While the Government Office for Science commissioned this review, the views are those of the author(s), are independent of Government, and do not constitute Government policy.
Copyright © 2010 The Royal SocietyThis is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
The total size of the world population is likely to increase from its current 7 billion to 8–10 billion by 2050. This uncertainty is because of unknown future fertility and mortality trends in different parts of the world. But the young age structure of the population and the fact that in much of Africa and Western Asia, fertility is still very high makes an increase by at least one more billion almost certain. Virtually, all the increase will happen in the developing world. For the second half of the century, population stabilization and the onset of a decline are likely. In addition to the future size of the population, its distribution by age, sex, level of educational attainment and place of residence are of specific importance for studying future food security. The paper provides a detailed discussion of different relevant dimensions in population projections and an evaluation of the methods and assumptions used in current global population projections and in particular those produced by the United Nations and by IIASA.
Keywords: world population, population increase, population decline, dimensions of population projections, age structure, level of educational attainment
While future trends in the number and composition of humans on this planet has been a topic of scientific enquiry and discussion for centuries and at least since Thomas Malthus entered the field of structured quantitative analysis, the first modern global population projection, which explicitly considered the age and sex structure of the population (the so-called cohort-component method), was carried out by Frank Notestein of the Princeton Office of Population Research in 1945 (Notestein 1945). At the national level, several population projections precede this first comprehensive global projection. Hajnal (1955) provides a good overview of these early population projections.
Frank Notestein subsequently became the first director of the then newly established United Nations Population Division. This unit began producing regular global population projections in the early 1950s. Between 1951 and 2008, the UN published 21 sets of estimates (past and current conditions) and projections (future) for all countries and territories of the world. Before 1978 these projections were revised approximately every 5 years; since then new revisions (called assessments and published in their World population prospects series) have been made every 2 years. The most recent UN assessments have a time horizon to 2050. Bongaarts (2009) gives a concise summary of the main results and implications of the UN projections. At irregular intervals the UN also publishes long-term population projections with time horizons from 2150 to 2300.
The World Bank started to produce independent population projections in 1978. These were always meant primarily for internal use in the Bank's development planning and were published as part of the series of World development reports. After 1984, the World Bank projections were revised approximately every two years and in most cases only one variant was published but with a long time horizon to 2150. Around 1995, the World Bank stopped publishing separate projections but presumably continued to use them for internal purposes for a number of years. The Washington-based Population Reference Bureau (PRB) publishes independent world population projections (population size only and a single scenario) every year as part of its annual World population data sheet. Since 2000 it has published projected population sizes for all countries and territories of the world for 2025 and 2050. In some cases the projections are identical to those of the UN and the US Census Bureau (USCB), but in some cases different country-specific information is used.
The USCB produces single scenario projections for all countries in the world as of 1985 with a varying time horizon. The World Population Programme of the International Institute for Applied Systems Analysis (IIASA) based outside Vienna (Austria) began producing global population projections at the level of 13 world regions in 1994. One of the purposes was to produce population projections as part of the Special Report on Emissions Scenarios (SRES) (Nakicenovic et al. 2000) that underlie the global emission scenarios used by the Intergovernmental Panel for Climate Change (IPCC). This was followed by three rounds of probabilistic projections at the level of 13 regions (which were all published in the pages of Nature: Lutz et al. 1997, 2001, 2008b). The IIASA projections come from a distinctly scientific background, which is illustrated by the fact that in its publication (e.g. Lutz et al. 2004a) much more space is used for justifying and discussing methods and assumptions than for presenting results. Recently, IIASA developed (in collaboration with Eurostat and the UK Office of National Statistics) a new approach for evaluating substantive arguments about alternative assumptions on possible future trends based on large numbers of questionnaires ascertaining the evaluation of alternative arguments by experts. This is currently being translated (together with Oxford University) into a new set of science-based population projections for all countries in the world by age, sex and level of education for publication in 2012.
