The analysis of complex leaf survival data

As an alternative to the standard methods of survival analysis, mortality data can be analysed with generalised linear models (GLM). While, for example, testing of nested designs or time-dependent covariates is difficult with the standard methods, the GLM approach offers a powerful tool to analyse mortality data even from complex experiments. To illustrate this, two different GLMs were fitted to leaf census data from a factorial experiment (mowing treatment crossed with fertiliser application) with goldenrod (Solidago altissima). In the first model we fitted a Weibull regression model to survival times which were interpolated from the interval counts. In the second model, the interval counts were directly analysed using the proportion of deaths per time interval. A specific continuous baseline hazard function is used to estimate mortality risk from survival times in the first model, whereas any function of time can be fitted to the mortality risk in the second model. Both analyses indicated that leaf mortality risk increased monotonically with leaf age. The analysis of interval counts, however, also revealed deviations from this pattern that could be attributed to seasonal fluctuations in microclimatic conditions: leaf mortality risk rose distinctly during drought periods but declined during cooler periods with less sunshine. Further the drought effect interacted with one of the two experimental treatments: the increase in leaf mortality was markedly higher in fertilised stands than in unfertilised stands towards the end of a prolonged drought period in July. Such time-dependent( covariates and their interactions with the experimental treatments could not be included in the Weibull regression model (nor in any other standard survival analysis). Considering the advantages, we strongly recommend to use interval counts rather than survival times in the analysis of complex survival data.