
NOTES: Regressions include state TANF policy parameters, state fixed effects, and a constant term.
Sample sizes are 193,450 person-months for SSI applications and 199,870 person-months for all applications.
. . . = not applicable; * = statistically significant at the 10 percent level; ** = statistically significant at the 5 percent level.
a. Includes DI only, SSI only, and concurrent DI and SSI applications.
b. Includes SSI only and concurrent DI and SSI applications.

The coefficient on the contemporaneous state unemployment rate on the hazard for any-program application (0.186) implies that, after controlling for the baseline unemployment rate, a 1 percentage point increase in the unemployment rate would lead to an increase of 0.186 in the natural log of the odds of DI or SSI application. Given the mean monthly application rate of 3 in 1,000, this translates to a 20 percent increase in the probability of application among those with recent job separations from 3.0 to 3.6 per thousand.11

The coefficient on the baseline state unemployment rate at the beginning of the unemployment spell for the hazard of any-program application (−0.117) is negative, supporting the theory that the pool of individuals who are jobless in periods of higher unemployment may be more employable and thus at lower risk of SSI application. This finding is consistent with previous research on application trends (Bound, Burkhauser, and Nichols 2003).

The coefficient on the indicator for duration of unemployment spell (−0.312 for any-program application, −0.348 for SSI application) is negative, suggesting that the risk of application declines with each additional month elapsed. As described above, there are a number of reasons why the hazard could increase with duration, such as depletion of assets and exhaustion of UI benefits. However, our empirical estimates of negative duration dependence likely reflect medical ineligibility for SSI and DI; that is, those who can apply will tend to do so, leaving the pool of remaining potential applicants less likely to apply over time.

Individual characteristics are associated with application risk in expected directions, both for any application for disability benefits (including DI-only applications) and for SSI applications specifically. The risk of any application for benefits is lower for married individuals and for women, but there is no statistical difference in rates between men and women in SSI application rates (implying that differential eligibility for DI plays a large role in gender differences). There are no significant differences in application risk by race. Those living in households with foreign-born individuals are significantly less likely to apply for benefits, which is consistent with post-1996 restrictions on immigrant receipt of SSI (Bitler and Hoynes 2013). Relative to those who did not graduate from high school, the risk of application is lower for those who did. Having baseline family income that is less than twice the federal poverty level significantly increases the risk of application. The SIPP panel fixed effects (with 1996 as the excluded category) show a large and significant decrease in any-program application risk in the 2008 panel (−0.783), which coincides with the Great Recession, suggesting that individuals who were unemployed during this recent, deep recession were less likely to apply for SSI or DI than were those unemployed in earlier years. This finding is consistent with a large shift observed in the characteristics of the population newly out of work during 2008 and subsequent years.12

The pattern of results for SSI application is similar to that for any-program application. In the model that controls for the baseline rate, the coefficient on the baseline unemployment rate for SSI application (−0.127) is negative and of similar magnitude as the coefficient for any-program application (−0.117), but is less precisely estimated. The coefficient on the contemporaneous unemployment rate (0.202) is positively and significantly associated with the risk of SSI application, and the coefficient on duration of unemployment (−0.348) is negative and statistically significant. The effects of demographic variables for SSI application are qualitatively the same as those for any-program application, with the exception of sex. Women have significantly lower risk for any-program application but show no significant difference from men for SSI application. This difference would be consistent with women having lower labor-force attachment than men and therefore being less likely to have amassed sufficient work history to qualify for DI, implying that sex differences are driven by DI application. The estimated relationship between low family income and application is much larger for SSI than for federal disability programs overall, which reflects the fact that SSI, unlike DI, is a means-tested program.

In Table 3, we present results from a number of alternative specifications to check the robustness of our findings to controls for unemployment duration and year of observation. All results in Table 3 are for the dependent variable of any-program application. The first column repeats, as applicable, the results from Table 2 of our preferred specification (controlling for both baseline and contemporaneous unemployment rates). The alternative specifications include replacing the log duration of unemployment variable with nonlinear controls for duration and controlling for time with calendar-year effects. The results for our main explanatory variables of interest—the unemployment rate variables (both baseline and contemporaneous)—are consistent with the results for the preferred specification. The nonlinear duration indicators are relatively consistent with the log specification, in that they show application probabilities that tend to decline as unemployment spells grow longer. Although there are reasons (discussed earlier) why application hazards might increase with spell duration, it is likely that selection plays a large role, such that the pool of individuals who have been unemployed longer and have not yet applied includes more people who will never apply. The pattern of year effects shows large decreases in the hazard of applying during the recession of the early 2000s, as well as during the Great Recession in 2008–2010.