O'Neill et al. (2001) published a comprehensive 85-page review entitled A guide to global population projections. In this article, they describe in considerable detail the projections of all five agencies discussed above. The paper includes a critical review of the methods used and of the ways that assumptions are defined by those five agencies. In addition, the results are compared in quite some detail. For this reason, we do not want to repeat the exercise of systematically comparing the approaches and results of the different projections, but rather focus on a few key challenges that point to the future.
In the next section we will discuss the important questions of what we call ‘dimensions’ of population projections. The basic idea is that any population is an assembly of individuals and each individual is different from any other individual. But which of the many relevant properties of individuals should we explicitly consider in population projections? The lowest dimensionality is chosen for the case that all individuals are taken as equal (homogeneous) and only a total growth rate is assumed for projecting population size. The cohort-component method breaks down by age and sex; this is a first step in the direction of increasing dimensionality by explicitly considering these individual properties. But there are many further properties that deserve consideration in population projections and which will be discussed in the following section.
We will also address the tedious but very important issue of how to deal with uncertainty in population projections. While a single best-guess population projection is likely to satisfy the expectations of a majority of users, disregarding the issues of uncertainty can be rather dangerous and therefore deserves serious attention. We will then turn to the one issue that has recently received the greatest public attention, namely, the questions if and when we will see an end of population growth followed by a possible population decline. In the final section of the paper we propose a specific existing population projection (the UN scenario of the IIASA Education Projections) for use in the Foresight Project. We present the results for the specified regions and countries in tabular and graphical form.
For many users the most important piece of information expected from population projections is the future total size of the population; for many policy and research questions other dimensions of the composition of the population are also of interest. The dimensions that will be explicitly discussed here are age, sex, rural/urban place of residence, educational attainment, labour force participation, parity status, household status and health status. All these dimensions have in common that they are properties of each individual member of the population which are, at least in theory, observable and measurable. They can hence be called observable sources of population heterogeneity. These dimensions are not only interesting in their own right for helping to answer specific questions relating to their future distributions in the population but to the extent that they are associated with different fertility, mortality and migration intensities, their changing distributions in the total population also impact on the projected future population size. In the following, we will discuss each of these dimensions individually, assess to what degree they are likely to impact on the course of overall population, discuss how they may be important in their own right in the context of studying food security, and review existing global level projections, which explicitly address this dimension.
It is important to note that in addition to these observable sources of population heterogeneity, there is still unobserved heterogeneity in every population which is hard to capture empirically. Theoretical considerations suggest that such unobserved heterogeneity can significantly impact future population dynamics (Vaupel & Yashin 1985), but there is little one can do about it except to be aware of the problem and be cautious about the validity of the conclusions drawn. Given this problem associated with hidden heterogeneity, it is even more important to explicitly measure and incorporate the observable sources of population heterogeneity wherever feasible and thus try to minimize the possible biases caused by overall heterogeneity.
Most users of population projections seem to be primarily interested in the total size of the population. From a substantive point of view, however, it is difficult to think of specific policy or research questions, where only changes in absolute size matter and where it is supposedly irrelevant whether one refers to newborn babies, young adults or frail elderly people. Presumably, the great interest in data on the sheer number of people has two main reasons. First, total population size can be seen as a first-order approximation of the scale of the population-related issues in particular, if it is assumed that the age patterns of the two populations to be compared do not differ greatly. Second, total population size is an important denominator of many frequently used indicators ranging from gross domestic product (GDP) per capita to food consumption per person to greenhouse gas emissions per person. In all those cases a separately derived total quantity is divided by total population size in order to produce an indicator that can be compared across populations.
Traditionally, population projections have been produced by simply taking total population size and making assumptions on the future growth rates of the population. This has been the standard method until the age-specific, so-called cohort-component method became the widely used standard after World War II. But for some specific applications, where the age structure is difficult to assess or where the population is simply too small to have meaningful age groups (as is sometimes the case with small area population projections), simple growth rate-based projections are still being used. But for national population projections, age-group-wise projections have clearly become the state-of-the-art because they allow differentiation between the behavioural components (fertility, mortality and migration) and embedded changes that are merely owing to age-structural effects. This is particularly important for countries that have gone through significant fluctuations in fertility and mortality trends and hence have irregular age structures, such as most European countries after the two world wars. But even for high fertility countries that traditionally had very stable and regular age pyramids, the assumed future fertility decline will be partly compensated in terms of population growth through the age-structural momentum, i.e. the fact that in the future larger numbers of potential mothers will enter the reproductive age and the number of births will continue to increase even if period fertility is at replacement level. This is another good reason for explicitly considering the age structure wherever possible.
