How Large are the Classification Errors in the Social Security Disability Award Process?
This paper presents an "audit" of the multistage application and appeal process that the U.S. Social Security Administration (SSA) uses to determine eligibility for disability benefits from the Disability Insurance (DI) and Supplemental Security Income (SSI) programs. We use a subset of individuals from the Health and Retirement Study who applied for DI or SSI benefits between 1992 and 1996, to estimate classification error rates under the hypothesis that applicants' self-reported disability
... tus and the SSA's ultimate award decision are noisy but unbiased indicators of a latent "true disability status" indicator. We find that approximately 20% of SSI/DI applicants who are ultimately awarded benefits are not disabled, and that 60% of applicants who were denied benefits are disabled. We also construct an optimal statistical screening rule that results in significantly lower classification error rates than does SSA's current award process. 1 2 and ρ @ λ © 3 4. The award error rate is estimated to be 21.7%, and the rejection error rate is estimated to be 59.9%. Table 5: Bayes Classification Errors using Alternative Models Model Error Type Award Rejection Type I Type II a . Equicorrelation case: ρ ad © 1 2; ρ aτ ρ dτ @ ρ ad σ τ 1 21.71% 59.94% 23.71% 67.67% b . Asymmetric correlation case with ρ ad © 1 2; ρ aτ © 4 ; ρ dτ © 2 σ τ © 9 19.34% 57.49% 21.44% 55.05% c . Asymmetric correlation case with ρ ad © 1 2; ρ aτ © 5 ; ρ dτ © 1 σ τ © 8 16.10% 52.03% 19.27% 41.60% Panels (b) and (c) show the effect of relaxing the equicorrelation assumption (and the structure of Assumption 3) in a direction that is favorable to the SSA. Panel (b) shows the results in the case where ρ dτ is lowered to © 2 and ρ aτ is increased to © 4 and σ τ is decreased to © 76.6%, with an award error rate of 23.2% and a rejection error rate of 53.5%; that is, Prã 1 © 7 66, Prd 0¤ã 1 © 2 32, and Prd 1 ¤ ˜a 0 © 5 35. 23 Although the magnitude of these classification error rates provides evidence of considerable noise in the DI award process, it is clear that the SSA award decisionã is not arbitrary. Figure 4 indicates thatã is an informative signal, positively correlated withd, that succeeds in partially differentiating disabled and 23 Notice that these numbers are slightly different from the ones in Table 4 due to the fact that in this figure we are not restricting the sample of applications to be in a one year window around an interview date. disability redesign initiative. SSA recognized that "unless SSA invests substantially more funds to research and development of the simplified disability determination methodology, the full benefits of the redesigned process