Likelihood Free Inference for Complex Mechanistic Models

This direction is merged into another direction: likelihood-free inference (Approximate Bayesian Computation). Its last version stays readable and citable here.

Version 1 of 157 papers60 conceptsfrozen 2026-10-04

In fields ranging from epidemiology to evolutionary biology, researchers often build "mechanistic models"—simulations that describe the step-by-step rules of a system. Whether it is the spread of a virus through a city or the growth of a tumor, these models are powerful because they are interpretable. However, they share a common flaw: they are notoriously difficult to "fit" to real-world data.

Standard statistical methods rely on a likelihood function, a mathematical formula that calculates the probability of observing specific data given a set of parameters. For complex simulations—like the agent-based models used by Xiaoyu Wang and Christopher Drovandi to study tumor growth—this function is "intractable." It is either impossible to write down or too computationally expensive to solve.

Likelihood-Free Inference (LFI), primarily through a technique called Approximate Bayesian Computation (ABC), bypasses this bottleneck by replacing formal equations with a simple "simulate-and-compare" loop.

The Intuition: Learning by Trial and Error

The core logic of ABC is intuitive: if a model is a good representation of reality, then the "correct" parameters should produce simulations that look like the real world.

Instead of calculating a likelihood, the ABC algorithm follows three steps:

  1. Sample: Pick a candidate set of parameters from a prior distribution.
  2. Simulate: Run the mechanistic model using those parameters to generate synthetic data.
  3. Compare: Check if the synthetic data is "close enough" to the observed data. If it is, keep the parameters; if not, reject them.

By repeating this millions of times, researchers like Yuval Itan and Mark Thomas have been able to reconstruct historical events, such as the origins of lactase persistence in Europe, by finding which ancestral migration patterns most likely produced the genetic diversity we see today.

Summarizing the Messy Reality

Comparing raw, high-dimensional data (like every individual's location in a pandemic) is computationally impossible. To make the comparison feasible, ABC uses summary statistics—low-dimensional "fingerprints" of the data.

For example, when Roxana Zeraati and Tatiana A. Engel developed a framework for estimating neural timescales, they didn't compare raw voltage traces. Instead, they compared the autocorrelation and power spectra of the simulated and observed signals. Similarly, Pieter Libin used the Hellinger distance to quantify the similarity between viral load distributions in HIV transmission models.

The choice of a "tolerance" threshold ($\epsilon$) is the final piece of the puzzle. If $\epsilon$ is too large, we accept bad parameters; if it is too small, we might never find a match. This trade-off defines the "approximate" nature of ABC.

Scaling Up with Sequential Monte Carlo

Simple "rejection sampling" is inefficient because most simulations are discarded. To solve this, researchers use Sequential Monte Carlo (SMC-ABC). Instead of sampling blindly, the algorithm maintains a "population" of parameter sets (particles) and iteratively refines them.

In each round, the tolerance $\epsilon$ is lowered, and the particles are "nudged" toward regions of the parameter space that produce better matches. David J. Warne and Antonietta Mira applied this to COVID-19 response data, allowing them to estimate how testing breadth and response speed varied across 158 countries in real-time.

Robustness vs. Speed

While newer "neural" simulation-based inference methods use deep learning to speed up the process, traditional ABC remains a gold standard for robustness. As Ryan P. Kelly and Adrianne L. Jenner argue in their guide to simulation-based inference, ABC is often better at handling "out-of-distribution" data—the messy, unpredictable noise common in biological systems.

Furthermore, the framework is not limited to Bayesian contexts. David J. Warne and Oliver J. Maclaren have extended these ideas into a frequentist framework called Generalised Likelihood Profiles, proving that the "simulate-and-compare" philosophy is a fundamental tool for modern science. By treating the simulation itself as the likelihood, researchers can finally bridge the gap between complex mechanistic theories and the empirical data that tests them.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_likelihood_free_inference_for_complex_me_4cbe6947,
  title = {Likelihood Free Inference for Complex Mechanistic Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/likelihood-free-inference-for-complex-mechanistic-models/rdr_4cbe6947341d4172900bf1fccbba000f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_likelihood_free_inference_for_complex_me_4cbe6947_v1,
  title = {Likelihood Free Inference for Complex Mechanistic Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/likelihood-free-inference-for-complex-mechanistic-models/rdr_4cbe6947341d4172900bf1fccbba000f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}