Table 3. Robustness checks of logistic regressions on hazard of filing any disability-program application, controlling for baseline and contemporaneous unemployment rates
Variable Log duration of unemployment spell and— Months of unemployment spell and—
Panel Year Panel Year
Coefficient z-statistic Coefficient z-statistic Coefficient z-statistic Coefficient z-statistic
State unemployment rate at the time of—
Unemployment onset (baseline) -0.117** -2.49 -0.127** -2.59 -0.096** -2.05 -0.108** -2.25
Application (contemporaneous) 0.186** 2.73 0.164** 2.41 0.162** 2.36 0.149** 2.20
Log duration of unemployment spell -0.312** -5.92 -0.302** -4.73 . . . . . . . . . . . .
Unemployment spell (months)
0–2 (reference category) . . . . . . . . . . . . . . . . . . . . . . . .
3–5 . . . . . . . . . . . . -0.151 -1.17 -0.158 -1.16
6–8 . . . . . . . . . . . . -0.408** -2.36 -0.430** -2.33
9–11 . . . . . . . . . . . . -0.380* -1.79 -0.415* -1.79
12–14 . . . . . . . . . . . . -0.444* -1.87 -0.467* -1.87
15–17 . . . . . . . . . . . . -0.763** -2.52 -0.763** -2.35
18–20 . . . . . . . . . . . . -0.866** -2.18 -0.834** -2.05
21–23 . . . . . . . . . . . . -1.475** -2.86 -1.410** -2.60
24–26 . . . . . . . . . . . . -1.341** -2.46 -1.260** -2.14
27–29 . . . . . . . . . . . . -1.552** -2.97 -1.459** -2.74
30–32 . . . . . . . . . . . . -0.426 -0.86 -0.322 -0.63
33–35 . . . . . . . . . . . . -1.558** -2.18 -1.455** -1.96
36–38 . . . . . . . . . . . . -2.816** -2.78 -2.678** -2.61
SIPP panel
1996 (reference category) . . . . . . . . . . . . . . . . . . . . . . . .
2001 -0.085 -0.42 . . . . . . -0.097 -0.48 . . . . . .
2004 0.035 0.19 . . . . . . 0.022 0.13 . . . . . .
2008 -0.783* -1.85 . . . . . . -0.811* -1.92 . . . . . .
Year
1996 . . . . . . -0.180 -0.58 . . . . . . -0.147 -0.45
1997 (reference category) . . . . . . . . . . . . . . . . . . . . . . . .
1998 . . . . . . -0.495** -2.44 . . . . . . -0.459** -2.23
1999 . . . . . . -0.365** -2.00 . . . . . . -0.290 -1.52
2000 . . . . . . -1.445 -1.43 . . . . . . -1.367 -1.33
2001 . . . . . . -0.677** -1.99 . . . . . . -0.644** -1.97
2002 . . . . . . -0.099 -0.33 . . . . . . -0.096 -0.32
2003 . . . . . . -0.609** -2.62 . . . . . . -0.545** -2.26
2004 . . . . . . -0.454 -1.64 . . . . . . -0.424 -1.50
2005 . . . . . . -0.215 -1.12 . . . . . . -0.214 -1.11
2006 . . . . . . -0.198 -0.91 . . . . . . -0.152 -0.68
2007 . . . . . . -0.504 -1.31 . . . . . . -0.424 -1.07
2008 . . . . . . -1.271* -1.72 . . . . . . -1.184 -1.57
2009 . . . . . . -0.950** -2.25 . . . . . . -0.949** -2.25
2010 . . . . . . -1.023* -1.90 . . . . . . -1.033* -1.90
SOURCE: Authors' calculations using SIPP 1996, 2001, 2004, and 2008 panels matched to Social Security administrative records.
NOTES: Regressions include state TANF policy parameters, state fixed effects, and a constant term.
Sample size is 199,870 application person-months.
. . . = not applicable; * = statistically significant at the 10 percent level; ** = statistically significant at the 5 percent level.

