Comparison of Testing Procedures for Detecting Short and Long Term Effects of Biomarkers for Cancer Screening

The local R package longscreenfun contains all functions used in simulations.

To replicate the results, please use the following steps:

1. Clone the GitHub repository via

   git clone https://github.com/SiddRoy/longbioscreen.git
   cd longbioscreen

2. Install the local R package in longscreenfun/

   Rscript -e 'devtools::install("longscreenfun", dependencies = TRUE)'

   • All exported functions are documented.
3. Run the worked examples below
4. Then run R scripts inside the subfolder to generate the results. As specified in supplement 2 of the simulation, fitting all three models takes less than 10 seconds per biomarker at the $N=600$ sample size.
   • Each subsection from section 3 has a corresponding subfolder to generate results. Folders 36-39 correspond to supplement sections and extra results referred to in the paper.
   • The full set of repetitions were split and implemented using SLURM. Variables Nrep0, n_per, determine the total number of repetitions and the number of repetitions done per split.

Running the simulations

For each aid, run both taskid=1 and taskid=2. The following replicates the conditional model simulations from Section 3.4.

cd 34_CM
for i in {1..250}
do
  Rscript CMsim.R --aid=$i --taskid=1 &
  Rscript CMsim.R --aid=$i --taskid=2 &
  wait
done   

The maximum array IDs and conservative runtime estimates per split used for each simulation are shown below.

Simulation folder	Script	Maximum aid	Total simulation blocks	Runtime per block
31_CP	CPsim.R	10	20	< 70 min
32_JM	JMsim.R	25	50	< 70 min
33_Cox	Coxsim.R	25	50	< 70 min
34_CM	CMsim.R	250	500	< 70 min
35_HDsim	HDsim.R	500	1,000	< 20 min
36_CM_xi2	CMsim.R	250	500	< 70 min
37_HDsim_bigN	HDsim.R	1,000	2,000	< 20 min ($N=2400$)
38_HDsim_rho	HDsim.R	2,000	4,000	< 20 min
39_HDsim_CS	HDsim.R	1,500	3,000	< 20 min

Outputs go to each folder's out/ directory.

For the low-dimensional simulations, the 17 rows are:

• rows 1-3: Cox estimate, standard error, and p-value
• rows 4-6: joint-model estimate, standard error, and p-value
• rows 7-13: conditional-model statistics
• row 14: Cox elapsed time in seconds
• row 15: joint-model elapsed time in seconds
• row 16: conditional-model elapsed time in seconds
• row 17: conditional-model initialization elapsed time in seconds

For the high-dimensional simulations, the 13 rows are:

• row 1: Cox p-values
• row 2: joint-model p-values
• rows 3-9: conditional-model estimates, standard errors, and p-values/statistics
• row 10: Cox elapsed time in seconds
• row 11: joint-model elapsed time in seconds
• row 12: conditional-model elapsed time in seconds
• row 13: conditional-model initialization elapsed time in seconds

Worked examples

Generate and fit one low-dimensional simulated dataset:

library(longscreenfun)

set.seed(1)
dat <- JM_sim_ISCP(N = 600)

cox_screen_pvalue(dat)
jm_screen_pvalue(dat)
jm_screen_pvalue(dat, jm_random_specs = c(b0 = "~1 | i", b01 = "~ tim | i"))
cm_screen_stats(dat)

Generate and analyze one high-dimensional dataset:

library(longscreenfun)

set.seed(1)
dat <- drawY100(th = 1)
prepared <- prepare_biomarker_screen(dat)

cox_screen_biomarkers(prepared, 1:5)
jm_screen_biomarkers(prepared, 1:5)
jm_screen_biomarkers(
  prepared, 1:5,
  jm_random_specs = c(b0 = "~1 | i", b01 = "~ tim | i"),
  jm_stats = c("est", "se", "p")
)
cm_screen_biomarkers(prepared, 1:5)