Particularly in the context of population ageing, the explicit interest in the population's changing age structure has recently increased significantly. While there has always been an interest by planners in the likely future number of school-age children or young adults trying to enter the labour market, the currently dominating questions relate to the sustainability of pension systems under conditions of rapid increase in the number of elderly. Closely related questions deal with the likely impact of ageing on healthcare cost or care for the elderly in general. There is significant concern that an increase in the mean age of the labour force may be associated with decreasing productivity in the future (Skirbekk 2008). These are all questions for which a projection of the population by age is essential. Much of this discussion, however, is based on fixed age categories and does not consider the fact that healthy life-expectancy tends to increase in parallel with total life-expectancy. Hence, a 65-year-old person today is typically quite different from one of the same age who lived several decades ago and who had a much shorter remaining life-expectancy. This idea has recently been translated into a redefinition of age, where age is not just measured as the time since birth but can alternatively be viewed as the expected time to death, in which case the population dynamics looks quite different (Lutz et al. 2008b; Sanderson & Scherbov 2008).
The distinction between men and women as a key dimension of population dynamics has been a feature of demography almost from the beginning. It has to do with the fact that fertility rates are almost exclusively measured with women. One of the reasons for this is that not all fathers are known and that the reproductive age range for women is shorter than for men and shows a clearer and partly biologically determined age pattern. Also, mortality rates tend to differ significantly between men and women. Except for some cases of extreme female discrimination, male mortality rates are typically higher at every single age. In addition to these demographic reasons there is, of course, significant substantive interest in the age-specific proportions of men and women in the population which matters for issues ranging from the marriage market to labour market to consumer preferences.
The formal model for doing population projections by age and sex was introduced by Cannan (1895) but it took until after World War II to have this cohort-component method spread around the world as the dominant way for doing population projections. Today, it is used by all national and international statistical agencies.
Place of residence has always been an important dimension since the very origin of demography when statistical information on households was collected for taxation purposes and conscriptions for the army. Associating people with a specific locality, however, requires that people have a more or less permanent residence and are not permanently on the move. This is one problem facing statisticians when trying to distinguish between urban and rural populations in countries (such as China), where millions of people (the so-called floating population) are constantly on the move between their home villages and the urban centres where they work but have no legal right of residence. Another problem with comparing reported proportions urban across countries and over time is that there is a multitude of changing national sets of definitions based partly on the size of the agglomeration, its statutory status, its population density or the proportion of the labour force that is working in agriculture. Since urbanization will be discussed elsewhere, it will not be covered here.
From a population projections point of view, the question whether a person lives in a rural or an urban community is not only seen in its own right but is also viewed as a relevant source of population heterogeneity. Almost universally women in urban areas have lower fertility than women in rural areas. Also in terms of international migration, new immigrants today generally move to urban areas except for some special cases where immigrant labour is concentrated in agriculture. With respect to mortality, the pattern is more complex. While the inhabitants of cities traditionally had higher mortality owing to a higher risk of infectious disease, this pattern virtually disappeared over the course of the twentieth century because healthcare and education were better in urban areas. But most recently there are indications of a reversal, with urban slums in some African cities having worse mortality conditions than even remote villages in the same country.
From a methodological point of view, the best model to project a population by age and sex for urban and rural areas separately is the so-called multi-state model, which was developed at IIASA in the 1980s (Rogers & Land 1982; Keyfitz 1985). It has become the state-of-the-art for multi-dimensional population projections in which separate fertility, mortality and migration rates are assumed for the different sub-populations as well as age- and sex-specific transition rates between sub-populations. Because of data limitations, such true multi-state projections by urban–rural place of residence exist only for selected groups of countries.