Table 4 shows the estimated hazard of any-program application on subpopulations stratified by sex and age.13 Although the unemployment-rate coefficients are greater for women than for men, the differences are not statistically significant. The relationship between economic conditions and application risk is not significant for the younger workers in our sample; but for those aged 45 or older, we observe a negative and significant effect of the baseline unemployment rate (−0.183) and a positive significant effect of the contemporaneous unemployment rate (0.251). The relationship may differ by age for several reasons. First, the long-term costs of leaving the labor force may be lower for older workers. Second, the composition of older and younger applicants may vary by type of disability and age of onset, with a smaller share of older applicants having childhood onset. Third, as noted earlier, the disability determination process introduces age-related discontinuities beginning at age 45, which Chen and van der Klaauw (2008) have shown to be associated with reduced labor supply.

Table 4. Logistic regression results: Hazard of filing any disability-program application, by selected population subgroup
Variable Overall Sex Age
Women Men 20–44 45–59
Coefficient z-statistic Coefficient z-statistic Coefficient z-statistic Coefficient z-statistic Coefficient z-statistic
State unemployment rate at the time of—
Unemployment onset (baseline) -0.117** -2.49 -0.131* -1.85 -0.070 -0.76 -0.027 -0.34 -0.183** -2.64
Application (contemporaneous) 0.186** 2.73 0.209** 2.38 0.143* 1.70 0.109 0.95 0.251** 3.04
SOURCE: Authors' calculations using SIPP 1996, 2001, 2004, and 2008 panels matched to Social Security administrative records.
NOTES: Regressions include state TANF policy parameters, state fixed effects, and a constant term.
Sample sizes (in person-months) are 199,870 for applications overall, 126,462 for women, 70,468 for men, 135,660 for applicants aged 20–44, and 61,366 for applicants aged 45–59.
* = statistically significant at the 10 percent level; ** = statistically significant at the 5 percent level.

Conclusion

Using SIPP data linked to SSA's 831 data file, we find that individuals who begin their unemployment spell in a time of high unemployment are less likely than those with job loss during a low-unemployment period to apply for SSI or DI, consistent with the idea that the characteristics of the newly unemployed vary with the business cycle. However, application risk among individuals with a recent job separation increases significantly when the state unemployment rate rises as their jobless period continues. In addition, omitting the baseline unemployment rate from the analysis leads us to substantially underestimate the relationship between SSI application and contemporaneous economic conditions.

Our findings suggest that recessions can have long-term fiscal implications for SSI. If the flow of allowances tracks that of applications, and if exits from SSI are rare, then extended periods of high unemployment may permanently expand SSI caseloads.14 Policymakers should account for that cost when considering programs to help at-risk or unemployed workers. The Congressional Budget Office (2014) estimates that the 2009 American Recovery and Reinvestment Act (ARRA) reduced the unemployment rate by between 0.4 and 2.0 percentage points during the third quarter of 2010. Our results suggest that a reduction in the current unemployment rate of 1 percentage point reduces SSI application among the recently unemployed by approximately 20 percent. However, stimulus spending such as that authorized by ARRA would also tend to restrict the pool of individuals at risk of applying for SSI by reducing the number of unemployed persons, so this estimate should be viewed as an upper bound on reductions in SSI application among the entire pool of the unemployed. The net effect on application would depend on reductions in separations and in the duration of unemployment. These results suggest on net that the ARRA dampened potential recession-induced increases in SSI (and DI) application. If so, the net benefits of federal aid during economic downturns may be underestimated, because even small changes in SSI application rates can have large budgetary consequences. Lindner and Nichols (2014) suggest that aid tied to labor-market attachment may reduce application rates, while increases in unconditional aid may increase application rates. Further research is needed to pinpoint the cyclical determinants of SSI application and the nature of the impacts of cyclical federal aid.

Appendix A

Listed below are the sources of the input data used in this analysis.

SSI application. This variable is coded using data from SSA's 831 data file, which are merged to the SIPP and are available for analysis at SSA through restricted access. The 831 file contains a record for all individuals who have ever applied for SSI or DI. We use variables noting date of application and type of application to identify whether an individual applied for SSI or DI in a given month.

