This paper analyzed the Social Security disability application, appeal, and award process using self-reported data from 12,652 individuals followed for the first three waves of the Health and Retirement Survey (HRS). There is obvious evidence of self-selection in application and appeal decisions: DI applicants are more likely to be black or hispanic, are more likely to be divorced or never married, have much lower education, fewer assets, and substantially lower earnings than non-applicants. However, our estimation results show that many of these socio-economic variables become statistically insignificant predictors of application and appeal decisions once we condition on richer measures of health and disability status. This suggests that most of the socio-economic differences between DI applicants and non-applicants are manifested through socio-economic differences in health and disability status, a result consistent with previous findings of Bound et al. (1995).
Among the health and disability status indicators, we have found that the applicant's self-assessed disability status is the most robust and important predictor of whether an individual will apply for DI benefits or appeal a rejection. We have used a simple indicator variable HLIMPW (equal to 1 if the respondent reports having a health limitation that prevents them from working), which we have interpreted as a measure of the applicant's private information about their ``true disability status''. We have shown that this private information constitutes an approximate sufficient statistic for predicting application, appeal, and award decisions, in the sense that relatively few other socio-economic variables or ``objective'' health status measures emerge as significant predictors of these decisions once we condition on the HLIMPW variable. Much of the previous literature has focused on the possibility that the significant predictive power of self-assessed health and disability indicators is a result of endogeneity bias resulting due to strategic misreporting on the part of respondents who use health problems or disability as a convenient rationalization for other problems they are experiencing in the labor market. In this paper we have assumed that this is not the case. In future research we plan to test this hypothesis, and if accepted use it to assess the magnitude of misclassifications (i.e., the size of Type I and II errors) in SSA disability award process.
The finding that HLIMPW is an
approximate sufficient statistic for
predicting disability application and appeal decisions
is especially important for future work on
estimation of structural dynamic programming (DP) models of individuals'
disability application and appeal
decisions.
In spite of recent improvements in hardware and solution methods,
the computational burden involved in estimating DP models still
grows quite quickly as the number of variables increases, so the feasibility
of this approach depends on our ability to discover relatively
low-dimensional vectors of state variables that adequately
capture the main factors influencing individuals'
decisions.
Our analysis of the disability award process using the HRS data has shown that the conditional probability of being awarded benefits are actually higher at the appeal stage (68%) than at the initial application stage (50%). In future work we will assess whether the higher award rate for appealed case is a result of a combination of large backlogs and excessive leniency at the ALJ stage as suggested in the GAO report discussed in the Introduction, or is a result of valid reversals due to excessive stringency and poor documentation of reasons for denials at the DDS stage. Our results suggest that the relatively large chances of waging a successful appeal could explain why nearly two thirds of rejected applicants choose to appeal. This has the effect of increasing the initial award rate at the DDS level from 50%, to an ``ultimate award rate'' of 72% when we account for the possibility of appeal and reapplication. However, our results also show that there is a substantial delay cost to appealing an initial rejection: the mean delay between application and receipt of benefits for ``first stage awardees'' is 4.6 months but the mean delay between application and receipt of benefits increases to 14.7 months for individuals who received benefits after one or more stages of appeal. This delay imposes a very high cost that deters many rejectees from appealing an initial denial.
As for the SSA decisions, we find that only a relatively small set of variables are good predictors of the accept/reject decisions at either the initial application stage or the appeal stage. It might initially seem surprising that self-assessed disability status is one of the most powerful predictors of award decisions despite the fact that the HLIMPW variable constitutes private information about the applicant that SSA does not observe. However, if this private information is correlated with other information that the SSA does observe, it is not difficult to see why HLIMPW should emerge as a significant predictor of award decisions. Given that one of our primary objectives in this paper is to model individuals' expectations of their chances of being awarded DI benefits, it is important to include all relevant information at their disposal, especially their private information about their own disability status.
Despite the fact that our results are based on self-reported data, we believe that we have been able to construct a fairly accurate model of SSA's disability award process. In particular, our estimates of award rates and delays at various stages of the application and appeal process are fairly close to estimates from administrative records, and our estimates of the success probabilities at each of the stages in the DDS' sequential disability evaluation procedure are quite close to previous estimates by Hu et al. (1997) and Lahiri et al. (1995) using SIPP data with matching Social Security administrative data. However to the extent that individuals are only interested in the ``bottom line'' of whether their application is accepted or rejected, our results suggest that a simple logit model that does not attempt to ``integrate out'' the unobserved basis for award or denial turns out to be a slightly more accurate predictor than a ``marginal probability model'' that captures the details of the sequential disability evaluation process currently used by the DDS.