The Practical Application of a Frequentist and Bayesian Method to Forecast Unwell-being in Hospice Patients

Background/aims: In hospice care, prediction models using longitudinal
data must take survival into account to prevent forecasting beyond
death. A joint model combines the linear mixed effects model with timeto-event model.
The aim of this study is to evaluate the practical application of a frequentist and a Bayesian approach to joint modelling for the prediction of
future unwell being of hospice patients.
Methods: A cross-sectional design was employed to assess the practical
application of both frequentist and Bayesian approaches to joint
modeling.
The primary outcome is practical application defined as:
1. Processing Time: Measured in minutes and seconds per imputation and analysis.
2. Congeniality: Assessing the appropriateness between imputation and analysis.
3. Software Implementation: Functionality within statistical
software.
Results: The Bayesian approach demonstrated a tenfold increase in computation time compared to the frequentist approach. Under the frequentist approach, multiple imputation and analysis were not congenial.
Furthermore, additional programming was necessary for multiple imputation, linear mixed-effect modeling, and joint modeling. In contrast, the
Bayesian approach required only a specific statement on covariates used
for imputation.
Both methods have yet to be fully integrated into current statistical software packages. The frequentist approach faces limitations: the multiple
imputation of multilevel data is not universally feasible across all specifications, and the analysis must be performed manually per imputed dataset and pooled afterwards. For the Bayesian approach the application of
the developed prediction model for other data is not yet implemented.
Conclusions: The Bayesian approach, despite its lengthy computational
requirements, better suits joint model analysis than the frequentist
approach. Further implementation of the Bayesian approach in statistical software is necessary to enable the validation and application of the
joint model for predicting future patients’ unwell-being