Unemployment rates. Bureau of Labor Statistics.

Maximum TANF (or Aid to Families with Dependent Children) benefit for a family of three. Data for 1997–2010 are from Urban Institute's Welfare Rules Database (Table IIA4). Multiple values were given for California, Massachusetts, and Wisconsin; for those states, we used the highest value.

Maximum SSI state supplement. Data for 2002–2010 are from the SSA publication State Assistance Programs for SSI Recipients and indicate the maximum state supplement available to an individual with a disability who lives alone. Data for 1999–2001 are from the 2004 edition of the SSA publication Consultative Examinations: A Guide for Health Professionals (known as the Green Book). Data for 1996–1998 come from various earlier editions of the Green Book, as collected by the University of Kentucky's Center for Poverty Research, with values converted to 2000 dollars.

Welfare reform variables. Inputs for earlier years were provided by Rebecca Blank and Jordan Matsudaira; those for later years were updated using the Urban Institute's Welfare Rules Database.

Unexpected deficit shock. Calculated as in Kubik (2003). Data on actual state expenditures and revenues (per capita) in year t are obtained from the National Association of State Budget Officers' State Fiscal Survey in year t + 1. Forecasted state expenditures and revenues in year t are obtained from State Fiscal Survey in year t − 1. Fiscal shock = (actual state expenditure − forecasted state expenditure) − (actual state revenue − forecasted state revenue).

Notes

1 Payment data are for all SSI recipients with a disability, including those aged younger than 18 and older than 64, and therefore do not exactly correspond with the caseload data, which include only recipients aged 18–64 with a disability (or blindness).

2 The description of the determination process that follows draws heavily from Wixon and Strand (2013).

3 Duggan, Kearney, and Rennane (2016) note that 45 states currently supplement benefits for some or all of their recipients.

4 The sums of these beneficiary counts and percentages do not equal the totals because of rounding.

5 Trends in allowances affect the flow onto the SSI rolls more directly than trends in applications do, and may therefore be more important from a budgetary perspective for SSA. However, we focus on application for several reasons. First, application provides a direct measure of the decision an individual makes in response to personal and economic factors, while allowance measures administrative judgments. Second, given the long application and appeal process, the low (31.4 percent) allowance rate (Rupp 2012), and the harm that time out of the labor force does to an individual's employment opportunities (Kroft, Lange, and Notowidigdo 2013), fluctuation in the application volume is of independent research interest.

6 Similarly, labor market conditions that affect aggregate application rates also affect the timeliness with which state Disability Determination Services process claims, and therefore could affect the average lag from first application to eventual receipt of benefits.

7 For a detailed discussion of measurement issues in matched data, see Davies and Fisher (2009).

8 Applications denied at stage 1 are not especially policy-relevant because in such cases the total social cost of application is minimal (in sharp contrast with applicants who remain out of work for months as they await a determination).

9 Using TANF benefit levels for a fixed family size is standard in the welfare literature, in part because actual family size could be endogenous to benefit levels.

10 The sample is not distributed evenly across the years because the SIPP panel sizes and employment outflows vary.

11 A logit coefficient of 0.186 translates to an increase in odds of 20.0 percent; the increase in probability is very close for low baseline probabilities but declines to zero as the baseline probability increases. For a baseline probability of 3 in 1,000, we can convert the coefficient to a marginal effect of a 1.0 percentage point increase in the unemployment rate on the probability of application by adding the coefficient estimate (0.186) to the natural log of the baseline odds (−5.806). We then exponentiate the sum to get the revised odds and back out the revised probability (0.0036), which is 22 percent higher than the baseline probability of 0.003.

12 The TANF policy variables described earlier are in most cases not statistically significant and are omitted from Table 2. Their inclusion does not significantly affect the unemployment rate results. Full results are available from the authors on request (lschmidt@williams.edu).

13 Because sex- and age-stratified results for SSI application resemble those in Table 3, they are omitted from Table 4.

14 However, initial allowance rates are negatively related to high unemployment (Rupp 2012), which would tend to diminish the fiscal implications.

References

Autor, David H., and Mark G. Duggan. 2003. “The Rise in the Disability Rolls and the Decline in Unemployment.” Quarterly Journal of Economics 118(1): 157–206.

