Alho and Spencer 1985 — Uncertain Population Forecasting

Summary

Alho and Spencer (1985) introduce the approximately linear propagation-of-error framework for assessing the accuracy of official U.S. Census Bureau (USBC) population projections. They show that the USBC high-low variant intervals, when interpreted as $67\%$ prediction intervals (PI), are systematically too narrow — especially at longer horizons — with fertility as the dominant uncertainty source. The paper also formalizes a mixed estimation approach that blends time-series data with expert judgment, anticipating the Bayesian prior-weighting strategy later developed in Lee's stochastic fertility framework.

Key Claims

The approximately linear model $V(t+1) = R(t)V(t)$ — where $V(t)$ is the forecast error covariance matrix at time $t$ and $R(t)$ is the cohort-component Leslie projection matrix — propagates errors through the full demographic projection system; three additive error sources: model misspecification (bias $\eta(t)$ ), parameter estimation error, and random variation.
Fertility is modeled as a first-order autoregressive (AR(1)) process for log age-specific rates; autocorrelation $\approx 0.71$ , cross-correlations across age groups averaging $0.88$ ; data window from 1957 onward (post-baby boom) to avoid structural break; bias parameter $a = 0.31$ for model misspecification.
Migration is the second-largest error source for 15-year population forecasts (after fertility); modeled log-normally; USBC high-low migration range $= 0.8$ probability interval; USBC assumes perfect temporal correlation of migration errors (produces intervals too wide at short horizons, potentially understated otherwise).
Mortality uncertainty is low except at ages 70+, where the USBC middle variant is used as the forecast mean.
Mixed estimation: expert judgment is incorporated as a pseudo-observation at time $n+m+1$ (beyond the data period); weight $\kappa$ controls the blend between data-driven and judgmental forecasts; $\kappa = 0$ recovers the pure time-series estimate, $\kappa \to \infty$ anchors fully to expert priors.
USBC intervals too narrow (Table 3): Alho-Spencer $67\%$ PI upper-to-birth-forecast ratio $= 1.096$ at $t=1$ , $1.315$ at $t=15$ ; USBC high-variant-to-medium-variant ratio $= 1.010$ at $t=1$ and $1.143$ at $t=15$ — USBC high-low is far too narrow for a $67\%$ PI interpretation at any horizon.
Total population 1995: Alho-Spencer point forecast $= 255.409$ M vs. USBC $= 259.631$ M; Alho-Spencer conservative upper $67\% = 267.802$ M vs. USBC high $= 268.834$ M.
Fertility dominates total forecast uncertainty at all horizons; migration is secondary at 15 years; mortality contributes negligibly.

Concepts Introduced or Extended

Stochastic Fertility Forecasting — foundational US application of stochastic error propagation; established fertility as dominant uncertainty source; introduced mixed estimation blending data with expert priors; set benchmark against which USBC variants were too narrow.

Entities Mentioned

Quotes

"The high and low series provide useful sensitivity analyses and are used here for that purpose, but they do not have a probabilistic interpretation."

"The dominant source of uncertainty in our forecasts is fertility."

My Take

Alho and Spencer (1985) is the methodological anchor of the entire Alho-Spencer-Lee stochastic demography program. It is the paper that first puts official US population forecast uncertainty on a probabilistic basis, and its core finding — USBC intervals are too narrow — has been confirmed in every subsequent validation study. The approximately linear model and mixed estimation framework are both underappreciated contributions: the former enables systematic error propagation without full Monte Carlo simulation, and the latter provides a principled way to blend expert judgment with time-series evidence that directly anticipates the Bayesian F* prior-weighting in Tuljapurkar and Boe 1999 — Validation, Probability-Weighted Priors, and Information in Stochastic Forecasts. The 1985 paper's constraint — using post-baby-boom data from 1957 — makes its fertility model conservative about long-run structural uncertainty, which Alho 1997 — Scenarios Uncertainty and Conditional Forecasts of the World Population later addresses through the naïve-error bound.