Methods & models strand abstracts

Applications of multilevel modelling: Monday 9 September, 1.30pm

A session with four short papers using multi-level modelling, followed by discussion and structured debate with a senior academic.

Household structure and child health in Botswana
Oleosi Ntshebe and Amos Channon, University of Southampton.

This study examines the association between household structure and child health in Botswana. Such relationships are little studied in lower income settings, which is the case in Botswana. Also, new household types are emerging in Botswana in the form of extensive lone parenting, non-marital childbearing and parental cooperation with nonresident fathers and other household members. Household structure is defined based on the character and the complexity of the family in which children are brought up. Two measures of child health are assessed: diarrhoea and acute respiratory infection. Data analyzed comes from nationally representative surveys: the 1996, 2007 Botswana Family Health Surveys (BFHS) and the 2000 Multiple Indicator Survey (MICS). Simple and logistic multilevel models serve to control for confounding variables associated with child health. By implication, this study provides important insights into family functioning and child health outcomes not only in Botswana but other countries, particularly in Africa, with similar demographic and social contexts.

Socio-cultural determinants of modern contraceptive use in West Africa
Megan Ledger, University of Southampton

Research question: What role do socio-cultural factors play in the geographic differences in modern contraceptive use in West Africa?

Methods: Multilevel modelling techniques will be used to investigate the influence of communication, socio-cultural values and identity and to identify within and cross-country variations in the use of modern contraceptives at the individual, community, national and international level.

Data: Demographic and Health Survey (DHS) data from seven West African countries are used in the analysis of this paper. They were chosen because they were either phase V or VI surveys and GPS data was also available for these countries.

Preliminary results: The analysis is currently being carried out but it is expected that there will be differences in the significant associations between the socio-cultural factors and use of modern contraceptives across and within the diverse countries and regions in the analysis. The hypothesis is that socio-cultural factors that are significantly associated with modern contraceptive use will vary between as well as within the seven countries in the analysis and that community aspects relating to particular socio-cultural environments will influence modern contraceptive use.

Potential applications: This paper aims to see if there are geographical patterns in the determinants of contraceptive use and how these relate to each other. A cross-national comparison of differences and similarities between population groups could be vital in informing the improvement and development of family planning interventions in this region.

The effect of family allowances on first births in Europe: cros-national longitudinal comparisons and critical junctures
Jonas Wood, Karel Neels, University of Antwerp

Fertility postponement and decline have contributed substantially to the acceleration of population ageing in later twentieth century Europe, which remains an important challenge for contemporary welfare states. As a result questions and debate arise on the impact of family policy on fertility. Using longitudinal microdata this article investigates the impact of family allowances on motherhood transitions in 13 selected European countries between 1970 and 2005. This contribution will present two contrasting approaches to assess the effect of allowances on first births. First a longitudinal cross-national comparative approach is taken in a multilevel framework. Second, focusing on critical junctures in family allowances by country, single-country analyses are performed and family allowances are operationalized with dummy-coded period variables. Controlling for individual level household and labour market position as well as aggregate-level yearly trends and unemployment rates, this contribution relates rises in cash benefits to higher first birth hazards for women aged over 30. However we also conclude that both approaches adopted in this research highlight the difficulties in assessing causal relations between macro-level allowances and micro-level childbearing decisions.

The association between individual and national-level characteristics in fertility and partnership: An application of two-level latent class models
Mark Lyons-Amos, University of Southampton

There is considerable variation in both partnership and fertility across the developed world. As a result, the policy context also varies considerably either reflecting or driving cross-national behavioural differences. However, the range of substantive evaluations of the interaction between policy and behaviour is limited to policy evaluations (e.g. Perelli-Harris and Sanchez Gassen). This analysis therefore aims to address this dearth of research by evaluating the association of demographic behaviour at the individual level with national level policies. While multilevel-models are a common means of examining the interaction individual and contextual level, the loss of information in using countries as cluster level units means that drawing substantive conclusions is often difficult. As such, I propose the use of a two-level latent class analysis, where individuals and countries are assigned to classes at both level 1 and level 2 (so called discrete random effects). This facilitates a more specific interpretation of the interaction between individuals and their national context. Level-1 latent classes are formed from growth curves for the probability of experiencing three key life events by a certain age: entry into first cohabitation, entry into first marriage and first birth. These demographic data are drawn from the Harmonized-Histories dataset, which contained standardised, retrospective demographic histories. Level-2 latent classes are formed from indicators from the GGP contextual database, incorporating policy indicators and social indicators relevant to family formation (e.g. alimony structures, child benefit). Level-1 and Level-2 classes are correlated to formally reflect the interaction of individual demographic behaviour and national level policy.

