Reformats recurrent event data (wide) into different models for survival analysis, but can also be used for simple survival analysis tables as well. The general format is how data tends to be collected. There is left and right censoring date, a labeled event column that contains the date of the event, and a censoring column for a final censoring event. The accepted parameter options are listed, with the type of table that will be generated:
traditional: Traditional survival table that has single censoring event (
trad)counting : Formally called the Andersen and Gill model (
ag). Counting process model assumes each event is independent and that a subject contributes to the risk set during the time under observation. Multiple events are treated as a new (but delayed) entry that is followed until the next event. This means subjects under observation are at risk for a second event, even without having had a prior event. There are thus no strata in the model.marginal: Formally called the Wei-Lin-Weissfield model, but more commonly known as a marginal model (
marginal). Marginal models assumes each event is a separate process. Each subject is at risk for all events. The time for an event starts at the beginning of follow-up for each subject. Thus, each risk period is considered a different strata (regardless of if subject had an event or not).conditional A: Formally called the Prentice, Williams, and Peterson total time model (
pwptt). Conditional A models order events by stratification, based on the number of events prior. All subjects are at risk for the left strata, but only those with a previous event are at risk for a successive event. The total time to event is used.conditional B: Formally called the Prentice, Williams, and Peterson gap time model (
pwpgt). Conditional B models also order events by strata (like conditional A), however the time to outcome is defined as the gap between the time of previous event.
Arguments
- data
A dataframe containing the subsequent parameters
- model_type
Model type that is indicated:
tradmakes traditional survival tableagmakes table with risk periods starting at time of prior event without conditional stratamarginalmakes table with risk periods from entry to censorship with strata per each eventpwpttmakes table with risk periods starting at time of prior event with conditional stratapwpgtmakes table with risk periods of each time interval between events, with conditional strata
- id
Column in dataframe that contains unique IDs for each row
- left
Column with left/enrollment dates
- right
Column with right/censoring time point, or right contact
- censor
Column that names if death/final censorship is known (0 or 1). The default is that, if no censorship information is given, that are no failure events at time of right contact.
censoris not required for recurrent event analysis, but is required for traditional survival tables.- event_dates
Vector of columns that contain event dates
Value
A data frame organized into a survival table format. Output options are in Details. Generally, the following columns are generated:
id: An ID column is created
start: A formatted start time, usually 0
stop: A formatted stop time, in days, from prior event
status: If event occurred or not
strata: Event strata that is being applied
Details
This function takes every event date, and creates several types of recurrent event tables. It orders the data chronologically for repeat events. Currently does normal (left event) and recurrent models (counting, marginal, and conditional A and B models). Further details can be found at IDRE.
For recurrent events, the final censoring event can include death, or can be ignored if its not considered a failure event.
For traditional survival analysis,
censoris required andevent_datesshould be left as NULL. The function will do the rest.
Performance: Importantly, for large datasets of recurrent data (>500 rows), this function will show significant slow-down since it uses an intuitive approach on defining the datasets. Future iterations will create a vectorized approach that should provide performance speed-ups.
Examples
# \donttest{
# Data
data("mims")
#> Warning: data set ‘mims’ not found
# Parameters
id <- "patid"
left <- "left_visit_date_bl"
right <- "ldka"
event_dates <- c("mi_date_1", "mi_date_2", "mi_date_3")
model_type <- "marginal"
censor <- "DEATH_CV_YN"
# Run analysis
out <- recur(
mims, model_type, id, left, right, censor, event_dates
)
#> Error in id %in% names(data): object 'mims' not found
# }