Analysis of Continuous and Discrete Time-to-Event Data Using Parametric Techniques

In this paper, the effect of discretization of time-to-event data on parameter estimates is investigated with the objective of finding out how discretization of nearly continuous or continuous survival data affects the outcome of the parameter estimates. Monte Carlo simulation was used to simulate data with different sample sizesfor the study. Discretisation of the simulated data was made. The parameters of the Weibull and the exponentially distributed models were estimated using maximum likelihood estimation techniques with the help of Davidon-fletcher-Powell optimization formula in MATLAB program for both the continuous and the discretized data. Using the two-sample Kolmogorov-Smirnov test, the hypothesis that the discrete and continuous samples come from the population with the same distribution could not be rejected for samples with sizes of less than 100 but rejected for sample of more than 100 sample sizes. It was also found out that discretization of survival data reduces their precision by increasing the parameter estimates. Researchers studying time-to-event data are therefore advised to avoid over discretization in order to reduce biasedness in the parameter estimates. Smaller counting units should be expressed as a proportion of the bigger counting unit used and in the event that there is no event in a given interval, they should resort into interpolation to find the missing value.