Cohort Component Method

Definition

The cohort component method is the standard actuarial and demographic technique for projecting age-structured population forward in time. Starting from a base population distributed across age groups, it applies separate age-specific vital rates (mortality, fertility, migration) to each cohort to generate future population stocks. All stochastic content enters through the forecasted vital rates; the projection itself is a deterministic identity.

Key Ideas

The central projection identity is deterministic conditional on vital rates — all uncertainty comes from the vital rate forecasts feeding into it.
Requires separate forecasting models for mortality, fertility, and migration, linked by the projection identity.
The age structure of the population emerges naturally from cohort survival — no aggregate assumptions needed.
Long-horizon uncertainty accumulates as successive cohorts are projected with uncertain rates.
Standard methodology at national statistical agencies (Social Security Administration (SSA), Eurostat, UN Population Division) for official population projections.

How It Works

Projection Identity

For a discrete-time, single-sex population with age groups $a = 0, 1, 2, \ldots, A$ :

Age 0 (newborns):

$P_t(0) = B_{t-1,t} - D_{t-1,t}(0) + M_{t-1,t}(0)$

where $B_{t-1,t}$ is births between $t-1$ and $t$ , $D_{t-1,t}(0)$ infant deaths, and $M_{t-1,t}(0)$ net migration of infants.

Ages $a \geq 1$ (surviving cohorts):

$P_t(a) = P_{t-1}(a-1) - D_{t-1,t}(a) + M_{t-1,t}(a)$

Deaths are derived from age-specific death rates: $D_{t-1,t}(a) = P_{t-1}(a-1)(1 - \exp(-m_{a,t}))$ , where $m_{a,t}$ is the central death rate for age $a$ in period $t$ . Births are obtained by applying age-specific fertility rates to the female population at risk: $B_{t-1,t} = \sum_a f_{a,t} \cdot P_t^F(a)$ .

Linking to Stochastic Vital Rate Forecasts

In the Meseguer (2010) application:

Mortality rates $m_{a,t}$ are forecasted by a 22-age-group Bayesian vector autoregression (BVAR) per gender.
Fertility rates $f_{a,t}$ are forecasted by a 7-age-group BVAR.
Migration forecasts are taken as external inputs (exogenous, scenario-based).

Given $S$ posterior draws $\{m_{a,t}^{(s)}, f_{a,t}^{(s)}\}$ from the BVAR posteriors, each draw generates a full population trajectory $\{P_t^{(s)}(a)\}_{t,a}$ via the projection identity. The ensemble of trajectories yields a complete forecast distribution for any population aggregate (total, by age group, dependency ratios, etc.).

Uncertainty Propagation: BVAR vs. Lee-Carter

The key advantage over the Lee-Carter framework is that BVAR uncertainty in all age groups enters simultaneously. Lee-Carter collapses the age dimension to a single factor $\kappa_t$ , so the only source of stochastic uncertainty is the variance of the random walk in $\kappa_t$ . The BVAR approach treats each age group as a separate variable with its own shock, capturing both own-age forecast uncertainty and correlated movements across ages — resulting in substantially wider and better-calibrated prediction intervals.

Why It Matters

Provides a principled way to translate stochastic vital rate forecasts into full population distributions — critical for pension solvency analysis, healthcare demand planning, and Solvency II longevity capital requirements.
Allows uncertainty from different sources (mortality vs. fertility vs. migration) to be attributed separately.
When combined with BVAR vital rate forecasts, it inherits the BVAR's well-calibrated predictive intervals, in contrast to the severe underestimation of uncertainty in Lee-Carter-based projections.

Open Questions

Migration is typically the hardest component to forecast and is usually treated as an exogenous scenario rather than a jointly modeled stochastic process, potentially understating total uncertainty.
The projection identity is deterministic conditional on vital rates — it does not capture demographic stochasticity (random births and deaths) that matters for small populations.
Long-run total fertility rate forecasts are extremely sensitive to assumptions about age-specific rate trends; small model differences compound over 50+ year horizons.