Face-based biometrics has become popular in many security applications due to the compromises made between reliability, social acceptance, and privacy. Many high performance face biometric systems now exist which rely on machine learnable features. However, when errors occur these features are hard to decipher in order to adjust for these errors. Through this work a method is presented which identifies nonmatching errors - potential false accepts. Covariates, or metadata values, are then examined with respect to how they interact with score outcomes and influence the likelihood of a false accept occurring. Last, an attempt is made to identify a connection between geometric and color features and score outcomes. Overall, this dissertation shows that the impostor distribution and nonmatching image pairs hold a plethora of information that can be used to improve biometric algorithms and systems.