Using Predisability Earnings as an Instrument for Disability Determination Outcomes

Summary

Researchers have consistently found that applications and awards for Social Security Disability Insurance (DI) benefits are affected by economic conditions, such as the unemployment rate, and by programmatic features, such as the weighted benefit formula. They have not agreed, however, on the size of these effects.

Fluctuations in the numbers of applications and awards are often thought to be the result of changes in incentives that are affected by program rules and by macroeconomic conditions (Rupp and Stapleton 1995). If this causality/effect result is true, a simple way to gauge the size of the incentive effect would be to measure the number of individuals who do not meet the eligibility standard relative to the total number of applicants.

The problem with using disability determination outcomes to explain labor force participation is that the decision to apply for benefits and the outcome of the decision are interrelated. On the one hand, the higher the allowance rate is, the more likely individuals are to apply. On the other hand, the rate of allowance will depend on characteristics of the applicants, such as their attachment to the labor force. Under these circumstances, standard techniques of estimation may produce a biased result. Failure to control for the correlation between labor force attachment and allowance status can lead to overestimating the extent to which beneficiaries would work in the absence of receiving DI benefits.

This paper uses the predisability earnings of DI applicants as an instrument for disability determination outcomes because those earnings represent labor force attachment; and the earnings affect applications but applications do not affect the earnings.¹ To validate this choice of instrument, the paper first shows that the disability determination outcomes are correlated with predisability earnings. It then shows fluctuations in predisability earnings profiles, for those who applied for DI benefits during the period from 1977 through 1997, as evidence of the incentive effects generated by the interaction of program rules and macroeconomic conditions.

Sources of Data

This paper constructs a series of records for yearly DI application covered under Social Security from 1977 through 1997 using Social Security Administration (SSA) program data that are matched to multiple panels of the Survey of Income and Program Participation (SIPP). It analyzes the predisability earnings of applicants, by year of their applications, after controlling for their demographic characteristics. Using this approach, instead of a regression-based approach that compares and contrasts applicants with nonapplicants in terms of individuals' characteristics, program rules, and economic conditions, this study can avoid most common controversies in this area, from sample restrictions to identification assumptions.²

Samples used in this study are drawn from the 1984, 1992, 1993, and 1996 SIPP panels. Historical earnings and DI applications and determination results are obtained from SSA's Summary of Earnings Record (SER), Master Beneficiary Record (MBR), and "831" disability files.

The SIPP is a national survey conducted by the U.S. Census Bureau, designed as a continuous series of national panels since 1984, with sample sizes ranging from 14,000 to 36,700 interviewed households. The SIPP is a multistage, stratified sample of the U.S. civilian, noninstitutionalized population. Panel duration ranges from 2½ to 4 years; each panel consists of six to twelve 4-month waves, depending on the panel duration. Each SIPP wave consists of both a core file and a topical module file. The core file contains the demographic, labor force, program participation, and income data designed to measure the economic situations of the individuals in the sample; these data are repeated at each interviewing wave. The topical modules are designed to provide more detailed coverage of topics such as assets, health and disability, and employment history that are not covered in the core files.

The SER contains annual summaries of individuals' Social Security earnings from 1951 to the present and total amounts earned between 1937 and 1950. The SER extract used here also contains sex, race, date of birth, and annual quarters of coverage (QCs). The MBR includes the data needed to generate Social Security benefit checks under the Old-Age, Survivors, and Disability Insurance (OASDI) program. A record is established whenever an individual applies for old-age or disability benefits and the application is adjudicated. The record reflects the final decision on the initial claim, including any denial at the initial, reconsideration, or hearing levels, and can be used to identify whether the person was allowed, whether the person was paid, and, if so, the benefit amount. However, it appears that records for some disability-denied applicants are overwritten by newer, age-related retirement application and payment information (if there is any).³

Hence this paper uses the SSA "831" disability file to identify DI applicants. An 831 record is established when the Disability Determination Services (DDS) renders an initial medical determination or a reconsideration decision for an individual applying for disability benefits covered under Social Security (Title II) and Supplemental Security Income (Title XVI). Thus, an individual can have more than one 831 record (resulting from multiple applications).⁴ Since this paper analyzes earnings and labor force activities during the period before filing for DI benefits, it uses the earliest observable application date.

