Spatial Microsimulation
- Introduced by
- Developed in the IPF/microsimulation tradition; synthesized for geography by Lovelace & Dumont
- Year
- 2016
- Type
- Method for generating and analysing synthetic individual-level populations within small areas
- Family
- Population synthesis / microsimulation
What does it do?
Spatial microsimulation is a family of techniques for generating realistic synthetic populations of individuals within small geographic areas, by combining detailed but geographically coarse survey microdata with geographically fine but aggregate census tables.
It estimates, for every neighbourhood, a population of individuals whose collective characteristics match the published margins — the right number of each age, sex, income, and tenure group — even though no survey directly samples individuals at that fine scale. Synthesized for the geographic community in Robin Lovelace and Morgane Dumont's 2016 book, it bridges the gap between rich individual data and small-area aggregates so that policy and behaviour can be modelled where people actually live.
The intuition behind the method
Censuses tell you how many people of each age or income band live in a neighbourhood, but never who has which combination of traits; surveys tell you the full profile of sampled individuals, but only for large regions. Spatial microsimulation marries the two by taking the survey people and cloning them into each neighbourhood in just the right proportions so the local totals come out correct — like dealing a deck of real individuals into each area until the counts of every category match the census. The outcome is a plausible roster of synthetic residents for every small area, each carrying a full set of attributes, on which you can then run what-if experiments: who would gain from a tax change, where fuel poverty concentrates, how a new policy lands street by street.
History
Microsimulation as a tool for socio-economic modelling traces to Guy Orcutt's 1957 vision of simulating individual households to study policy, and the iterative proportional fitting algorithm itself dates to Deming and Stephan in 1940. The spatial branch grew from the 1990s as geographers, notably in the work of researchers such as Ballas, Clarke, Williamson, and others, adapted reweighting and combinatorial-optimisation techniques to estimate populations for census output areas and to model the local incidence of taxes, health, and transport policy. Robin Lovelace and Morgane Dumont's 2016 book Spatial Microsimulation with R consolidated the methods, codified the IPF and combinatorial-optimisation workflows, and made reproducible open-source implementation the norm. The field now overlaps heavily with synthetic population generation for agent-based and activity-based transport models.
When to use it
Use spatial microsimulation when you need individual-level data at a fine geographic scale that no single dataset provides — typically because surveys lack spatial detail and censuses lack individual records — and you have survey microdata plus small-area constraint tables that share variables. It is the standard approach for estimating small-area distributions of income, health, or deprivation, for modelling the spatial incidence of policies, and for building synthetic populations that seed agent-based or transport models. It is less appropriate when the variable of interest is absent from both the survey and the constraints, when the survey is unrepresentative of the local population, or when relationships between attributes vary spatially in ways the global survey cannot capture. Strong reliance on the survey's joint distributions is its key assumption.
Limitations and considerations
- Relies on the assumption that the survey's joint relationships between attributes hold across all small areas.
- Can only synthesize variables present in both the survey microdata and the constraint tables.
- Iterative proportional fitting yields fractional weights that require integerisation, introducing approximation.
- Validation is hard because the true small-area microdata it estimates are, by definition, unobserved.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Creates individual-level microdata at small-area scales where no direct individual-level data exist.
- Reproduces published census margins exactly (or closely), grounding the synthetic population in official totals.
- Enables spatial what-if policy analysis, revealing who is affected by a reform and where impacts concentrate.
- Produces synthetic populations that seed agent-based, transport, and land-use models requiring individual agents.
Applications
- Estimating small-area distributions of income, poverty, and deprivation from survey and census data.
- Modelling the spatial incidence of tax, benefit, and welfare reforms across neighbourhoods.
- Building synthetic populations to seed agent-based, land-use, and activity-based transport models.
- Projecting small-area demand for health and social care from individual-level risk profiles.
- Assessing fuel poverty, obesity, or smoking prevalence at fine geographies for targeted intervention.
