This repository contains the validation tests for the CompAREdesign R package, specifically for evaluating the functionality and accuracy of the time-to-event functions.
This repository includes the following folders:
data: The output of the validation tests in Rdata formatfigures: The figures resulting for the validation tests.scripts: The R scripts used to perform the validation tests.table: The table summarizing the test results in a structured format (CSV file).
The validation tests cover 72,576 different scenarios, systematically evaluating the accuracy of the time-to-event functions implemented in the package. The tests focus on:
- Functionality verification across different parameter settings.
- Identification of numerical instability issues.
- Execution time assessment.
- Minor precision corrections were required to address numerical errors caused by values close to zero in certain denominators.
- The functions proved robust, with only 0.1% of cases yielding unstable results.
- The total execution time for all scenarios was 52.67 hours.
Next, we will provide the graphical outputs of the test. The complete results are in table/validation_results.csv
Asymptotic Relative Efficiency (ARE) according to probabilities of observing the event.
- As
$p0_e1$ increases the ARE decreases. This is logical because the more prevalent the first endpoint is, the less is the need of the second component. - The ARE takes greater values when the value of the
$p0_e2$ is high.
ARE according to the cause-specific HRs of the components.
- ARE is almost always greater than 1 when
$HR_e2<HR_e1$ - ARE is usually greater than 1 when
$HR_e2=HR_e1$
ARE according to shape parameters of the Weibull distributions for the time to event of the components.
- Shape parameters of the weibull distribution for the time to event for the components have less impact than other input parameters on the ARE.
ARE according to the degree/type/measure of association between components.
- Association structures have less impact than other input parameters on the ARE.
Sample size of the composite endpoint (
- As the probabilities of observing the event (
$p0_e1$ ,$p0_e2$ ) increase, the sample size for the design with composite endpoint ($SS_CE$ ) decreases. - It seems that there are some unstable results for large probabilities of observing the first component (
$p0_e1 \approx 0.9$ )
SS_CE according to the cause-specific HRs of the components.
- For moderate effect sizes, that is close to 1 (
$HR \approx 0.9$ ) the$SS_CE$ is larger.
SS_CE according to shape parameters of the Weibull distributions for the time to event of the components.
- Shape parameters of the weibull distribution for the time to event for the components have less impact than other input parameters on the sample size.
- For
$beta_\e1=0.5$ , there are larger sample sizes.
SS_CE according to the degree/type/measure of association between components.
- Association structures have less impact than other input parameters on the sample size.
- For the combination of
$tau=0.7$ $rho_type='Kendall'$ and$copula='Clayton'$ , there are larger/unstable sample sizes.
Geometric Average Hazard Ratio (gAHR) according to probabilities of observing the event.
- gAHR is almost always below 1.
- Unstable results for large probabilities of observing the event.
gAHR according to the cause-specific HRs of the components.
- When the HR of both components is the same, The gAHR usually matches that value.
gAHR according to shape parameters of the Weibull distributions for the time to event of the components.
gAHR according to the degree/type/measure of association between components.
- Clayton copula provides more unstable results than the other two copulas.
To reproduce the tests, clone this repository and run the scripts in the scripts/ directory.