Creating boxplot based on some conditions - r

Question

Data given are a sample of cholesterol levels taken from 24 hospital employees who were on a standard American diet and who agreed to adopt a vegetarian diet for 1 month. Serum-cholesterol measurements were made before adopting the diet and 1 month after.
Subject Before After Difference
1 1 195 146 49
2 2 145 155 -10
3 3 205 178 27
4 4 159 146 13
5 5 244 208 36
6 6 166 147 19
7 7 250 202 48
8 8 236 215 21
9 9 192 184 8
10 10 224 208 16
11 11 238 206 32
12 12 197 169 28
13 13 169 182 -13
14 14 158 127 31
15 15 151 149 2
16 16 197 178 19
17 17 180 161 19
18 18 222 187 35
19 19 168 176 -8
20 20 168 145 23
21 21 167 154 13
22 22 161 153 8
23 23 178 137 41
24 24 137 125 12
Now here is the question I am trying to answer. Some investigators believe that the effects of diet
on cholesterol are more evident in people with high rather than low cholesterol levels. If you split the data  according to whether baseline cholesterol is above or below the median, can you comment descriptively on this issue?
Now, I am thinking of creating boxplot based on two categories here. I wish to use dplyr for data manipulation here. So, I will create a new column based on if Before is less than or greater than median of Before. So, I will have a new character vector with "high" for high Before cholesterol and low for low Before cholesterol. And, then I will do a boxplot of Difference based on the categorical new column. So, here is my code. I call the original data set as df2.
df2 %>%
mutate(new_col = if_else(Before < median(Before), "low", "high")) %>%
group_by(new_col) %>%
ggplot(aes(x= new_col, y=Difference)) +
geom_boxplot()
And following is the boxplot I get
So, based on this, I conclude that investigators are right and effects of diet on cholesterol are more evident in people with high rather than low cholesterol levels. I want to know if this can be done more effectively.

Answer 1

This is more a statistical plan question rather than a programming question, therefore it would belong more to stats.stackexchange than StackOverflow.
Anyway, categorizing a variable depending on the median is not the recommended way of visualizing associations, as you are suppressing a lot of information. You can read about this in this very good article by Peter Flom.
It is better to keep all the points and apply some spline or smoothing algorithm.
For instance, you could consider something like this:
ggplot(df2, aes(x= Before, y=Difference)) +
geom_point() +
geom_smooth()
Here, the relationship is clearly seeable, while keeping all the information you want.
If you really have to generate subgroups, you could also try something like this:
df2 %>%
mutate(new_col = if_else(Before < median(Before), "low", "high")) %>%
ggplot(aes(x= Before, y=Difference, group=new_col, color=new_col)) +
geom_point() +
geom_smooth(span=3) #try some other values here
However, using the median is still not a very good idea, especially with that amount of data points. You might want to assess the functional form of the relationship, but that would need a specific question on stats.stackexchange.com.

Answer 2

not really an answer, but more of a different approach in visualisation of the data..
library( data.table )
library( ggplot2 )
DT.melt <- melt( DT, id.vars = "Subject", measure.vars = c( "Before", "After" ) )
ggplot() +
geom_line( data = DT.melt,
aes( x = variable, y = value, group = Subject ) ) +
geom_line( data = DT.melt[, .(mean = mean(value)), by = variable ],
aes( x = variable, y = mean, group = 1 ), color = "red", size = 2 ) +
labs( x = "", y = "" )
sample data used
DT <- fread(" Subject Before After Difference
1 195 146 49
2 145 155 -10
3 205 178 27
4 159 146 13
5 244 208 36
6 166 147 19
7 250 202 48
8 236 215 21
9 192 184 8
10 224 208 16
11 238 206 32
12 197 169 28
13 169 182 -13
14 158 127 31
15 151 149 2
16 197 178 19
17 180 161 19
18 222 187 35
19 168 176 -8
20 168 145 23
21 167 154 13
22 161 153 8
23 178 137 41
24 137 125 12")