This article was downloaded by: [ECU Libraries]

On: 24 April 2015, At: 14:35

Publisher: Routledge
Informa Ltd Registered in England and Wales Registered Number: 1072954
Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Mathematical Population Studies:

An International Journal of
Mathematical Demography
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/gmps20

Population forecasting: Do simple

models outperform complex
models?
a
Andrei Rogers
a
Population Program , University of Colorado , Campus Box
484, Boulder, CO, 80309–0484, USA
Published online: 21 Sep 2009.

To cite this article: Andrei Rogers (1995) Population forecasting: Do simple models
outperform complex models?, Mathematical Population Studies: An International Journal of
Mathematical Demography, 5:3, 187-202, DOI: 10.1080/08898489509525401

To link to this article: http://dx.doi.org/10.1080/08898489509525401

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information
(the “Content”) contained in the publications on our platform. However, Taylor
& Francis, our agents, and our licensors make no representations or warranties
whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and
views of the authors, and are not the views of or endorsed by Taylor & Francis. The
accuracy of the Content should not be relied upon and should be independently
verified with primary sources of information. Taylor and Francis shall not be liable
for any losses, actions, claims, proceedings, demands, costs, expenses, damages,
and other liabilities whatsoever or howsoever caused arising directly or indirectly in
connection with, in relation to or arising out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden.
 Terms & Conditions of access and use can be found at http://www.tandfonline.com/
page/terms-and-conditions
Downloaded by [ECU Libraries] at 14:35 24 April 2015
 Mathematical Population Studies © 1995 OPA (Overseas Publishers Association)
1995, Vol. 5(3), pp. 187-202 Amsterdam B.V. Published under license by
Reprints available directly from the publisher Gordon and Breach Science Publishers SA.
Photocopying permitted by license only Printed in Malaysia

POPULATION FORECASTING: DO SIMPLE MODELS

OUTPERFORM COMPLEX MODELS?*
ANDREI ROGERS
Population Program, Campus Box 484, University of Colorado,
Boulder, CO 80309-0484, USA
Downloaded by [ECU Libraries] at 14:35 24 April 2015

February 20, 1995

This paper reviews the growing literature on population forecasting to examine a curious paradox: de-
spite continuing refinements in the specification of models used to represent population dynamics, simple
exponential growth models, it is claimed, continue to outperform such more complex models in forecast-
ing exercises.
Shrinking a large complex model in order to simplify it typically involves two processes: aggregation
and decomposition. Both processes are known to introduce biases into the resulting representations of
population dynamics. Thus it is difficult to accept the conclusion that simple models outperform complex
models. Moreover, assessments of forecasting performance are notoriously difficult to carry out, because
they inevitably depend not only on the models used but also on the particular historical periods selected
for examination. For example, the accuracy of the Census Bureau's forecasting efforts apparently has
improved during the past two decades. How much of this improvement is due to improved methods and
how much to the decreased variability in the components of change? Clearly one needs to introduce the
dimension of "degree of difficulty" into each assessment.
This paper reviews some of the recent debate on the simple versus complex modeling issue and links
it to the questions of model bias and distributional momentum impacts.

KEY WORDS: Population geography, population forecasting, population models.

Communicated by Andrei Rogers.

1. INTRODUCTION
During the past half a century, the U.S. Bureau of the Census has been produc-
ing population projections that over time have become both methodologically more
sophisticated and demographically more detailed. Yet this added complexity has
not invariably led to increased accuracy (Long, 1993). Nevertheless, the "complex"
cohort-component method, used (in various implementations) by the Bureau since
1945, has become the dominant population projection model virtually everywhere,
because it takes advantage of the built-in momentum of age structure effects. It
is this age momentum, for example, that creates concern about the expected rising
growth rate of the 65 year and older population after 2010, when the Baby Boomers
will begin to join the elderly population. However, despite this advantage over

*The author thanks Robert McNown and Dennis Ahlburg for helpful discussions and comments.

187
 188 A. ROGERS

'Tor short-term forecasts, simple projection methods, such as assuming constant geometric growth,
are at least as good as complicated ones." (Cohen, 1986b, p. 122).
"A number of studies have concluded that simple extrapolation techniques produce short- to medium-
term forecasts of total population that are at least as accurate as those produced by more sophisticated
techniques . . . " (Smith and Sincich, 1988, p. 463).
"So for short-term, total population projections simple geometric projection gives more accurate re-
sults than the more complicated component method." (Stoto, 1983, p. 18).
"Elaboration of projection methodology has not resulted in any great increase in the precision of the
projections . . . " (Siegel, 1972, p. 51).
"A general law seems to be at work: More complex, subtle, or elegant techniques give no greater
accuracy than simple, crude, or naive ones." (Pant and Starbuck, 1990, p. 442).
FIGURE 1. A selection of quotations regarding the performance of projection models.

simple aggregate exponential growth models,

Downloaded by [ECU Libraries] at 14:35 24 April 2015

"There seems little correlation between methodological innovation and the accuracy of the
projections. In most cases the constant growth scenario does as well or better than the more
complex models in predicting future total growth rates." (Long, 1993, p. 10)