Characteristics of the Sample

This paper combined four SIPP panels with extracts of the SER, MBR, and "831" disability files to construct the sample of yearly DI applicant pools for 1977 through 1997. The sample includes DI applicants who are aged 35 through 60 at their first application. The sample excludes those under age 35 because their disability incidence rates are low and they are unlikely to have sufficient past work activity for analysis here. The sample also excludes those over 60 because they are nearing early retirement ages. The resulting sample size is 7,368 DI applicants: 4,943 allowed and 2,425 denied.⁵ The SIPP provides the demographic characteristics (sex, race and ethnicity, and education) for the sample.⁶

This paper reports the distribution of DI applicants by age at earliest time of application for the sample in Table 1 and the distribution of applicants by filing year in Table 2. The aggregated allowance rate of the sample is 67.1 percent.⁷ The sample shows that both the number of applicants and the percentage of allowed applicants increase as age at first application increases. Results presented in Table 2 and Chart 1 are consistent with historical trends: the number of applicants declined from 1979 through 1983, began to increase in 1984, and increased rapidly from 1990 through 1993.⁸

Chart 1.
Number of Disability Insurance applications and awards, 1977–1997

Show as table

Table equivalent for Chart 1. Number of Disability Insurance applications and awards, 1977–1997, (thousands)

Year	All applicants	Denied a	Allowed
1977	1,235.2	666.3	568.9
1978	1,184.7	720.3	464.4
1979	1,187.8	771.1	416.7
1980	1,262.3	865.7	396.6
1981	1,161.2	809.4	351.8
1982	1,019.8	722.6	297.2
1983	1,019.3	707.8	311.5
1984	1,036.7	674.7	362.0
1985	1,066.2	688.8	377.4
1986	1,118.4	701.5	416.9
1987	1,108.9	693.1	415.8
1988	1,017.9	608.4	409.5
1989	984.9	559.3	425.6
1990	1,067.7	599.7	468.0
1991	1,208.7	672.3	536.4
1992	1,335.1	698.5	636.6
1993	1,425.8	790.6	635.2
1994	1,443.8	811.9	631.9
1995	1,338.1	692.3	645.8
1996	1,279.2	654.9	624.3
1997	1,180.2	592.8	587.4

SOURCE: The Annual Statistical Report on the Social Security Disability Insurance Program, 2000 (Social Security Administration; Office of Policy; Office of Research, Evaluation, and Statistics).

a. Technical denials are included.

Demographic characteristics of the whole sample and for applicants who were allowed and denied are reported in Table 3. The mean age of the sample at first application is 48.7, and 74.1 percent of the sample have 12 years of education or less (that is, high school graduate or less). Allowed applicants tend to be older and to have slightly more education than the denied. Differences in these characteristics between the two groups are much smaller than the standard errors of the characteristics, indicating that the predisability earnings differentials may not be attributable to differences in their demographic characteristics.

Descriptive Analyses

Generally, there are two difficulties in using the Social Security earnings data. First, the earnings of workers not covered by Social Security are recorded as zeros. So zeros can indicate that the worker either has no earnings or is not covered by Social Security.⁹ Zero earnings in this study indicate that the worker has no earnings, because the sample consists of only DI applicants who are covered by Social Security. Zero earnings are recorded in years both before labor market entry (the first year of working) and after withdrawal from the labor market. The zero earnings after retirement also do not affect this analysis, because the sample includes only individuals who applied for DI from ages 35 through 60. This age restriction also eliminates most zero earnings occurring before the first year of working.¹⁰ Additionally, this study treats earnings in years before the first nonzero earnings as "missing" rather than as zero. Any zero earnings between the first year with nonzero earnings and the year when the individual reaches age 62 are treated as zeros.

The second difficulty in using the Social Security earnings data arises because recorded Social Security earnings are capped at the annual taxable maximum.¹¹ Fortunately, this censoring appears not to be a significant problem for the sample, as less than 5 percent of the sample reached the taxable maximum. Furthermore, any potential bias from ignoring the taxable maximum would not weaken the findings from this study, because the percentage of those who reached the maximum among the allowed exceeds that among the denied.¹²

Observed earnings around the time of DI applications are affected by the interaction of applicants' self-selection behavior and the substantial gainful activity (SGA) requirement.¹³ Generally, these earnings are considered to be endogenous; thus, they contain no useful information. This study examines each individual's annual earnings in the 3 to 7 years before his or her earliest DI application. Table 4 reports summary statistics of annual earnings in the years from 7 years before through 7 years after the application year.¹⁴ Because applications were filed from 1977 through 1997, this study deflated (or inflated) the annual earnings to 1990 values using national average annual wages.¹⁵ Table 4 contrasts the earnings of the allowed and those of the denied during the predisability period. Both mean and median earnings of the allowed are consistently higher, by approximately $5,000, than those of the denied over all the preapplication years. For example, in the 5 years before application, the median earnings for the allowed and the denied are $15,792 and $9,453, respectively. As expected, the earnings fall sharply around the year of application regardless of allowance status.