How it works
How it works
The process begins by assembling, for each small area, the aggregate constraint tables — the marginal totals from the census that the synthetic population must reproduce, such as counts by age band, sex, employment status, and housing tenure. These margins, denoted m_ac for area a and constraint c, are the targets the synthesis aims to hit, and they must be chosen to overlap with variables present in the survey microdata. The quality and number of constraints directly determine how tightly the synthetic population is pinned down.
Formula
\{m_{ac}\}:\ \text{small-area } a \text{ marginal totals for constraint } cThe most common synthesis engine is iterative proportional fitting, which assigns each survey individual a weight and repeatedly rescales those weights so that, taken in turn, the weighted totals match each census margin. After cycling through all constraints to convergence, the weights describe how many times each survey respondent should be replicated in the area to satisfy every margin simultaneously. Because weights are continuous, an integerisation step is usually applied to turn them into whole synthetic individuals that can be enumerated and modelled.
Formula
w^{(k+1)} = w^{(k)}\,\frac{m_{c}}{\sum w^{(k)}}\quad\text{(IPF: iterate over each margin)}An alternative to reweighting is to treat synthesis as a selection problem: choose a combination of whole survey individuals whose aggregated characteristics best match the constraints. Methods such as simulated annealing search over possible selections to minimize the discrepancy between the synthetic and target margins, while approaches like GREGWT use generalized regression weighting. These combinatorial-optimisation methods directly yield integer populations and can handle complex constraint sets, at the cost of more computation than iterative proportional fitting.
Formula
\min_{\text{selection}} \sum_c \bigl\| \hat{m}_c - m_c \bigr\|\quad\text{(simulated annealing / GREGWT)}Before use, the synthetic population is validated by checking how well its aggregated characteristics reproduce the constraints, ideally including held-out tables not used in fitting, and by examining internal and external consistency. Once trusted, the micro-populations are the analytical payoff: because every synthetic individual carries a full attribute profile and a location, analysts can estimate small-area distributions of variables the census never crossed, and run what-if policy scenarios — a tax reform, a benefit change — to see who is affected and where the impacts concentrate spatially.
Formula
\text{compare } \hat{m}_c \text{ vs held-out tables; analyse micro-populations}Common pitfalls
- Choosing constraints that do not overlap with survey variables, so the margins cannot be matched.
- Including too many or redundant constraints, which can prevent convergence or overfit to noisy small-area totals.
- Skipping integerisation or doing it carelessly, producing fractional or inconsistent individuals that break downstream models.
- Validating only against the constraints used in fitting, which guarantees a good fit and hides genuine error.
Frequently asked questions
How does spatial microsimulation differ from generic microsimulation?
Generic microsimulation models the evolution or behaviour of a population of individuals or households, often without explicit geography, to study policy or dynamics over time. Spatial microsimulation adds the geographic dimension: its defining task is to synthesize individual-level populations for many small areas so that each area's totals match local census constraints, attaching a location to every synthetic individual. In practice spatial microsimulation is the population-synthesis front end that produces the georeferenced agents which a behavioural or dynamic microsimulation then operates on.
What is the difference between iterative proportional fitting and combinatorial optimisation here?
Iterative proportional fitting reweights survey individuals, repeatedly rescaling continuous weights until the weighted totals match every census margin, and then integerises those weights into whole people. Combinatorial optimisation instead searches directly for a set of whole survey individuals — using simulated annealing, genetic algorithms, or generalized regression weighting like GREGWT — whose aggregate matches the constraints. IPF is fast and deterministic but needs integerisation, whereas combinatorial methods yield integer populations natively and handle complex constraints, at higher computational cost.
How can a synthetic population be validated when the true small-area microdata are unknown?
Because the individual-level small-area data being estimated are unobserved, validation relies on indirect checks. Internal validation confirms that the synthetic population reproduces the constraints used to build it; external validation compares aggregated synthetic outputs against tables held out of the fitting or against an independent dataset for the same area. Analysts also examine plausibility of joint distributions and, where any ground-truth microdata exist for a subset of areas, benchmark against them. No single test is decisive, so credible studies report several complementary validation measures.
Related methods
Sources
Lovelace, R., & Dumont, M. (2016). Spatial Microsimulation with R. Chapman and Hall/CRC, Boca Raton.
ISBN: 9781498711548Copied.