Similar conclusions have been reached in the literature on population forecasting,

as evidenced by the quotations set out in Figure 1.
There are a few challenges to this conclusion. Keyfitz (1981, p. 589), for example,
found that the ex-ante errors of simple growth models that assumed a future con-
tinuation of the rate of increase of the past five years produced root-mean-square-
errors that were twice as large as those produced by cohort component models fore-
casting the same national population. And Beaumont and Isserman (1987, p. 1008),
commenting on Smith's (1987) argument in favor of simple models, call the claim
a shibboleth of population forecasting and observe that little rigorous evidence has
been put forward in favor of simple models.
This paper reviews the debate over forecasting performance, considers the ad-
vantages held by complex cohort-component models over the corresponding simple
models derived by aggregation or decomposition, and concludes with a number of
observations that need to be examined further. Since assessments of forecasting
performance in the literature (such as those quoted in Figure 1) are expressed in
terms of such ill-defined words as simple, complex, naive, sophisticated, crude, and
elegant, it becomes necessary to address the issue of semantics from the start.
How are simple models different from complex models? And how are naive mod-
els different from sophisticated models? Smith and Sincich (1992, p. 496) offer a
useful typology:
"We define a naive model as one in which future population values depend solely on past pop-
ulation values, whereas a sophisticated model is one in which population change is expressed
as a function of changes in economic and/or other variables. Under this definition, trend and
ratio techniques are naive but their mathematical forms may be either simple (e.g. linear ex-
trapolation) or complex (e.g. ARIMA time series models). Cohort-component techniques in-
volve complex, disaggregations of population change, but they may be either sophisticated or
naive, depending on whether they use structural models for projecting fertility, mortality and/or
migration. Structural models are by definition sophisticated, but they may be either stochastic
or deterministic and the models themselves may vary considerably in terms of complexity."
 POPULATION FORECASTING 189

Because this two-by-two typology seems to be becoming accepted by others and

is adequate for the purposes of this discussion, it is adopted in the remainder of
this paper, although I personally prefer and will use extrapolative instead of naive
and causal instead of sophisticated. To clarify the typology and give it substance,
consider its possible use to characterize different existing forecasting models of the
urban and rural parts of a nation's population.
The United Nations regularly issues forecasts of the urban and rural populations,
and associated urbanization levels, for all of the countries of the world. The core of
their method is the application of a simple exponential growth model to the urban-
rural ratio of a nation, with a growth rate that is an extrapolation into the future
of the most recently observed urban-rural growth rate difference. This procedure is
known as the United Nations URGD method. (United Nations, 1984). It can easily
be demonstrated that an exponential evolution of the urban and rural populations
Downloaded by [ECU Libraries] at 14:35 24 April 2015

contained in such a ratio is equivalent to assuming that the percent urban curve
follows a logistic function over time, with the URGD being the logistic growth
rate.
If the UN method of forecasting urbanization is an illustration of the simple ex-
trapolative model in the two-by-two typology, what is an example of a simple causal
model? Here we may turn to any number of studies that have sought to associate
urbanization levels with levels of economic development as measured, for example,
by GNP per capita (e.g., Chenery and Syrquin, 1975; Ledent and Rogers, 1986).
Multiregional cohort-component models of urban growth and urbanization oc-
cupy the complex extrapolative cell of our typology matrix (Rogers, 1985). Incorpo-
rating demographic accounting principles involving fertility, mortality, and migration
rates for each age-sex-cohort in the two subpopulations, the methodology may be
said to be complex because, although the arithmetic is simple and transparent, the
projection process becomes somewhat involved. Of course, the arithmetic can be
made less simple by the specification of nonlinear dynamics, for example, by incor-
porating the "gravity" model of rural and urban outmigration rates formulated by
Sam Preston and published as a footnote in the United Nations (1980, p. 10) study:
Patterns of Urban and Rural Population Growth.
Finally, a classic example bf a complex causal model of urban growth is offered
by the computable general equilibrium model developed by Kelley and Williamson
(1984) in their study: What Drives Third World City Growth? In it they argue that
urbanization and the growth of cities can only be understood by embedding both
processes in a general equilibrium framework, because economic growth not only
influences urbanization but is, in turn, influenced by it. Within the typology set
out above, this paper considers the relative forecasting performances of simple and
complex extrapolative models.

2. THE ASSESSMENT OF MODEL PERFORMANCE

Most assessments of the forecasting performance of alternative models have focused
on the differences between observed population totals and growth rates and those
 190 A. ROGERS

predicted by the particular models being evaluated. Excellent examples of such

studies are two papers written over a decade ago by Keyfitz (1981) and Stoto
(1983), respectively, and a more recent review and update by Grummer-Strawn and
Espenshade (1991). All three efforts focus on the difference between the average
annual growth rate implied by the projection and the one actually observed sub-
sequently:
Ar = Tp-r0 (1)

This statistic has the virtues of being independent of population size and of the
length of the projection interval, whereas the difference between the absolute num-
bers of people depends on population size, and percent error doesn't take into ac-
count the length of the time period over which the projection is carried out.
Keyfitz (1981) examined distributions of Ar of some 1,100 projections—develop-
ed by different agencies, at different times, and for different countries—with their
Downloaded by [ECU Libraries] at 14:35 24 April 2015

corresponding realizations. Comparing these projections, generated by cohort-com-

ponent methods, with the corresponding projections produced by the simple expo-
nential growth model, he concluded that the former extrapolative models outper-
formed those based on simple geometric increase.
"Aside from assessing projection methods in absolute terms, I have examined how they com-
pare with naive methods. One way of doing so is simply to find the rate of increase of the
preceding five years and carry the population forward into the future with that rate of in-
crease. The departure of that rate from that materialized shows a root-mean-square error of
0.90 ... as against the 0.48 of actual forecasts. Judgment and the techniques of demography do
indeed diminish forecast errors." (Keyfitz, 1981, p. 590)

Stoto (1983), on the other hand, examining many of the same projection efforts,
and also focusing on Ar as the indicator of accuracy, came to the opposite conclu-
sion:
"The U.N. data allow us to evaluate this technique for each of the 24 regions ... For these
data the simple geometric projection technique has been almost unbiased and has a standard
deviation equal to or smaller than the more complex methods." (Stoto, 1983, p. 18).

