Enable Left-Censoring for Random Slope Model (#291)

R/LongitudinalModel.R

@@ -27,7 +27,6 @@ LongitudinalModel <- function(
     name = "<Unnamed>",
     ...
 ) {
-
     base_stan <- read_stan("base/longitudinal.stan")
 
     stan_full <- decorated_render(

inst/stan/base/longitudinal.stan

@@ -55,7 +55,24 @@ parameters {
 }
 
 transformed parameters {
+    //
+    // Source - base/longitudinal.stan
+    //
+    vector[Nta_total] Ypred_log_lik = rep_vector(0, Nta_total);
+
     {{ stan.transformed_parameters }}
+
+    //
+    // Source - base/longitudinal.stan
+    //
+    log_lik += csr_matrix_times_vector(
+        Nind,
+        Nta_total,
+        w_mat_inds_all_y,
+        v_mat_inds_all_y,
+        u_mat_inds_all_y,
+        Ypred_log_lik
+    );
 }
 
 model {

inst/stan/lm-gsf/model.stan

@@ -66,31 +66,17 @@ transformed parameters{
         lm_gsf_psi_phi[ind_index]
     );
 
-    log_lik += csr_matrix_times_vector(
-        Nind,
-        Nta_obs_y,
-        w_mat_inds_obs_y,
-        v_mat_inds_obs_y,
-        u_mat_inds_obs_y,
-        vect_normal_log_dens(
-            Yobs[obs_y_index],
-            Ypred[obs_y_index],
-            Ypred[obs_y_index] * lm_gsf_sigma
-        )
-    );
 
+    Ypred_log_lik[obs_y_index] = vect_normal_log_dens(
+        Yobs[obs_y_index],
+        Ypred[obs_y_index],
+        Ypred[obs_y_index] * lm_gsf_sigma
+    );
     if (Nta_cens_y > 0 ) {
-        log_lik += csr_matrix_times_vector(
-            Nind,
-            Nta_cens_y,
-            w_mat_inds_cens_y,
-            v_mat_inds_cens_y,
-            u_mat_inds_cens_y,
-            vect_normal_log_cum(
-                Ythreshold,
-                Ypred[cens_y_index],
-                Ypred[cens_y_index] * lm_gsf_sigma
-            )
+        Ypred_log_lik[cens_y_index] = vect_normal_log_cum(
+            Ythreshold,
+            Ypred[cens_y_index],
+            Ypred[cens_y_index] * lm_gsf_sigma
         );
     }
 }

inst/stan/lm-random-slope/model.stan

@@ -35,18 +35,18 @@ transformed parameters {
 
     vector[Nta_total] Ypred = lm_rs_ind_intercept[ind_index] + lm_rs_rslope_ind .* Tobs;
 
-    log_lik += csr_matrix_times_vector(
-        Nind,
-        Nta_total,
-        w_mat_inds_all_y,
-        v_mat_inds_all_y,
-        u_mat_inds_all_y,
-        vect_normal_log_dens(
-            Yobs,
-            Ypred,
-            rep_vector(lm_rs_sigma, Nta_total)
-        )
+    Ypred_log_lik[obs_y_index] = vect_normal_log_dens(
+        Yobs[obs_y_index],
+        Ypred[obs_y_index],
+        rep_vector(lm_rs_sigma, Nta_obs_y)
     );
+    if (Nta_cens_y > 0 ) {
+        Ypred_log_lik[cens_y_index] = vect_normal_log_cum(
+            Ythreshold,
+            Ypred[cens_y_index],
+            rep_vector(lm_rs_sigma, Nta_cens_y)
+        );
+    }
 }
 
 

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

@@ -60,31 +60,16 @@ transformed parameters{
         lm_sf_psi_kg[ind_index]
     );
 
-    log_lik += csr_matrix_times_vector(
-        Nind,
-        Nta_obs_y,
-        w_mat_inds_obs_y,
-        v_mat_inds_obs_y,
-        u_mat_inds_obs_y,
-        vect_normal_log_dens(
-            Yobs[obs_y_index],
-            Ypred[obs_y_index],
-            Ypred[obs_y_index] * lm_sf_sigma
-        )
+    Ypred_log_lik[obs_y_index] = vect_normal_log_dens(
+        Yobs[obs_y_index],
+        Ypred[obs_y_index],
+        Ypred[obs_y_index] * lm_sf_sigma
     );
-
     if (Nta_cens_y > 0 ) {
-        log_lik += csr_matrix_times_vector(
-            Nind,
-            Nta_cens_y,
-            w_mat_inds_cens_y,
-            v_mat_inds_cens_y,
-            u_mat_inds_cens_y,
-            vect_normal_log_cum(
-                Ythreshold,
-                Ypred[cens_y_index],
-                Ypred[cens_y_index] * lm_sf_sigma
-            )
+        Ypred_log_lik[cens_y_index] = vect_normal_log_cum(
+            Ythreshold,
+            Ypred[cens_y_index],
+            Ypred[cens_y_index] * lm_sf_sigma
         );
     }
 }

