Add unscaled variance for SF model

R/LongitudinalSteinFojo.R

@@ -36,6 +36,8 @@ NULL
 #'
 #' @param sigma (`Prior`)\cr for the variance of the longitudinal values `sigma`.
 #'
+#' @param scaled_variance (`logical`)\cr whether the variance should be scaled by the expected value
+#' (see the "Statistical Specifications" vignette for more details)
 #' @param centred (`logical`)\cr whether to use the centred parameterization.
 #'
 #' @export
@@ -51,12 +53,14 @@ LongitudinalSteinFojo <- function(
 
     sigma = prior_lognormal(log(0.1), 1),
 
+    scaled_variance = TRUE,
     centred = FALSE
 ) {
 
     sf_model <- StanModule(decorated_render(
         .x = read_stan("lm-stein-fojo/model.stan"),
-        centred = centred
+        centred = centred,
+        scaled_variance = scaled_variance
     ))
 
     # Apply constriants

R/SimLongitudinalSteinFojo.R

@@ -18,6 +18,8 @@ NULL
 #' @param link_identity (`number`)\cr the link coefficient for the SLD Identity contribution.
 #' @param link_growth (`number`)\cr the link coefficient for the log-growth parameter contribution.
 #' @param link_shrinkage (`number`)\cr the link coefficient for the log-shrinkage parameter contribution.
+#' @param scaled_variance (`logical`)\cr whether the variance should be scaled by the expected value
+#' (see the "Statistical Specifications" vignette for more details)
 #'
 #' @slot sigma (`numeric`)\cr See arguments.
 #' @slot mu_s (`numeric`)\cr See arguments.
@@ -31,6 +33,7 @@ NULL
 #' @slot link_identity (`numeric`)\cr See arguments.
 #' @slot link_growth (`numeric`)\cr See arguments.
 #' @slot link_shrinkage (`numeric`)\cr See arguments.
+#' @slot scaled_variance (`logical`)\cr See arguments.
 #'
 #' @family SimLongitudinal
 #' @name SimLongitudinalSteinFojo-class
@@ -50,7 +53,8 @@ NULL
         link_ttg = "numeric",
         link_identity = "numeric",
         link_growth = "numeric",
-        link_shrinkage = "numeric"
+        link_shrinkage = "numeric",
+        scaled_variance = "logical"
     )
 )
 
@@ -69,7 +73,8 @@ SimLongitudinalSteinFojo <- function(
     link_ttg = 0,
     link_identity = 0,
     link_growth = 0,
-    link_shrinkage = 0
+    link_shrinkage = 0,
+    scaled_variance = TRUE
 ) {
 
     if (length(omega_b) == 1) omega_b <- rep(omega_b, length(mu_b))
@@ -89,7 +94,8 @@ SimLongitudinalSteinFojo <- function(
         link_ttg = link_ttg,
         link_identity = link_identity,
         link_growth = link_growth,
-        link_shrinkage = link_shrinkage
+        link_shrinkage = link_shrinkage,
+        scaled_variance = scaled_variance
     )
 }
 
@@ -144,7 +150,8 @@ sampleObservations.SimLongitudinalSteinFojo <- function(object, times_df) {
         dplyr::mutate(mu_sld = sf_sld(.data$time, .data$psi_b, .data$psi_s, .data$psi_g)) |>
         dplyr::mutate(dsld = sf_dsld(.data$time, .data$psi_b, .data$psi_s, .data$psi_g)) |>
         dplyr::mutate(ttg = sf_ttg(.data$time, .data$psi_b, .data$psi_s, .data$psi_g)) |>
-        dplyr::mutate(sld = stats::rnorm(dplyr::n(), .data$mu_sld, .data$mu_sld * object@sigma)) |>
+        dplyr::mutate(sld_sd = ifelse(object@scaled_variance, .data$mu_sld * object@sigma, object@sigma)) |>
+        dplyr::mutate(sld = stats::rnorm(dplyr::n(), .data$mu_sld, .data$sld_sd)) |>
         dplyr::mutate(
             log_haz_link =
                 (object@link_dsld * .data$dsld) +