Because the earnings (not logged) distributions are highly skewed, this study reports percentage distributions of individuals in specific earnings ranges in Table 5: zero earnings; $1 to $3,000; $3,001 to $6,000; $6,001 to $12,000; $12,001 to $18,000; $18,001 to $24,000; and $24,001 or more. In the 5 years before the first application, 6.15 percent of the allowed have no labor earnings, and 21.38 percent earn $500 or less per month (the monthly maximum SGA amount in 1990). By contrast, 17.18 percent of the denied have no earnings, and 38.83 percent earn $500 or less per month (data not shown in table).¹⁶ Similar differences between the allowed and the denied are observable in all years throughout the preapplication period.

Chart 2 shows the lower end of the earnings distribution: percentages of zero earners; workers with earnings below $3,000; below $6,000; below $12,000; and below $18,000. Significantly higher percentages of the denied are observed in all of these categories, indicating that the denied have lower predisability earnings than the allowed. It is neither surprising nor disturbing to find that the previous earnings of the allowed applicants were higher than those of the denied applicants. To the extent that individuals' applications for disability benefits depend on the generosity of the benefits that could be available to them, the progressivity of the benefit formula will cause low-wage earners to be more likely than high earners to apply for benefits. Because successful applicants must earn less than the SGA threshold, the opportunity cost of applying is higher for high earners than for low earners with similar health conditions. The statistically significant differentials in predisability earnings between the allowed and the denied confirm that the denied are more likely to apply because of economic incentives. Thus, differentials in predisability earnings do not stem from decisions on eligibility but reflect other individual characteristics that affect the application decision.

Chart 2.
Distribution of allowed and denied Disability Insurance applicants at the lower end of the earnings distribution, by earnings level at years before and after application

Show as table

Table equivalent for Chart 2. Distribution of allowed and denied Disability Insurance applicants at the lower end of the earnings distribution, by earnings level at years before and after application (percent)

Earnings range	-7	-6	-5	-4	-3	-2	-1	0	1	2	3	4	5	6	7
	Allowed
Workers with earnings below—
$0	6.54	6.52	6.15	7.28	8.44	10.54	15.13	32.49	72.59	80.71	82.16	81.64	82.1	81.99	81.84
$3,000	14.34	14.11	13.41	14.52	16.17	19.32	26.12	55.09	86.63	89.18	89.09	88.64	87.75	87.8	87.64
$6,000	21.78	22.12	21.38	23.14	23.98	27.03	35.3	68.02	91.53	92.09	91.65	91.09	90.58	90.79	90.39
$12,000	38.94	39.6	38.64	40.26	41.92	44.95	53.93	82.93	95.76	95.69	95.03	94.69	94.09	94.02	93.96
$18,000	55.27	56.11	56.58	56.73	58.08	60.69	68.25	90.23	97.64	97.2	97.19	96.85	96.64	96.48	96.04
	Denied
Workers with earnings below—
$0	14.11	16.09	17.18	18.16	20.97	24.67	30.98	42.9	47.73	48.82	46.55	47.51	47.96	49.19	50.24
$3,000	26.52	26.85	28.53	30.83	33.93	36.43	45.79	65.8	64.07	60.38	57.85	57.49	56.93	58.85	59.37
$6,000	36.86	37.44	38.93	40.28	43.13	45.71	56.14	76.81	73.39	68.95	66.34	66.16	64.31	65.83	65.46
$12,000	55.56	55.73	56.97	58.27	59.93	62.58	71.65	88.69	85.31	81.8	79.74	79.11	77.77	76.38	76.02
$18,000	70.08	70.46	70.47	71.31	72.85	74.75	82.05	94.14	92.28	89.65	88.44	88.2	87.36	85.71	86.2

SOURCE: Author's tabulations using the 1984, 1992, 1993, and 1996 Survey of Income and Program Participation matched with the Social Security Administration's Summary Earnings Record, Master Beneficiary Record, and "831" disability files.

Yearly quarters of coverage (QCs) earned can be used as an alternative measure of labor force attachment to compare the allowed with the denied (Bound 1989).¹⁷ Table 6 reports both mean and median QCs for the allowed and the denied. The mean and median total QCs (during the 10-year period before application) for the allowed are 34.8 and 39.0, respectively. The denied tend to have earned fewer QCs than the allowed—the mean and median of the denied are 29.8 and 33.0, respectively. The table also reports both means and medians of years in which at least one QC and all four QCs were earned. Although these differences in QCs earned are not as dramatic as are the earnings levels, differences are still apparent.¹⁸

Regression Analyses

The following regression analyses quantify both the predisability earnings differential between the allowed and the denied and the average predisability earnings of individuals in the applicant pool by their application years, after controlling for their demographic characteristics. The dependent variable is predisability earnings (7 to 3 years before application) in 1990 constant values (in thousands of dollars).¹⁹ These analyses use both ordinary least squares (OLS) and Tobit on the pooled data.²⁰ These regressions control for demographic variables such as age at application, education, and race/ethnicity and time relative to the application year.²¹ Regression analyses here are designed to show how much of the variation in predisability earnings can be explained by the ex-post disability determination outcomes and by the calendar year in which applications were filed.