Bailey, Michelle Stegman, and Jeffrey Hemmeter. 2015. “Characteristics of Noninstitutionalized DI and SSI Program Participants, 2013 Update.” Research and Statistics Note No. 2015-02. Washington, DC: SSA.

Benítez-Silva, Hugo, Moshe Buchinsky, Hiu Man Chan, Sofia Cheidvasser, and John Rust. 2004. “How Large Is the Bias in Self-Reported Disability?” Journal of Applied Econometrics 19(6): 649–670.

Benítez-Silva, Hugo, Richard Disney, and Sergi Jiménez-Martín. 2010. “Disability, Capacity for Work, and the Business Cycle: An International Perspective.” Economic Policy 25(63): 483–536.

Berkowitz, Edward D., and Larry DeWitt. 2013. The Other Welfare: Supplemental Security Income and U.S. Social Policy. Ithaca, NY: Cornell University Press.

Bitler, Marianne, and Hilary Hoynes. 2013. “Immigrants, Welfare Reform, and the US Safety Net.” In Immigration, Poverty, and Socioeconomic Inequality, edited by David Card and Steven Raphael, 315–380. New York, NY: Russell Sage Foundation.

Black, Dan, Kermit Daniel, and Seth Sanders. 2002. “The Impact of Economic Conditions on Participation in Disability Programs: Evidence from the Coal Boom and Bust.” American Economic Review 92(1): 27–50.

Bound, John, Richard V. Burkhauser, and Austin Nichols. 2003. “Tracking the Household Income of SSDI and SSI Applicants.” In Research in Labor Economics, Volume 22: Worker Well-Being and Public Policy, edited by Solomon W. Polacheck, 113–158. Oxford: Elsevier Science Ltd.

Bureau of Labor Statistics. 2017. “Labor Force Statistics from the Current Population Survey.” https://www.bls.gov/cps/tables.htm.

Chen, Susan, and Wilbert van der Klaauw. 2008. “The Work Disincentive Effects of the Disability Insurance Program in the 1990s.” Journal of Econometrics 142(2): 757–784.

Coe, Norma B., Kelly Haverstick, Alicia H. Munnell, and Anthony Webb. 2011. “What Explains State Variation in SSDI Application Rates?” CRR Working Paper No. 2011-23. Chestnut Hill, MA: Center for Retirement Research at Boston College. http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1970314.

Congressional Budget Office. 2014. “Estimated Impact of the American Recovery and Reinvestment Act on Employment and Economic Output in 2013.” Washington, DC: CBO. http://www.cbo.gov/sites/default/files/cbofiles/attachments/45122-ARRA.pdf.

Davies, Paul S., and T. Lynn Fisher. 2009. “Measurement Issues Associated with Using Survey Data Matched with Administrative Data from the Social Security Administration.” Social Security Bulletin 69(2): 1–12.

Duggan, Mark, Melissa S. Kearney, and Stephanie Rennane. 2016. “The Supplemental Security Income (SSI) Program.” In Economics of Means-Tested Transfer Programs in the United States, Volume II, edited by Robert A. Moffitt. Chicago, IL: University of Chicago Press.

Guo, Xuguang (Steve), and John F. Burton, Jr. 2012. “The Growth in Applications for Social Security Disability Insurance: A Spillover Effect from Workers' Compensation.” Social Security Bulletin 72(3): 69–88.

Huynh, Minh, Kalman Rupp, and James Sears. 2002. “The Assessment of Survey of Income and Program Participation (SIPP) Benefit Data Using Longitudinal Administrative Records.” SIPP Working Paper No. 238. Washington, DC: Census Bureau.

Kroft, Kory, Fabian Lange, and Matthew J. Notowidigdo. 2013. “Duration Dependence and Labor Market Conditions: Evidence from a Field Experiment.” Quarterly Journal of Economics 128(3): 1123–1167.

Kubik, Jeffrey D. 2003. “Fiscal Federalism and Welfare Policy: The Role of States in the Growth of Child SSI.” National Tax Journal 56(1): 61–79.

Lindner, Stephan. 2016. “How Do Unemployment Insurance Benefits Affect the Decision to Apply for Social Security Disability Insurance?” Journal of Human Resources 51(1): 62–94.

