Application possibilities of survival analysis for time-to-event data in animal science

Abstract

The application efficiency of several statistical methods was tested based on an open field behavior test of mice. The examined trait was the duration time until the animals approached the experimenter’s hand. The available time was fixed in 300 seconds. There were monitored 80 mice belonging to two different species of the Mus genus in equal proportion. Besides, male and female and young and adult animals have also represented the evaluated groups in equal proportions. The data of the examined trait was analyzed with Generalized Linear Models, Kaplan-Meier survival curves, and with Cox Proportional hazard model. The applied statistical procedure provided completely discordant results. According to the GLM results none of the examined factors (species, sex, and age) had significant effects on the examined variable. On the contrary, all factors proved to be significant using a procedure based on the survival analysis. Kaplan-Meier survival curves indicated a higher proportion of individuals successfully approaching the experimenter’s hand in all of the compared groups representing different species sexes and ages, respectively. The estimated Cox regression coefficients were significant indicating the significant effects of the species, sex, and age on the investigated trait. Based on the estimated Hazard ratios the probability that in the next time unit a successful approach of the experimenter’s hand would occur is three times more likely for one species than the other, twice as much for males and the juveniles than for the females and for the adults. Based on the present study it could be concluded that the successful approach of the experimenter’s hand by the mice is clearly “time to event data” thus it is suitable to be analyzed with survival analysis procedures. It could be concluded that applying conventional GLM was not adequate because due to the lack of the successful approach and approach times the censored data should not be used and thus the sample size would largely be reduced.