Enhanced Comparison Examples

- Added example of log-logistic distribution
- Added examples of how to calculate AIC / BIC / LOO
- Added examples of how to calculate survival quantities + plots
- Added brms examples

NAMESPACE

@@ -59,6 +59,8 @@ S3method(extractVariableNames,DataSurvival)
 S3method(generateQuantities,JointModelSamples)
 S3method(getParameters,Link)
 S3method(getParameters,LinkComponent)
+S3method(getParameters,StanModel)
+S3method(getParameters,default)
 S3method(initialValues,JointModel)
 S3method(initialValues,Link)
 S3method(initialValues,LinkComponent)

R/StanModel.R

@@ -68,6 +68,7 @@ as.list.StanModel <- function(x, ...) {
 # getParameters-StanModel ----
 
 #' @rdname getParameters
+#' @export
 getParameters.StanModel <- function(object) object@parameters
 
 

R/defaults.R

@@ -10,6 +10,7 @@ NULL
 
 
 #' @rdname getParameters
+#' @export
 getParameters.default <- function(object) {
     if (missing(object) || is.null(object)) {
         return(NULL)

design/examples/loglogistic.R

@@ -0,0 +1,157 @@
+
+
+
+library(dplyr)
+library(flexsurv)
+library(survival)
+library(cmdstanr)
+library(posterior)
+library(bayesplot)
+library(here)
+
+
+dat <- flexsurv::bc |>
+    as_tibble() |>
+    mutate(arm = "A", study = "S", pt = sprintf("pt-%05d", 1:n()))
+
+
+#
+#
+# Have specified no covariates in this models so that they are comparable
+# JMpost uses a PH model where as the others use AFT thus adding covariates
+# would result in different models. Without covariates they are just fitting
+# the base distribution which should be identical.
+#
+#
+
+
+
+
+################################
+#
+# Flexsurv parametric Regression
+#
+
+
+mod_flex <- flexsurvreg(
+    Surv(recyrs, censrec) ~ 1,
+    data = dat,
+    dist = "llogis"
+)
+mod_flex
+
+logLik(mod_flex)
+AIC(mod_flex)
+BIC(mod_flex)
+
+
+################################
+#
+# Bayesian Weibull Regression
+#
+
+mod <- cmdstan_model(
+    stan_file = here("design/examples/loglogistic.stan"),
+    exe_file = here("design/examples/models/loglogistic")
+)
+
+design_mat <- model.matrix(~ 1, data = dat)
+
+
+stan_data <- list(
+    n = nrow(dat),
+    design = design_mat,
+    p = ncol(design_mat),
+    times = dat$recyrs,
+    event_fl = dat$censrec
+)
+fit <- mod$sample(
+    data = stan_data,
+    chains = 2,
+    parallel_chains = 2,
+    refresh = 200,
+    iter_warmup = 1000,
+    iter_sampling = 1500
+)
+vars <- c(
+    "beta_design",
+    "alpha_0",
+    "beta_0"
+)
+fit$summary(vars)
+
+
+# Log Likelihood
+log_lik <- fit$draws("log_lik", format = "draws_matrix") |>
+    apply(1, sum) |>
+    mean()
+log_lik
+
+# AIC
+k <- 2
+-2 * log_lik + k * (stan_data$p + 1) # +1 for the scale parameter
+
+# BIC
+((stan_data$p + 1) * log(stan_data$n)) + (-2 * log_lik)
+
+
+
+################################
+#
+# JMpost
+#
+
+devtools::load_all()
+# library(jmpost)
+
+jm <- JointModel(
+    survival = SurvivalLogLogistic()
+)
+
+jdat <- DataJoint(
+    subject = DataSubject(
+        data = dat,
+        subject = "pt",
+        arm = "arm",
+        study = "study"
+    ),
+    survival = DataSurvival(
+        data = dat,
+        formula = Surv(recyrs, censrec) ~ 1
+    )
+)
+
+mp <- sampleStanModel(
+    jm,
+    data = jdat,
+    iter_warmup = 1000,
+    iter_sampling = 1500,
+    chains = 2,
+    parallel_chains = 2
+)
+
+vars <- c(
+    "sm_logl_lambda",
+    "sm_logl_p"
+)
+
+x <- mp@results$summary(vars)
+
+c(
+    "scale" = 1 / x$mean[1],
+    "shape" = x$mean[2]
+)
+
+
+# Log Likelihood
+log_lik <- mp@results$draws("log_lik", format = "draws_matrix") |>
+    apply(1, sum) |>
+    mean()
+log_lik
+
+# AIC
+k <- 2
+-2 * log_lik + k * 2
+
+# BIC
+(2 * log(nrow(dat))) + (-2 * log_lik)

design/examples/loglogistic.stan

@@ -0,0 +1,51 @@
+
+
+
+data {
+    int<lower=1> n;                          // Number of subjects
+    int<lower=0> p;                          // Number of covariates (including intercept)
+    vector<lower=0>[n] times;                // Event|Censor times
+    array[n] int<lower=0, upper=1> event_fl; // 1=event 0=censor
+    matrix[n, p] design;                     // Design matrix
+}
+
+transformed data {
+    // Assuming that the first term is an intercept column which
+    // will conflict with the alpha_0 term so remove it
+    matrix[n, p-1] design_reduced;
+    if (p > 1 ) {
+        design_reduced = design[, 2:p];
+    }}
+
+parameters {
+    vector[p-1] beta_design;
+    real<lower=0> alpha_0;
+    real<lower=0> beta_0;
+}
+
+transformed parameters {
+    vector[n] alpha;
+    if (p == 1) {
+        alpha = rep_vector(alpha_0, n);
+    } else {
+        alpha = alpha_0 .* exp(design_reduced * beta_design);
+    }
+
+    // Likelihood
+    vector[n] log_lik;
+    for (i in 1:n) {
+        if (event_fl[i] == 1) {
+            log_lik[i] = loglogistic_lpdf(times[i] | alpha[i], beta_0);
+        } else {
+            log_lik[i] = log(1 - loglogistic_cdf(times[i] | alpha[i], beta_0));
+        }
+    }
+}
+
+model {
+    // Priors
+    beta_design ~ normal(0, 3);
+    alpha_0 ~ lognormal(log(2), 1);
+    beta_0 ~ lognormal(log(2), 1);
+    target += sum(log_lik);
+}