[Paper Review] Non parametric Bayesian approach to LR assessment in case of rare haplotype match

Giulia Cereda|arXiv (Cornell University)|Jun 28, 2015

Bayesian Methods and Mixture Models11 references4 citations

TL;DR

This paper proposes a nonparametric Bayesian method using the two-parameter Poisson-Dirichlet distribution to estimate the likelihood ratio (LR) in forensic DNA analysis when a suspect's Y-STR profile is absent from the reference database—a rare type match. By modeling population profile frequencies nonparametrically and discarding profile identity information, the method provides a robust LR estimate that outperforms naive or add-constant estimators, especially in low-sample regimes, with asymptotic normality enabling efficient computation via MLE plug-in approximation.

ABSTRACT

The evaluation of a match between the DNA profile of a stain found on a crime scene and that of a suspect (previously identified) involves the use of the unknown parameter $p=(p_1, p_2, ...)$, (the ordered vector which represents the proportions of the different DNA profiles in the population of potential donors) and the names of the different DNA types. We propose a Bayesian non parametric method which considers $P$ as a random variable distributed according to the two-parameter Poisson Dirichlet distribution, and discard information about names of DNA types. The ultimate goal of this model is to evaluate DNA matches in the rare type case, that is the situation in which the suspect's profile, matching the crime stain profile, is not one of those in the database of reference.

Motivation & Objective

• Address the fundamental challenge in forensic genetics where a suspect’s DNA profile is absent from reference databases, making standard frequency-based likelihood ratio (LR) estimation unreliable.
• Overcome the limitations of empirical frequency estimators (e.g., naive or add-constant) in rare type match scenarios, particularly for Y-STR profiles with vast potential haplotype diversity.
• Develop a Bayesian nonparametric model that treats the infinite-dimensional vector of population profile frequencies as a random variable governed by the two-parameter Poisson-Dirichlet distribution.
• Improve LR estimation accuracy by discarding profile identity information, reducing nuisance parameters and enhancing precision in high-dimensional, sparse data settings.
• Provide a computationally feasible approximation to the true LR using maximum likelihood estimates of hyperparameters, leveraging asymptotic normality of the log-likelihood.

Proposed method

• Model the unknown population profile frequencies as a random probability vector P following a two-parameter Poisson-Dirichlet distribution (PD(α, θ)), allowing for infinite, unknown types.
• Use a hierarchical prior with hyperpriors on the concentration parameter θ and the discount parameter α, enabling flexible modeling of population diversity and rare type probabilities.
• Reduce the observed data by discarding profile names, retaining only the count of distinct profiles and their frequencies, thus simplifying inference while preserving key probabilistic structure.
• Represent the model via a Chinese restaurant process (CRP) for intuitive understanding of the random partitioning of profiles, facilitating simulation and theoretical analysis.
• Derive the LR using a novel theoretical result that allows computation under agreement on part of the data distribution, enabling elegant derivation of the posterior odds ratio.
• Approximate the LR using maximum likelihood estimates (MLE) of α and θ, and apply a transformation φ = n(1−α)/(n+1+θ) to achieve asymptotic Gaussianity, enabling efficient MCMC or analytical approximation.

Experimental results

Research questions

• RQ1 How can the likelihood ratio be reliably estimated in the rare type match problem when the suspect’s Y-STR profile is absent from the reference database?
• RQ2 Can a nonparametric Bayesian model with a Poisson-Dirichlet prior provide a more accurate and robust LR estimate than traditional empirical or add-constant estimators in sparse data settings?
• RQ3 What is the impact of discarding profile identity information on the precision and accuracy of LR estimation in high-dimensional, low-coverage forensic data?
• RQ4 Under what conditions does the MLE of the hyperparameters α and θ yield a consistent and efficient approximation to the true LR when the true profile frequencies are unknown?
• RQ5 How does the asymptotic normality of the log-likelihood function for the transformed parameters φ and θ support practical computation and uncertainty quantification in forensic LR assessment?

Key findings

• The proposed method significantly reduces estimation error in rare type matches compared to naive or add-constant estimators, especially when database sizes are small.
• The posterior distribution of the transformed parameters (φ, θ) is approximately Gaussian, enabling efficient approximation of the expected LR via MLE plug-in: LR ≈ (n + 1 + θ_MLE)/(1 − α_MLE).
• Experiments on a real Y-STR database (Purps et al., 2014) show that the model fits the data well, with the log-likelihood function exhibiting power-law behavior consistent with the PD(α, θ) model.
• In controlled simulations using synthetic data from a true PD(α, θ) process, the method’s LR estimates closely track the true LR (LR|p), with minimal bias when hyperparameters are correctly estimated.
• The error between the true LR (when p is known) and the estimated LR (when p is unknown) is minimized when both model and parameter estimation errors are controlled, demonstrating robustness.
• The MLE of α is found to be consistent under certain regularity conditions, and the observed Fisher information for α grows with sample size, suggesting potential for consistent estimation of α even in high-dimensional settings.