Summary Statistic Selection in Approximate Bayesian Computation

30 concepts26 papers

In Bayesian inference, we evaluate the likelihood of observing our data given a set of parameters, $p(x \mid \theta)$. But what happens when evaluating this likelihood is mathematically intractable or computationally impossible? This is the domain of simulation-based inference, most notably Approximate Bayesian Computation (ABC).

The core mechanism of ABC is elegantly simple: instead of calculating a likelihood, you sample candidate parameters from a prior, simulate a synthetic dataset from your model, and check if the simulated data matches your real-world observations. If it matches closely enough, you keep the parameter; if it doesn't, you reject it.

However, this simple mechanism hides a profound geometric trap known as the curse of dimensionality. If your raw data consists of thousands of data points, the probability that a simulated dataset will exactly match your observed data is virtually zero. To make matching feasible, researchers must compress high-dimensional raw data into a low-dimensional vector of summary statistics, $S(x)$.

This compression introduces a fundamental, unresolved trade-off that dictates whether an approximate posterior will accurately mirror reality or fail entirely.


The Fundamental Tension: Sufficiency vs. Dimensionality

The performance of ABC hinges entirely on the choice of summary statistics. As Mertens and Voß (2018) and Bharti (2022) articulate, researchers are caught in a strict mathematical tension between information loss and the curse of dimensionality:

  1. Too Many Statistics (The Curse of Dimensionality): If you include too many summary statistics to ensure you don't miss anything, the dimensionality of $S(x)$ explodes. In high-dimensional spaces, distance metrics lose their contrast—every simulated point becomes roughly equidistant from the observed data. As a result, you require an impractical, often impossible number of simulations to find matches within a reasonable tolerance.
  2. Too Few Statistics (Information Loss): To make matching efficient, you might compress the data down to just a few basic metrics (like the mean or variance). However, if these statistics are not sufficient for the parameters of interest, you discard critical information.

When summary statistics lack sufficiency, the true data-generating process cannot be uniquely identified. As Filippi et al. (2013) demonstrate, using non-sufficient statistics directly inflates credible intervals and produces misleading, overly broad posterior distributions.

[High-Dimensional Data] ───(Compression)───► [Summary Statistics S(x)]
                                                      │
             ┌────────────────────────────────────────┴────────────────────────────────────────┐
             ▼                                                                                 ▼
   Too Many Statistics                                                                Too Few Statistics
   (High Dimensionality)                                                             (Information Loss)
   • Distance metrics fail.                                                          • Discards critical information.
   • Simulations rarely match.                                                       • Non-sufficient for parameters.
   • Requires infinite simulation budget.                                            • Inflated credible intervals & bias.

The severity of this choice is underscored by Chaudhuri (2024) and Ulzega et al. (2016), who note that standard ABC methods remain highly sensitive to this manual selection process, which often operates as an unsolved, ad-hoc challenge in practical applications. If your choice of statistics fails to capture the underlying data-generating process, the entire inference pipeline breaks down, yielding biased posteriors (Cornuet et al., 2009; Zuluaga et al., 2015).


The Higher Hurdle of Model Selection

The challenge intensifies when moving from estimating parameters within a single model to performing model selection (determining which of several competing models generated the data).

A common misconception is that a summary statistic that is sufficient for parameter estimation within a model will also be sufficient for discriminating between different models. As Didelot et al. (2011) and Filippi et al. (2013) point out, this is rarely true.

For instance, a statistic that perfectly captures the scale of a distribution within Model A might completely fail to capture the structural, qualitative differences that separate Model A from Model B. Fearnhead et al. (2013) show that without explicit mathematical sufficiency for the model choice task itself, the reduced data will lack the information necessary to correctly identify the true model structure, rendering approximate model comparisons highly unreliable.


Emerging Pathways to Sufficiency

To move away from arbitrary, hand-crafted statistics, recent literature focuses on automating the extraction of low-dimensional, highly informative representations.

Neural Information Maximization

Instead of relying on human intuition, modern frameworks use deep learning to learn optimal representations directly from simulated data. Chen et al. (2021) frame the search for summary statistics as a mutual information maximization problem. By training neural networks to maximize the information between the compressed data representations and the underlying parameters, they can automatically discover approximate sufficient statistics. Similarly, Dyer et al. (2022) leverage Graph Neural Networks (GNNs) to automatically ingest complex, relational microdata from agent-based models, bypassing hand-crafted summaries entirely and avoiding catastrophic information loss.

Component-wise and Copula Factorizations

When dealing with high-dimensional parameter spaces, finding a single joint summary statistic vector is daunting. To combat this, Clarté et al. (2020) propose a component-wise approach (ABC-Gibbs), breaking the inference down into smaller steps where low-dimensional summary statistics only need to be sufficient for individual parameter components. In a similar spirit of decomposition, Li et al. (2017) utilize Gaussian copulas to approximate complex high-dimensional posteriors by focusing on simpler, lower-dimensional bivariate relationships among summary statistics.

Summary-Free Discrepancies

An alternative paradigm seeks to bypass summary statistics altogether by changing how data distributions are compared. Rather than reducing data to a vector and computing a Euclidean distance, researchers use global discrepancy measures. For example, Legramanti and Durante (2025) analyze summary-free ABC methods using Rademacher complexity to establish uniform concentration bounds for posteriors based directly on data discrepancies.

Human-in-the-Loop Optimization

When simulation budgets are tight and automated methods lack data to converge, researchers have introduced active learning frameworks. Bharti (2022) developed a sequential Bayesian experimental design framework that puts a domain expert in the loop, using human intuition to guide the active selection of summary statistics and protect the model against catastrophic failures caused by misspecification.

Ultimately, whether through neural compression, structural decomposition, or summary-free discrepancies, solving the summary statistic selection problem is what transforms ABC from a heuristic matching game into a mathematically rigorous engine for scientific discovery.

Key Foundational Work

Adaptive approximate Bayesian computation (2009) — contributed 1 concept elements to this direction.
COMPUTER PROGRAMS: <scp>on</scp>e<scp>samp</scp>: a program to estimate effective population size using approximate Bayesian computation (2007) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_summary_statistic_selection_sensitivity__fb70c790,
  title = {summary statistic selection (sensitivity to summary statistics)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/summary-statistic-selection-sensitivity-to-summary-statistics/rdr_fb70c79030f44babb74c9acf41d8a04d},
  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_summary_statistic_selection_sensitivity__fb70c790_v1,
  title = {summary statistic selection (sensitivity to summary statistics)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/summary-statistic-selection-sensitivity-to-summary-statistics/rdr_fb70c79030f44babb74c9acf41d8a04d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake