[WIP] GAM Tract Profile Tutorial

examples/howto_examples/gam-tract-profiles-example.Rmd

@@ -0,0 +1,101 @@
+---
+title: "AFQ-GAMS-Tutorial"
+author: "Jason Yeatman"
+date: "2024-03-30"
+output: html_document
+---
+
+```{r setup, include=FALSE}
+knitr::opts_chunk$set(echo = TRUE)
+```
+
+## GAMs Tutorial
+
+First we'll pull the Sarica data and merge nodes.csv (diffusion data) with subjects.csv (participant meta data)
+
+```{r, warning=FALSE, message=FALSE}
+library(dplyr)
+library(mgcv)
+library(gratia)
+library(ggplot2)
+df <- read.csv('https://raw.githubusercontent.com/yeatmanlab/Sarica_2017/gh-pages/data/nodes.csv')
+df <- inner_join(df, read.csv('https://raw.githubusercontent.com/yeatmanlab/Sarica_2017/gh-pages/data/subjects.csv'))
+mutate(df, class = factor(class), subjectID = factor(subjectID))-> df
+
+```
+
+```{r, message=FALSE, warning=FALSE}
+# Fit a simple GAM with a random effect of subject and smoothers that 
+# differ by class (control or ALS)
+g1 <- gam(md ~ s(nodeID,class, bs='fs') + s(subjectID, bs= 're'), data=df)
+# Get smooth estimates
+dfs <- smooth_estimates(g1) %>% add_confint(coverage=.68)
+
+# Remove random effects and plot
+filter(dfs, type != 'Random effect') %>%
+ggplot(aes(x=nodeID,y=est, color=class)) +
+  geom_line()  +
+  geom_ribbon(aes(ymin = lower_ci, ymax=upper_ci),alpha = 0.2 )
+
+```
+
+```{r, message=FALSE, warning=FALSE}
+# The above plot looks good but now let's let the curves vary across # subjects too. What I notice here is that it dramatically effects the standard error
+g2 <- gam(md ~ s(nodeID,class, bs='fs') + s(nodeID, subjectID, bs= 're'), data=df)
+# Get smooth estimates
+dfs <- smooth_estimates(g2) %>% add_confint(coverage=.68)
+
+# Remove random effects and plot
+filter(dfs, type != 'Random effect') %>%
+ggplot(aes(x=nodeID,y=est, color=class)) +
+  geom_line()  +
+  geom_ribbon(aes(ymin = lower_ci, ymax=upper_ci),alpha = 0.2 )
+```
+
+```{r, message=FALSE, warning=FALSE}
+# Yet another way to do the same thing and I don't get the difference
+g3 <- gam(md ~ s(nodeID,by=class, bs='fs', k=10) + s(subjectID, bs= 're'), data=df)
+# Get smooth estimates
+dfs <- smooth_estimates(g3) %>% add_confint(coverage=.68)
+
+# Remove random effects and plot
+filter(dfs, type != 'Random effect') %>%
+ggplot(aes(x=nodeID,y=est, color=class)) +
+  geom_line()  +
+  geom_ribbon(aes(ymin = lower_ci, ymax=upper_ci),alpha = 0.2 )
+```
+
+Now calculate differences between the smooths for the different factors and plot out the difference an 95% CI
+
+```{r}
+ds<-difference_smooths(g3, smooth="s(nodeID)")
+ggplot(ds,aes(x=nodeID,y=diff)) +
+  geom_line()  +
+  geom_ribbon(aes(ymin = lower, ymax=upper),alpha = 0.2 ) +
+  geom_hline(yintercept=0)
+```
+
+Now look at continuous measures along the tract profile
+
+```{r}
+# Yet another way to do the same thing and I don't get the difference
+g4 <- gam(md ~  s(nodeID,ALSFRSbulbar, bs='fs', k=10) + s(subjectID, bs= 're'), data=df)
+dfs <- smooth_estimates(g4) %>% add_confint(coverage=.68)
+#filter(dfs,smooth=='s(nodeID)') %>%
+#  ggplot(aes(x=nodeID,y=est)) +
+#  geom_line()  +
+#  geom_ribbon(aes(ymin = lower_ci, ymax=upper_ci),alpha = 0.2 )
+
+# 2d plot showing how tract profiles vary as a function of AFS bulbar
+filter(dfs,smooth=='s(nodeID,ALSFRSbulbar)') %>%
+ggplot(aes(x=nodeID, y=ALSFRSbulbar))+
+  geom_raster(aes(fill=est))
+
+# Now plot separate tract profiles by bulbar
+dfs %>% mutate(ALSFRSbulbar.factor=factor(ALSFRSbulbar)) %>% filter(ALSFRSbulbar== sort(unique(dfs$ALSFRSbulbar))[2] | ALSFRSbulbar== sort(unique(dfs$ALSFRSbulbar))[99]) %>% 
+  filter(smooth=='s(nodeID,ALSFRSbulbar)') %>%
+  ggplot(aes(x=nodeID,y=est,color=ALSFRSbulbar.factor)) +
+  geom_line(show.legend = FALSE) +
+  geom_ribbon(aes(ymin = lower_ci, ymax=upper_ci),alpha = 0.2 )
+
+```