Alho 1997 — Scenarios Uncertainty and Conditional Forecasts of the World Population

Summary

Alho (1997) argues that official high-low population scenarios carry implicit probabilistic content that can and should be made explicit by combining volatility-based (naïve forecast) error assessments with expert judgment. Applied to the 1994 International Institute for Applied Systems Analysis (IIASA) world population forecast, the high-low interval ($8{,}317$–$10{,}736$ million for 2030) functions as roughly an $85\%$ prediction interval (PI), with only $2\%$ probability of falling below the low scenario and $13\%$ probability of exceeding the high. The paper then uses the conditional forecast framework — treating a scenario as a conditional forecast of the population given some policy variable $Z$ is achieved — to show that the competing United Nations (UN) (1993) forecast's narrower high-low interval implies implausible assumptions about the success of member-state demographic policies, with at most $51\%$ coverage under a neutral uncertainty model.

Key Claims

• Naïve/base-line forecast error bound: A naïve (random-walk) fertility forecast assumes future total fertility rate (TFR) = current TFR. For log TFR in industrialized countries, errors follow $e_t \sim N(0, t\sigma_e^2)$ with $\sigma_e \approx 0.05$–$0.10$; Alho uses $\sigma_e = 0.08$. This implies $u/p \approx 1.05$ at $t = 1$ and $u/p \approx 1.40$ at $t = 15$ (where $u$ is the upper interval bound and $p$ is the central forecast value), meaning the expected $67\%$ PI width is $\approx 5\%$ at one year and $\approx 40\%$ at 15 years. Any expert forecast should achieve a smaller ex ante mean squared error (MSE) than the naïve forecast; therefore the naïve error distribution is a conservative upper bound on ex ante forecast uncertainty.
• Scenarios vs. forecasts: Scenario-based projections avoid probabilistic commitment by design, but users invariably interpret the middle scenario as a point forecast and the high-low range as an uncertainty interval. Alho argues this is correct and that forecasters should quantify what probability content their high-low ranges actually carry.
• IIASA 1994 world population forecast:
  • The IIASA high-low fertility interval for industrialized countries ($[1.3, 2.1]$ for Western Europe; $[1.4, 2.3]$ for North America) corresponds to roughly $36$–$38\%$ prediction intervals under the volatility-based model.
  • The mortality high-low is interpreted as ${\sim}50\%$ PI based on Vaupel and Lundström's subjective assessment ($\log e_0(2065) \sim N(\log 92, 0.12^2)$).
  • Propagating these through the cohort-component model for 12 regions: world population in 2030 is approximately $\log T \sim N(\log 9746, 0.083^2)$, giving median $\approx 9{,}746$M (above the IIASA central scenario of $9{,}482$M due to log-normal skewness).
  • $P(T < 8{,}317\text{M}) = 2\%$ (the IIASA low scenario is at the 2nd percentile).
  • $P(T > 10{,}736\text{M}) = 13\%$ (the IIASA high scenario is at the 87th percentile).
  • Therefore the IIASA high-low interval has $\approx 85\%$ probability content (not the intended 80–90%).
  • If regional forecast errors are positively correlated ($6$ effective independent regions instead of $12$), the predictive variance doubles and prediction intervals widen by $\sqrt{2} \approx 1.41$, reducing probability content to $\approx 70\%$.
• Conditional forecast framework: A scenario can be viewed as a conditional forecast: $Y$ is the demographic process without intervention; $Z = a + be_Y + e$ is a policy control variable; the controlled version $Y_Z = Y - Z$ has $\text{var}(Y_Z) = (1-b)^2 \text{var}(Y) + \text{var}(e)$. The parameter $b$ represents how effectively the policy reduces the variance of the demographic process; $a$ represents the mean shift.
• UN 1993 vs. IIASA 1994 (for 2025):
  • IIASA central for 2025: $8{,}955$M; high-low $[8{,}093, 9{,}871]$.
  • UN central for 2025: $8{,}472$M; high-low $[7{,}852, 9{,}080]$.
  • Under the IIASA predictive model, the UN high-low for 2025 has only $51\%$ probability content.
  • To reconcile the UN and IIASA forecasts as both valid $85\%$ intervals, the UN must be conditioning on policies achieving $b = 0.39$ (i.e., reducing $80\%$ of variance); if $95\%$ interval, $b = 0.55$. These values are implausibly high given governments' historical track records with fertility policy.
  • Conclusion: The UN forecast is either a conditional forecast on unrealistically optimistic policy assumptions, or it simply understates uncertainty.
• Official forecasts have not improved since Whelpton (1940s): The cohort-component methodology and the arguments used to justify assumptions have not fundamentally changed in 50 years. Whelpton produced some of the most erroneous forecasts of the 20th century by missing the baby boom — even while having seen signs of rising fertility in his survey data, he judged them short-term fluctuations.

Concepts Introduced or Extended

• Stochastic Fertility Forecasting
• SSA Mortality Forecasting

Entities Mentioned

• Juha M. Alho

Quotes

"Current official population forecasts differ little from those that Whelpton made 50 years ago either in the cohort-component methodology used or in the arguments used to motivate the assumptions. However, Whelpton produced some of the most erroneous forecasts of this century."

"The probability that the world population in the year 2030 will be less than the low scenario of 8317 million [is] only about 2%... The probability that the world population will exceed the high scenario of 10 736 million is about 13%."

"It is only about 51% that the high-low interval of a recent United Nations (UN) forecast will contain the true population in the year 2025."

My Take

The paper's most important contribution is the conditional forecast reinterpretation of scenario differences. Rather than treating the IIASA/UN disagreement as a purely epistemic dispute about demographic dynamics, Alho shows it implies quantifiable assumptions about policy effectiveness — assumptions that can be checked against governments' historical track records. The verdict (implausibly high $b$ values required) is a crisp way to demonstrate that the UN forecast understates uncertainty rather than reflecting genuine conditional information.

The naïve/base-line error bound approach is methodologically elegant: it requires only historical data and no model judgment beyond the random walk assumption, yet it produces defensible conservative intervals that can evaluate official forecasts. This is the fertility-side analogue to what Soneji and King (2012) do for mortality: compare official intervals to a data-driven alternative and find the official intervals too narrow. See SSA Mortality Forecasting.

The regional independence assumption is the main limitation the paper acknowledges. If IIASA researchers applied common modeling frameworks across all 12 regions, positive error correlation is likely, which pushes the true probability content of the high-low interval closer to $70\%$ than $85\%$.

Connection to this wiki: Alho (1997) connects the stochastic forecasting literature (Lee-Carter, Tuljapurkar-Boe) to the broader world population context. The volatility-based error approach introduced in Alho (1990) and extended with Spencer (1985, 1990) is the alternative benchmark that Alho uses throughout his career to evaluate official forecasts — a methodology that has been applied to US mortality (Alho and Spencer 1990) and US Census Bureau fertility (Alho and Spencer 1985). See Stochastic Fertility Forecasting.