Events and identity in probabilistic models of legal evidence

• Pages: 157-180
• Date: 01 July 2026
• Published date: 01 July 2026
• Author: Marcello Di Bello
• Subject Matter: Derecho Procesal

Quaestio facti. Revista Internacional sobre Razonamiento Probatorio / International Journal on Evidential Legal Reasoning

Año 2026 11 pp. 157-180 DOI: 10.33115/udg_bib/qf.i11.23278

Quaestio facti. Revista Internacional sobre Razonamiento Probatorio

Quaestio facti. International Journal on Evidential Legal Reasoning

Sección: Conjeturas y refutaciones

2026 l 11 pp. 157-180

Madrid, 2026

DOI: 10.33115/udg_bib/qf.i11.23278

Marcial Pons Ediciones Jurídicas y Sociales

© Marcello Di Bello

ISSN: 2604-6202

Recibido 5/5/26 | Aceptado 25/5/26 | Publicado online: __/__/2026

Editado bajo licencia Reconocimiento 4.0 Internacional de Creative Commons

EVENTS AND IDENTITY IN PROBABILISTIC MODELS

OF LEGAL EVIDENCE

Marcello Di Bello

Arizona State University

ABSTRACT: is paper examines whether Bayesian networks are expressive enough to model reason-

ing with evidence in legal cases. Bayesian networks can represent many familiar patterns of eviden-

tial reasoning, including inferences from evidence to hypotheses, cumulative support from multi-

ple items of evidence, and chains of inferences linking intermediate hypotheses to ultimate guilt.

But other evidential inferences common in legal cases are more dicult to model. Focusing on a

real criminal case, the paper distinguishes between identity- and event-level inferences. Event-level

inferences show why certain actions amount to guilty conduct, while identity-level inferences link

the defendant to those actions. e challenge for Bayesian network models of legal evidence is to

represent how identity- and event-level inferences combine and reinforce one another. Meeting

this challenge requires extending Bayesian networks beyond a purely propositional language.

KEYWORDS: bayesian networks; likelihood ratio; probability; criminal law.

1. INTRODUCTION

Can the evidence in a legal case be modeled probabilistically? Following Mackor

(2026), this question is best divided into more specic subquestions. One concerns

the target of the probabilistic modeling: are we asking whether individual items of

evidence can be modeled probabilistically, or is the question about modeling the

totality of the evidence in a case? Another subquestion concerns who is supposed to

carry out the probabilistic analysis. Should the analysis be assigned to experts who

158 MARCELLO DI BELLO

Quaestio facti. Revista Internacional sobre Razonamiento Probatorio / International Journal on Evidential Legal Reasoning

Año 2026 11 pp. 157-180 DOI: 10.33115/udg_bib/qf.i11.23278

possess domain-specic knowledge, or should it remain with judges and jurors, per-

haps with the assistance of court-appointed experts?

ere is little doubt that individual items of evidence can be—and often are—

evaluated probabilistically by experts. e most familiar examples come from foren-

sic science (Taroni et al., 2014). Suppose that genetic material matching the defend-

ant is found at the crime scene. A forensic expert will typically oer the following

probabilistic analysis of the match: they will assess how probable that nding would

be if the defendant were the source of the genetic material, as compared with how

probable it would be if someone else were the source. Expressed as a ratio, this com-

parison is known as the likelihood ratio:

P ( match | defendant is the source)

P ( match | defendant is not the source)

is ratio quanties the degree of support that the match evidence provides for

the source hypothesis: the greater the ratio, the stronger the support.

Forensic experts have also begun to apply probabilistic analyses beyond source-lev-

el propositions to activity-level propositions (Taylor et al., 2018). e question of

interest is often not merely whether the defendant is the source of a trace recovered

at the crime scene, but how the trace came to be there, for instance, whether it got

there during a violent confrontation between the defendant and the victim. How to

carry out a probabilistic analysis of trace evidence for activity-level propositions is a

topic of ongoing discussion in the literature (Stacey et al., 2025).

A typical legal case, however, will consist of many pieces of evidence bearing on

dierent propositions, not a single item of evidence bearing on a single proposition,

such as a source- or activity-level hypothesis. So it is natural to ask whether the to-

tality of the evidence in a legal case can be modeled probabilistically. at is a more

complicated question. Which also makes the institutional question—who should

carry out the probabilistic analysis?—more complicated. Experts may have the tech-

nical competence to build probabilistic models of individual items of evidence, but

they lack the institutional authority to model a legal case as a whole. Judges and

jurors, by contrast, have the authority to assess the totality of the evidence—that is

what they are tasked with doing at trial—but they usually lack the technical compe-

tence required to construct a probabilistic model.

In this paper, I am going to set aside the institutional question. I will focus in-

stead on whether it is in principle possible to model a case as a whole in probabilistic

terms. A number of probabilistic models of entire legal cases have been developed.

ese models all rely on Bayesian networks (more on these soon). Examples include

analyses of the Sacco and Vanzetti case (Kadane & Schum, 1996), the Anjum mur-

ders (Vlek et al., 2014), the Simonshaven case (Fenton et al., 2020), and a recent su-

permarket robbery case in the Netherlands (Hampson & Leeuwen, 2025). ese ex-

amples suggest that probabilistic methods can be extended beyond individual items