tests/testthat/_snaps/JointModelSamples.md

@@ -11,6 +11,7 @@
         
             Variables:
                 Ypred[2800]
+                Ypred_log_lik[2800]
                 beta_os_cov[3]
                 lm_rs_ind_intercept[400]
                 lm_rs_ind_rnd_slope[400]

tests/testthat/test-LongitudinalRandomSlope.R

@@ -193,3 +193,105 @@ test_that("Random Slope Model can recover known parameter values", {
     )
     expect_gt(lmer_cor, 0.99)
 })
+
+
+
+test_that("Random Slope Model left-censoring works as expected", {
+    ## Generate data with known parameters
+    set.seed(739)
+
+    jlist <- SimJointData(
+        design = list(
+            SimGroup(250, "Arm-A", "Study-X"),
+            SimGroup(150, "Arm-B", "Study-X")
+        ),
+        survival = SimSurvivalExponential(1 / 100),
+        longitudinal = SimLongitudinalRandomSlope(
+            times = c(-200, -150, -100, -50, 0, 1, 100, 125, 200, 300, 350, 400, 500, 600),
+            intercept = 30,
+            sigma = 3,
+            slope_mu = c(1, 3),
+            slope_sigma = 0.2
+        ),
+        .silent = TRUE
+    )
+
+
+    # Trash the data for negative values
+    # As the data is censored this shouldn't impact the sampled predictions
+    dat_lm <- jlist@longitudinal
+    negative_index <- dat_lm$sld < 0
+    dat_lm$sld[negative_index] <- runif(sum(negative_index), -999, -100)
+
+    jm <- JointModel(
+        longitudinal = LongitudinalRandomSlope(
+            intercept = prior_normal(30, 2),
+            slope_sigma = prior_lognormal(log(0.2), sigma = 0.5),
+            sigma = prior_lognormal(log(3), sigma = 0.5)
+        )
+    )
+
+    jdat <- DataJoint(
+        subject = DataSubject(
+            data = jlist@survival,
+            subject = "pt",
+            arm = "arm",
+            study = "study"
+        ),
+        longitudinal = DataLongitudinal(
+            data = dat_lm,
+            formula = sld ~ time,
+            threshold = 0
+        )
+    )
+
+    mp <- run_quietly({
+        sampleStanModel(
+            jm,
+            data = jdat,
+            iter_sampling = 200,
+            iter_warmup = 400,
+            chains = 1,
+            refresh = 0,
+            parallel_chains = 1
+        )
+    })
+
+
+    vars <- c(
+        "lm_rs_intercept",       # 30
+        "lm_rs_slope_mu",        # 1 , 3
+        "lm_rs_slope_sigma",     # 0.2
+        "lm_rs_sigma"            # 3
+    )
+
+    pars <- mp@results$summary(vars)
+
+
+    ## Check that we can recover main effects parameters
+    z_score <- (c(30, 1, 3, 0.2, 3) - pars$mean) / pars$sd
+    expect_true(all(abs(z_score) < qnorm(0.95)))
+
+
+    ## Check that we can recover random effects parameters
+    pars <- suppressWarnings({
+        mp@results$summary("lm_rs_ind_rnd_slope")$mean
+    })
+
+    ## Extract real random effects per patient
+    ## We store them as (random effect + mean)
+    ## thus need to subtract mean for comparison
+    ## to nle4
+    group_mean <- c(1, 3)
+
+    ## Check for consistency of random effects with lmer
+    mod <- lme4::lmer(
+        sld ~ time:arm + (time - 1 | pt),
+        jlist@longitudinal
+    )
+    lmer_cor <- cor(
+        lme4::ranef(mod)$pt$time,
+        pars - group_mean[as.numeric(jlist@survival$arm)]
+    )
+    expect_gt(lmer_cor, 0.99)
+})