Methods: migration in models: Monday 9 September, 4.45pm

Estimating annual migration flows by age and sex for subnational geographies in the UK, 2001-2011 
Paul Norman, Nik Lomax, John Stillwell, Philip Rees; University of Leeds

Data on migration between local government areas is an important component of annual estimates of population. Changes in migration levels at this subnational geography are indicators of social change and provide challenges for resource allocation. Outside of census year, in the absence of a population register in the UK, there is a need to estimate migration between areas using proxy data sources. One such source is derived from patient registers when people move house and inform their doctor of their change of address. Unfortunately, whilst data are available annually, the geographical and age-sex information is not sufficient to provide direct counts of migrants in the required detail. This methodological paper will demonstrate how Iterative Proportional Fitting (IPF) can be used to update base year local level origin-destination flows to subsequent years where only in/out totals are available and how nonlinear regression can be used to estimate age-sex schedules from grouped and / or sparse age-specific migration rates. An interim step making the estimation process more efficient is to group together local areas which have similar migration rates by age using k-means classification. A problem overcome here is that the data availability and specification is different in England & Wales from that in Scotland and also in Northern Ireland so producing a UK-wide migration matrix is especially challenging. The result of this work is an annual time-series of counts of migrants by age and sex between origins and destinations for local areas across the UK for 2001 to 2011.

Combining Internal Migration data sources in England
Rebecca Newell, ONS/ESRC/University of Southampton

A new methodology for combining internal migration flow data from the 2001 Census and NHSCR data sources is presented in this paper. Central to the research is the development of an integrated Bayesian modelling framework, capable of combining the strengths from each of the two data sources to produce harmonised estimates of internal migration flows at the regional level in England. The methodology also provides a mechanism for describing and presenting the uncertainty in the estimates. As internal migration is the primary mechanism behind population redistribution in England, this work has the potential to better understand the quality and limitations of the available data.

Redevelopment of the migration assumptions setting methodology for the UK’s national population projections
Fern Leather, Office for National Statistics

A review was conducted by Southampton University in 2012 with the objective of evaluating the migration assumptions setting methodology for the Office for National Statistics’ (ONS) national population projections and suggesting improvements. Following this, ONS have redeveloped the assumption setting methods for the 2012-based national population projections to incorporate gross flows rather than net migration, use of a Rogers-Castro curve with a student peak to smooth age distributions, and the use of formal time series models for extrapolation. These changes also prepare the assumption setting methods for the inclusion of rates-based flows in the future and could ultimately facilitate a move to probabilistic forecasting. The paper will detail the new methods and set out the improvements compared with the old methodology, which have brought the UK’s migration assumption setting closer in line with academic discourse and current best practice in official population projections.

Examining the Role of International Migration in Global Population Projections
Guy J. Abel, Wittgenstein Centre for Demography and Global Human Capital, Vienna Institute of Demography of the Austrian Academy of Sciences; Nikola Sander, Wittgenstein Centre for Demography and Global Human Capital, Vienna Institute of Demography of the Austrian Academy of Sciences; Samir K.C. Wittgenstein Centre for Demography and Global Human Capital, International Institute for Applied Systems Analysis (IIASA)

