Codex/cran 0.0.1 review

.github/workflows/check-reproducible.yaml

@@ -11,6 +11,11 @@ jobs:
   R-CMD-check-renv:
     runs-on: ubuntu-latest
 
+    env:
+      GITHUB_PAT: ${{ secrets.GITHUB_TOKEN }}
+      R_KEEP_PKG_SOURCE: yes
+      RENV_CONFIG_REPOS_OVERRIDE: https://cloud.r-project.org
+
     steps:
       - uses: actions/checkout@v5
 
@@ -19,7 +24,10 @@ jobs:
       - uses: r-lib/actions/setup-r@v2
         with:
           r-version: "4.4.2"
-          use-public-rspm: true
+          # renv is loaded from .Rprofile before setup-renv can install it.
+          # Use CRAN directly so that bootstrap does not depend on a
+          # platform-specific Posit Package Manager endpoint.
+          use-public-rspm: false
 
       - uses: r-lib/actions/setup-renv@v2
 

DESCRIPTION

@@ -1,34 +1,36 @@
 Package: clonecensorweighting
-Title: Infrastructure for Clone-Censor-Weighting Analyses
+Title: Clone-Censor-Weight Methods for Target Trial Emulation
 Version: 0.0.0.9000
 Authors@R: c(
     person("Sang Ho", "Park", email = "shstat1729@gmail.com", role = c("aut", "cre")),
     person("Youngrok", "Lee", role = "aut")
     )
-Author: Sang Ho Park [aut, cre],
-    Youngrok Lee [aut]
-Maintainer: Sang Ho Park <shstat1729@gmail.com>
-Description: Provides a starter package for clone-censor-weighting workflows
-    used in target trial emulation. The package currently includes data
-    ingestion helpers, a minimal data cloning routine, and utilities for
-    constructing survival responses.
+Description: Implements clone-censor-weight analyses for emulating target
+    trials with grace-period treatment strategies using observational
+    time-to-event data. Provides tools to clone subjects across strategies,
+    apply strategy-specific artificial censoring, estimate inverse probability
+    of censoring weights using pooled logistic or Cox models, fit weighted
+    outcome models, and obtain subject-level bootstrap confidence intervals by
+    repeating the complete analysis. The methods are described by Maringe et
+    al. (2020) <doi:10.1093/ije/dyaa057> and Gaber et al. (2024)
+    <doi:10.1002/cam4.70461>.
 License: MIT + file LICENSE
+URL: https://github.com/CausalInferenceLab/clonecensorweighting
+BugReports: https://github.com/CausalInferenceLab/clonecensorweighting/issues
 Encoding: UTF-8
 Depends:
     R (>= 4.4.0)
 Imports:
     dplyr,
-    purrr,
     readr,
     survival,
     tibble,
-    tidyr,
     rlang,
     glue
 Suggests:
     knitr,
     rmarkdown,
-    testthat (>= 3.0.0)
+    testthat (>= 3.2.0)
 Config/testthat/edition: 3
 Roxygen: list(markdown = TRUE)
 RoxygenNote: 8.0.0

NAMESPACE

@@ -2,7 +2,6 @@
 
 export(apply_logics)
 export(clone_arms)
-export(clone_censor_weighting)
 export(create_censoring_logics_A)
 export(create_final_data)
 export(create_policy_A)

NEWS.md

@@ -1,5 +1,19 @@
 # clonecensorweighting 0.0.0.9000
 
+## Correctness
+
+* Fixed Cox censoring probabilities to evaluate the cumulative baseline hazard
+  as a step function at every interval start rather than assigning zero when an
+  interval start did not exactly match a censoring-event time.
+* Added a configurable natural cubic spline for interval start time in
+  pooled-logistic censoring models (`time_spline_df`; `NULL` retains the
+  linear-time default).
+* Changed `emul_estimate_bootstrap()` to resample the original subjects and
+  repeat cloning, artificial censoring, person-time expansion, censoring-model
+  estimation, weighting, and outcome-model estimation in every replicate.
+* Removed the development scaffold `clone_censor_weighting()` from the public
+  API.
+
 ## Refactoring
 
 * Split the final censoring-probability training data workflow into focused
@@ -20,8 +34,8 @@
   inverse probability of censoring weights.
 * Added `emul_estimate()` as the public interface for Cox, logistic, and
   Kaplan-Meier analyses of the emulated trial.
-* Added `emul_estimate_bootstrap()` as the public interface for bootstrap
-  confidence intervals for two-arm Cox and logistic analyses.
+* Added `emul_estimate_bootstrap()` as the public interface for full-workflow
+  bootstrap confidence intervals for two-arm Cox and logistic analyses.
 * Kept formula construction, cumulative uncensoring, baseline predictor
   extraction, predictor/weight normalization, arm binding, and coefficient
   extraction as internal helpers.

