FROM PUPPY TO PARTNER: IDENTIFYING EARLY PREDICTORS OF GUIDE DOG QUALIFICATION OUTCOMES

Abstract

This project analyzes data from the South African Guide Dogs Association (2014–2025) to identify the characteristics that influence whether a dog successfully graduates into guiding or service work, is selected for breeding, or is withdrawn from the program. The original dataset included 866 dogs across multiple breeds and mixed broods. To narrow the research question and ensure stronger internal validity, a pivot table was used to isolate purebred Labrador Retrievers, the most common breed in guide dog programs, resulting in a final analytical sample of 466 dogs. Focusing on a single breed reduces confounding effects, increases model consistency, and allows clearer inference about predictors of success. Guide dog training is costly and resource-intensive, making early predictions of outcomes essential for effective breeder selection, program planning, and animal welfare. The dataset contains demographic, developmental, radiographic, temperament, and early training indicators, along with a five-category outcome variable. Extensive data cleaning was completed, including handling missing values, removing redundant variables, resolving near-perfect collinearity, and constructing composite hip and elbow scores. Exploratory data analysis assessed distributions, feature correlations, and outcome associations. Two modeling approaches were then implemented: a standard multinomial logistic regression (MLR) for interpretability and a regularized Elastic Net multinomial model to improve predictive accuracy and reduce overfitting. Results indicate that radiographic health scores, birth-year cohort effects, and early developmental measures are among the strongest predictors of outcome categories. The MLR model revealed clear relationships between orthopedic indicators and withdrawal, while the Elastic Net model produced a more stable set of predictors with superior model fit. These findings provide actionable insights for breeding, early screening, and resource allocation within guide dog programs. Future work will extend these models using larger, multi-breed datasets and longitudinal behavioral measures.