Multilevel Approximate Bayesian Computation
Also known as: multilevel ABC, hierarchical ABC, multi-level ABC, ABC for hierarchical models
Multilevel Approximate Bayesian Computation (multilevel ABC) extends simulation-based Bayesian inference to hierarchically structured data. When the likelihood is intractable and observations are nested within groups, it replaces direct likelihood evaluation with simulations at each level of the hierarchy, accepting parameter draws whose simulated summary statistics are close to the observed ones.
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 multilevel ABC when data have a hierarchical structure (e.g., individuals nested within groups, repeated measures, schools within districts) and the likelihood function is intractable or too expensive to evaluate — as is common in agent-based models, complex population dynamics, or mechanistic biological models. It is also appropriate when classical multilevel MCMC mixes poorly because the posterior is highly multimodal or the parameter space is irregular. Do not use it when a tractable likelihood exists and standard multilevel MCMC or variational inference is feasible: ABC introduces additional approximation error from the epsilon threshold and the choice of summary statistics, and it typically requires very large numbers of simulations, making it computationally expensive.
Strengths & limitations
- Handles intractable likelihoods in hierarchical settings where MCMC is infeasible.
- Works with any forward-simulable generative model regardless of likelihood complexity.
- Naturally propagates uncertainty across levels of the hierarchy.
- Can be parallelised easily because simulations are independent draws.
- Compatible with regression adjustment and SMC refinements to reduce approximation error.
- Approximation quality depends critically on the choice of summary statistics — poor summaries yield biased posteriors.
- Very high simulation cost: thousands to millions of model evaluations are typically needed.
- The epsilon tolerance introduces a systematic approximation; the posterior is exact only as epsilon tends to zero.
- Curse of dimensionality: performance degrades rapidly as the number of parameters or levels grows.
Frequently asked
What makes ABC 'approximate'?
ABC replaces the exact likelihood with a comparison of summary statistics within a tolerance epsilon. The resulting posterior is an approximation that equals the true posterior only when epsilon is zero and the summaries are sufficient. In practice both conditions are violated, so posterior estimates carry an additional bias beyond ordinary Monte Carlo error.
How is multilevel ABC different from standard hierarchical MCMC?
Hierarchical MCMC requires an evaluable likelihood at each step; it is exact (up to MCMC error) when it converges. Multilevel ABC replaces likelihood evaluation with forward simulation, making it applicable when the likelihood is intractable, but at the cost of the epsilon-approximation and much higher simulation counts.
How do I choose summary statistics for hierarchical data?
Include statistics that capture variation both within and between groups — for example, group means, within-group variances, and cross-level correlations. Learned neural summary statistics (e.g., via neural posterior estimation) have shown strong performance when manual choice is difficult.
Can I combine multilevel ABC with sequential Monte Carlo?
Yes. SMC-ABC propagates a population of particles through a sequence of decreasing tolerances, which is particularly beneficial in multilevel settings because it avoids the high rejection rates of naive ABC rejection sampling when epsilon is small.
When should I prefer multilevel variational inference instead?
If the likelihood is tractable but posteriors are complex, multilevel variational inference is faster and avoids the simulation overhead of ABC. Choose ABC only when the likelihood genuinely cannot be evaluated, and prefer variational inference or MCMC otherwise.
Sources
- Beaumont, M. A., Zhang, W., & Balding, D. J. (2002). Approximate Bayesian computation in population genetics. Genetics, 162(4), 2025–2035. DOI: 10.1093/genetics/162.4.2025 ↗
- Jasra, A., Singh, S. S., Martin, J. S., & McCoy, E. (2012). Filtering via approximate Bayesian computation. Statistics and Computing, 22(6), 1223–1237. link ↗
How to cite this page
ScholarGate. (2026, June 3). Multilevel Approximate Bayesian Computation. ScholarGate. https://scholargate.app/en/bayesian/multilevel-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
- Bayesian Hierarchical Model with Missing DataBayesian↔ compare
- Hierarchical Bayesian InferenceBayesian↔ compare
- Markov Chain Monte CarloSimulation↔ compare
- Multilevel Bayesian InferenceBayesian↔ compare
- Sequential Monte CarloBayesian↔ compare