inst/stan/lm-stein-fojo/model.stan

@@ -63,13 +63,21 @@ transformed parameters{
     Ypred_log_lik[subject_tumour_index_obs] = vect_normal_log_dens(
         tumour_value[subject_tumour_index_obs],
         Ypred[subject_tumour_index_obs],
-        Ypred[subject_tumour_index_obs] * lm_sf_sigma
+        {%- if scaled_variance -%}
+            Ypred[subject_tumour_index_obs] * lm_sf_sigma
+        {% else %}
+            rep_vector(lm_sf_sigma, n_tumour_obs)
+        {%- endif -%}
     );
     if (n_tumour_cens > 0 ) {
         Ypred_log_lik[subject_tumour_index_cens] = vect_normal_log_cum(
             tumour_value_lloq,
             Ypred[subject_tumour_index_cens],
-            Ypred[subject_tumour_index_cens] * lm_sf_sigma
+            {%- if scaled_variance -%}
+                Ypred[subject_tumour_index_cens] * lm_sf_sigma
+            {% else %}
+                rep_vector(lm_sf_sigma, n_tumour_cens)
+            {%- endif -%} 
         );
     }
 }

tests/testthat/test-LongitudinalSteinFojo.R

@@ -265,6 +265,11 @@ test_that("Can recover known distributional parameters from a SF joint model", {
     expect_true(all(dat$ess_bulk > 100))
 })
 
+
+
+
+
+
 test_that("Can recover known distributional parameters from a SF joint model with growth link", {
 
     skip_if_not(is_full_test())
@@ -431,6 +436,9 @@ test_that("Can recover known distributional parameters from a SF joint model wit
 
 
 
+
+
+
 test_that("Quantity models pass the parser", {
     mock_samples <- .JointModelSamples(
         model = JointModel(longitudinal = LongitudinalSteinFojo(centred = TRUE)),
@@ -454,6 +462,11 @@ test_that("Quantity models pass the parser", {
 
 
 
+
+
+
+
+
 test_that("Can generate valid initial values", {
 
     pars <- c(
@@ -500,3 +513,163 @@ test_that("Can generate valid initial values", {
     expect_true(all(vals > 0))
 
