Hierarchical Approximate Bayesian Computation
Also known as: hierarchical ABC, ABC for hierarchical models, multilevel ABC, population ABC
Hierarchical ABC is a likelihood-free Bayesian inference method designed for multilevel data structures in which individual-level parameters are themselves drawn from a population-level distribution. By combining simulation-based rejection sampling with hierarchical pooling, it recovers both within-group and between-group posterior distributions without requiring a tractable likelihood function.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use hierarchical ABC when the data have a clear nested structure (individuals within groups, repeated measures within subjects, species within genera) and the mechanistic model has no tractable likelihood. It is particularly valuable in systems biology, ecology, and pharmacokinetics where simulator-based models are standard and groups differ but share common biology. Avoid it when a likelihood can be written down and evaluated efficiently — full Bayesian hierarchical MCMC will be more accurate and far faster. Also avoid it when summary statistics that capture all relevant information cannot be identified, or when the number of hierarchical levels is very deep, as acceptance rates collapse rapidly with dimensionality.
Strengths & limitations
- Enables principled Bayesian inference for complex mechanistic models with intractable likelihoods in a multilevel setting.
- Partial pooling across groups stabilises estimates for data-sparse individuals while still allowing group-specific variation.
- Returns full posterior distributions for both individual and population parameters, quantifying uncertainty at every level.
- Flexible: compatible with any forward-simulation model regardless of structural complexity.
- Can be embedded within SMC-ABC to adaptively reduce the tolerance ε and improve efficiency.
- Computationally intensive: a large number of simulations is required, especially as dimensionality or the number of groups grows.
- Posterior quality depends critically on the choice and sufficiency of summary statistics — insufficient statistics lead to biased posteriors.
- Acceptance rates drop steeply with the number of parameters, making high-dimensional hierarchical problems impractical without variance-reduction strategies.
- Tolerance ε introduces a bias-variance trade-off: small ε is accurate but slow; large ε is fast but approximate.
- The method approximates the posterior; credible intervals are exact only in the limit ε → 0.
Frequently asked
How does hierarchical ABC differ from running ABC separately for each group?
Independent per-group ABC fits treat each group in isolation, requiring enough data within each group to identify its parameters. Hierarchical ABC pools information across groups via a shared population distribution, so groups with sparse data borrow strength from the population and individual estimates are regularised — exactly the same advantage that hierarchical MCMC offers over separate group-level regressions.
What summary statistics should I use?
Ideally, statistics that are jointly sufficient for the model parameters at both individual and population levels. In practice, domain knowledge is the best guide. Candidate statistics should be validated by checking that the ABC posterior recovers ground-truth parameters in simulated experiments. Automatic summary-statistic selection methods (e.g. semi-automatic ABC by Fearnhead & Prangle) can help when intuition is limited.
Can hierarchical ABC scale to large numbers of groups?
With many groups the joint parameter space grows, and acceptance rates in naive rejection-ABC collapse. SMC-ABC or MCMC-ABC kernels with population-level proposals substantially improve scaling. Still, beyond a few dozen groups with complex individual models, the computational burden becomes prohibitive and likelihood-based alternatives (e.g. integrated nested Laplace approximation or hierarchical variational inference) should be considered.
How do I choose the tolerance ε?
Start with a pilot run using a generous ε to obtain accepted samples quickly, then tighten ε progressively — or use an adaptive SMC-ABC schedule that reduces ε automatically. The final ε should be small enough that the ABC posterior is visually stable across a further reduction in ε, serving as a practical convergence check.
Is there software implementing hierarchical ABC?
Yes. The pyABC library (Python) supports hierarchical population models within an SMC-ABC framework. ABCpy (Python) and the R package EasyABC also offer hierarchical extensions. Stan and PyMC do not implement ABC directly but can be combined with external simulators via custom likelihood approximations.
Sources
- Toni, T. & Stumpf, M. P. H. (2010). Simulation-based model selection for dynamical systems in systems and population biology. Bioinformatics, 26(1), 104–110. DOI: 10.1093/bioinformatics/btp619 ↗
- Wilkinson, R. D. (2013). Approximate Bayesian computation (ABC) gives exact results under the assumption of model error. Statistical Applications in Genetics and Molecular Biology, 12(2), 129–141. DOI: 10.1515/sagmb-2013-0010 ↗
How to cite this page
ScholarGate. (2026, June 3). Hierarchical Approximate Bayesian Computation. ScholarGate. https://scholargate.app/en/bayesian/hierarchical-approximate-bayesian-computation
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Approximate Bayesian ComputationSimulation↔ compare
- Hierarchical Bayesian InferenceBayesian↔ compare
- Hierarchical Markov Chain Monte CarloBayesian↔ compare
- Sequential Monte CarloBayesian↔ compare