The United Nations Population Division publishes the only urbanization projections on a global level, with updates every two years. The most recent revision (United Nations 2008) estimates the population living in urban areas in 229 countries and territories of the world for the period 1950–2005 and projects the same for 2010–2050 in five year intervals. The basic method of estimation and projection is still the same as when it was developed in the 1970s with only a few minor modifications. The main source of information used in the projection is the urban–rural growth difference at the two latest available time points (mostly inter-census). The difference is then assumed to follow a logistic path and is estimated based on the past experience of many countries and the expectation that the urbanization process will slow down as the level of urbanization grows. In other words, this is not a multi-state model, but only the application of assumed overall proportions urban to national level projections. This model by definition assumes continuous increases in the proportion urban and is unable to model possible ups and downs. This model is isomorphic to the literacy projection model that has been used by UNESCO for decades, but which was recently replaced by a cohort-component model.
Like the other dimensions discussed above, education is also an important source of population heterogeneity and bears a significant weight of its own. Almost universally more educated people have lower mortality and there is sufficient evidence that this is a real effect and not just owing to selectivity. Also for all populations that are still in the process of demographic transition, more educated women have lower fertility. These educational differentials can be very significant. The Demographic and Health Survey for Ethiopia, for instance, shows that women without any formal education have on average six children, whereas those with secondary education have only two (see http://www.measuredhs.com). Significant differentials can be found in most populations of all cultural traditions. Only in a few modern societies does the strongly negative association give way to a U-shaped pattern in which the most educated women have a somewhat higher fertility than those with intermediate education. But globally, the education differentials are so pervasive that education may well be called the single most important observable source of population heterogeneity after age and sex (Lutz et al. 1999). There are good reasons to assume that during the course of a demographic transition the fact that higher education leads to lower fertility is a true causal mechanism, where education facilitates better access to and information about family planning and most importantly leads to a change in attitude in which ‘quantity’ of children is replaced by ‘quality’, i.e. couples want to have fewer children with better life chances.
In terms of measuring education, it is important to distinguish between stock and flow variables. Most education programmes and studies are concerned with the process of schooling itself, including many important topics that range from the construction of schools and the training of teachers to the organization of schools, the quality of teaching and the content of curricula. The extent to which these schooling efforts cover the entire population are usually measured through enrolment rates as published for most countries by UNESCO and other agencies. But for many of the social and economic benefits of education it does not directly matter how many people are in school at any given point in time, but rather how many have completed their education and use their acquired skills in the labour force. Hence, the main interest is not on the flow (school enrolment) but rather on the changing stock (human capital). The stock can be measured by categories of highest educational attainment (here the UNESCO-ISCED categories have become the standard) as well as by mean years of schooling. Since the original empirical data mostly come in terms of attainment and the calculation of mean years of schooling requires additional country-specific assumptions, the former is often preferred. Also, considering the full attainment distribution provides information about inequalities that is hidden in the case of only using an average such as the mean years of schooling.
Because of the strong association between female education and fertility, future changes in the composition of the female population by educational attainment make a big difference. Since many developing countries have seen major improvements in school enrolment rates of both girls and boys in those countries, the future women of reproductive age are bound to be more educated than today's. This great momentum and rather easy predictability of the educational attainment distribution is a consequence of the fact that education is typically acquired early in life and then remains virtually constant along cohort lines. If we know how many 10-year-old girls with some primary education there are today, we know (when considering differential mortality and migration) how many 50-year-olds with at least primary education there will be in 40 years' time. Given the fact that in most countries the younger cohorts are better educated than the older ones, further declines in fertility as well as mortality are virtually pre-programmed. In a few African countries where school enrolment rates have actually declined over the past two decades, this has led to a stalled fertility decline and partly worsening mortality conditions.