What accounts for the opposite conclusions reached? Both scholars adopted the
same indicator of accuracy. But Keyfitz used a fixed base period of five years (1950-
1955) and focused on the root-mean-square error, whereas Stoto adopted a "float-
ing" base period, which at times was a five-year base period and at other times a
ten-year period, and he focused on the standard deviation as the indicator of accu-
racy. Although these differences are unlikely to totally explain the opposite conclu-
sions reached, they undoubtedly had a significant impact and illustrate the need for
a fair competition, or tournament, in which a common set of procedures are applied
throughout, particularly with regard to the choice of length of base period and of
the error index. For example,
"In fact, the choice of length of base period may be just as important as the choice of method
for the accuracy of a projectioa" (Beaumont and Isserman, 1987, p. 1005).

With respect to selection of an index of error, it is instructive to consider the

literature on alternative error measures and their relative effectiveness in choosing
accurate forecasting methods (see, for example, Armstrong and Collopy, 1992, and
 POPULATION FORECASTING 191

Ahlburg, this issue). The popularity of the root-mean-square error, for example, is
sobering in light of the wide recognition among forecasters of its unreliability.
Observed ex-post forecast errors, of coursé, depend not only on the projection
methodology used but also on the particular historical periods selected for exami-
nation. As John Long points out, Census Bureau projections have had difficulties in
anticipating periods of rapid rises or rapid falls in fertility. " . . . the first 25 years
of Census Bureau forecasting activity corresponded to the 'baby boom' and conse-
quent 'baby bust'... projections made before 1955 and between 1966 and 1970 were
uniformly worse than the naive assumption of constant growth rates. Projections
made by the Census Bureau for 1955 through 1966 and after 1970 were generally
better than the naive assumption." (Long, 1987, p. 7 and p. 9).
The accuracy of the Census Bureau's national population forecasting efforts ap-
parently has improved during the past decade and a half. How much of this im-
provement is due to improved projection methods and how much is due to the
Downloaded by [ECU Libraries] at 14:35 24 April 2015

decreased variability in the components of change?

If a population is experiencing close to stable growth, then even the simplest
model will perform about as well as complicated alternatives in forecasting that
population's evolution. To conclude in such instances that simple models outperform
complex models obviously is not a rigorous test. And to accord this performance
the same weight as one in which the assessment involved a population experiencing
an unexpected baby boom is unfair. Obviously one needs a way of introducing the
dimension of "degree of difficulty" into each assessment In diving competitions,
the degree of difficulty for a swan dive is considerably lower than for a double
somersault. The assessment of the diver's performance is weighted by that degree
of difficulty. Perhaps an analogous weighting should enter assessments of forecasting
performance.
The problem of comparing performance in the face of several different degrees of
difficulty can be relaxed somewhat by comparing forecast performance over many
historical periods. But the input requirements for such a comparison are rarely
available. Keilman (1990) suggests a statistical method for controlling for degree
of difficulty by applying an age-period-cohort model that strives to separate the
contributions of the confounded effects to forecast accuracy of ex-post observed
demographic trends and of the forecasting methodology.
Keilman's method is dependent on the availability of accurate demographic data,
describing a relatively long history of population evolution, together with a signifi-
cantly large number of associated efforts to forecast that evolution. That serendip-
itous combination, though not limited to the Netherlands, is not found in many
countries today. A possible alternative "fair test" of the performance of alternative
simple and complex extrapolative models would be to apply them to simulated pop-
ulation evolutions generated by Monte Carlo methods that select randomly among
a wide range of hypothetical schedules of fertility, mortality, and migration to create
particular histories evolving from particular demographic regimes. Such a "tourna-
ment" would test the performance of alternative models fairly, and it would illumi-
nate the contributions made by the various components of the projection method.
In particular, it could reveal the particular conditions that tend to be associated
with "wins" for simple models over complex models.
 192 A. ROGERS

3. WHEN SIMPLE MODELS OUTPERFORM COMPLEX MODELS

The simple growth models used by Keyfitz (1981), Stoto (1983), Smith (1987) and
others are general in nature and could be used as readily to project national in-
comes or automobile sales. They do not take advantage of an important attribute
of human populations, namely, that individuals age one year at a time. And they do
not consider the fundamental demographic events and associated accounting iden-
tities that underlie population change. How is it, then, that in short-run projections
they often produce reasonable results and, on a number of occasions, results that
are more accurate than those produced by more complex age-specific extrapolative
models, which are built on the basic demographic accounting relationships familiar
to us all?
A possible, though not thoroughly convincing, answer is that no hard evidence
exists to support the conclusion that simple models have outperformed complex
models. For example, Beaumont and Isserman argue that:
Downloaded by [ECU Libraries] at 14:35 24 April 2015

"The literature on accuracy does not constitute a test ... comparisons are few and hardly
constitute a rigorous test. Thus the accuracy argument in favor of continuing to use simple
methods for now at least, must be the weak one, based only on the fact that no rigorous tests
exist" (Beaumont and Isserman, 1987, p. 1008)
Although the above remark was directed at a comparison of simple extrapolative
models versus complex causal models, it is also true, though perhaps to a lesser
degree, of a comparison versus complex extrapolative models. Nevertheless, since
even Census Bureau forecasters such as Long (1993) have admitted that simple
models have outperformed their complex cohort-component models in the past, it
is instructive to examine the conditions under which this counter-intuitive result
might arise.
We begin by clarifying our terminology, which is illustrated in Figure 2. The base
period extends from the initial base year, the first year for which observed demo-
graphic data are used to estimate the model's parameter(s), to the launch year, the
last year for which observed data are available and the "jump-off" point for the
forecasting exercise. The target year is the moment in time for which the forecast
has been made, and the forecast period is the time horizon across which the forecast
is being carried out, i.e., the target year minus the launch year. The forecasted pop-
ulation is the ex-ante total predicted for the target year; the observed population
is the ex-post total that is observed during that year; and the confidence interval
that spans the lower and upper values straddling the forecasted total is an ex-ante
statement regarding the expected accuracy of the population forecast.
Studies of the accuracy of past Census Bureau population forecasts present a
clear message: simple models have outperformed complex models at major turning
points in U.S. demographic trends. For example, the forecasters did not anticipate
the Baby Boom, and after it began they expected it to continue. Thus their early
forecasts were too low and their later ones were too high. Simple models that av-
eraged past trends produced more accurate forecasts in such situations (Ahlburg,
1982).
At times of relative stability in demographic trends, on the other hand, the com-
plex (cohort-component) models outperformed the simple models. For example,
during the relatively stable periods of high fertility following 1957 and 1963 and of
 POPULATION FORECASTING 193

