RMSTpowerBoost Pipeline

Power & sample size planning with Restricted Mean Survival Time (RMST)

This page is the “how-to” for the app. Follow the 3-step workflow below: choose data → map variables & pick a model → run analytical or repeated calculations → export reports.

Inputs

Pilot data (.csv) or simulated data

Models

IPCW · Stratified · GAM · Dependent censoring

Outputs

Power curves · KM plots · Summaries · Reports

Upload / Generate pilot data

Provide a dataset with time, status, arm (plus covariates).

Map columns & choose RMST model

Select the time/event/treatment columns, then pick the model family and τ.

Run analysis & export

Compute power or sample size (analytical or repeated), then download HTML/PDF.

Quick start (the 90‑second path)

Step 1 (left panel): choose Upload and upload a .csv pilot dataset.
When the analysis panel appears, map Time-to-Event, Status (1=event), and Treatment Arm (1=treat).
Choose Target Quantity: Power (enter sample sizes) or Sample Size (enter target power).
Set RMST L (τ) and α. Pick Analytical (fast) or Repeated (simulation-based).
Click Run. Review plots and tables, then download the report.

Data requirements

Your pilot dataset must contain at least three columns that you will map in Step 2: time (positive follow-up time), status (1 = event, 0 = censored), and arm (treatment indicator; 1 = treated). Additional covariates are optional but strongly recommended for adjusted/robust planning.

Recommended coding

time: numeric, in days (or your study unit), strictly > 0.
status: 0/1. If your file uses “Yes/No”, convert to 0/1 before upload.
arm: 0/1 or a 2-level factor. The app treats 1 as the treatment arm.

Common gotchas

Missing values in mapped columns will usually break model fitting.
Non-binary status (e.g., 1/2) needs recoding.
More than two treatment levels: collapse to two arms before upload.

Minimal example schema

Column	Type	Meaning
`time`	numeric	follow-up time
`status`	0/1	1 = event, 0 = censored
`arm`	0/1	1 = treated, 0 = control
`x1, x2, …`	mixed	baseline covariates (optional)
`strata`	factor	needed only for stratified models

Step 1. Data Upload or Generate

Upload pilot data

Use this when you have real study or pilot data and want the app to learn empirical survival/censoring patterns.

Accepted: .csv (the uploader is configured for CSV).
Keep the file “rectangular”: one row per subject, one column per variable.
If you have dates, pre-compute durations before upload.

Generate pilot data (built-in simulator)

Use this when you are prototyping or running scenario planning without a pilot dataset.

Build covariates (continuous/categorical), then set the event-time model (AFT/PH/PWExponential).
Choose allocation ratio, treatment effect, target censoring, and seed (optional).
Click Generate. Data, Summary, and KM Plot tabs refresh on each generate run.
After confirming covariates, panel 1B opens and panel 1A folds (it stays available).

Simulator notes (what the controls mean)

Allocation (a:b): randomized ratio, e.g. 1:1, 2:1.
Treatment β (arm): effect of the treatment indicator in the chosen event-time model (sign interpretation depends on AFT vs PH).
Target censoring: desired censoring proportion; used to calibrate censoring distribution.
Transform (continuous only): applied after generation as (x - a) / b.

Step 2: Map variables & pick the RMST model

Once data exist, the “Model & Analysis” panel becomes available. First map the required columns: Time-to-Event, Status, and Treatment Arm. Some models require additional inputs: stratified models need a Stratification Variable; dependent censoring needs covariates for the censoring Cox model.

Model guide (when to use what)

RMST model	Use when…	Extra mapping
Linear IPCW Model	you want a straightforward adjusted RMST comparison with inverse-probability-of-censoring weights	none
Additive Stratified Model	effects differ by strata and you want additive structure across strata	`strata` required
Multiplicative Stratified Model	you prefer multiplicative structure across strata	`strata` required
Semiparametric (GAM) Model	nonlinear covariate effects are plausible and you want flexible adjustment	none
Dependent Censoring Model	censoring may depend on covariates and you want an explicit censoring model	choose censoring model covariates

Choosing τ (RMST L)

Pick τ within the reliable support of follow-up. Practically: choose τ near a clinically meaningful horizon (e.g., 6/12 months), but not so large that almost everyone is censored before τ. If in doubt, start with a percentile of observed follow-up (e.g., 70–80%).

Step 3: Power vs Sample size, Analytical vs Repeated

Target quantity

Power: provide a comma-separated list of candidate sample sizes (per arm / per stratum).
Sample Size: provide a target power; the app searches for a suitable N.

Calculation method

Analytical: fastest; best for iterative planning and sensitivity analysis.
Repeated: simulation-based; set replications R (≥ 500 recommended) and an optional seed.

What “Repeated” does (high level)

For each candidate sample size, the app repeatedly simulates/bootstraps realizations under the pilot-data mechanism, recomputes the test statistic, and estimates power as the rejection proportion at level α.

Higher R reduces Monte Carlo error (but increases runtime).
Use a seed when you want reproducibility across runs or reports.

Where to look for results

Tabs

Data: current dataset and export tools
Summary: data summary first, then histograms
KM Plot: Kaplan–Meier curves
Analysis: key results, power/sample-size tables, analysis summary, and power plot

Tables & logs

Key results summary
Analysis summary and data summary
Console/run log (what happened, warnings, failures)

Reports

Download HTML (most robust) and PDF (requires LaTeX such as TinyTeX/MiKTeX) after a successful run.

Step 4. Analysis

Run the analysis after mapping inputs. Results populate the plots, tables, and run log, and the report downloads become available.

Troubleshooting (fast fixes)

“Please map Time/Status/Arm column”

Map the correct columns in Step 2. The app will default to the first columns if it can’t detect names like time, status, arm.

Status coding looks wrong (power is nonsense)

Ensure status = 1 indicates the event (not censoring).
If your event indicator is reversed, fix it before upload (or create a new column and map that one).

Dependent censoring model fails

Pick covariates for the censoring Cox model. Avoid including treatment; the app excludes it by design.

PDF report fails to render

Use the HTML report first. If you need PDF, install TinyTeX/MiKTeX and re-run; LaTeX errors will appear in the log.