Summary
Gelber, Moore, and Strand (2016) is the companion paper to their 2017 mortality study, using the same Regression Kink Design (RKD) at the bend points of the disability insurance (DI) benefit formula — but here applied to the upper bend point (where the marginal replacement rate shifts from 32% to 15%) and to earnings as the outcome rather than mortality. Using administrative data on all new DI beneficiaries 2001–2007 (n = 610,271 around the upper bend point), they find that a $1 increase in DI payments causes a $0.20 decrease in beneficiaries' annual earnings — an estimate of the pure income effect of DI income on labor supply. This 20-cent income effect is nearly identical to the total crowdout estimated by Maestas, Mullen, and Strand (MMS; 2013) and French and Song (2014), implying that income effects dominate the DI work disincentive and substitution effects from the Substantial Gainful Activity (SGA) threshold are small.
Key Claims
- Main result: −20 cents per dollar. Baseline linear RKD estimate: −$0.2028 per $1 increase in annual DI (p<0.01, standard error [SE] = 0.0224). Range across linear–cubic specifications with and without covariates: −0.19 to −0.27. Remarkably stable across years: Year 1 −$0.23, Year 2 −$0.20, Year 3 −$0.18, Year 4 −$0.20.
- This is the income effect, not the total effect. The RKD variation (slope change in the Primary Insurance Amount–Average Indexed Monthly Earnings [PIA-AIME] formula) affects only the size of the benefit check, not the SGA threshold or program eligibility rules. It therefore cleanly isolates the income channel. The substitution channel — the SGA notch that creates an incentive to stay below SGA — is held constant. This is the key design feature distinguishing this paper from MMS (2013) and French/Song (2014), which estimate the total effect of receipt vs. non-receipt.
- Substitution effect appears small. Only 0.9% of beneficiaries complete the Trial Work Period in any given year; only 0.4% have eligibility terminated for work. The SGA threshold binds for very few beneficiaries in practice. Combined with the finding that the 20-cent income effect nearly accounts for the full 18–19-cent crowdout found by MMS and French/Song, this implies the substitution channel contributes little to the observed work disincentive.
- Null at the lower bend point. No significant earnings effect at the lower bend point (contrast with Gelber/Moore/Strand 2017, which finds large mortality effects precisely at the lower bend point). The null is statistically distinguishable from the upper bend point estimate (p<0.01). The lower bend point's earnings effect is confounded by Supplemental Security Income (SSI) interactions and family maximum rules, but the internal contrast supports the identifying assumption.
- Placebo test confirms causality. Four years before DI entry, earnings show no kink at the upper bend point. Four years after entry, the kink is sharp. The effect appears only post-DI, ruling out pre-existing trends as an explanation.
- Heterogeneity. Women: −35.5 cents (vs. men: −22.9 cents). Younger workers <45: −32.2 cents (vs. ≥45: −18.8 cents). Disability Determination Services (DDS)-initial-allowed: −22.8 cents (vs. Administrative Law Judge (ALJ)-hearing-allowed: −15.4 cents). By disability: circulatory: −28.4 cents; mental: −26.6 cents; musculoskeletal: −14.0 cents; cancer: not significant.
- Extensive margin. Effect on probability of positive earnings: −1.29 percentage points (pp) per $1,000 DI (linear); smaller and insignificant in higher-order polynomial specs, suggesting the intensive margin (how much people earn conditional on working) accounts for most of the effect.
- Policy arithmetic. A 3% benefit cut (e.g., chain-weighted cost-of-living adjustment [COLA]) for a 10-year beneficiary (~$638/year) raises earnings by ~$128/year, yields ~$32 in federal tax revenue (at 25% marginal rate), and returns only ~$2 to the DI fund. Behavioral response offsets to benefit cuts are real but fiscally small.
Concepts Introduced or Extended
- Regression Kink Design — applies RKD to the PIA-AIME upper bend point; isolates the income effect of benefit amount on labor supply independent of program status
- Causal Effects of DI Receipt — isolates the income effect component of the total DI work disincentive; establishes that income effects dominate and SGA substitution effects are small
- Disability Insurance Program — income-effect estimate implies benefit cuts reduce earnings; substitution channel (SGA) binds for very few beneficiaries
Entities Mentioned
Quotes
"Our preferred estimate is that an increase in DI payments of one dollar causes an average decrease in beneficiaries' earnings of twenty cents."
"This suggests that the income effect represents an important factor in driving DI-induced reductions in earnings."
"Only 0.9% of DI beneficiaries are TWP completers in 2012, and only 0.4% have their eligibility terminated due to substantial work."
My Take
The key insight is in the comparison with MMS (2013) and French/Song (2014): those papers find ~18–19 cents crowdout per DI dollar from the total effect of receipt; this paper finds ~20 cents from the income effect alone. If those three estimates are measuring what they claim, then the substitution effect of the SGA threshold is essentially zero — the entire work disincentive of DI comes from the fact that it raises unearned income, not from the strategic incentive to stay below SGA. This is an important finding for welfare analysis: income effects are not "distortionary" in the conventional sense (they don't create deadweight loss from misallocation of labor), whereas substitution effects do. A program that reduces work primarily through income effects is less distortionary than one that reduces work primarily through substitution effects, even if the magnitude of the labor supply reduction is the same.
The companion paper structure — same authors, same design, same formula, different bend point and different outcome — is an unusually tight research architecture. The lower bend point identifies mortality effects with no earnings effect; the upper bend point identifies earnings effects. Together they suggest that DI money at the bottom of the income distribution extends life (via consumption of necessities), while DI money at the top of the distribution reduces earnings (via income effects on labor supply), with both being expressions of the same underlying income-mortality and income-leisure relationships.
One limitation: the upper bend point sample (average annual DI of $21,276) is above the lower bend point sample (average $8,543 in the 2017 paper). The income-earnings relationship being estimated here applies to a somewhat higher-income segment of beneficiaries than the income-mortality relationship. Whether the income effect is the same at the very bottom of the distribution — where the mortality effect is largest — is an open question.