Methodology

Overview

The main demographic model forecasts:

• population by gender, individual age and district for each year over 50 years based on forecasts of domestic migration rates, international immigration, overseas emigration rates, births (fertility rates) and deaths (mortality rates)
• the number of people in the labour force, by gender and individual age and district based on forecasts of labour force participation rates.

  The model focusses on individual ages rather than population age-groups because data on population by age-group overlooks important differences in the size of age-specific populations. For example, for many districts, the number of 17 year-olds in the population has a significant effect on the number of people likely to emigrate in the next 12 months. For other districts the number of people aged 21 in district can have marked impacts, because of high outflows of tertiary graduates.

Our modelling includes extensive use of sub-models for estimating population characteristics, due to important gaps in official statistics, particularly in regard to households and dwellings.

Date range

Our model produces forecasts from 2020 to 2054, at five year intervals.

Population model

Simulates demographic characteristics of every New Zealander

Our model begins with current population base using Stats NZ data. Then we age the population, adjusting it for births, deaths and internal and external migration.

Our demographic forecasts are based on our existing district demographic and economic forecasting model (DDEM). DDEM forecasts the number of people in New Zealand, by gender, individual age, and local council area over a 50-year period.

To forecast local characteristics, such as dwelling demand by small area unit or meshblock, we use a subsidiary model that focusses on a subset of districts (in this case those of the eight local councils). This two-stage method ensures our models do not become unnecessarily unwieldy as we explore the specifics of local areas.

The advantage of starting by forecasting the populations of all districts in New Zealand, simultaneously, is that the forecasts are consistent in the sense that domestic migration into one district depends on the size and age composition of the population of other districts, as discussed below.

The DDEM is a probabilistic simulation model. The model is simulated hundreds of times, with each simulation producing a separate population forecast. In each simulation the key assumptions in the model are varied randomly but in proportion to observed variation in the values of these factors (parameters).

Our methods differ from other demographic projections we are aware of, such as those of Stats NZ. We take care to decompose key drivers of net migration by:

• treating inward migration and outward migration separately, to capture distinctly different long-term trends in inward and outward migration and thus trend increases in net migration
• differentiating inward migration by source country
• incorporating economic predictors that are associated with push and full factors, focussing on:
  • domestic economic conditions (proxied by the unemployment rate)
  • economic conditions in Australia (proxied by the unemployment rate)
  • domestic economic conditions relative to the rest of the world.

Our models take account of the fact that economic conditions both influence international migration (driving net migration up when times are good and down when times are bad) and are influenced by net migration (with high net migration increasing domestic demand).

We fit a range of models and conducted variable selection tests across a range of model specifications. Ultimately, we chose to average forecasts across 5 separate models.

Fertility

Age-specific fertility rates, by district, are modelled and forecast with a model that decomposes trends in fertility rates into an age-specific component and year- or cohort-specific components.

The model takes account of relationships between fertility rates at different ages. It accounts for shifts in the timing of births, amongst individuals, when fertility is delayed. And it accounts for changing trends in numbers of births per female, age of first birth and time between births. This is a conventional and widely used approach to forecasting fertility rates.

There are limits to the amount and quality of data available on district-level age-specific fertility rates. We overcome this by constructing age-specific district fertility rate data based on regional and national age-specific fertility rates. These estimates are not official data, but they are accurate predictors of numbers of births by district.

Our probabilistic projections make use of the observed variation in births by district, relative to predicted births, to calibrate distributions from which we take samples to simulate uncertainty in our projections.

Mortality

Our model of district-, age- and sex-specific mortality rates is methodologically similar to the one used for fertility rates, based on decomposing mortality trends into age-specific and year- specific components. In addition, the model accounts for trends in the difference between female and male mortality rates. This ensures that there is logical consistency between forecast male and female mortality rates. This model has been used by Stats NZ.

