ALTERNATIVE APPROACHES TO FORECASTING MIGRATION: FRAMEWORK AND ILLUSTRATIONS Philip Rees 1, Nikolas Lomax 1 and Peter Boden 2 1 School of Geography, University of Leeds, Leeds LS2 9JT 2 Edge Analytics Ltd, Leeds Innovation Centre, Leeds LS2 9DF Contact: p.h.rees@leeds.ac.uk, 0113 343 3286 Acknowledgements: Peter Boden: for permission to present this work, commissioned by Edge Analytics National Records of Scotland: Some ideas have been developed for a consultancy on the design of the Scottish Sub-National Population Projections, 2014-Based Office for National Statistics: Some ideas originate from recent work by ONS on the design of the (UK) National Population Projections, 2014-based Economic and Social Research Council: some of the ideas will be used in our project (Secondary Data Analysis Initiative, Phase 2) - Rees P, Wohland P, Norman P and Lomax N (2015) Evaluation, Revision and Extension of Ethnic Population Projections NewETHPOP. ESRC, Grant Ref ES/L013878/1, 1/1/2015 to 31/5/2016
Background There is convergence towards co-production of knowledge between official statistical agencies and academic researchers in the development of population projections UN 2014-based World Population Prospects projections incorporate stochastic elements (fertility, mortality) resulting from a collaboration of the UN Population Division and a team from the University of Washington led by Adrian Raftery ONS population projections team, led by Denise Williams are exploring the recommendations on Migration Assumptions in National Projections prepared by Jakub Bijak, University of Southampton ONS and the Universities of Southampton and Leeds collaborated in producing A Conceptual Framework for Population and Migration Statistics, led by James Raymer, Australian National University
Aim and Outline Aim: to fill the gap in knowledge about the best way to handle migration in population projection models Outline: Population Accounts Embedding Migration Flows Should Migration be Modelled as a Flow or Rate? Examples of Recent Projection Models placed within the Accounts and Flow/Rates Framework Which is the Best Model: Horses for Courses
Population Accounting Framework Country 1 of interest Other countries From Region 1 Region 2 Region n Country 2 Country m Deaths Totals 1 ݎ 1 ݎ ܧ 1 2 ݎ ܧ ݎ 1 ݎ ܯ 2 ݎ 1 ݎ ܯ 1 ݎ Region r1 + 1 ݎݐ ܦ 2 ݎ 2 ݎ ܧ 2 2 ݎ ܧ ݎ 2 ݎ ܯ 2 ݎ 1 ݎ 2 ݎ ܯ Region r2 + 2 ݎݐ ܦ : : : : : : : : + ݎݐ ݎ ܦ ݎ ܧ 2 ݎ ܧ ݎ 2 ݎ ݎ ܯ 1 ݎ ݎ ܯ Region rn + 2 ݐ 2 ܦ 2 ܯ 2 ݎ 2 ܫ 2 ݎ 2 ܫ 1 ݎ 2 ܫ Country 2 : : : : : : : : Country m + ݐ ܦ 2 ܯ ݎ ܫ 2 ݎ ܫ 1 ݎ ܫ + ܤ 0 ܤ 2 ܤ ݎ ܤ 2 ݎ ܤ 1 ݎ ܤ Births Totals 2 ݎ+ 1 ݎ+ 1 +ݐ 1 +ݐ ݎ+ 1 +ݐ + 2 1 +ݐ + ++ + ܦ 1 +ݐ
Key features Moves (events) used, not Transitions (migrants) Diagonal terms are just accounting Residuals (balances) Four quadrants: Top-left = inter-regional migrations within a Top-right = emigrations from regions to outside countries Bottom-left = immigrations to regions from outside countries Bottom-right = inter- migrations within the rest of the world Practice: Regions in size down to municipalities, but not small areas Countries usually collapsed to Rest of World category But there is considerable political interest in flows to and from selected countries
2.1 Assume flows continue at historic levels (h =time horizon of projections ) Generally regarded as far too simple and unlikely Often formulated as constant net migration flows
h =ݐ 1 =ݐ 0 =ݐ 2.2 Assume flows vary according to judgment or a time series model ݎ ݎ h =ݐ ݎ h =ݐ ݎ h =ݐ h =ݐ ݎ ݎ ݎ ݎ 0 =ݐ 1 =ݐ ݎ 0 =ݐ ݎ 1 =ݐ ݎ ݎ 1 =ݐ 0 =ݐ 1 =ݐ 0 =ݐ This model is often used for International Migration flows. Experts are asked to judge what future flows are likely Or a simple time series model is used (e.g. linear with ceiling or floor, exponential smoothing, auto-regressive moving average model, logistic function with ceiling or floor) Use of a full explanatory model of migration flows rarely attempted, mainly because such a model requires forecasts of the determinant variable
2.3 Assume flows are rates multiplied by a population at risk Transmission rates, time constant or varying input to a projection model ݎ ܣ / ݎ ݎ ܯ = ݎ ݎ ݐ ݎ ܣ ݎ ݎ ݐ = ݎ ݎ ܯ ݎ ܣ / ݎ ܧ = ݎ ݐ ݎ ܣ ݎ ݐ = ݎ ܧ ܣ / ݎ ܫ = ݎ ݐ ܣ ݎ ݐ = ݎ ܫ ܣ / ܯ = ݐ ܣ ݐ = ܫ Admission rates, time constant or varying input to a projection model ݎ ܣ / ݎ ݎ ܯ = ݎ ݎ ݎ ܣ ݎ ݎ = ݎ ݎ ܯ ݎ ܣ / ݎ ܧ = ݎ ݎ ܣ ݎ = ݎ ܧ ܣ / ݎ ܫ = ݎ ܣ ݎ = ݎ ܫ ܣ / ܯ = ܣ = ܯ Rates are used to reflect the influence of changing populations at risk. Rates are often used in internal migration models because it is assumed that there is no constraint acting on the flow. Flows are preferred for international flows because these usually have controls. A pure transmission rate model using the Rest of the World population at risk leads to an immigration explosion. Such a model is useful only to demonstrate that policy controls work. Using a model with emigration rates and immigration flows means emigration increases as the regional populations grow and net international migration falls.
