Microsimulation — Individual-Level Policy Modelling
- Introduced by
- Guy Orcutt (concept, 1957); modern tax-transfer frameworks developed through EUROMOD and related projects
- Year
- 1957
- Type
- Policy simulation / computational social science
What does it do?
Microsimulation is a computational method that simulates policy effects by operating directly on a population of individual micro-units — households, firms, patients — and applying rules to each unit according to its own demographic, economic, and behavioural characteristics.
Developed conceptually by Guy Orcutt in 1957, it has become the standard tool for evaluating tax reform, pension systems, and health policy before implementation.
The intuition behind the method
Most policy questions are really questions about heterogeneous people, not about averages. A flat-tax reform affects a low-income single parent very differently from a high-income couple, and those differences matter for understanding who wins, who loses, and what the aggregate cost will be. Microsimulation captures this by keeping the population as a dataset of individual records and running the policy rules record by record — exactly as a tax authority would — rather than feeding aggregate totals into a macro-model. The result is a rich, unit-by-unit picture of distributional impact that aggregate models cannot produce.
History
The conceptual foundations of microsimulation were laid by Guy Orcutt in a 1957 American Economic Review article, where he argued that economic models should operate on the individual households and firms that actually make decisions, rather than on aggregate quantities. Static tax-benefit microsimulation expanded through the 1980s and 1990s with government models in many countries. The EUROMOD model (Sutherland & Figari, 2013) established a cross-national, harmonised tax-benefit microsimulation for European Union countries. Dynamic microsimulation — which ages the population through time using demographic and behavioural transition equations — developed in parallel, with models such as LIAM2 providing open-source infrastructure. The Handbook of Microsimulation Modelling (O'Donoghue, 2014) provides a comprehensive survey of both branches.
When to use it
Microsimulation is appropriate when you need distributional, unit-level analysis of a policy change and have access to a large, representative micro-dataset (typically at least 1 000 records, often tens of thousands). It is the natural choice for tax-benefit analysis, pension reform, health financing, and social insurance design. A static model applies when the reform is discrete and behavioural response can be ignored or treated as a sensitivity case; a dynamic model is needed when the time path of ageing, employment, or health transitions matters. The method is not suited to settings where only aggregate data are available, where the population of interest has no individual-record representation, or where behavioural responses are theoretically central but cannot be parameterised from evidence.
Limitations and considerations
- Requires a high-quality, large-scale micro-dataset; results are only as representative as the underlying data.
- Behavioural responses are difficult to model reliably; static models assume no response, which can be a strong assumption for large reforms.
- Elasticity parameters needed for dynamic models must be imported from external studies and introduce additional uncertainty.
- Building and validating a microsimulation model is resource-intensive: it demands programming expertise, deep knowledge of the policy rules, and careful data management.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Produces full distributional impact profiles — not just mean effects but who gains and who loses, by income, age, gender, or any other characteristic in the data.
- Handles policy heterogeneity naturally: complex, nonlinear tax-benefit rules that cannot be approximated in a macro-model are applied exactly at the unit level.
- Results are directly interpretable by policymakers: fiscal cost, headcount poverty change, and Gini coefficient shifts are reported in familiar units.
- Dynamic models can project the evolution of a population and test policy sustainability over long horizons.
Applications
- Estimating the distributional and fiscal effects of income tax reform across household income groups before legislation is enacted.
- Projecting the long-run costs and benefit-adequacy of pension system reforms as the population ages.
- Assessing the poverty and inequality impact of social benefit redesigns, including means-testing changes and benefit caps.
- Evaluating health financing reforms — such as co-payment changes or insurance coverage expansions — on patient out-of-pocket costs by demographic group.
- Producing baseline distributional statistics for economic inequality research using large administrative or survey datasets.
How it works
How it works
A representative micro-dataset is assembled — usually a large-scale household survey or an administrative register — and sampling weights are preserved to ensure the synthetic population represents the true population. Variables needed by the policy rules (income components, family structure, employment status, health indicators) are cleaned, harmonised, and validated. For dynamic models, the population is 'aged' forward using demographic transition probabilities before the policy is applied.
Each policy instrument (tax schedule, benefit entitlement, contribution rate) is translated into explicit computational rules. In static microsimulation the rules reflect the current law and the counterfactual reform, and no behavioural response is assumed. In dynamic models, labour-supply elasticities, saving responses, or health-seeking behaviour are also parameterised and linked to the rule set.
The rule engine processes every record in the micro-dataset: it computes taxes paid, benefits received, or outcomes experienced under baseline and reform scenarios for each unit. For dynamic models, the engine steps through discrete time periods, updating unit characteristics at each period according to transition probabilities and applying the prevailing rules. Stochastic models draw random numbers to resolve transitions; deterministic models apply expected-value weights.
Unit-level outputs are weighted and aggregated to produce population-level estimates: distributional tables (by income decile, age group, region), fiscal cost or saving, poverty and inequality indicators, and — in dynamic models — projected paths over time. Sensitivity checks vary key behavioural parameters to bound the uncertainty in the estimates.
Common pitfalls
- Failing to preserve or correctly apply sampling weights leads to estimates that do not represent the true population — always propagate weights through every aggregation step.
- Treating a static model as if it captures behavioural response: the 'morning-after' assumption (no behaviour change) is a deliberate simplification, not a finding; communicate this clearly in any report.
- Using population data without checking representativeness: survey non-response and administrative coverage gaps can bias distributional results for specific subgroups.
- Conflating model precision with model accuracy — a microsimulation can produce very precise wrong answers if the behavioural parameters or policy rules are misspecified.
- For dynamic models, ignoring the compounding uncertainty of long projection horizons; always report sensitivity ranges around central projections.
Frequently asked questions
What is the difference between static and dynamic microsimulation?
A static model applies new policy rules to a population snapshot and computes the 'morning-after' distributional change, assuming behaviour is unchanged. A dynamic model advances the population forward through time using demographic transition probabilities (birth, death, employment change, health deterioration) and can incorporate behavioural responses such as labour supply adjustments. Static models are simpler and less data-demanding; dynamic models are needed when the time path of change matters, for example in pension or long-term care projections.
How large does the micro-dataset need to be?
The statwise registry specifies a practical minimum of 1 000 micro-units, but meaningful distributional analysis — especially for small subgroups — typically requires tens of thousands of records. Administrative registers with near-complete population coverage are preferred over small household surveys, whose subgroup estimates can have wide confidence intervals even after weighting.
How are behavioural responses incorporated?
In static models they are not: the 'first-order' or 'morning-after' effect is reported without adjustment. To add behavioural response, analysts layer in labour-supply elasticities, savings rates, or health-seeking parameters estimated from the literature or from natural experiments. Each parameter introduces additional uncertainty, so sensitivity analysis across plausible parameter ranges is essential.
Is microsimulation the same as agent-based modelling?
They share the principle of operating on individual units rather than aggregates, but they differ in emphasis. Microsimulation typically uses a fixed empirical population and deterministic (or statistically calibrated) policy rules to estimate distributional outcomes; agent-based models emphasise emergent collective behaviour arising from agent interactions and tend to use stylised, not empirically representative, populations. For tax-benefit and policy costing work, microsimulation is the standard framework.
Related methods
Sources
O'Donoghue, C. (Ed.) (2014). Handbook of Microsimulation Modelling. Emerald.
Open DOI referenceCopied.Li, J. & O'Donoghue, C. (2013). A Survey of Dynamic Microsimulation Models: Uses, Model Structure and Methodology. International Journal of Microsimulation, 6(2), 3–55.
Open publication referenceCopied.