DI Classification Errors

Definition

Disability Insurance (DI) classification errors are adjudication mistakes in the Social Security disability determination process. A Type I error (false rejection) occurs when a genuinely disabled applicant is denied benefits; a Type II error (false award) occurs when a non-disabled applicant receives benefits. These error types map directly to the fundamental insurance-incentive tradeoff in DI design: tightening eligibility to reduce Type II errors necessarily increases Type I errors, and vice versa. Low and Pistaferri (2019, 2020) argue that Type I errors are the dominant inefficiency in the U.S. DI system, with a false rejection rate estimated around 37–54% at first contact for claimants who self-report severe, permanent, work-preventing impairments.

Key Ideas

Type I error (false rejection): applicant is truly disabled but is denied DI. Economically costly to the individual: no benefit income, and typically no return to work either.
Type II error (false award): applicant is not disabled by the program's standard but receives DI. Economically costly to the program: benefits paid to someone who could work.
Trade-off: any change in adjudication stringency that reduces one error type raises the other. Optimal policy weighs the welfare costs of each.
Gender gap in Type I errors (Low and Pistaferri 2019): women with severe, permanent, work-limiting impairments are $\approx 21.5$ percentage points (pp) more likely to be falsely rejected than observationally equivalent men. The gap is at the vocational stage (Step 4/5 — residual functional capacity for other work), not at medical assessment. The Social Security Administration (SSA) overestimates women's residual functional capacity.
Scale: The overall DI first-contact award rate is $\approx 37\%$ . Among those with severe, permanent, work-preventing self-reported impairments, Type I error rates are approximately 55% for women and 34% for men (Low and Pistaferri 2019).
Lasting consequences: rejected work-limited women do not return to work (2% employment rate in 3 years post-rejection vs. 19% for rejected non-limited women), confirming genuine errors rather than SSA using superior information.

How It Works

The SSA Sequential Determination Process produces errors when the administrative signal (Disability Determination Services [DDS] evaluator assessment based on medical records, occupational history, and functional capacity forms) diverges from the applicant's true disability status. Four channels can produce higher Type I errors for a demographic group:

True underlying health difference: the group's impairments are less objectively severe on average ( $\lambda_L < 0$ ).
Reporting threshold difference: the group perceives and reports impairments using a different severity scale — e.g., lower "pain threshold" causing over-reporting of borderline cases ( $\lambda_L^* < 0$ ).
Application threshold difference: the group has lower costs of applying, shifting the applicant pool toward marginal cases ( $\lambda_A^* < 0$ ).
SSA evaluation bias: DDS evaluators systematically apply a stricter standard to the group regardless of objective health ( $\lambda_{SSA} < 0$ ).

Low and Pistaferri (2019) use disability vignettes from the 2007 Health and Retirement Study (HRS) to separately identify these channels. For women, channels 1–3 are ruled out (all parameters are positive or insignificant). Only channel 4 — supply-side bias — is statistically confirmed. Vignette respondents rate female hypothetical individuals as 1.7 pp less likely to be disabled (name randomized, same description), approximating a randomized audit of evaluator bias.

Structural Estimates: Low and Pistaferri (2015)

Low and Pistaferri (2015) provide the most granular quantitative estimates of Type I and II errors by age and health state using a structural life-cycle model estimated on Panel Study of Income Dynamics (PSID) data (male heads, high school [HS] or less, 1984–2008). Their three health states — L=0 (healthy), L=1 (moderately work-limited), L=2 (severely work-limited) — allow separate identification of the false-rejection and false-award margins.

Type I error rates (false rejections of severely disabled):

Age group	Type I error rate at first application
Workers under 45 with severe work limitation (L=2)	67%
Workers over 45 with severe work limitation (L=2)	37%

The young-severely-disabled are particularly poorly insured: they face the highest rejection rates and have the fewest assets to self-insure during the gap.

Type II error rates (false awards to moderately disabled):

$\approx 17\%$ of first applications from moderately disabled (L=1) workers are awarded benefits.
Award error (fraction of DI recipients not severely disabled): approximately 12%.

Application elasticity by health state:

Health state	Application elasticity to benefit generosity
All (overall)	0.62
Moderately disabled (L=1)	2.22
Severely disabled (L=2)	0.018

The asymmetry is diagnostic: false applications are almost entirely concentrated among the moderately disabled, who are responsive to incentives. The severely disabled are near-inelastic — they apply regardless of program parameters because they cannot work.

Nonemployment elasticity: 0.056, at the low end of the reduced-form literature (range 0.06–0.93).

Nature of false claimants: predominantly individuals who recovered health while on DI and failed to exit — not people who were never disabled.

Food stamps interaction: food stamps are a substitute for DI among the moderately disabled (reducing false applications) but a complement for the severely disabled (enabling application by raising the consumption floor). A 10% increase in food stamp generosity raises welfare by 0.9% of consumption, more than any DI design change, while reducing false DI applications.

Why It Matters

Program efficiency: if Type I errors dominate in welfare cost, the correct policy response is to loosen the eligibility screen — not tighten it. This inverts the common policy narrative around DI "abuse."
Gender equity: the 21.5 pp excess rejection rate for women is an adjudication disparity unexplained by health, self-reporting, or application behavior differences. Gender-blind evaluation forms are a direct policy lever.
Labor market implications: falsely rejected disabled individuals do not recover to productive employment. The social cost of Type I errors is therefore the full welfare loss of income denial, not offset by labor supply gains.
Racial disparities: the Type I/II error framework also applies to racial allowance gaps. Mitchell and Thompson (2025) find that Black–White gaps are largely compositional (diagnostic mix, age at application), not adjudication errors — a contrast with the gender gap, which is a genuine Type I error.

Open Questions

Does the gender gap in Type I errors persist through the Administrative Law Judge (ALJ) appeals stage? Low and Pistaferri (2019) study first-contact decisions only.
How large are Type I errors for other demographic groups (race, age, education level)?
Does the expedited vocational assessment procedure (piloted 1999, extended nationwide 2012) affect the gender gap by altering how Step 4 is adjudicated?
Would gender-blind application forms reduce the gap, or would it shift to implicitly gendered occupational documentation?