Time constant rates or time varying? Traditional multi-regional models which simply use time constant outmigration rates have been criticised: Application of constant migration rates leads to convergence of the regional system to a stable equilibrium where regional populations may grow or decline but region-age-sex shares are constant But migration rates will respond to changing conditions at the destination as well as at the origin There is a huge literature on Spatial Interaction Models that represents origin to destination flows as a product of origin characteristics, destination characteristics and O-D impedance factors (e.g. distance). In the short term constant out-migration rates models perform better than SIMs. In the medium to long term the difficulties caused by convergence to stable equilibrium kick in
A Statistics Canada solution Patrice Dion (Statistics Canada) has proposed a simple method of introducing destination influences on O-D flows The adjustment consists of modifying the out-migration rates, for each year projected, on the basis of the average out-migration rates and population sizes observed during the reference period and on the basis of the population sizes at time t, i.e., at the beginning of the year to be projected. Hence, the out-migration rate between t and t+1 ( ௧ǡ௧ ଵ ) is modified as follows: ௧ǡ௧ ଵ (8.1) where is the average rate observed during the reference period, ௧ is the size of the population of destination, and is the average size of the population of destination during the reference period.
Handling migrations in regions and countries Key to Figures 1.2 to 1.5 Internal flows in system of interest External flows in system of interest Outside system of interest 1.2 Single region system SINGLE REGION DESTINATIONS MIGRATION FLOWS Regions in a Other s of countries world ORIGINS 1 2 n + 2 m + 2 c + 1 Regions in a 2 n + 2 m + Other s of countries world + 2 c Many LAD models Model for LAD and neighbours DEMIFER Model for EU Regions and Countries 1.3 Regions used in a POPGROUP application 1.4 Multi-region system within a whole REGIONS IN POPGROUP DESTINATIONS MIGRATION Regions in a Regions in a Countries in a Other s of FLOWS countries world ORIGINS 1 2 k k+1 n + 2 m + 2 c + 1 Regions in a Regions in a 2 k k+1 n + 2 m + 2 c Other s of countries world + MANY REGIONS DESTINATIONS MIGRATION FLOWS Regions in a Other s of countries world ORIGINS 1 2 n + 2 m + 2 c + 1 Regions in a 1.5 Multi-region system in a of countries 1.6 Multi- system in all s of countries MANY REGIONS, COUNTRIES DESTINATIONS MIGRATION FLOWS Regions in a Other s of countries world ORIGINS 1 2 n + 2 m + 2 c + 1 Regions in a 2 n + 2 m + Other s of countries world + 2 c 2 n + 2 m + Other s of countries world + MANY COUNTRIES 2 c DESTINATIONS MIGRATION FLOWS Regions in a Other s of countries world ORIGINS 1 2 n + 2 m + 2 c + 1 Regions in a 2 n + 2 m + Other s of countries world + 2 c ONS SNPP model WIC World Country Model
Summary of Advice It is better to use gross migration flows rather than net. It is better to use gross migration rates than flows, as long as there are no constraints on the number of migrants who can be received. If there are constraints, then a model of future in-migration or immigration totals should be developed. Migration admission rates can be used to source migrants by origin. A trade-off needs to be considered between more detail (when input variables are harder to estimate) and less detail (easier to estimate). The bi-regional model has less detail than the multi-regional model but gives results that are close. The three faces version of the multi-regional model is a reasonable approximation to the full model. Different models may be needed for the different sets of flows (internal, international, or intermediate e.g. within Europe). The best choice for internal migration (migration transmission rate in a bi-regional or multi-regional model) may not be the best choice for international migration. A world model of inter- migration may exaggerate migration to richer destinations which may be subject to controls.