Mortality rates are estimated using data on district mortality rates for the Census years 1995-1997, 2000-2002, 2005-2007, and 2012-2014 and updated to 2018 using national estimates.

Migration

Forecasts migration by applying net migration rates by age, gender and territorial authority (TA).

Rather than simply adopting a number for net migration, we apply propensities to each TA, age and gender group. Immigration is based on historically observed propensities for immigrants to arrive in each TA.

Our model pays careful attention to international migration trends and cycles. This is because international migration is both a significant driver of population growth and the largest source of uncertainty in demographic projections. This is true even for smaller population centres as increases in national net migration cause ripple effects throughout the country over time.

One of the key benefits of paying careful attention to international migration is that our analysis of international migration can be used to:

• quantify the extent of uncertainty in our projections
• explain our demographic projections and sources of uncertainty in the projections.

Domestic migration

Careful analysis of domestic migration, between regions or districts, is as important as international migration. This is true for all areas of New Zealand. In summary, our methods ensure that modelled migration flows are internally consistent and can be explained and decomposed into key drivers (sources and destinations of migrants). Data on domestic migration is not very good, so logically and internally consistent analysis takes on special importance.

Our methods capture three key aspects of domestic migration not always dealt with well in other forecasts:

1. The interdependence of domestic migration, between areas, that is:
   • when a person arrives in a district, they must have left another district; and
   • net domestic migration is zero by definition.
2. Domestic migration varies significantly between individual ages, especially in terms of young people moving to attend university and moving again to take up their first “real” job
3. Longer-term, increases in population growth through, for example, increased international migration, causes a rise in domestic migration flows. This is best understood by considering the case of population growth in Auckland:
   • population growth puts pressure on local resources, pushing up the cost of living and increasing rates of outward migration
   • there is a consistent net loss of people from Auckland, to domestic migration, especially at older ages: when Auckland grows other areas grow (with a lag).

We model all domestic migration flows in and out of all districts of New Zealand using age-specific and sex-specific rates of outward migration, by district. Origin-destination matrices capture where people are likely to go if they migrate. That is, we start with peoples’ decisions to leave an area and then determine inflows of migrants into other areas based on the sum of all migrants from all districts.

Families-households

Simulates families based on the number of each type and average number of children. We use living arrangement type rates based on age and gender to allocate people to family types. We simulate every family in New Zealand and then produce household forecasts, allowing for multi-family households.

We model household formation using entirely conventional methods (the same as those used by Stats NZ) involving tables of ‘Living Arrangement Type Rates’. These tables describe the probability (or propensity) that an individual of a given sex and age will be in one of six living arrangements (including living in a non-private residence) and their position in those arrangements (i.e. whether a child in the family or an adult in the family).

The living arrangement tables do not account for the number of families that are in multi-family households. This is calculated using Census data on the types of families (e.g. couples, sole-parent families, two-parent families) that typically form multi-family households).

Our model also differentiates household formation rates by household size and district, based on Census data on propensities to form household sizes and types. That is, we augment the living arrangement type rate tables to include measures of the number of children and adults by household type and people’s ages and genders.

Dwelling demand is forecast using a housing location demand model of the sort typically integrated within land-use and transport models. The advantages of using this sort of model are that:

• it is consistent with other planning and investment models, such as strategic transport models, that rely on forecasts of the location of population growth
• it incorporates, as inputs, scenarios or forecasts that can be drawn from other planning processes or forecasts, such as business land demand
• it incorporates development restrictions, and these can be varied by scenario
• it takes account of socio-economic and lifecycle effects, such as people prefer to live away from town in retirement or near university in their late teens and early twenties.

Spatial interactions

The model is known as a spatial interaction model because it takes account of both push and pull factors – the attractiveness of areas and the costs of choosing to live there – and also the idea that people’s choice of where to live includes consideration of a wide range of possible locations. That is, location decisions are assumed to be like people looking at a menu of possible places and weighing the relative merits of each area before choosing where to live.