Aside from its effects on population dynamics, there is also significant interest in knowing the educational structure of the population for a broad range of social and economic development concerns. Based on a newly reconstructed set of educational attainment distributions by age and sex (Lutz et al. 2007a) for most countries back to 1970, it has recently been shown that indeed the improvement of educational attainment in the working age population has been the most consistent and significant driver of economic growth around the world (Lutz et al. 2008a). But improving education by age and sex has also been shown to matter for countries transitioning to democracy and more liberal rights (Abbasi-Shavazi et al. 2008; Lutz 2009b). It has already been mentioned that education is perhaps the single most important health determinant. For the question of food security, it has long been shown that the basic education of the agricultural labour force is a key factor in agricultural production (Hayami & Ruttan 1971). As the set of population–education–development–agriculture (PEDA) models commissioned by the UN Economic Commission for Africa for a number of African countries shows, when including education in an agricultural production function, it turns out to be one of the key determinants in reducing malnutrition and food insecurity (Lutz et al. 2004b).
IIASA has recently produced population projections by age, sex and four levels of educational attainment for 120 countries following different scenario assumptions with respect to alternative future education trends and different education-specific fertility trends (KC et al. 2010). These projections are based on a true multi-state model, i.e. individuals are assumed to be subjected to different mortality and fertility levels depending on their educational attainment level. Hence, even identical trends in education-specific fertility can lead to different overall fertility levels in the case of differing education trends. We will return to this issue when discussing the recommended combination of these education projections with the UN population projections by age and sex.
There are several other dimensions of population that are considered relevant either in their own right or as further important sources of heterogeneity. The parity distribution of the female population, i.e. the distribution of women by the number of children they have already given birth to, is an important determinant of fertility in the near-term future. Highly important for economic and labour force consideration is the distribution of the population by age, sex and labour force participation rates. While there are several national projections of future labour force participation, consistent global projections do not yet exist. Another very relevant dimension in the context of global population ageing and the associated needs for care for the elderly is the projection of the population by health status. While significant progress has been made in producing internationally comparable indicators of disability and its impact on activities of daily life, up to now this has not yet led to the production of global projections which cross-classify health status with age and sex. Alternatively, one could try to classify people according to certain disease risks or infection statuses (such as HIV-positive) if the data are available, and a certain disease is of specific interest. Finally, an important field of analysis and projections deals with the living arrangements/household composition of people. Here, some progress in the direction of global projections has recently been made in the context of trying to project future greenhouse gas emissions and the importance of demographic/household factors in climate change mitigation (van Imhoff & Keilman 1991; Jennings et al. 2004; Jiang & O'Neill 2009).
In the field of social and economic trends—as in all fields that deal with human behaviour—the future can never be projected with certainty. In the context of the cohort-component model of population projection that differentiates by age and sex, all three components of population change (i.e. future fertility, mortality and migration trends) are uncertain. For each of these three components of change, the levels as well as the specific age patterns (by sex) are uncertain, but the level is considered to be of primary importance.
How to deal with uncertainty in population forecasting was the topic of a recent special issue of the International statistical review (Lutz & Goldstein 2004). It summarizes the state-of-the-art in a field of research that has recently gained more interest both among demographers as well as users. Based on this analysis, we will distinguish between four different ways that are commonly used to deal with the difficult issue of uncertainty: (i) ignore uncertainty and publish only one projection; (ii) define alternative probability-free scenarios; (iii) publish high, medium and low variants that are supposed to cover a ‘plausible range’; and (iv) produce fully probabilistic projections that give quantitative information about the range of uncertainty.
For many users it is sufficient to have only one ‘best-guess’ forecast. This is the projection which is seen as the most likely population trajectory from today's perspective, based on the best knowledge available today. Such a projection can give some orientation about the direction into which things are moving. It is sufficient for many applications with a relatively short time horizon and where the costs associated with a deviation from the projected mean are negligible. But for many planning purposes these costs can be rather significant. For example, if there is a plan to build a new primary school and the projection should tell the local authorities how many 5–10-year-olds there are likely to be in the next 10–20 years in a certain school district, such projections are associated with significant uncertainties and any deviations from the best-guess projection can turn out to be rather costly. This is even more so for projections of the number of people with pension entitlements in a given country. In terms of global level projections, those presented by the USCB, the World Bank and the Population Reference Bureau provide only single projections for all countries in the world.
