Fits one Cox model per feature (`Surv(time, status) ~ feature`, or `Surv(start, time, status) ~ feature` for counting-process data), optionally in parallel, and returns one consolidated table with hazard ratios, confidence intervals and a multiple-testing adjusted p-value per feature. A feature whose model fails is reported with its message instead of stopping the screen.
Usage
batch_cox_single(
data,
time = "time",
status = "status",
features,
time_start = NULL,
adjust_method = "BH",
exponentiate = TRUE,
conf_level = 0.95,
parallel = TRUE
)Arguments
- data
A data frame.
- time, status
Names of the follow-up time and event columns (event 1/`TRUE`, censored 0/`FALSE`).
- features
Character vector of feature columns.
- time_start
Optional name of a start-time column for start-stop (counting process) data; `time` is then the stop time.
- adjust_method
Any method of [stats::p.adjust()] (default `"BH"`), or `"none"`.
- exponentiate
If `TRUE` (default) report hazard ratios with their limits; otherwise raw coefficients (log hazard ratios).
- conf_level
Confidence level (default 0.95).
- parallel
If `TRUE` (default) and the `furrr` package is installed, features are fitted with `furrr::future_map()` under the current [future::plan()] (sequential unless you set a plan). Without `furrr` the fits run one after the other.
Value
A tibble of class `cbe_batch_cox` ordered by `p_adj`, with `feature`, `term`, `estimate`, `conf.high`, `std.error`, `statistic`, `p.value`, `p_overall`, `p_adj`, `n`, `n_events`, `status` (`"ok"`, `"not estimable"` for a coefficient `coxph` could not estimate, or the error message) and `warning` (the warnings `coxph` raised for the feature, joined with `"; "`, otherwise `""`). Look at `warning` before trusting a result: "coefficient may be infinite" means perfect separation, and the hazard ratio and its p-value are then not meaningful.
Rows, p-values and adjustment
A numeric or two-level feature gives one row; a factor with `k` levels gives `k - 1` rows (one per non-reference level). `p.value` is the Wald p-value of that coefficient. `p_overall` is the p-value of the feature as a whole (the Wald test of all its coefficients, `summary(fit)$waldtest`) and is the one adjusted: `p_adj = p.adjust(p_overall, adjust_method)` over **features** (each feature counted once). Features that failed or have no p-value are left out of the adjustment, so `p_adj` is adjusted over the fitted features only.
Examples
lung <- survival::lung
lung$status <- lung$status - 1
batch_cox_single(lung, "time", "status", c("age", "sex", "ph.ecog", "wt.loss"), parallel = FALSE)
#> # A tibble: 4 × 14
#> feature term estimate conf.low conf.high std.error statistic p.value
#> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 ph.ecog ph.ecog 1.61 1.29 2.01 0.113 4.20 0.0000269
#> 2 sex sex 0.588 0.424 0.816 0.167 -3.18 0.00149
#> 3 age age 1.02 1.00 1.04 0.00920 2.03 0.0419
#> 4 wt.loss wt.loss 1.00 0.989 1.01 0.00608 0.217 0.828
#> # ℹ 6 more variables: p_overall <dbl>, p_adj <dbl>, n <int>, n_events <dbl>,
#> # status <chr>, warning <chr>