TimeDepFrail: Time-Dependent Shared Frailty Cox Models in R

TimeDepFrail is the ultimate R package for fitting and analyzing Time-Dependent Shared Frailty Cox Models. These models extend the traditional Shared (Gamma) Frailty Cox Models by incorporating a time-dependent frailty component, making it a robust tool for studying how unexplained heterogeneity in data evolves over time.

This package implements the methods discussed in “Centre-Effect on Survival After Bone Marrow Transplantation: Application of Time-Dependent Frailty Models” by C.M. Wintrebert et al. (2004).

Installation

You can install the development version of the package from GitHub:

{r, eval=FALSE} devtools::install_github("alessandragni/TimeDepFrail")

Dataset data_dropout

The data_dropout dataset is used to exemplify the package. It tracks the academic progress of students enrolled in 2012 over three academic years (six semesters). This dataset aims to explore the factors leading to student dropout.

The dataset is composed of four variables: - Gender: Categorical covariate indicating gender (Male or Female). - CFUP: Numeric covariate representing the standardized number of credits or CFUs (Credito Formativo Universitario) passed by the student in the first semester. - time_to_event: The time (in semesters) when a student decides to drop out. A value greater than 6.0 means the student did not drop out during the follow-up period. - group: Categorical variable representing the student’s course of study, with 16 levels from CosA to CosP.

Students are followed for a maximum of 6 semesters (3 academic years), from the start of lectures until they drop out or the follow-up ends.

Model execution

To fit a Time-Dependent Shared Frailty model, the following elements are required: - dataset as data.frame, e.g. data_dropout - time_axis vector: The time intervals for which the model is applied. For example, in the data_dropout dataset, no events occur in the first semester, so the time_axis starts at the end of the first semester (t = 1) and ends at the end of the third year (t = 6). - categories_range_min and categories_range_max vectors: Provide minimum (categories_range_min) and maximum (categories_range_max) bounds for each parameter category to constrain the optimization. - formula object: Specify the relationship between time-to-event, covariates, and group. For the clustering variable (group), it must be provided as cluster(group) in the formula.

Once these elements are prepared, you can call the desired model using the AdPaikModel() function. While PowParModel() and StocTimeDepModel() are also available, they are secondary models with room for performance improvements.

For full examples, refer to the Examples/ModelsApplication.R script.

Additionally, for guidance on selecting model parameters such as time_axis, categories_range_min and categories_range_max, we recommend basing these choices on insights gained after fitting a Time-Unvarying Shared Frailty model. You can find a relevant example in the ExamplesTimeUnvarying.R.

Analyzing results

Several built-in methods are available to analyze the results of the fitted model: - Baseline Hazard Step-Function: plot_bas_hazard() - Frailty Standard Deviation/Variance: plot_frailty_sd() - Posterior Frailty Estimates: plot_post_frailty_est() - Model Summary: summary()

These methods provide insightful visualizations and summaries to help you interpret your model results effectively.

Furthermore, also a support function suitable for the choice of the range of parameters and analysis of the 1D log-likelihood is available, AdPaik_1D().

To be aware of

• The AdPaikModel model is optimized for fast computation although estimating certain coefficients (e.g., Male versus Female) may vary slightly in computational time. Note that changing the reference category (e.g., using Male as the baseline) alters the coefficient estimates but not the overall log-likelihood or model fit. Users should choose reference categories based on interpretability rather than performance.
• The ‘Centre-Specific Frailty Model with Power Parameter’ (PowParModel) is slower than the AdPaikModel, but it produces coherent and expected results.
• The ‘Stochastic Time-Dependent’ Centre-Specific Frailty Model’ (StocTimeDepModel) is computationally heavy and may not converge easily.

Authors and maintainers of the code

Alessandra Ragni (alessandra.ragni@polimi.it), Giulia Romani (giulia.romani@mail.polimi.it), Chiara Masci (chiara.masci@polimi.it).