Advances in projecting international migration have been hindered by a lack of adequate data. Consequently, international projection-making agencies commonly use simplistic assumptions of net-migration measures derived as residuals from demographic accounting. However, past net migration can be often volatile and are known to introduce inaccuracies when projecting populations (Rogers, 1990). This paper presents a set of global population projections to 2060, focusing on two alternative assumptions of international migration. Assumptions on rates of other demographic factors, namely fertility and mortality, are held constant. In the first projection, we set up the future migration in each country to mirror that of the United Nations Population Division (UNPD) , where “Projected levels of net migration are generally kept constant over the next decades. After 2050, it is assumed that net migration will gradually decline.” (UNPD 2011, p. 12, paragraph C.1). In the second projection, we use a first-of-a-kind set of estimated 5-year bilateral migration flows by sex developed from Abel (2013). The net migration of the estimates within these flow tables matches those of the UNPD. The bilateral flow table estimates are further disaggregated by age using a standard Rogers-Castro migration schedule, and then summed over rows and columns to obtain immigration and emigration rates by age and sex. These estimates are used as base data in the projection model, where immigration and emigration rates are assumed to remain constant up to 2060. Our results highlight differences in the future level of populations around the globe and numbers of migrant flows between the net migration projection model and the immigration and emigration projection model.

Methods: Variations in outputs – Tuesday 10 September, 9.00am

The future inequality of population estimates: how will we measure and how should we adjust?
Steve Smallwood, Office for National Statistics

As with every Census the 2011 Census has increased our understanding of how population estimates drift over intercensal periods. It has also given us insights into how population as measured by the Census compares with population figures generated from other data sources. Alongside this knowledge lessons were learnt from improved quality assurance of both 2011 Census and mid-year estimates. Other work has been carried out attempting to consider uncertainty around population estimates and also to consider the application of plausibility ranges around population. This is all in the shadow of the potential for major changes in population systems if the Beyond 2011 programme results in major changes to the demographic system. This presentation will show some of the work that will be carried out to help communicate the quality of mid-year estimates. It will also discuss issues around whether, if information from that work suggests strongly that a mid-year estimate is drifting, some form of adjustment should be made.

On the decomposition of life expectancy and limits to life
Les Mayhew & David Smith, Cass Business School

Life expectancy is a measure of how long people are expected to live and a widely used measure of human development. Variations in life expectancy reflect not only the process of ageing but also the impacts of epidemics, wars, economic recession. Since 1950 their influence in the most developed countries has waned and life expectancy is growing at an unprecedented pace. As a result it has become more difficult to forecast long run trends or possible upper limits to life expectancy. However, accurate population projections are needed for consideration of the social and economic consequences in fields such as health and social care and pension provision. In this paper we compare life expectancy in developed countries and the prospects for further advancement at both a macro and micro level. New methods are presented for comparing past advancements in life expectancy and also future prospects using data from five developed economies. Specifically, we consider life expectancy in ten year age intervals rather than over remaining life and show how natural ceilings in life expectancy can be used for extrapolating future trends.

How confident can we be in the projections of the older population in the UK?
Philip Rees, University of Leeds

This paper responds to a request by Lord Filkin, the Chair of the House of Lords Selective Committee on Public Service and Demography, to supply an assessment of confidence in the projected numbers of older persons for the UK. The paper reviews current methods of establishing confidence through probabilistic projection carried out by academics, national statistical offices and international statistical agencies. However, no UK probabilistic projections could deliver reasonable confidence intervals, required by the Committee by the end of 2012. Therefore, an alternative method was proposed that used a set of recent plausible projections of the UK population to form a distribution of outcomes that could deliver population ranges with known probability. Nineteen projections were pooled and used to compute confidence bands around a median projected population for each age group. The UK population aged 65+, 10.2 millions in 2010, is projected to 18.3 million in 2050, with a 95 confidence that it will lie between 21.3 and 15.7 million. We can be more confident about the numbers for the younger elderly (aged 65-84) than for the older old (aged 85+). We cannot be very confident about the future numbers of centenarians. The population of centenarians is projected to be 242 thousand in 2050 compared with 12 thousand in 2010 but the 95% confidence band stretches from 426 thousand to 59 thousand. Discussion in the final part of the paper addresses the issue of how much confidence we can have in the confidence band estimates.

