Soneji and King 2012 — Statistical Security for Social Security

Summary

Soneji and King (2012) reverse-engineer the Social Security Administration's (SSA) Office of the Chief Actuary's (OACT) previously undocumented mortality forecasting process and replace it with a formal Bayesian statistical model that incorporates demographic smoothness priors and risk-factor covariates (smoking, obesity, each lagged 25 years). They find that SSA's qualitative, expert-judgment-based method systematically overestimates Social Security solvency: substituting their mortality forecasts into the SSA solvency simulator produces negative annual net income 2–3 years earlier, a trust fund balance $730 billion lower by 2031, and program costs 0.66 percentage points higher as a share of taxable payroll. The key mechanism is that SSA's method underestimates future longevity gains — particularly from historically declining smoking rates — so it understates the number of years over which benefits will be paid. Their model produces expected ages at death closer to SSA's high-cost scenario than its intermediate scenario.

Key Claims

• SSA's mortality forecasting process is opaque and unreplicable. The OACT selects 70 "ultimate rates of decline" (5 age groups × 2 sexes × 7 causes of death) through subjective expert judgment, with the final choices approved by Old-Age, Survivors, and Disability Insurance (OASDI) trustees (political appointees). The reasoning behind specific choices is not public. This is the first paper to document the process in sufficient detail for replication.
• Linear extrapolation + qualitative judgment violates known demographic regularities. Age profiles of mortality should be smooth across adjacent age groups and smooth over time. SSA's method produces non-smooth forecasts after age 65 (Fig. 6 vascular disease example). A key demographic regularity — that mortality rates in adjacent age groups are nearly identical — is baked into the Bayesian prior rather than imposed informally.
• The Bayesian alternative formally incorporates two major known risk factors. Cohort smoking prevalence and cohort obesity prevalence, each lagged 25 years, enter as covariates. The 25-year lag is supported by the life-course literature on exposure timing and mortality outcomes and by empirical validation of lag robustness (King and Soneji 2011a). If the priors or covariates are unsupported by data, the model automatically down-weights them.
• Declining smoking raises longevity, increasing Social Security (SS) costs. Smokers die earlier than non-smokers, so they represent a net financial gain to the Trust Funds: they pay payroll taxes but collect fewer lifetime benefits. As smoking rates fall, this actuarial gain diminishes. SSA's intermediate-cost scenario does not fully capture this mechanism, leading to underestimated future benefit outlays.
• Soneji/King model projects longer lives than SSA's intermediate assumption. At age 0 in 2030, Soneji and King forecast male expected age at death between 78.9 and 79.5 years; SSA intermediate projects 77.6 years. The Soneji/King forecast is close to SSA's high-cost scenario (79.4 years), which is the worst-case for solvency.
• The solvency implications are substantial. Compared with SSA intermediate projections in 2031: annual net income is −$217 billion vs. −$128 billion (SSA); trust fund balance is $1.87 trillion vs. $2.67 trillion (SSA); cost rate is 17.70% vs. 17.05% of taxable payroll. Net income turns negative 2–3 years earlier under their model (2023–2024 vs. 2026–2027 for SSA).
• Out-of-sample validation favors the Bayesian model. Forecasting cancer and all-cause mortality 1980–2002, then comparing to observed 2006 data: mean absolute error is 0.005 (Soneji/King) vs. 0.011 (SSA) for cancer-specific mortality; 0.022 vs. 0.028 for all-cause mortality.
• The process may be vulnerable to politicization. Because OASDI trustees (political appointees) must approve the 70 ultimate rates of decline, the forecasting process is structurally exposed to political influence on assumptions. Formal statistical methods reduce — though do not eliminate — this vulnerability by making assumptions transparent and auditable.

Concepts Introduced or Extended

• SSA Mortality Forecasting

Entities Mentioned

• Samir Soneji
• Gary King
• Office of the Chief Actuary
• Social Security Administration

Quotes

"By improving only mortality forecasting methods, we predict three fewer years of net surplus, $730 billion less in Social Security Trust Funds, and program costs that are 0.66% greater for projected taxable payroll by 2031 compared with SSA projections."

"More important than specific numerical estimates are the advantages of transparency, replicability, reduction of uncertainty, and what may be the resulting lower vulnerability to the politicization of program forecasts."

My Take

The paper's main contribution is methodological: it opens the black box of OACT's mortality forecasting and shows that a formal statistical model outperforms expert judgment both in accuracy (out-of-sample validation) and transparency. The solvency finding — that SSA's optimism stems largely from underweighting the longevity gains from declining smoking — is counterintuitive but well-grounded: smokers are actuarially profitable for SS, so their exit from the population raises future costs. The limitation is that the model was calibrated on 1980–2006 data, before deaths of despair (drug overdose, suicide, alcohol) began reversing the longevity improvements the model projects forward. The paper was written in 2012, before Case and Deaton (2015) documented the reversal — so the longevity optimism in the Bayesian model, while better than SSA's intermediate scenario, may itself be overstated in hindsight given post-2013 mortality trends.