Population Totals Upper Confidence

Bound \
A

Forecast Population I Confidence

Totals / Interval
Ar-r, -r.
Observed Population
Totals

Lower Confidence
Downloaded by [ECU Libraries] at 14:35 24 April 2015

Bound

Initial Base Year Launch Year Target Year

i>Calendar Years
FIGURE 2. Illustration of the terminology.

low fertility following 1974, 1976, and 1982, the Bureau's complex models outper-
formed the simple models (Long, 1987). The reduced variability in fertility rates
allowed the cohort-component method to take advantage of its ability to incorpo-
rate age momentum effects by tracing the impacts of changing age compositions in
the childbearing ages on aggregate fertility levels.
It appears, then, that complex models have outperformed simple models in times
of relative stable demographic trends, when the degree of difficulty has been rel-
atively low, and have been outperformed by simple models in times of significant
unexpected shifts in such trends, when the degree of difficulty has been relatively
high. Why has this been so? One would expect the opposite to have been the case.
Smith and Sincich (1992, p. 505) point to the difficulty of anticipating future trends
in the basic demographic components of change as a possible reason:
"We believe the cohort-component projections were no more accurate than the trend and
ratio projections because forecasting fertility, mortality, and migration rates is just as difficult
as forecasting changes in total population (probably more so). This difficulty counteracts the
advantages of disaggregation and the stability the age-sex structure adds to cohort-component
projections. Will the application of time series techniques ... improve the forecast accuracy
of current and future cohort-component projections? ... We do not think so ... We believe
it is unlikely that more complex approaches to extrapolating past trends in cohort-component
models will lead to any significant improvements in forecast accuracy."

But why should it be easier to forecast changes in the evolution of the annual
aggregate growth rate, which by definition is a function of the basic components
 194 A. ROGERS

of demographic change, than to forecast the underlying changes in the evolution of

those components? Moreover, in the assessments of Keyfitz (1981), Stoto (1983),
and Smith and Sincich (1992) no trend extrapolation of the aggregate growth rate
was carried out. In each instance an average across the preceding five or ten-year
base period was adopted and assumed to remain fixed across the forecasting pe-
riod. This suggests the following conjecture: when simple models have outperformed
complex models, it has been a consequence of a serendipitous averaging of aggre-
gate trends during times of relative demographic instability.
Because an average growth rate is more conservative in its growth impact than
is an extrapolation of past trends (whether increasing or decreasing), its use gener-
ally leads to a lower error for times that experience a turning point. For example a
simple projection in 1945 of the future U.S. population, using the average fertility
regime that prevailed during the preceding five years, would underproject that pop-
ulation less than one based on an extrapolation of the decline. Conversely, a similar
Downloaded by [ECU Libraries] at 14:35 24 April 2015

projection exercise carried out a dozen years later would find the simple model
overprojecting the U.S. population less than one based on an extrapolation of the
Baby Boom fertility. However, since neither the proponents of simple or complex
models can anticipate such changing times, it behooves a public provider of fore-
casts, such as the Census Bureau, to "expect the unexpected," adopt the complex
model, and strive to improve the quality of the input assumptions, rather than resort
to simple growth models, even in those few situations that only need an accurate
forecast of total population.
Of course, most uses of population forecasts require accuracy for subgroups of
the total population. Plans for future schools and of future nursing homes, after all,
depend on forecasts of different population subgroups. And in such efforts, cohort-
component models generally outperform simple growth models (Long, 1993).
Finally, the accuracy of a forecast is not the only dimension on which forecasting
models should be judged. Long (1987) suggests the additional dimensions of face
validity, internal consistency, and level of detail. Face validity refers to the reason-
ableness and "believability" of the model and assumptions that were used to gener-
ate the forecast. Internal consistency refers to the inclusion in the modeling process,
of accounting mechanisms for ensuring that standard demographic identities are sat-
isfied. And, of course, level of detail refers to the disaggregations needed to satisfy
the particular needs of the users of the forecasts. All three point to the desirability
of age-sex disaggregated cohort-component models, which we have defined as com-
plex extrapolative models. Griffith Feeney identifies yet another powerful reason
for adopting complex over simple forecasting models:

"The ... reason ... has to do with the outcome of the forecast If the forecast is unsuccessful,
as so many are, there may be an inquiry of sorts. Should it turn out that a simple method was
used, criticism is likely to be intense. Surely a more sophisticated approach would have yielded
a better result, or at least a higher probability of a better result. The producer is at risk of
appearing lazy, ignorant, incompetent." (Feeney, 1988, p. 5).

Simplifying complex models is an art. Regrettably, there are only a few research
studies that have addressed the issues surrounding this activity. As a result, few
"rules-of-thumb" have been developed to aid professional demographers entrusted
 POPULATION FORECASTING 195

with the task of regularly issuing population forecasts. What is clear from the lim-
ited research findings now available is that the process of simplifying a demographic
projection model comes with a price in the form of bias and an absence of built-in
momentum effects.

