Completes the fertility of birth cohorts that have not yet reached the end of the reproductive age span, using the method of Myrskyla, Goldstein & Cheng (2013): the age-specific trend observed over the last few years is allowed to continue for a few more years and is then frozen.
Arguments
- data
Data frame in the long format with the columns
age,yearandfx(age-specific fertility rates for periods, i.e. calendar years).- n
Numeric. Number of periods (years) to forecast. By default
NULL, which forecasts just far enough to complete the youngest cohort that has at least one observed rate.- trend
Numeric. Number of the last observations used to estimate the age-specific trend \(b\). By default,
5as in the original article.- extrap
Numeric. Number of periods for which the trend is allowed to continue before the drift is set to zero. By default,
5, the variant with the best forecast accuracy in the original article.extrap = 0gives the "freeze rates" benchmark andextrap = Infan unlimited linear trend.- nsim
Numeric. Number of simulated paths of \(K\) used for the prediction intervals. By default,
10000as in the original article.nsim = 0skips the simulation and returns point forecasts only.- alpha
Numeric. The level of uncertainty. By default,
alpha = 0.05for 95% prediction intervals.- age.int
Numeric. Width of the age interval of
fx. By defaultNULL, which takes it fromage.
Value
A list with three components. fx is a data frame with the observed
and forecast age-specific fertility rates by age and year, with the
columns fx.obs, fx.hat, conf.low and conf.high. cf is a data frame
with completed fertility by birth cohort, with the columns cf (the point
forecast), conf.low, conf.high, last.age (the oldest age observed for
that cohort, so last.age == max(age) marks the cohorts that are fully
observed) and obs.share (the share of completed fertility that comes from
observed rates). model is a list with the fitted quantities A, b,
K, delta, se.delta, sigma and the settings joy, trend, extrap.
Details
Following Myrskyla et al. (2013), period fertility is written as
$$F(t) = A + bK_t, \ \ \ K_t = K_{t-1} + \delta + \epsilon_t$$
where \(F(t)\) is the vector of age-specific fertility rates in period
\(t\), \(A\) is the age schedule of fertility in the last observed period
(the jump-off year), \(b\) is the vector of age-specific trends estimated
over the last trend observations, and \(K_t\) is a scalar time index that
equals zero in the jump-off year by construction.
\(K_t\) is recovered for the last trend periods by regressing the residual
\(F(t) - A\) on \(b\), and the drift \(\delta\) and the innovation
variance are estimated from that short series. The forecast keeps the drift
for extrap periods and sets it to zero afterwards, so that
\(K_{t+j} = \min(j, \mathrm{extrap})\delta\). This is the balance the
authors strike between freezing the rates immediately, which ignores current
trends, and extrapolating them indefinitely, which is not credible over a
long horizon.
Prediction intervals are obtained by simulation, as in the original article: the drift is drawn from a normal distribution centred on \(\delta\), the shocks \(\epsilon_t\) are drawn from a normal distribution with the estimated innovation variance, each path of \(K\) yields one age-period fertility surface, and cohort fertility is read off its diagonal. Forecast rates are truncated at zero.
Completed fertility of cohort \(c\) is the sum of the rates that this cohort
experiences over the whole reproductive age span, taken from the observed part
of the surface where available and from the forecast part otherwise. The
column obs.share reports how much of it is observed, so that cohorts resting
mostly on the forecast can be treated with caution.
In a systematic comparison of 20 methods on the Human Fertility Database this
method had the highest overall forecast accuracy, and it was one of only four
that consistently outperformed the naive freeze-rates benchmark, which is
available here as extrap = 0 (Bohk-Ewald et al., 2018). Note, however, that
its prediction intervals were found to be too narrow in that comparison, as
was the case for most of the methods evaluated.
References
Myrskyla, M., Goldstein, J. R., & Cheng, Y. A. (2013). New Cohort Fertility Forecasts for the Developed World: Rises, Falls, and Reversals. Population and Development Review, 39(1), 31-56. doi:10.1111/j.1728-4457.2013.00572.x
Bohk-Ewald, C., Li, P., & Myrskyla, M. (2018). Forecast accuracy hardly improves with method complexity when completing cohort fertility. Proceedings of the National Academy of Sciences, 115(37), 9187-9192. doi:10.1073/pnas.1722364115
See also
tfr() for the period total fertility rate, tatfr() for the
tempo-adjusted total fertility rate and fert.approx() for fitting
parametric fertility curves, which are meant for smoothing rather than for
forecasting.
Examples
age <- 15:49
year <- 2000:2019
d <- expand.grid(age = age, year = year)
# postponement: fertility falls at young and rises at older ages
d$fx <- 0.09 * exp(-0.5 * ((d$age - (26 + 0.1 * (d$year - 2000)))/5.5)^2)
fc <- fertcast(d, nsim = 200)
head(fc$cf)
#> cohort cf conf.low conf.high last.age obs.share
#> 1 1985 1.347734 1.347728 1.347740 34 0.8846792
#> 2 1986 1.346754 1.346746 1.346763 33 0.8456042
#> 3 1987 1.345097 1.345087 1.345108 32 0.7986763
#> 4 1988 1.342652 1.342641 1.342666 31 0.7441168
#> 5 1989 1.339323 1.339309 1.339340 30 0.6826992
#> 6 1990 1.335038 1.335022 1.335056 29 0.6157467
