An R package for analyzing survival using continuous-time open capture-recapture models

Capture–recapture software packages have proven to be very powerful tools for analysing factors affecting survival in wild populations. However, all such packages are limited to discrete-time protocols. Appropriate survival analysis tools are still lacking for data acquired from continuous-time protocols. We have developed a statistical method and propose an r package for analysing such data based on an extension of classical survival analysis models incorporating an inhomogeneous Poisson process for modelling capture histories. First, data were simulated from a continuous-time protocol. These data were used to (i) compare survival estimation biases of discrete- and continuous-time approaches and (ii) investigate the performance and accuracy of our r package for four types of covariates: factors varying between individuals (like sex), in time (like climatic factors), both in time and between individuals (like physical condition) and age (as a categorical factor). Secondly, the r package has been applied to a real data set for survival analysis of cats in the Kerguelen archipelago (regrouping 682 cats over 20 years) as an illustrative example. Results of the simulated data analysis show that the method performs better than its discrete-time counterpart for analysing data acquired from continuous-time protocols. It provides unbiased parameter estimates for all parameters except those that vary both in time and between individuals – which is not surprising, since in our case, these factors were not updated in continuous time (i.e. only upon capture). When applied to the Kerguelen cat data set, the results suggest that survival is lower in juveniles than in adults and subadults, varies between study sites and increases with physical condition, and this latter effect being more important in females than in males. Sex, season, temporal linear trend in survival and the NDVI vegetation index were also tested but were not found to be significant. However, confidence intervals were too large (due to a low recapture rate) for excluding such effects. Further analyses are still needed for rigorous covariate testing in this context.
In conclusion, continuous-time approaches – such as that presented in this paper – should be preferred when data acquired from continuous-time protocols is analysed.