Sampling variability in forensic likelihood-ratio computation: A simulation study.

Area of Science:

• Forensic Science
• Biometrics
• Pattern Recognition

Background:

• Forensic biometrics increasingly uses likelihood-ratios (LR) to evaluate evidence.
• LR methods convert biometric system scores into a measure of evidence strength.
• Score-to-LR conversion is sensitive to the training data used.

Purpose of the Study:

• To empirically quantify the impact of sampling variability on LR computation methods.
• To investigate how training set size and score distribution shape affect LR accuracy.
• To assess the performance of LR methods across different score values.

Main Methods:

• Developed a simulation framework to model score distributions and sampling variability.
• Varied probability density function shapes, training set sizes, and target scores.
• Empirically evaluated common LR computation methods under these varying conditions.

Main Results:

• Sampling variability in training data can considerably affect LR computation outcomes.
• The impact of variability is more significant with smaller training datasets.
• LR computation methods exhibit performance variations depending on score distribution and the specific score being evaluated.

Conclusions:

• Sampling variability is a critical factor in the reliability of forensic biometric LR calculations.
• Careful consideration of training data characteristics is essential for robust LR estimation.
• The choice of LR method and the evaluated score significantly influence results.