Benítez-Silva et al 1999 — An Empirical Analysis of the Social Security Disability Application, Appeal, and Award Process

Summary

Using the first three waves of the Health and Retirement Study (HRS; $n = 13{,}142$ ), this paper provides the first systematic empirical decomposition of the full Disability Insurance (DI) "game tree": the individual's application decision, the Social Security Administration's (SSA) initial award/rejection via Disability Determination Services (DDS), the individual's appeal decision, and the SSA's appeal determination. Binary logit and marginal probability models are estimated separately for each node. The central finding is that a single self-reported indicator — HLIMPW, equal to 1 if the respondent has "a health limitation that prevents working altogether" — functions as an approximate sufficient statistic across all three decision nodes, carrying the largest marginal effect at each stage. The paper also quantifies the option value of appeal: the appeal option raises the effective award rate from $46\%$ to $73\%$ , but triples the duration from application to benefit receipt ( $5$ months $\to$ $15$ months for those who must appeal).

Key Claims

First-stage DDS award rate: $45.9\%$ of applications are accepted without appeal.
Ultimate award rate: $72.5\%$ when the appeal option is accounted for ( $.459 + .541 \times .683 \times .719 = .725$ ).
Appeal rate: $68.3\%$ of rejected applicants appeal; $31.7\%$ do not — despite $52.2\%$ of non-appealers reporting HLIMPW $= 1$ (health limitation prevents work), suggesting delay costs deter even severely impaired individuals.
Duration: Median application-to-award is $4$ months for first-stage awardees vs. $13$ months for those who must appeal (means: $5$ months vs. $15$ months).
HLIMPW as approximate sufficient statistic: Marginal effects — application: $+8.8$ percentage points (pp); appeal: $+19.7$ pp; government first-stage award: $+9.8$ pp (binary logit), $+13.8$ pp (marginal probability model). After conditioning on HLIMPW, most socio-economic predictors (race, marital status) lose significance.
30% incentive-compatibility problem: $30\%$ of DI applicants report HLIMPW $= 0$ — they do not consider themselves disabled. This group is concentrated in the lowest income decile, where the progressive benefit formula generates after-tax replacement rates exceeding $100\%$ .
Age-62 cliff: Application probability drops sharply at age 62+ (marginal effect $-4.1$ pp) despite DI benefits being $\approx 20\%$ higher than early retirement — the "hassle cost" of DI application exceeds the expected utility gain once Old-Age and Survivors Insurance (OASI) early retirement becomes available.
Income and wealth effects: Negative and significant on application probability even after conditioning on health, consistent with incentive compatibility failure for workers facing high replacement rates.
DDS five-stage sequential process (Fig. 1): Initial DDS determination (mean $5$ months) $\to$ Reconsideration ( $2$ months, $50\%$ accept, $48\%$ of denials request) $\to$ Administrative Law Judge (ALJ) ( $9$ months, $75\%$ accept, $75\%$ of reconsideration denials request) $\to$ Appeals Board ( $3$ months, $30\%$ accept, $18\%$ of ALJ dispositions) $\to$ Federal Court (unknown).
Binary logit vs. marginal probability model: The binary logit fits slightly better by Bayesian Information Criterion (BIC) / Akaike Information Criterion (AIC); both models produce comparable predicted acceptance distributions (average acceptance probabilities $55.0\%$ and $53.4\%$ , respectively).
20% of disabled non-applicants in the HRS sample never applied over 1990–1996; $58\%$ of this group were ineligible — confirming that barriers to application are partly legal (non-eligibility), not only attitudinal.

Concepts Introduced or Extended

Entities Mentioned

Quotes

"The appeal option increases the award probability from 46% to 73%. However, this comes at the cost of significant delays: the duration between application and award is over three times longer for those who are awarded benefits after one or more stages of appeal."

"An individual's self-assessed disability status emerges as one of the most powerful predictors of application, appeal, and award decisions."

"Over 30% of DI applicants in our sample do not consider themselves disabled [and are] in the lowest income decile: for this group, the temptation to apply for DI benefits can be quite high."

"HLIMPW 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."

My Take

This is the foundational empirical paper for understanding DI application and appeal as a sequential dynamic process. Three points stand out for this wiki's focus.

The HLIMPW result is the most important methodological contribution. Identifying a single binary indicator that functions as an approximate sufficient statistic across application, appeal, and award decisions validates the reduced-form approach in all subsequent work that uses self-reported disability measures. It also grounds the Autor-Duggan (2003) framework: if HLIMPW predicts application so powerfully, the conditional applicant pool is observable via self-report and its size is directly tied to health impairment prevalence among workers.
The 30% incentive-incompatibility finding is the earliest direct HRS-based quantification of what Autor and Duggan (2003) later theorized as the incentive compatibility failure of the progressive benefit formula. Replacement rates exceeding 100% for low-income workers make application rational even when individuals do not self-identify as disabled. This is the micro-behavioral foundation for DI Replacement Rate and the structural DI growth literature.
The appeal process quantification motivates the entire subsequent literature on application costs and delay costs. The $31.7\%$ non-appeal rate — despite $52.2\%$ of non-appealers being HLIMPW $= 1$ — is direct evidence that delay deters genuinely disabled individuals. This set up the Autor-Maestas-Mullen-Strand (2015) "decay effect" finding that processing delay reduces employment by $0.47$ pp per month. The B-S et al. delay finding is the empirical precursor.

Limitations: The paper uses self-reported HRS data without SSA administrative match, cannot distinguish Social Security Disability Insurance (SSDI) from Supplemental Security Income (SSI) applications, and collapses the multi-stage appeal process into a single node due to HRS data constraints. The paper explicitly defers structural dynamic programming (DP) modeling to future work (see Benítez-Silva, Buchinsky, and Rust 2003). The endogeneity of HLIMPW (rationalization bias) is acknowledged but not corrected here.