The definition of ‘alternative scenarios’ is a frequently chosen approach that goes beyond a single projection and presents the user with several possible future trajectories. The terminology is somewhat inconsistent and variants (see next paragraph) are often called scenarios. Generally, scenarios are defined as consistent sets of assumptions. The notion originally came from the world of theatre, where scenes show consistent pictures of our possible setting. In demography, this is typically done by combining different assumptions about future trends in fertility, mortality and migration in a way that should ‘not be impossible’. Hence, scenarios are generally considered free of probabilities with the only criterion being internal consistency. The user is given a broad range of what could be theoretically possible without being told what is likely to happen and what probability range is covered by the interval between the highest and lowest scenario considered. At the global level, the most frequently used population scenarios are those produced in the context of the SRES scenarios of the IPCC (Nakicenovic et al. 2000), which in turn are based on combinations of projections produced by IIASA and the United Nations Population Division. Eurostat also presents its population projections in the form of different scenarios and an increasing number of European national statistical agencies follow this practice.
The presentation of ‘high, medium and low variants’ goes a step further in providing the user with some range of uncertainty that can be interpreted intuitively. For many decades, this approach has been the common practice of the UN population projections as well as a large majority of national statistical offices. The variants are mostly generated by assuming three alternative fertility trajectories, which are then combined with identical mortality and migration trajectories. They are typically presented to the user as providing a ‘plausible range’ of future population trends, hence providing a first approach to uncertainty although the user is not specifically told what is meant by ‘plausible’. Does it mean 100 per cent of all possible cases, 80 per cent or only 50 per cent? A further shortcoming of this approach is that it only covers the uncertainty associated with future fertility trends and does not include uncertainties associated with future mortality or migration. Since fertility is the most important determinant of long-term population growth, this can be viewed as an acceptable simplification when interested primarily in total population size. When the interest is in population ageing, however, or age-specific indicators such as the future proportion of the population above age 80, this can result in a major underestimation of uncertainty. A final problem with this approach arises when aggregating national projections to regional or global ones. The global high variant of the UN projections is calculated as the sum of all national level high variants, hence assuming that there is a perfect correlation among all national trends. Each country simultaneously experiences the highest plausible fertility trajectory. Such perfect correlation is highly unlikely, however, and a higher than expected trajectory in one part of the world may be partly offset by diverging trends in other parts of the world. Hence, what is considered as a plausible range at the national level is not necessarily plausible at the global level.
Only ‘fully probabilistic’ projections can avoid such problems. These are based on pre-defined uncertainty distributions over time for each of the three components which are stochastically combined in large numbers (typically several thousands) of cohort-component projections with the individual random draws being subject to assumed autocorrelations and, in the case of multi-regional projections, correlations among the regions. In most projections, the components are assumed to be uncorrelated between each other. As the field of probabilistic population projections has grown, there have been essentially three traditions in defining the uncertainty distributions: one based on time series analysis (which is only applicable to countries that have long empirical series of data and where no structural discontinuities are expected); one based on the analysis of the errors of old projections (ex post error analysis combined with the assumption that future errors will essentially be the same as past projection errors); and one based on the evaluation of expert arguments resulting in sets of subjective probability distributions. While more experts as well as national and international agencies have entered this field of probabilistic projections, some convergence among these different approaches seems to be ongoing. It is also interesting that the initiative for doing such projections does not always come from the producers but rather from the users. For example, in the case of the UK Office of National Statistics, such projections were recently produced upon request by the financial planning authorities. At the global level, IIASA (Lutz et al. 1997, 2001, 2008b) has produced the only probabilistic population projections at the level of 13 world regions.
From a theoretical perspective, fully probabilistic projections are superior in avoiding the problems of other approaches as discussed above. They provide the users with the most comprehensive and detailed information in terms of the demographic ‘risks’ that can be combined with a cost function. But this approach is criticized by some as conveying a false sense of precision. It is indeed true that this approach requires more detailed explicit assumptions in terms of the full distributions and correlations than the other common approaches. Yet, the proponents of probabilistic projections argue that they only try to make otherwise implicit assumptions explicit and extensive sensitivity analyses show that the overall results tend to be very robust to changes in the specific shapes of the assumed uncertainty distributions. But at this stage it is clear that a universally accepted approach to quantitatively describing the uncertainty of population projections has not yet been developed and more research in this field is needed.