R/censoring_estimation.R

@@ -12,6 +12,10 @@
 #' @param id Column name for the subject identifier.
 #' @param time_start Column name for interval start time.
 #' @param time_stop Column name for interval stop time.
+#' @param time_spline_df Degrees of freedom for the natural cubic spline of
+#'   `time_start` in pooled-logistic censoring models. Set to an integer of at
+#'   least `2` to use a spline. The default `NULL` uses a linear time term.
+#'   Ignored when `method = "Cox"`.
 #' @param eps Small probability floor to avoid division by zero.
 #'
 #' @returns A named list of clone data frames with censoring probability
@@ -51,6 +55,7 @@ estimate_censoring <- function(
   id = "id",
   time_start = "Tstart",
   time_stop = "Tstop",
+  time_spline_df = NULL,
   eps = 1e-6
 ) {
   method <- match.arg(method)
@@ -67,6 +72,9 @@ estimate_censoring <- function(
   }
   .assert_clone_columns(clones, required_columns)
   .assert_probability_floor(eps)
+  if (method != "Cox") {
+    .assert_spline_df(time_spline_df, "time_spline_df")
+  }
 
   arms <- names(clones)
   res <- vector("list", length = length(arms))
@@ -88,7 +96,8 @@ estimate_censoring <- function(
   pooled_formula <- make_censoring_formula(
     backtick_name(censoring),
     predictors,
-    time_var = time_start
+    time_var = time_start,
+    time_spline_df = time_spline_df
   )
 
   for (arm in arms) {
@@ -112,16 +121,17 @@ estimate_censoring <- function(
         survival::basehaz(ms_cens, centered = FALSE)
       )
       names(base_hazard) <- c("hazard", "t")
+      hazard_index <- findInterval(dat[[time_start]], base_hazard$t)
+      baseline_hazard <- numeric(nrow(dat))
+      has_hazard <- hazard_index > 0L
+      baseline_hazard[has_hazard] <-
+        base_hazard$hazard[hazard_index[has_hazard]]
 
       res[[arm]] <-
         dat |>
-        dplyr::mutate(lin_pred = .env[["lin_pred"]]) |>
-        dplyr::left_join(
-          base_hazard,
-          by = stats::setNames("t", time_start)
-        ) |>
         dplyr::mutate(
-          hazard = dplyr::coalesce(.data$hazard, 0),
+          lin_pred = .env[["lin_pred"]],
+          hazard = .env[["baseline_hazard"]],
           P_uncens = exp(-.data$hazard * exp(.data$lin_pred))
         )
     } else {
@@ -161,7 +171,8 @@ estimate_censoring <- function(
         numerator_formula <- make_censoring_formula(
           backtick_name(censoring),
           numerator_data$predictors,
-          time_var = time_start
+          time_var = time_start,
+          time_spline_df = time_spline_df
         )
         fit_num <- stats::glm(
           numerator_formula,
@@ -200,15 +211,30 @@ estimate_censoring <- function(
 make_censoring_formula <- function(
   response,
   predictors = NULL,
-  time_var = NULL
+  time_var = NULL,
+  time_spline_df = NULL
 ) {
-  terms <- unique(c(
-    normalize_predictors(time_var),
-    normalize_predictors(predictors)
-  ))
-  terms <- backtick_name(terms)
+  time_var <- normalize_predictors(time_var)
+  predictors <- backtick_name(normalize_predictors(predictors))
 
-  formula <- stats::reformulate(terms, response = response)
+  time_term <- backtick_name(time_var)
+  if (length(time_var) > 0L && !is.null(time_spline_df)) {
+    .assert_spline_df(time_spline_df, "time_spline_df")
+    time_term <- paste0(
+      "splines::ns(",
+      backtick_name(time_var),
+      ", df = ",
+      as.integer(time_spline_df),
+      ")"
+    )
+  }
+  terms <- unique(c(time_term, predictors))
+
+  if (length(terms) == 0L) {
+    formula <- stats::as.formula(paste(response, "~ 1"))
+  } else {
+    formula <- stats::reformulate(terms, response = response)
+  }
   environment(formula) <- parent.frame()
   formula
 }

R/data.R

@@ -1,6 +1,7 @@
 #' Simulated lung cancer patients data by Maringe et al. (2020)
 #'
-#' The dataset is from supplmentary files of [Maringe et al. (2020)](https://doi.org/10.1093/ije/dyaa057).
+#' The dataset is from the supplementary files of Maringe et al. (2020),
+#' \doi{10.1093/ije/dyaa057}.
 #' The dataset is a set of 200 simulated lung cancer patients. 
 #' These patients are followed up for a year following their cancer diagnosis:
 #' 106 of them received surgery within six months of their diagnosis and 
@@ -22,15 +23,16 @@
 #'   \item{charlson}{Charlson's comorbidity index}
 #'   \item{emergency}{route to diagnosis}
 #' }
-#' @source <https://doi.org/10.1093/ije/dyaa057>
+#' @source \doi{10.1093/ije/dyaa057}
 #' @examples
 #' data(lungcancer)
 "lungcancer"
 
 
 #' 13 types of patient records in Maringe et al. (2020)
 #'
-#' The dataset is from Figure 2 in [Maringe et al. (2020)](https://doi.org/10.1093/ije/dyaa057).
+#' The dataset is from Figure 2 in Maringe et al. (2020),
+#' \doi{10.1093/ije/dyaa057}.
 #' The dataset illustrates all possible censoring mechanisms with 13 types of 
 #' patients records that could be seen in the cancer registry data, when 
 #' allowing at most one treatment (surgery in this example) for each patient
@@ -46,7 +48,7 @@
 #'   \item{death}{observed event at the latest follow-up, 1: dead, 0: alive}
 #'   \item{followup}{observed follow-up time (time to death or time to latest followup with 1 year (365 days) at maximum)}
 #' }
-#' @source <https://doi.org/10.1093/ije/dyaa057>
+#' @source \doi{10.1093/ije/dyaa057}
 #' @examples
 #' data(patients13)
-"patients13"
+"patients13"