As seen in Chart 1, applications and awards for DI benefits underwent three distinct phases: both applicants and awards decreased from 1979 through 1983; they began to increase in 1984; and they increased rapidly from 1990 through 1993. These phases coincided with administrative tightening of enrollment in 1980 through 1983 and the recession in 1991 and 1992. Hence, the regression analyses here can establish a linkage between these application growth patterns and the predisability earnings profiles of yearly applicant pools, particularly when it is focused on the decrease in applications and awards in 1980 through 1983 and the fast growth in 1990 through 1993.²²

Table 7 presents both OLS and Tobit regression results that quantify the differential of the predisability earnings between the allowed and the denied. Coefficients of age and education variables, including the second-order term of age, are statistically significant in both OLS and Tobit regressions. Men tend to earn more than women. Estimates of the time dummies show that the earnings differentials decrease in the years close to application. Perhaps the most interesting finding in these regression results is the coefficient of the dummy variable representing benefit eligibility (allowed). The OLS estimate indicates that the allowed earn $3,717 more than the denied during the 7 to 3 years before the first DI application. The Tobit estimate shows that the allowed earn $4,518 more than the denied.²³

The statistically significant predisability earnings differential between the allowed and the denied confirms the validity of using the predisability earnings as an instrument for disability determination outcomes in detecting the effect of induced applicants on fluctuations in applications. More specifically, the severity of disability, measured by disability determination outcomes, is positively correlated with predisability earnings among DI applicants. This result implies that conclusions drawn from postdenial labor force activity, without considering individuals' labor force activity before application, may misrepresent the counterfactual labor force activity of beneficiaries in the absence of the program.

In studying the labor supply disincentive effect of the DI program, for example, Bound (1989) considered the denied as a control group for beneficiaries. In doing so, he assumed that (1) the denied are healthier than the allowed, and (2) characteristics of the denied, besides health, which are different from those of the allowed, will not lead the denied to low labor force performance. Based on these assumptions, the labor force activity of the denied can be interpreted as an upper-bound estimate for the labor force activity of the allowed in the absence of the program. Bound concluded that only 40 percent to 50 percent of DI beneficiaries would be working in the absence of the program, since fewer than 50 percent of the denied return to work after being denied.

Bound's estimate will overstate the effect of the disability program on labor force participation, however, if the rejected applicants earn systematically less or are less attached to the labor force than are the allowed applicants. The result presented here suggested that, indeed, rejected applicants have lower average earnings than do allowed applicants during the preapplication period. Adjusting for this differential between the allowed and the denied, the labor supply reduction caused by the DI program appears to be at least 10 percent smaller than Bound's estimate, because work-participation rates for the denied are at least 10 percent lower than for the allowed throughout the preapplication period.

Although the above result shows a significant earnings differential between the allowed and the denied, the differential may arise as a result of the time aggregation. To assess whether the earnings differential occurs consistently over the study period, the regression in Table 8 replaces the benefit eligibility dummy with 21 dummies that are obtained from the interaction of application calendar year and benefit eligibility. Thus, coefficients of these interacted dummies (application year and being allowed) capture earnings differentials by year of application. The relative-time dummies and other demographic variables also are included in the analysis, and, as expected, their coefficients remain similar to the estimates reported in Table 7. Coefficient estimates of the interacted dummies are all positive and statistically significant at the 1 percent level.²⁴ Coefficients of these dummies show interesting and meaningful fluctuations. While estimates of the dummies of 1980 through 1983 are the largest, estimates of the dummies of 1990 through 1994 are the smallest. These results indicate that the predisability earnings differentials are inversely correlated with the growth of DI applications.²⁵

Last, to investigate how the average predisability earnings of the yearly applicant pool changes in both the fast-growing and the administrative-tightening periods, application-year dummies (1977–1996) were included as explanatory variables. Thus, coefficients of these dummies represent the relative earnings level of yearly applicant pools.²⁶ Results presented in Table 9 indicate that predisability earnings are the highest among applicants in the 1981–1983 period and that they are the lowest among applicants during the 1990–1994 period.²⁷