4. SPECIFICATION BIAS AND MOMENTUM: SIMPLIFYING

COMPLEX EXTRAPOLATIVE MODELS BY MEANS OF
AGGREGATION AND DECOMPOSITION
Shrinking a large complex extrapolative model in order to simplify it typically in-
volves two processes: aggregation and decomposition (Rogers, 1976). Aggregation
seeks to reduce the scale of a model by a consolidation along its three fundamen-
tal dimensions of population characteristics (e.g., age, sex, race), time units (e.g.,
Downloaded by [ECU Libraries] at 14:35 24 April 2015

five-year instead of one-year projection subintervals) and spatial units (e.g., the 9
Census Divisions instead of the 50 states). Decomposition seeks a partitioning of
the total population system that exploits the possibility of treating parts of the sys-
tem separately from the rest. Using age-specific net migration rates, in place of
the corresponding origin-destination-specific migration rates, is an example of de-
composition. Consolidating such net migration rates into a single crude rate is an
example of aggregation.

4.1. Aggregation Bias and Decomposition Bias

Demographers disaggregate data because they believe that this will lead to the iden-
tification of stable and consistent relationships and measures that can be extrapo-
lated into the future. Simplicity is hidden in the aggregate data they argue, as they
search for a disaggregation of "the population into sub-classes for which change
over time was so regular that prediction would become a formality." (Brass, 1974,
p. 542).
Demographer's also disaggregate data to minimize "the ruses of heterogeneity"
(Vaupel and Yashin, 1985). Cross-sectional comparisons of demographic rates are
known to depend on the variables on which the comparison is conditioned. In math-
ematical demography this dependence is often illustrated by various examples of
what is called Simpson's Paradox—an apparent contradiction of two comparisons
that arises as a consequence of the stratification of populations into two or more
subgroups and the consequent reversal of the rank ordering of those populations
on the variable of interest. As Joel Cohen rightly observes such counterintuitive
reversals may occur in comparisons of a wide array of demographic process:
"Simpson's paradox can occur in any comparisons of probabilities, rates, or measurements that
are weighted averages of component probabilities, rates, or measurements from subgroups.
Thus Simpson's paradox is a potentially widespread phenomenon." (Cohen, 1986a, p. 33).

Most illustrations of Simpson's paradox have focused on cross-sectional compar-

isons of measures and weighted averages that do not depend on selection over time.
Recently, Vaupel and Yashin (1985) broadened this perspective to include the dy-
namics of selectivity in the presence of unobserved heterogeneity, dynamics which
 196 A. ROGERS

bring about changes over time in the relative weights themselves. Both kinds of
comparisons may be used to demonstrate the potential bias that aggregation can in-
troduce into the projected population dynamics of heterogeneous populations. Sim-
ple models in such applications tend to produce more selection biases than com-
plex models, in projections carried out with fixed rates. If they outperform complex
models in environments with changing rates, then their superiority surely must be
a consequence of a serendipitous combination of offsetting changes in demographic
behavior.
Decomposition is often a complement to aggregation in efforts to simplify a com-
plex model. The simplification permits the representation of multiregional popula-
tion dynamics "one region at a time." Smith and Sincich (1992), for example, use
such a simplification when they adopt a simple exponential growth model to repre-
sent the population dynamics of the 50 U.S. stales.
The notion that the dynamics of a system of n multiple interacting subpopulations
Downloaded by [ECU Libraries] at 14:35 24 April 2015

can be analyzed profitably by decomposing the system into n independent models of

each subpopulation's evolution over time, dies hard. Despite over a quarter century
of published work on the dynamics of multistate populations, one still finds innu-
merable articles in prominent journals that ignore this literature and continue to
follow the "unistate fallacy."
Describing the U.S. 50-state population system as a simultaneous system of 50
interdependent equations is an example of a multiregional representation of the na-
tional population. Describing the same system with 50 independent single equations
illustrates the corresponding uniregional representation. Projections carried out with
both representations will differ because a decomposition bias is present in the latter
approach, a bias that arises from representing directional migration streams as net
flows. Thus, for example, Florida's year 2000 population was projected to increase
to 17.4 million by the Census Bureau in 1983 using a net migration model. Five
years later the corresponding projected total for Florida dropped to 15.4 million,
when the Bureau adopted a multiregional perspective.
Decomposition bias is an unrecognized contributor to the controversy between
the Census Bureau and a number of State Demographers regarding the plausibil-
ity of the Bureau's recent state projections. The Bureau's 1988 projections for the
first time reflect a multiregional perspective, whereas those produced by most State
Demographers continue to be based on the uniregional perspective.1 The resulting
differences arising out of decomposition bias has led to disagreements regarding
future state population totals (Rogers and Woodward, 1991).
Since net-migration-based representations produce a decomposition bias, and be-
cause simple exponential growth models must adopt that form of specification, it is
difficult to understand why such simple models should outperform the more com-
plex multiregional cohort-survival models that deal with migration flows, except for
reasons related to the serendipitous cancellations of errors.

l
A few states such as Florida, have adopted a minimum multiregional perspective in which net migration
is replaced by in- and out-migration. This results in a biregional model in which the Rest of the United
States is the second region. This representation should reduce decomposition bias.
 POPULATION FORECASTING 197