Methods: Digging deeper - Wednesday 10 September, 9.00am

Identifying age, period and cohort effects: Sarah Connor versus the Terminator
David Voas, University of Essex and Siobhan McAndrew, University of Manchester

How and why do people and societies change? Demographers and other social scientists often wrestle with the question of whether shifts in attitudes and behaviour are associated with age or the life course, historical period, or birth cohort / generation. The problem of disentangling age, period and cohort effects has become a staple issue in longitudinal analysis. New statistical approaches to age-period-cohort analysis (e.g. cross-classified multilevel models, intrinsic estimator, maximum entropy and Deaton-Paxson) have appeared in the past decade. Despite (or perhaps because of) their technical impenetrability, many scholars use them relatively uncritically. They have come to be seen as state-of-the-art tools that should be used in preference to less sophisticated methods. The avoidance of substantive assumptions in favour of a kind of robotic neutrality is regarded as a virtue. Using religious change in Great Britain and the United States as empirical illustrations, we argue that these methods are no more likely to produce correct solutions than approaches that explicitly call on human fallibility. We propose an alternative strategy based on visualisation and an iterative inductive-deductive approach.

Synthetic data estimation for the UK Longitudinal Studies: an introduction to the SYLLS project
Adam Dennett, University College London; Belinda Wu, University College London; Gillian Raab, University of St. Andrews

The England and Wales Longitudinal Study (LS), Scottish Longitudinal Study (SLS) and Northern Ireland Longitudinal Study (NILS) are incredibly rich micro-datasets linking census and other health and administrative data (births, deaths, marriages, cancer registrations) for individuals and their immediate families across several decades. Whilst unique and valuable resources, the sensitive nature of the information they contain means that access to the microdata is restricted, limiting the user base. The SYLLS project will develop synthetic data which mimics the real longitudinal data but crucially will not be subject to the same access restrictions as the national LSs. In this paper we will introduce two different but complementary methods that we will be adopting to generate the synthetic data – microsimulation and multiple imputation. Microsimulation will be used to generate a synthetic LS ‘spine’, mimicking the full population of individuals in the LSs but for a limited set of core variables, transitioning between 1991 and 2001. Multiple Imputation will be used to generate bespoke synthetic data extracts which match precisely the requirements of individual research projects. This paper will report on the methodological progress to date, issues and prospects for the new synthetic datasets.

Agent-based models and statistical emulators in demography: An example of modelling social care for the ageing UK
Eric Silverman, Jakub Bijak, Jason Hilton, Jason Noble, University of Southampton

Statistical demography is concerned with populations of statistical individuals, whose life courses can be inferred from empirical information. In contrast, agent-based models study simulated individuals, for whom certain behavioural rules are assumed. We wish to bring these two approaches closer together by proposing a method to analyse rule-based outcomes statistically. For that purpose, we utilise Gaussian process emulators – statistical models of the base model – to analyse the impact of selected parameters on model outputs. Emulators permit a statistical analysis of model properties and help select plausible parameter values, despite non-linearities and feedbacks in agent-based models which preclude a direct statistical analysis. We also undertake a sensitivity analysis to assess the relative importance of different parameters. The discussion is illustrated by a multi-state agent-based model of the ageing UK population, which aims to examine the interaction between population change and the cost of social care. The model captures the basic processes which affect the demand for and supply of social care, including fertility, mortality, health status, and partnership formation and dissolution. The mortality and fertility rates in this population are drawn from UK population data for 1951–2011 and Lee-Carter forecasts until 2050. Results show that the cost of social care in the UK is expected to rise significantly as the population continues to age. An in-depth sensitivity analysis performed using statistical emulators confirms that the level of care need within the population and the age of retirement have the most profound impact on the projected cost of social care.

Embedding economic models of migration and household formation in a housing market model for localities and regions in New Zealand
Glen Bramley, Chris Leishman and David Watkins, Heriot Watt University

Traditional extrapolative projections of population and migration remain popular as a basis for planning housing provision in many countries. However, it can be argued that these processes of demographic change are significantly influenced by, as well as influencing, economic and housing market conditions at the level of local and sub-regional jurisdictions. This suggests that these processes should be treated as endogenous within models that seek to forecast housing needs and outcomes and test the impact of varying policies and economic conditions. Such a perspective, informed by UK experience and research, is brought to bear in a study commissioned by the national state housing agency to develop a demand and supply forecasting model for New Zealand. This paper would describe the method and findings for the specific models developed to forecast local migration flows and household formation rates by age, using 5-yearly Census data for territorial local authorities. It will go on to outline how these are incorporated in the wider economically-oriented forecasting system. System simulations are examined particularly in terms of deviations between the resulting population and household forecasts and the conventional projections provided by the national statistics agency. These highlight the importance of housing supply constraints and he potential spillovers between areas.