diff --git a/ccn2019/ccn2019.rev3.Rmd b/ccn2019/ccn2019.rev3.Rmd
index 61a178e..d8c501e 100644
--- a/ccn2019/ccn2019.rev3.Rmd
+++ b/ccn2019/ccn2019.rev3.Rmd
@@ -163,7 +163,9 @@
 
 ```{r rfe_feature_selection}
 #f <- rt_cat ~ n + t + s + v + l + vl + sl + tl + ul + ll
-f <- rt_cat ~  n + t + s + v + l + vl + sl + tl + ul + ll
+#f <- correct ~ n + t + s + v + l + vl + sl + tl + ul + ll + rt_cat
+f <- rt ~ n + t + s + v + l + vl + sl + tl + ul + ll# + rt_cat
+#f <- a ~ n + t + v + l + vl + sl + tl + ul + ll# + rt_cat
 
 splt <- split_data(f, seqs)
 
@@ -176,7 +178,6 @@
 message("Optimal Variables: ")
 rfeProfile$optVariables
 
-message("Model-Free Variable Importance")
 # ROC for each categories
 filterVarImp(as.data.frame(splt$train.x), splt$train.y)
 
@@ -191,13 +192,25 @@
 
 In addition to the model comparasion plotd, he final output of this section is a single ROC plot that compares new and old models. The same chunk can be used for other categorical models, including RT_CAT and penalized models.
 
+Note: uncomment desired `old.f` and `new.f` formulas to run respective analysis.
+
 ```{r comapre_correct_models}
 
-old.f <- rt_cat ~ n + t + v
-new.f <- rt_cat ~ n + t + tl + l + sl + vl + sl + tl + ul + ll
+#1. atypical local features lead to longer response time.
+#old.f <- rt_cat ~ n + t + v + rt_cat
+#new.f <- rt_cat ~ n + t + tl + l + sl + vl + sl + tl + ul + ll + rt_cat
 
-old.f <- correct ~ n + t + v + rt_cat
-new.f <- correct ~ n + t + tl + l + sl + vl + sl + tl + ul + ll + rt_cat
+old.f <- correct ~ n + t
+new.f <- correct ~ n + t + rt_cat
+
+#old.f <- correct ~ n + t + tl + l + sl + vl + sl + tl + ul + ll
+#new.f <- correct ~ n + t + tl + l + sl + vl + sl + tl + ul + ll + rt_cat
+
+#2. local features provide a better prediction for the correctness of choice
+#old.f <- correct ~ n + t + v
+#new.f <- correct ~ n + t + tl + l + sl + vl + sl + tl + ul + ll
+
+#3. Longer 
 
 ctrl <- trainControl(method="cv",
                      number=5, 
@@ -207,7 +220,8 @@
                      savePredictions = T,
                      summaryFunction=twoClassSummary)
   
-train_model <- function(f, splt, ctrl) {
+train_model <- function(f, data, ctrl) {
+  splt <- split_data(f, data, .balance = T)
   model <- train(splt$train.x,
                splt$train.y, 
                method = "pls",
@@ -225,11 +239,13 @@
   )
 }
 
-splt <- split_data(f, seqs)
+old.model <- seqs %>%
+  #filter(rt_cat == "high") %>%
+  train_model(old.f, ., ctrl)
 
-  
-old.model <- train_model(old.f, splt, ctrl)
-new.model <- train_model(new.f, splt, ctrl)
+new.model <- seqs %>%
+  #filter(rt_cat == "high") %>%
+  train_model(new.f, ., ctrl)
 
 bwplot(resamples(list(old=old.model$model, new=new.model$model)))
 densityplot(resamples(list(old=old.model$model, new=new.model$model)))
@@ -240,9 +256,8 @@
 confusionMatrix(new.model$test.y, new.model$test.observed_y)
 plot(varImp(new.model$model, scale = F, useModel = F))
 
-
 roc(old.model$test.observed_y,
-    old.model$test.y_prob$mid,
+    old.model$test.y_prob$YES,
     legacy.axes=T,
     plot = T,
     lwd=2,
@@ -257,7 +272,7 @@
     ci = T)
 
 roc(new.model$test.observed_y,
-    new.model$test.y_prob$mid,
+    new.model$test.y_prob$YES,
     legacy.axes=T,
     add = T,
     plot = T,
@@ -275,7 +290,7 @@
 ```
 
 
-## Compare RT models (high vs average RT)
+## RT Analysis
 
 ```{r comapre_rt_models}
 f <- rt_cat ~ n + t + v