Application of Survival Analysis and Multistate Modeling to Understand Animal Behavior: Examples from Guide Dogs

Abstract

Epidemiology is the study of patterns of health-related states or events in populations. Statistical models developed for epidemiology could be usefully applied to behavioral states or events. The aim of this study is to present the application of epidemiological statistics to understand animal behavior where discrete outcomes are of interest, using data from guide dogs to illustrate. Specifically, survival analysis and multistate modeling are applied to data on guide dogs comparing dogs that completed training and qualified as a guide dog, to those that were withdrawn from the training program. Survival analysis allows the time to (or between) a binary event(s) and the probability of the event occurring at or beyond a specified time point. Survival analysis, using a Cox proportional hazards model, was used to examine the time taken to withdraw a dog from training. Sex, breed, and other factors affected time to withdrawal. Bitches were withdrawn faster than dogs, Labradors were withdrawn faster, and Labrador × Golden Retrievers slower, than Golden Retriever × Labradors; and dogs not bred by Guide Dogs were withdrawn faster than those bred by Guide Dogs. Multistate modeling (MSM) can be used as an extension of survival analysis to incorporate more than two discrete events or states. Multistate models were used to investigate transitions between states of training to qualification as a guide dog or behavioral withdrawal, and from qualification as a guide dog to behavioral withdrawal. Sex, breed (with purebred Labradors and Golden retrievers differing from F1 crosses), and bred by Guide Dogs or not, effected movements between states. We postulate that survival analysis and MSM could be applied to a wide range of behavioral data and key examples are provided.

Keywords: animal behavior; epidemiology; guide dogs; multistate models; survival analysis.

Figure 1

Neutral image and visual analog scale used to score behavior in questionnaire about 5-month-old potential guide dogs.

Figure 2

Survival curve of the entire guide dogs span for withdrawn from training and working for behavioral reasons (dotted line represents confidence intervals).

Figure 3

Survival curves focused on training phase of guide dogs span, for all dogs exiting training for any reason (black line) and dogs exiting training due to behavioral withdrawals (dotted line represents confidence intervals).

Figure 4

Multistate model of transitions between states of guide dog training and working from retrospective data. The arrows indicate the direction of movements between states and the numbers on the arrows the number of dogs moving between states. The total dog years spent in each state across the dataset is written in each state box, with the mean years per dog in brackets.

Figure 5

Multistate model of transitions between states of guide dog training, from behavioral questionnaire data. The arrows indicate the direction of movements between states and the numbers on the arrows the number of dogs moving between states. The total dog years spent in each state across the dataset is written in each state box, with the mean years per dog in brackets.

Figure 6

Age and date of dogs in guide dog training, with stage and exits indicated from the behavioral questionnaire sample. Each line represents one dog. Red lines show dogs in puppy walking, green lines show dogs in training, black dots show dogs exiting training for behavioral reasons, with the absence of black dots indicating continuation within Guide Dogs.