Robust Approximate Bayesian Computation
Also known as: Robust ABC, robust ABC inference, outlier-robust ABC, robust likelihood-free inference
Robust ABC extends standard Approximate Bayesian Computation to handle outliers, model misspecification, and sensitivity to summary statistic choice. By replacing conventional distance measures with robust alternatives — such as composite scores, trimmed statistics, or synthetic likelihoods — it protects posterior inference from being distorted by atypical observations or an imperfect simulator.
Key highlights
- Protects posterior inference from distortion by outliers and model misspecification.
- Requires no analytical likelihood — only the ability to simulate data from the model.
- Composite score and synthetic-likelihood variants retain efficiency close to that of standard ABC under clean data.
- Compatible with sequential Monte Carlo and adaptive tolerance schemes for computational efficiency.
- Applicable to complex scientific simulators in ecology, epidemiology, genetics, and engineering.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use robust ABC when the data may contain outliers or the simulator is only a rough approximation of the true process, and you need posterior inference that does not collapse around those anomalies. It is particularly suited to ecology, epidemiology, and engineering problems where the generative model is complex but imperfect and data quality is uneven. Do not use robust ABC as a routine replacement for standard ABC if the data are clean and the simulator is well validated; the extra complexity is unnecessary. Also avoid it when the sample size is very small (fewer than ~30 observations), because robust summary statistics are themselves imprecise with little data.
Strengths & limitations
- Protects posterior inference from distortion by outliers and model misspecification.
- Requires no analytical likelihood — only the ability to simulate data from the model.
- Composite score and synthetic-likelihood variants retain efficiency close to that of standard ABC under clean data.
- Compatible with sequential Monte Carlo and adaptive tolerance schemes for computational efficiency.
- Applicable to complex scientific simulators in ecology, epidemiology, genetics, and engineering.
- Choosing an appropriate robust summary statistic or composite score requires domain knowledge and often pilot experimentation.
- Robust summaries may discard information, slightly widening posteriors compared with an optimal non-robust approach.
- Computational cost is higher than standard ABC because robust distance calculations are more expensive per iteration.
- Theoretical guarantees on posterior consistency under misspecification are model-specific and not always available.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does robust ABC differ from standard ABC?
Standard ABC uses simple summary statistics and a Euclidean or other basic distance. Robust ABC replaces these with outlier-resistant summaries — composite scores, trimmed statistics, or robust synthetic likelihoods — so that a few anomalous observations or a slightly wrong simulator do not distort the accepted posterior sample.
What is a composite score function in this context?
A composite score is derived from a partial or pseudo-likelihood rather than the full likelihood. It summarises the data in a way that separates the contribution of each observation, allowing outlier contributions to be down-weighted without discarding the observation entirely.
Can robust ABC be combined with Sequential Monte Carlo?
Yes. SMC-ABC sequences a series of decreasing tolerances ε, propagating a population of particles with importance weights. Swapping the standard distance for a robust one requires only changing the acceptance criterion; the SMC machinery remains the same, yielding an adaptive robust ABC posterior efficiently.
When should I prefer robust Bayesian inference over robust ABC?
If you can write down an analytical (or approximately analytical) likelihood, robust Bayesian inference methods — such as heavy-tailed likelihoods or power-likelihood approaches — are usually more efficient than ABC. Use robust ABC when no tractable likelihood exists and you must rely on simulation.
How do I choose the tolerance ε?
Start with a pilot run to understand the distribution of robust distances, then select ε at a quantile (e.g., 1–5th percentile of pilot distances). Perform sensitivity analysis across several ε values; if the posterior changes substantially, the approximation is not yet stable. Sequential schemes that adapt ε automatically reduce this guesswork.
Sources
- 1.Ruli, E., Sartori, N. & Ventura, L. (2016). Approximate Bayesian computation with composite score functions. Statistics and Computing, 26(3), 679–692.
- 2.Frazier, D. T., Drovandi, C. & Nott, D. J. (2020). Robust Approximate Bayesian Inference with Synthetic Likelihood. Journal of Computational and Graphical Statistics, 30(4), 958–976.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Robust Approximate Bayesian Computation. ScholarGate. https://scholargate.app/bayesian/robust-approximate-bayesian-computation