Lindner, Stephan, and Austin Nichols. 2014. “The Impact of Temporary Assistance Programs on Disability Rolls and Re-employment.” in Safety Nets and Benefit Dependence: Research in Labor Economics, Vol. 39, edited by Stéphane Carcillo, Herwig Immervoll, Stephen P. Jenkins, Sebastian Königs, and Konstantinos Tatsiramos, 219–258. Bingley (UK): Emerald Group Publishing Limited.

Meyer, Bruce D., Wallace K. C. Mok, and James X. Sullivan. 2009. “The Under-Reporting of Transfers in Household Surveys: Its Nature and Consequences.” NBER Working Paper No. 15181. Cambridge, MA: National Bureau of Economic Research.

Mueller, Andreas I. 2012. “Separations, Sorting and Cyclical Unemployment.” IZA Discussion Paper No. 6849. Bonn: Institute for the Study of Labor (IZA).

Mueller, Andreas I., Jesse Rothstein, and Till M. von Wachter. 2016. “Unemployment Insurance and Disability Insurance in the Great Recession.” Journal of Labor Economics 34(S1): S445–S475.

Rupp, Kalman. 2012. “Factors Affecting Initial Disability Allowance Rates for the Disability Insurance and Supplemental Security Income Programs: The Role of the Demographic and Diagnostic Composition of Applicants and Local Labor Market Conditions.” Social Security Bulletin 72(4): 11–35.

Rupp, Kalman, and David Stapleton. 1995. “Determinants of the Growth in the Social Security Administration's Disability Programs—An Overview.” Social Security Bulletin 58(4): 43–70.

Rutledge, Matthew S., and April Yanyuan Wu. 2014. “Why Do SSI and SNAP Enrollments Rise in Good Economic Times and Bad?” CRR Working Paper No. 2014-10. Chestnut Hill, MA: Center for Retirement Research at Boston College.

Schmidt, Lucie, and Purvi Harendra Sevak. 2004. “AFDC, SSI, and Welfare Reform Aggressiveness: Caseload Reductions versus Caseload Shifting.” Journal of Human Resources 39(3): 792–812.

Social Security Administration. 2017a. Annual Statistical Supplement to the Social Security Bulletin, 2016. SSA Publication No. 13-11700. Washington, DC: SSA. https://www.socialsecurity.gov/policy/docs/statcomps/supplement/2016/index.html.

———. 2017b. SSI Annual Statistical Report, 2015. SSA Publication No. 13-11827. Washington, DC: SSA. https://www.socialsecurity.gov/policy/docs/statcomps/ssi_asr/2015/index.html.

Soss, Joe, and Lael R. Keiser. 2006. “The Political Roots of Disability Claims: How State Environments and Policies Shape Citizen Demands.” Political Research Quarterly 59(1): 133–148.

SSA. See Social Security Administration.

Stapleton, David, Kevin A. Coleman, Kimberly A. Dietrich, and Gina A. Livermore. 1998. “Econometric Analyses of DI and SSI Application and Award Growth.” In Growth in Disability Benefits: Explanations and Policy Implications, edited by Kalman Rupp and David Stapleton, 31–92. Kalamazoo, MI: W. E. Upjohn Institute for Employment Research.

Stapleton, David, Michael Fishman, Gina Livermore, and David Wittenburg. 1999. Policy Evaluation of the Overall Effects of Welfare Reform on SSA Programs: Final Report. Falls Church, >VA: Lewin Group.

Waidmann, Timothy, John Bound, and Michael Schoenbaum. 1995. “The Illusion of Failure: Trends in the Self-Reported Health of the U.S. Elderly.” Milbank Memorial Fund Quarterly 73(2): 253–288.

Wittenburg, David, John Tambornino, Elizabeth Brown, Gretchen Rowe, Mason DeCamillis, and Gilbert Crouse. 2015. “The Child SSI Program and the Changing Safety Net.” ASPE Research Brief. Washington, DC: Department of Health and Human Services, Office of the Assistant Secretary for Planning and Evaluation.

Wixon, Bernard, and Alexander Strand. 2013. “Identifying SSA's Sequential Disability Determination Steps Using Administrative Data.” Research and Statistics Note No. 2013-01. Washington, DC: SSA.