 })
+
+
+
+
+
+
+test_that("Unscaled variance SF mode pass the parser", {
+    jm <- JointModel(
+        longitudinal = LongitudinalSteinFojo(centred = FALSE, scaled_variance = FALSE),
+        survival = SurvivalLogLogistic(),
+        link = linkDSLD()
+    )
+    x <- as.StanModule(jm)
+    expect_stan_syntax(x)
+})
+
+
+
+
+
+
+test_that("Can recover known distributional parameters from unscaled variance SF model", {
+
+    skip_if_not(is_full_test())
+
+    sim_params <- list(
+        sigma = 0.4,
+        mu_s = log(c(0.15, 0.3)),
+        mu_g = log(c(0.4, 0.25)),
+        mu_b = log(60),
+        omega_b = 0.1,
+        omega_s = c(0.1, 0.1),
+        omega_g = c(0.2, 0.2),
+        link_ttg = 0,
+        link_dsld = 0,
+        link_growth = 0,
+        lambda = 2,
+        lambda_cen = 1 / 9000,
+        beta_cat_b = -0.1,
+        beta_cat_c = 0.5,
+        beta_cont = 0.3
+    )
+
+    set.seed(2338)
+    ## Generate Test data with known parameters
+    jlist <- SimJointData(
+        design = list(
+            SimGroup(200, "Arm-A", "Study-X"),
+            SimGroup(200, "Arm-B", "Study-X")
+        ),
+        longitudinal = SimLongitudinalSteinFojo(
+            times = c(1, 50, 100, 150, 200, 300, 400, 500, 600, 700) / 365,
+            sigma = sim_params$sigma,
+            mu_s = sim_params$mu_s,
+            mu_g = sim_params$mu_g,
+            mu_b = sim_params$mu_b,
+            omega_b = sim_params$omega_b,
+            omega_s = sim_params$omega_s,
+            omega_g = sim_params$omega_g,
+            link_ttg = sim_params$link_ttg,
+            link_dsld = sim_params$link_dsld,
+            link_growth = sim_params$link_growth,
+            scaled_variance = FALSE
+        ),
+        survival = SimSurvivalExponential(
+            time_max = 4,
+            time_step = 1 / 365,
+            lambda = sim_params$lambda,
+            lambda_cen = 1 / 9000,
+            beta_cat = c(
+                "A" = 0,
+                "B" = sim_params$beta_cat_b,
+                "C" = sim_params$beta_cat_c
+            ),
+            beta_cont = sim_params$beta_cont
+        ),
+        .silent = TRUE
+    )
+
+    # nolint start⁠
+    ### Diagnostics helpers
+    # plot(survival::survfit(Surv(time, event) ~ 1, data = jlist@survival))
+    # median(jlist@survival$time)
+    # nolint end
+
+
+    jm <- JointModel(
+        longitudinal = LongitudinalSteinFojo(
+            mu_bsld = prior_normal(log(60), 0.5),
+            mu_ks = prior_normal(log(0.2), 0.5),
+            mu_kg = prior_normal(log(0.3), 0.5),
+            omega_bsld = prior_lognormal(log(0.1), 0.5),
+            omega_ks = prior_lognormal(log(0.1), 0.5),
+            omega_kg = prior_lognormal(log(0.1), 0.5),
+            sigma = prior_lognormal(log(0.5), 0.5),
+            centred = TRUE,
+            scaled_variance = FALSE
+        )
+    )
+
+    jdat <- DataJoint(
+        subject = DataSubject(
+            data = jlist@survival,
+            subject = "subject",
+            arm = "arm",
+            study = "study"
+        ),
+        longitudinal = DataLongitudinal(
+            data = jlist@longitudinal,
+            formula = sld ~ time,
+            threshold = 5
+        )
+    )
+
+
+    set.seed(2213)
+    mp <- run_quietly({
+        sampleStanModel(
+            jm,
+            data = jdat,
+            iter_warmup = 750,
+            iter_sampling = 750,
+            chains = 2,
+            parallel_chains = 2
+        )
+    })
+
+    summary_post <- function(model, vars, exp = FALSE) {
+        dat <- model$summary(
+            vars,
+            mean = mean,
+            q01 = \(x) purrr::set_names(quantile(x, 0.01), ""),
+            q99 = \(x) purrr::set_names(quantile(x, 0.99), ""),
+            rhat = posterior::rhat,
+            ess_bulk = posterior::ess_bulk,
+            ess_tail = posterior::ess_tail
+        )
+        if (exp) {
+            dat$q01 <- dat$q01 |> exp()
+            dat$q99 <- dat$q99 |> exp()
+            dat$mean <- dat$mean |> exp()
+        }
+        dat
+    }
+
+    dat <- summary_post(
+        as.CmdStanMCMC(mp),
+        c(
+            "lm_sf_mu_bsld", "lm_sf_mu_ks", "lm_sf_mu_kg",
+            "lm_sf_sigma", "lm_sf_omega_bsld", "lm_sf_omega_kg", "lm_sf_omega_ks"
+        )
+    )
+    true_values <- c(
+        sim_params$mu_b, sim_params$mu_s, sim_params$mu_g,
+        sim_params$sigma, sim_params$omega_b, sim_params$omega_g, sim_params$omega_s
+    )
+    expect_true(all(dat$q01 <= true_values))
+    expect_true(all(dat$q99 >= true_values))
+    expect_true(all(dat$ess_bulk > 100))
+})

tests/testthat/test-SimLongitudinalSteinFojo.R

@@ -53,7 +53,7 @@ test_that("SimLongitudinalSteinFojo works as expected", {
         names(res_obvs),
         c(
             "subject", "arm", "study", "psi_b", "psi_s", "psi_g", "time",
-            "mu_sld", "dsld", "ttg", "sld", "log_haz_link"
+            "mu_sld", "dsld", "ttg", "sld_sd", "sld", "log_haz_link"
         )
     )
 })

vignettes/statistical-specification.Rmd

@@ -185,7 +185,10 @@ Where:
 
 * $\eta_{\theta i}$ is a random effects offset on parameter $\theta$ for subject $i$
 
-
+If using the unscaled variance parameterisation then the following alternative formulation is used:
+$$
+y_{ij} &\sim N(SLD_{ij},\ \sigma^2)
+$$
 
 ### Derivative of the SLD Trajectory Link