add casebase

.gitignore

@@ -1,3 +1,4 @@
 
 .DS_Store
 .Rproj.user
+.Rhistory

survival-analysis/README.Rmd

@@ -1,14 +1,16 @@
-
+---
+always_allow_html: true
+---
 
 ```{r, load_libs, message = FALSE, warning = FALSE, echo = FALSE, results = 'hide'}
 library(prodlim)
 library(survival)
 library(flexsurv)
+library(casebase)
 library(rpart)
 library(pec)
 library(ipred)
 library(randomForestSRC)
-library(ranger)
 library(mboost)
 library(tidymodels)
 library(cli)
@@ -105,6 +107,7 @@ fitw <-
               dist = "weibull")
 times <- 
   summary(fitw, ovarian[1:2,], type = "mean") %>% 
+  map(~ setNames(.x, c("est", "lcl", "ucl"))) %>% 
   bind_rows() %>% 
   select(.pred_time = est) %>% 
   as_tibble()
@@ -195,6 +198,42 @@ summary(flxs_mod, test_pred, type = "survival", t = .times, tidy = TRUE)
 
 The `flexsurvspline()` function has the same interface. 
 
+## `casebase::`[`fitSmoothHazard`](https://rdrr.io/cran/casebase/man/fitSmoothHazard.html)
+
+The main fitting function is `fitSmoothHazard`:
+
+```{r, cb_fit}
+library(casebase)
+cb_mod <-
+  fitSmoothHazard(status ~ time + (X1 + X2) ^ 2,
+                  data = train_dat)
+```
+
+Note that `time` is modeled explicitly and with flexibility:
+
+```{r, cb_splines}
+library(splines)
+cb_sp_mod <-
+  fitSmoothHazard(status ~ bs(time) + (X1 + X2) ^ 2,
+                  data = train_dat)
+```
+
+And we can also do penalized estimation:
+
+```{r, cb_glment}
+cb_glmnet_mod <-
+  fitSmoothHazard(status ~ time + (X1 + X2) ^ 2,
+                  data = train_dat, family = "glmnet")
+```
+
+Predictions for the cumulative incidence or survival probabilities are obtained using `absoluteRisk`: 
+
+```{r, cb_pred}
+absoluteRisk(cb_mod, newdata = test_pred, time = .times, type = "CI")
+absoluteRisk(cb_mod, newdata = test_pred, time = .times, type = "survival")
+```
+
+(Note that the argument `type` is only available in the development version.)
 
 ## Cox PH Models
 
@@ -389,34 +428,6 @@ although these values do not bracket the observed times:
 summary(train_dat$time)
 ```
 
-### `ranger:::`[`ranger`](https://rdrr.io/cran/ranger/man/ranger.html)
-
-```{r, ranger_fit}
-library(ranger)
-ranger_mod <- ranger(Surv(time, status) ~ X1 + X2, data = train_dat, num.trees = 1000)
-```
-
-The `predict` function generates an object with class `"ranger.prediction"`. 
-
-```{r, ranger_pred}
-ranger_pred = predict(ranger_mod, test_pred)
-str(ranger_pred)
-```
-
-The `unique.death.times` slot has the unique death times:
-
-```{r, }
-all(sort(unique(train_dat$time)) == ranger_pred$unique.death.times)
-```
-
-The `survival` slot contains a matrix of predicted survival curves, with a row for each test case and a column for each death time. The `chf` slot is the same for the cumulative hazard function:
-
-```{r, ranger_surv}
-ranger_pred$survival[, 1:5]
-ranger_pred$chf[, 1:5]
-```
-
-
 # Model Prediction Scorecard
 
 In most cases, the open circle means that this type of prediction is facilitated _indirectly_. For example, using "tricks" related to `survfit()`, methods in the `pec` package, or other means. 
@@ -431,15 +442,15 @@ pred_types <-
     ~Package, ~Function, ~`pred event time`, ~`survival probs`, ~`linear pred`,
     "`survival`",        "`survreg`",        yep,  nope,  yep,
     "`flexsurv`",        "`flexsurvreg`",    yep,   yep,  yep,
+    "`casebase`",        "`fitSmoothHazard`",    nope,   yep,  yep,
     "`survival`",        "`coxph`",         yep, yep,  yep,
     "`glmnet`",          "`glmnet`",        kinda, kinda,  yep,
     "`mboost`",          "`*boost`",        kinda, kinda,  yep,
     "`rpart`",           "`rpart`",          yep, kinda, nope,
     "`ipred`",           "`bagging`",        yep, kinda, nope,
     "`party`",           "`ctree`",         yep,  kinda, nope,
     "`party`",           "`cforest`",       yep,  kinda, nope,
-    "`randomForestSRC`", "`rfsrc`",         yep,  kinda, nope,
-    "`ranger`",          "`ranger`",        kinda,  yep, nope
+    "`randomForestSRC`", "`rfsrc`",         yep,  kinda, nope
   )
 
 pred_types %>%