4.2. Distributional Momentum in Population Dynamics

Since age-specific rates of demographic events vary in a predictable way with age,
the current age distribution of a population tells us something about demographic
changes that the population experienced in the past, as well as those that are likely
to occur to it in the future. The likely future changes are commonly referred to as
age momentum effects.
A particularly powerful illustration of the projection impact of age momentum ef-
fects occurs when the age distribution of a rapidly growing population is favorable
to further increase. Whenever an initial population distribution differs from the sta-
tionary distribution that would arise were the current fertility regime to immediately
drop to bare replacement level, a "momentum" is associated with that population's
projection, its magnitude defined by the ratio by which the ultimate stationary pop-
ulation exceeds the current one. Typically, if the initial age distribution is that of
Downloaded by [ECU Libraries] at 14:35 24 April 2015

a "young" population whose high fertility level is assumed to immediately drop to

bare replacement level, then population growth will nevertheless continue on for
some time (about 60-70 years) before zero population growth (i.e., stationarity) is
achieved.
An analogous result holds with respect to an initial divergence from the stationary
spatial distribution. For example, India's urban population would continue to grow
for several generations, even if fertility levels throughout India were miraculously to
drop immediately to bare replacement levels. Age momentum would be one source
of continued growth; spatial momentum, in the form of continued rural-to-urban
migration, would be another.
The importance of spatial momentum can be illuminated further by considering
how different spatial alternatives for reducing national fertility to bare replacement
level might affect the resulting population dynamics (Rogers and Willekens, 1978).
Imagine how India's urban population growth would evolve if the rates of childbear-
ing in urban and in rural sectors of the country were immediately reduced to bare
replacement levels (i.e., a net reproduction rate of unity for each urban- or rural-
born cohort). Now contrast this evolution to zero population growth with the one
arising from an alternative fertility reduction scheme in which urban and rural fer-
tility rates were reduced proportionately to immediately arrive at a national fertility
regime that is at bare replacement level (i.e., a reduction after which the rural net
reproduction rate still would be above unity but the corresponding urban rate would
be below unity). Clearly, the population momentum impacts of the two alternative
scenarios are different, and that difference is a reflection of spatial momentum ef-
fects.2
Age momentum and spatial momentum effects are suppressed when an aggrega-
tion across all age groups and a decomposition of all regional subpopulations are
carried out simultaneously. Simple exponential growth models, therefore, carry no
momentum impacts in their forecasting applications, and the loss of that impact can

2
As a further illustration of the effects of decomposition bias, the reader may wish to consider how one
would deal with the twin notions of bare replacement level fertility and zero population growth if the
urban and rural populations were each to be represented by a uniregional cohort-component model.
 198 A. ROGERS

0J032
PAST BIRTH RATES

1920-1925 1930-1335 1940-1945 1950-1955

1915-1920 1925-1930 1935-1940 1945-1950
YEAR
Downloaded by [ECU Libraries] at 14:35 24 April 2015

0.04

CURRENT AND FUTURE ELDERLY

0.03
GROWTH RATES
0X)2

o:
0.01

-0.01
1985-1990 1995-2000 „„„2005-2010 2015-2020
1980-1985 1990-1995 2000-2005 2010-2015

YEAR
FIGURE 3. Past birth rates and current and future elderly growth rates (Source: Rogers and Woodward,
1988).

produce serious biases which contribute to inaccurate forecasts. Consider, for exam-
ple, the problem of projecting the U.S. elderly population with simple and complex
models. The 66 year-old population of next year will consist of the survivors of this
year's 65 year-olds, who were 64 years old last year, 63 years old the year before
that, and, of course, who were born 65 years ago. Figure 3 reflects this relationship
between today's elderly and yesterday's births in its plot of the curve of crude birth
rates from 1915 onward and the corresponding curve of the elderly population's
annual growth rate 65 years later, i.e., from 1980 onward, including a projection to
2020. Not surprisingly, the latter curve is almost a perfect mirror image of the
former. Cohort component models have such a relationship embedded in their for-
mal dynamics; simple exponential growth models do not. Thus the former will anti-
cipate the forthcoming turning point in the elderly growth rate, whereas the latter
will not.
 POPULATION FORECASTING 199

5. CONCLUSION
Although population forecasting seems to many people the principal mission of de-
mography, the topic has not been a major preoccupation of the discipline.
"Ask anyone outside the profession what he thinks demographers do and he will give a much
bigger place to forecasting than the editors of our journals give it space on their pages. Our
most distinguished demographers do not put their major efforts into forecasting ... we assign
the highest professional standing to those who derive relations among variables in application
to past data, in short who can most convincingly explain the past" (Keyfitz, 1985, p. 60).

The inadequate scholarly attention to population forecasting noted by Keyfitz is

beginning to change. The subfield is experiencing a renaissance of sorts as more
and more scholars are exploring effective ways of combining " . . . recent develop-
ments in probabilizing the classical cohort component projection model... with the
methodologies and concerns of the forecasting mainstream." (Ahlburg and Land,
Downloaded by [ECU Libraries] at 14:35 24 April 2015