This sort of model has a very long history, dating back to the 1970s, and is typically a core component of the most advanced modern models of urban development and transport and land use interaction modelling.

In complex models of urban development, the housing location model is solved jointly (simultaneously) with business land decisions and travel demand. As a result, these sorts of models are typically very complex and take a long time to calibrate to local conditions.

On their own, however, housing location models are reasonably straightforward and produce logically consistent forecasts that can be explained in terms of key factors in urban development. The housing location demand model predicts housing demand based on:

• household type and size (number of people)
• preferences for different types of dwelling, by household type or size
• income
• location of employment opportunities
• costs of distance (i.e. travel costs)
• costs and constraints on development by location, for example, available land, development restrictions and land prices.

To calibrate the model and for forecasting purposes, we liaised with local councils and used existing HBA measures of development constraints and available land. By default, we can use physical constraints and ignore development constraints for forecast purposes but only if this is the preference of client team members.

The housing location model also incorporates a view of employment growth by location and we can discuss including Councils’ business land demand forecasts as the core input for this purpose, so that the dwelling forecasts are consistent with business land demand forecasts.

To translate housing demand into dwellings and vacancies, we assume that developers always meet new demand in the year that it appears. This simplifying assumption means that vacancy rates can only fall below the base year value if a combination of the following occurs:

• an area becomes less desirable due to e.g.
  • a change in characteristics of the population (e.g. the population becomes wealthier or older or younger)
  • increased transport costs (which can only occur by scenario)
• population growth slows or becomes negative.

The mix of potential housing, by type, is set to reflect existing compositions, but development restrictions can be set so that only dwellings of certain types are available for development in an area – and hence will tend to appeal to households that are interested in that type of housing.

Income

Estimates personal income based on distributions for qualification, age and employment status. We estimate personal incomes from people’s characteristics and our knowledge of income distributions. We adjust incomes up or down to reflect local territorial authority characteristics.

Our demographic forecasts will produce forecasts of:

• labour force numbers, based on a model of age-specific labour force participation rates
• unemployment, based on a district’s differences relative to national rates and NZ Treasury forecasts and long-term projections
• employment, following directly from labour force numbers and unemployment rates, by location of residence and location of employment (two different measures)
• median incomes by household type and counts of households, by type, above and below median incomes based on typical relationships between district income growth (employment multiplied adjusted for total factor productivity growth) and household income growth, for households above and below the median.

Housing continuum

Tests whether families are eligible for public housing and allocates households to housing tenure. We focus not just on people on the register but people and families who are eligible for social housing.

Challenges with gaps in official statistics

A key methodological challenge is the considerable uncertainty about where we are, in addition to uncertainty about what the future holds. Official data on demographics is patchy at the district level and at lower levels of spatial detail. We have already noted some of these gaps, in terms of fertility rates by districts, and how we deal with these gaps.

The most pronounced data gap is in understanding/estimating the number of households by district and by small area unit because “no subnational dwelling or household estimates are available, apart from the subnational household estimates at 30 June 1996, 2001 and 2006 derived for subnational family and household projections”.

While the Census provides some data on counts of people and counts of households it undercounts the true population and number of households in the Wellington region. For example, the 2013 Census counted approximately 176,000 households in the Wellington region while Stats NZ estimated that the true count was approximately 185,000 - 5% or 9,000 more households than the Census count.

Subnational family and household projections are not likely to be released until the second half of 2021. So, official estimates of the number of households in the Wellington Region in 2018 will not be available until the second half of 2021.

We have two solutions to this problem. First, we conduct our own estimates of the number of households per district. Second, we use our probabilistic modelling to reflect uncertainty about the current number of households as well as future changes.

The numbers of dwellings by district and by small area unit and meshblock is only available at each Census. The solution to this is to use Council information on building consents, completions and work in progress to provide informed estimates of the number and type of dwellings within the region. We assume that the Council project partners will be willing to assist with this.