A final source of uncertainty that often tends to be disregarded in population projections is the uncertainty about current conditions, i.e. the population size and structure as well as the levels of fertility and mortality in a given country in the starting year of the projections. While this tends to be well documented in industrialized countries, there are major gaps of information, in particular, in Africa and parts of Asia. For many countries the information is only based on sample surveys and for some countries even such information is not available. While the UN Population Division makes heroic efforts to come up with population estimates for all countries in the world, the fact that many numbers referring to the recent past have to be adjusted in every new assessment round as new information becomes available shows how difficult this task is. But demographers have become so used to having a precise point estimate for any demographic indicator, even in countries with very fragmentary empirical information, that they often forget about the uncertainty of those indicators referring to the starting conditions when making projections. A recent study on past projections in five southeast Asian countries (Khan & Lutz 2008) showed that the errors introduced by incorrect information about starting conditions was in some cases even higher than the error owing to incorrect assumptions about the future. Lutz et al. (2007b) expanded the concept of probabilistic population projections to explicitly include the uncertainty about starting conditions in addition to the uncertain future paths for the three components. This is even relevant at the level of world population projections because of the great uncertainty of current fertility levels in the world's most populous country, China, as will be discussed in more detail below.
There has been much discussion recently about the long-term outlook for world population and in particular whether the global population is likely to stabilize or even decline in size after a peak during the second half of the twenty-first century. This discussion was, in part, triggered by an article published in Nature in 2001 entitled The end of world population growth (Lutz et al. 2001), which indicated that there was an 80–90% chance that the world population would reach a peak before 2100 and start to decline thereafter. Whether or not such a peak will actually occur and at what level and how steep the following decline will be is primarily a function of the assumptions made about long-term fertility levels in different parts of the world.
When it comes to global population projections beyond 2050, there are only the occasionally published UN long-term projections as well as the IIASA world population projections for comparison. In 2004, the UN Population Division published a study entitled World population to 2300 (United Nations 2004) in which different long-term scenarios were presented for all countries in the world by extending their usual projections to 2050 by another 250 years. While life-expectancy was assumed to continue to increase over the entire period although at a decelerating rate, the fertility scenarios defined were all very close to replacement level. Of the three long-term fertility assumptions, the low one assumed universal convergence to 1.85, the high one to 2.35 and the medium one to whatever the replacement level might be (somewhat below 2.1 under good mortality conditions). It is interesting to note that the fertility level of 1.85, which in the projections to 2050 is assumed to be the convergence level of the medium variant (with the low variant being 1.35) for the long-range projections, is the lowest fertility scenario presented. But even these very small differences in the long-term fertility levels produce significantly different long-term global population sizes: by 2100, the resulting populations are 5.5 (low), 9.1 (medium) and 14.0 (high) billion, and by 2200, the differences further increase to 3.2, 8.5 and 21.2 billion, respectively. Hence, the medium scenario results in some sort of population stabilization in the very long run, but only because global long-term fertility is assumed to remain constant exactly at the replacement level of fertility—a level that is defined as the one producing long-term stationarity. Minor deviations to the lower side will produce significant population decline and to the higher side will result in substantial long-term increases.
Lutz & Scherbov (2008) recently published another set of long-term global projections, which extends the IIASA probabilistic population projections—which go to 2100—further into the future by defining scenarios covering a wider range of possible future fertility levels. Selected findings are shown in figure 1 and were discussed in a recently published editorial in the Journal of the Royal Statistical Society entitled Towards a world of 2–6 billion well-educated and therefore healthy and wealthy people (Lutz 2009a). The extension scenarios start in 2080 (the year until which the assumptions for the probabilistic projections were defined); some scenarios continue from the level of the median of the projected distribution; others from the upper and lower bounds of the projected 90 per cent uncertainty range ( figure 1 ).