1992, p. 296). The resulting increased attention has generated a number of issues,
only a few of which have been touched on in this paper, which has focused primar-
ily on the issue of the relative performance of simple (small) models as opposed to
complex (large) models.
Large scale models, in several disciplines, have had their ups and downs over the
past few decades. They have been criticized for .being overly complicated, opaque,
atheoretical, multipurposeful, structurally inflexible, beyond validation, hypercom-
prehensive, data hungry, mechanical, expensive, and often exceeding the modeler's
span of control (e.g., Alonso, 1968; Lee, 1973; Arthur and McNicoll, 1975; Pittenger
and Schroeder, 1985). Several authors have pointed particularly to the lack of the-
ory needed to support the models, arguing that the amount of theory is nowhere
near sufficient to support such efforts, thereby limiting the uses to which the models
can be put (Lee, 1973; Arthur and McNicoll, 1975). Perhaps Alonso (1968, p. 252)
put it best when he observed:
"I am questioning whether we have arrived at the design of skyscrapers but we have only
lumber for construction material. If we do, we had better bund low to the ground while we
improve upon our materials."
None of the above, however, urged the abandonment of,modeling activities;
rather, each argued for "simpler" models. And this debate no doubt has influenced
the thinking of those working in the field of population forecasting, according fur-
ther popularity to the "simple is beautiful" position.
But as Keyfitz has pointed out, models that are not based on causal theory still
do make a contribution to population forecasting:
'Tor policy purposes causal knowledge is essential; for forecasting it is desirable, of course,
but the forecast is not necessarily a failure if the causal mechanism remains undiscovered.
Observed regularities serve perfectly well for forecasting as long as they continue to hold, and
such successes as there have been in forecasting are based on observed regularities ... What
we can seek with some hope of success is a statistical procedure that will marginally reduce the
error made in past forecasts ... Pending the discovery of a truly behavioral way of estimating
the future, we cannot afford to be ashamed of extrapolating the observed regularities of the
past." (Keyfitz, 1982, p. 747).

And in extrapolating the regularities of the past, models that are too simple are
likely to be successful only by chance. Their consolidation of past age-sex-specific
 200 A. ROGERS

patterns introduces aggregation and decomposition biases. Their neglect of standard

demographic accounting relationships introduces inconsistencies (e.g., non-zero to-
tals for interregional migration flows aggregated to the national level.) And their
lack of age-specific detail denies them the built-in advantage that demography has
over other disciplines engaged in forecasting, namely, that individuals age one year
at a time. Thus, for example, tomorrow's elderly have already been born, and we are
therefore in a relatively good position to predict the population aged 65 and over in
the year 2050 from predicted patterns of future mortality and the number of today's
adolescents and teenagers. (Of course, the predicted patterns of mortality need to
be reasonably accurate).
In conclusion, a number of observations are suggested by this review of the sim-
ple versus complex models debate. They are listed here, in no particular order, to
stimulate further informed debate on the issue.
First, whether simple models outperform complex models is an empirical issue
Downloaded by [ECU Libraries] at 14:35 24 April 2015

that depends on the particular historical period observed and the degree of demo-
graphic variability exhibited during this period. In consequence, it is imperative to
somehow control for the relative degree of difficulty associated with each historical
period. Rigorously developed forecasting tournaments could be a useful vehicle for
introducing such control.
Second, paralleling the degree of difficulty of the phenomenon is the degree of
robustness of the model adopted to forecast that phenomenon. Although many ar-
gue that "there is no single best model for all occasions," I find that proposition
to be unconvincing, because forecasters usually cannot anticipate the likely occa-
sions that are in prospect. They are in the position of couples at a ballroom dancing
competition, drawing dancing assignments out of a hat. To be best at waltzing does
them no good, if the selection is to dance a tango. Thus of what ex ante use are
findings such as: " . . . exponential extrapolation was found to be most accurate for
rapidly growing or declining areas, whereas linear extrapolation was most accurate
for moderately growing areas." (Isserman, 1977, p. 247)
Third, accuracy is a multidimensional notion for which aggregate indices, such as
root-mean-square errors, are an inadequate measure. Such indices are subject to
the ruses of heterogeneity and Simpson's Paradox. For example, they weight equally
the errors contributed by populations of vastly different sizes. And they are highly
susceptible to compensating errors generated by fertility, mortality, and migration
forecasts that err in ppposing directions, thereby according a spurious sense of ac-
curacy for the forecasting exercise.
Fourth,-the simple versus complex classification has been viewed as a dichotomy
when in fact it is a continuum. If, simple exponential growth models with no age or
locational disaggregation represent one end of the continuum, among the extrapola-
tive linear forecasting models, and detailed multiregional growth models the other,
then what can be said of the relative accuracy of models that lie in between these
extremes? Should their performance indices lie in between those of the two defining
the continuum? And if they do not (which is likely), then what conclusions can one
legitimately draw about the simple versus complex model issue?
Finally, model performance is a multifaceted concept that involves much more
than forecasting accuracy alone. As Long (1993) points out, additional attributes
 POPULATION FORECASTING 201

such as transparency, utility, and face validity all play an important role in the pre-
sentation of official population forecasts. Even though simple models may have pre-
dicted last year's population more accurately, would one bet one's earnings that they
will do so again for next year's population? Here one is reminded of studies carried
out on alternative market investment strategies:
"Academics report that while some people do beat the market, it's temporary and random,
for proof, they offered up chimpanzees who picked stocks by throwing darts at a list and
outperformed sophisticated investors." (Shapiro, 1994, p. 4)

Despite the superior performance of the chimpanzees, who would follow that invest-
ment strategy when charged to invest a significant sum of money?
Much of demographic analysis has focused on the appropriate specification and
accurate measurement of the dependent variables at the center of population fore-
casting activities. This body of work has helped to identify regularities that may
have been obscured by earlier inappropriate representations and measurements.
Downloaded by [ECU Libraries] at 14:35 24 April 2015

The modeling strategy that this suggests for population forecasting efforts may be
summarized in the motto: forecast only changing behavior, taking appropriate advan-
tage of well-established regularities and accounting relationships, and do not forecast
"de novo" relationships that are stable enough to not need forecasting. Exploiting ob-
served regularities in the relative age patterns of demographic rates and adopting
the standard age-specific demographic accounting equations are prominent exam-
ples of such an approach, and it is a strategy that still offers the best hope for
"marginally reducing the error made in past forecasts."

