Can we access all data columns in a custom ggplot2's stat? - r

Question

I would like to implement diagnostics for Cox proportional hazards model with ggplot2, by creating new stats functions and ggproto objects. The idea is to benefit from grouping (by color, facet_grid, etc.) for conditional computation of desired statistics (e.g. martingale residuals).
In the example below, the null model is refitted and martingale's computed for each cofactor level. The issues are:
Fit is done based on data argument provided internally for the ggproto's compute_group method, not the actual dataset. This means, that dataframe is stripped of columns not explicitly defined in aes so it's up to the user to provide at least "time to event" and "event indicator" columns each time. (unless they're hardcoded in a wrappper for stat_... function). Can compute_group somehow access corresponding rows from the original dataset?
Rownames for compute_group's data are unique per PANEL, which means they do not reflect the actual rownames of original dataset supplited to ggplot, and we can no longer exclude e.g. incomplete cases omitted by full model, unless explicitly specifying/hardcoding some id variable. The question is same as above.
Can a geom layer access the y's computed by another custom stat, which is also plotted? Consider a scatterplot of residuals with a smoother, such as. ggplot(data, aes(x = covariate, time = time, event = event)) + stat_funcform() + geom_smooth(aes(y = ..martingales..)), where ..martingales.. are actually computed by stat_funcform.
.
# Load & Set data ---------------------------------------------------------
require(data.table)
require(dplyr)
require(ggplot2)
require(survival)
set.seed(1)
dt <- data.table(factor = sample(c(0, 2), 1e3, T),
factor2 = sample(c(0, 2, NA), 1e3, T),
covariate = runif(1e3, 0, 1))
dt[, `:=`(etime = rexp(1e3, .01 * (.5 + factor + covariate)),
ctime = runif(1e3, 0, 1e2))]
dt[, `:=`(event = ifelse(etime < ctime, 1, 0),
time = ifelse(etime < ctime, etime, ctime))]
survfit(Surv(time, event) ~ factor + I(covariate > .5), data = dt) %>%
autoplot(fun = 'cloglog')
# Fit coxph model ---------------------------------------------------------
fit <- coxph(Surv(time, event) ~ factor + covariate, data = dt)
# Define custom stat ------------------------------------------------------
StatFuncform <- ggproto(
"StatFuncform", Stat,
compute_group = function(data, scales, params) {
head(data) %>% print
data$y <- update(params$fit, ~ 1, data = data) %>%
resid(type = 'martingale')
return(data)
}
)
stat_funcform <- function(mapping = NULL, data = NULL, geom = "point",
position = "identity", na.rm = FALSE, show.legend = NA,
inherit.aes = TRUE, ...) {
layer(
stat = StatFuncform, data = data, mapping = mapping, geom = geom,
position = position, show.legend = show.legend, inherit.aes = inherit.aes,
params = list(na.rm = na.rm, ...)
)
}
# Plot --------------------------------------------------------------------
# computation unsuccessfull - no 'time' and 'event' variables
ggplot(data = dt,
mapping = aes(x = covariate,
color = factor(event))) +
stat_funcform(params = list(fit = fit)) +
facet_grid(~ factor, labeller = label_both, margins = TRUE)
# ok - 'time' and 'event' specified explicitly
ggplot(data = dt,
mapping = aes(x = covariate, time = time, event = event,
factor2 = factor2,
color = factor(event))) +
stat_funcform(params = list(fit = fit)) +
facet_grid(~ factor, labeller = label_both, margins = TRUE)
I'm aware there are already a few packages implementing diagnostics for survival analysis (e.g. survminer) or ggplot2's autoplot methods, but they're rather providing wrappers instead of taking advantage of default ggplot2() + stat_ semantics.