Approximate Bayesian Computation with Measurement Error
Also known as: ABC with measurement error, ABC-ME, likelihood-free inference with measurement error, simulation-based inference under measurement error
Approximate Bayesian Computation with measurement error (ABC-ME) extends the standard ABC likelihood-free framework to settings where observed data are themselves noisy or imprecisely recorded. By explicitly incorporating a measurement-error kernel into the acceptance step, ABC-ME targets the correct posterior over model parameters even when the true data-generating process cannot be directly observed.
Key highlights
- Handles arbitrarily complex forward models where the likelihood cannot be written or evaluated.
- Explicitly accounts for measurement noise, preventing the parameter bias that naive ABC would incur.
- Requires only the ability to simulate from the model, not to evaluate any analytic density.
- Wilkinson's (2013) result provides a formal theoretical guarantee: ABC-ME is exact under the assumed measurement-error model.
- Naturally integrates with sequential Monte Carlo or regression-adjustment ABC extensions for better efficiency.
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 ABC-ME when (1) the likelihood is intractable or too expensive to evaluate, (2) observed data are known to carry measurement noise (sensor error, self-report bias, rounding), and (3) you need posterior inference rather than point estimates. It is particularly valuable in systems biology, ecology, and epidemiology where complex forward models simulate error-free outcomes but field measurements are noisy. Do not use it when the likelihood is tractable — exact Bayesian or MCMC methods are more efficient. Avoid it when informative summary statistics cannot be constructed, as the resulting posterior may lose critical information about the parameters.
Strengths & limitations
- Handles arbitrarily complex forward models where the likelihood cannot be written or evaluated.
- Explicitly accounts for measurement noise, preventing the parameter bias that naive ABC would incur.
- Requires only the ability to simulate from the model, not to evaluate any analytic density.
- Wilkinson's (2013) result provides a formal theoretical guarantee: ABC-ME is exact under the assumed measurement-error model.
- Naturally integrates with sequential Monte Carlo or regression-adjustment ABC extensions for better efficiency.
- Computationally intensive: each accepted sample requires at least one forward simulation, and acceptance rates can be very low.
- Posterior quality is sensitive to the choice of summary statistics; insufficient statistics discard information and can lead to biased posteriors.
- The tolerance epsilon is difficult to calibrate: too small causes near-zero acceptance rates, too large blurs the posterior.
- Requires the measurement-error distribution to be at least partially specified; misspecified error models propagate into the posterior.
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 ABC-ME differ from standard ABC?
Standard ABC implicitly assumes the observed data are exact; when they are not, accepted parameters may be systematically wrong. ABC-ME treats the observation noise explicitly, either by building it into the forward simulation or by using a tolerance calibrated to the noise level, ensuring the accepted posterior targets the correct distribution under the assumed measurement-error model.
Do I need to know the measurement-error distribution exactly?
You need at least a reasonable approximation. Wilkinson's (2013) result guarantees correctness under the assumed noise model; if the noise model is misspecified, the posterior will be biased by the misspecification. In practice, sensitivity analyses over plausible noise magnitudes are recommended.
How should I choose the tolerance epsilon?
The tolerance must be large enough to accommodate both the ABC approximation error and the variance introduced by measurement noise. A common approach is to set epsilon at a quantile of the distribution of discrepancies observed in pilot runs. Sequential Monte Carlo (SMC-ABC) adaptively reduces epsilon over iterations, which is more principled than a fixed choice.
What summary statistics work best?
Sufficient statistics are ideal but rarely available for complex models. In practice, use domain-knowledge-driven summaries (e.g. mean, variance, autocorrelation at key lags) and validate that they are informative for the parameters of interest. Regression-based post-processing (Beaumont et al. 2002) can reduce the sensitivity to the choice of epsilon.
Can ABC-ME scale to high-dimensional parameter spaces?
Standard ABC scales poorly beyond about 10-15 parameters due to the curse of dimensionality in summary-statistic space. Combining ABC-ME with SMC or using neural posterior estimation (simulation-based inference) can partially address this limitation in higher dimensions.
Sources
- 1.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.
- 2.Beaumont, M. A. (2010). Approximate Bayesian computation in evolution and ecology. Annual Review of Ecology, Evolution, and Systematics, 41, 379-406.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Approximate Bayesian Computation with Measurement Error. ScholarGate. https://scholargate.app/bayesian/approximate-bayesian-computation-with-measurement-error