REFERENCES
Ahlburg, D. A. (1982) How accurate are the U.S. Bureau of the Census projections of total live births?
Journal of Forecasting 1: 365-374.
Ahlburg, D. A., and Land, K. C. (1992) Population forecasting: Guest editors' introduction. International
Journal of Forecasting 8: 289-299.
Alonso, W. (1968) Predicting best with imperfect data. Journal of the American Institute of Planners 34
(4): 248-255.
Armstrong, J. S., and Collopy, E (1992) Error measures for generalizing about forecasting methods:
Empirical comparisons. International Journal of Forecasting 8 (1): 99-111.
Arthur, W. B., and McNicoll, G. (1975) Large-scale simulation models in population and development:
What use to planners? Population and Development Review 1 (2): 251-265.
Beaumont, P. M., and Isserman, A. M. (1987) Comment. Journal of the American Statistical Association
83 (400): 1004-1009.
Brass, W. (1974) Perspectives in population prediction: Illustrated by the statistics of England and Wales.
Journal of the Royal Statistical Society A 137 (4): 532-583.
Chenery, H., and Syrquin, M. (1975) Patterns of Development, 1950-1970, New York: Oxford University
Press.
Cohen, J. E. (1986a) An uncertainty principle in demography and the unisex issue. The American Statis-
tician 40: 32-39.
Cohen, J. E. (1986b) Population forecasts and confidence intervals for Sweden: A comparison of model-
based and empirical approaches. Demography 23 (1): 105-126.
Feeney, G. (1988) Comparative structure of low fertility in Japan and the United States. Paper presented
at the HASA Conference on Future Changes in Population Age Structure. Sopron, Hungary, October
18-21, 1988.
Grummer-Strawn, L., and Espenshade, T. J. (1991) Evaluating the accuracy of U.S. population projection
models, Chapter 9 in C. F. Citro and E. A. Hanushek (eds.), Improving Information for Social Pol-
icy Decisions: The Uses of Microsimulation Modeling 2: 305-329, Washington, DC: National Academy
Press.
 202 A. ROGERS

Isserman, A. M. (1977) The accuracy of population projections for subcountry areas. Journal of the
American Institute of Planners 43 (3): 247-259.
Keilman, N. W. (1990) Uncertainty in National Population Forecasting: Issues, Backgrounds, Analyses,
Recommendations, Amsterdam: Swets and Zeitlinger.
Kelley, A. C., and Williamson, J. G. (1984) What Drives Third World City Growth?. Princeton, NJ: Prince-
ton University Press.
Keyfitz, N. (1981) The limits of population forecasting. Population and Development Review 7: 579-593.
Keyfitz, N. (1982) Can knowledge improve forecasts? Population and Development Review 8: 729-751.
Keyfitz, N. (1985) Demography in the twenty-first century: The uses of forecasting. In Proceedings of
the IUSSP International Population Conferences I: 59-81, Liege: International Union for the Scientific
Study of Population.
Ledent, J., and Rogers, A. (1986) Assessing the United Nations urbanization projections for the Asian
Pacific, in R. Woods and P. Rees (eds.), Population Structures and Models, London: Allen and Unwin,
367-389.
Lee, D. B., Jr. (1973) Requiem for large-scale models. Journal of the American Institute of Planners 39:
163-178.
Long, J. (1987) The accuracy of population projection methods at the U.S. census bureau. Paper pre-
Downloaded by [ECU Libraries] at 14:35 24 April 2015

sented at the Annual Meeting of the Population Association of America, Chicago, Apr. 29-May 2,
1987.
Long, J. (1993) The appropriate level of complexity in official population projections. Paper presented at
the Colorado Workshop on Population Forecasting, September 1993, Boulder, CO.
Pant, P. N., and Starbuck, W. H. (1990) Innocents in the forest: Forecasting and research methods. Jour-
nal of Management 16: 433-460.
Pittenger, D. B., and Schroeder, E C. (1985) Limits to the size of deterministic demographic forecasting
system. Paper presented at the annual meeting of the Population Association of America, March 28-
30, 1985, Boston, MA.
Rogers, A. (1976) Shrinking large-scale population-projection models by aggregation and decomposition.
Environment and Planning 8: 515-541.
Rogers, A. (1985) Regional Population Projection Models, Beverly Hills, Sage.
Rogers, A., and Willekens, E (1978) The spatial reproductive value and the spatial momentum of zero
population growth. Environment and Planning A 10 (5): 503-518.
Rogers, A., and Woodward, J. (1988) The sources of regional elderly population growth: Migration and
aging-in-place. Professional Geographer 40 (4): 450-459.
Rogers, A., and Woodward, J. A. (1991) Assessing state population projections with transparent multire-
gional demographic models. Population Research and Policy Review 10: 1-26.
Shapiro, A. (1994) The evolution of investing. Hemispheres, United Airlines Magazine, March issue, 4-5.
Siegel, J. S. (1972) Development and accuracy of projections of population and households in the United
States. Demography 9: 51-68.
Smith, S. K. (1987) Tests of forecast accuracy and bias for county population projections. Journal of the
American Statistical Association 81: 991-1003.
Smith, S. K., and Sincich, T. (1992) Forecasting state and household populations. Evaluating the forecast
accuracy and bias of alternative population projections for states. International Journal of Forecasting
8 (3): 495-508.
Stoto, M. (1983) The accuracy of population projections. Journal of the American Statistical Association
78: 13-20.
United Nations (1980) Patterns of Urban and Rural Population Growth, Population Studies, No. 68. New
York: Department of International Economic and Social Affairs, United Nations.
United Nations (1984) Population Projections: Methodology of the United Nations, Population Studies, No.
83. New York: Department of International Economic and Social Affairs, United Nations.
Vaupel, J. W., and Yashin, A. I. (1985) Heterogeneity's ruses: Some surprising effects of selection in
population dynamics. The American Statistician 39: 176-185.