条形图，条形宽度可变，作为 x 轴上的日期范围

Question

I wish to make a bar graph where the response variable (weight change) is measured over time periods of different length, defined by a start and an end date. The width of the bars should correspond to the length of the period. A small example of my data:

wtchange.data <- structure(list(start.date = structure(1:3, .Label = c("2015-04-01", 
    "2015-04-15", "2015-04-30"), class = "factor"), end.date = structure(1:3, .Label = c("2015-04-15", 
    "2015-04-30", "2015-05-30"), class = "factor"), wtchange = c(5L, 
    10L, 15L), se = c(1.2, 2.5, 0.8)), .Names = c("start.date", "end.date", 
    "wtchange", "se"), class = "data.frame", row.names = c(NA, -3L
    ))

wtchange.data
#   start.date   end.date wtchange  se
# 1 2015-04-01 2015-04-15        5 1.2
# 2 2015-04-15 2015-04-30       10 2.5
# 3 2015-04-30 2015-05-30       15 0.8

wtchange.data$start.date <- as.Date(wtchange.data$start.date)
wtchange.data$end.date <- as.Date(wtchange.data$end.date)

Attempting to use geom_bar :

library(ggplot2)
ggplot(wtchange.data, aes(x = start.date, y = wtchange)) +
  geom_bar(stat = "identity", color = "black") +
  geom_errorbar(aes(ymin = wtchange-se, ymax = wtchange+se), width = 1)

(not allowed >2 links with <10 reputation, so can unfortunately not show the first plot)

The main problem is that when aesthetics of the plot area are defined ( x = start.date, y = wtchange ), I can use only one variable ( start.date in this example) for the x-axis, but I really need to somehow use both start.date and end.date to delimit bar widths corresponding to each period. The graph should look something like this (drawn in Paint):

A secondary problem is that the bars should touch without gaps, but I am not sure if it is even possible, given that the bars have to be of different widths, so you cannot set one bar width for all bars. Would it be possible to set width for each bar manually?

---

Edit: Thank you Henrik for the links. I have made some further progress. I calculated date midpoints for centering the bars at:

wtchange.data$date.midpoint <- wtchange.data$start.date +
(wtchange.data$end.date - wtchange.data$start.date)/2

And then calculated period lengths for using as bar widths:

wtchange.data$period.length <- wtchange.data$end.date - wtchange.data$start.date

The updated graph code is now:

ggplot(wtchange.data, aes(x = date.midpoint, y = wtchange)) +
  geom_bar(stat = "identity", color = "black", width = wtchange.data$period.length) +
  geom_errorbar(aes(ymin = wtchange-se, ymax = wtchange+se), width = 1)

The only problem remaining is that there still is a small gap between bars in one place. I guess this is due to the way R rounds date difference calculation to the nearest number of days?

Answer 1

You are right: it's the calculation of difference between end and start dates which is the reason for the gap. We need to use numeric periods instead of difftime (see explanation below) when calculating the width and the midpoint.

# length of periods, width of bars as numeric
df$width <- as.numeric(df$end.date - df$start.date) 

# mid-points
df$mid <- df$start.date + df$width / 2

# dates for breaks 
dates <- unique(c(df$start.date, df$end.date))

ggplot(df, aes(x = mid, y = wtchange)) +
  geom_bar(stat = "identity", color = "black", width = df$width) +
  geom_errorbar(aes(ymin = wtchange - se, ymax = wtchange + se), width = 1) +
  scale_x_date(breaks = dates)

---

Corresponding geom_rect code:

# mid-points
df$mid <- df$start.date + as.numeric(df$end.date - df$start.date) / 2

# dates for breaks 
dates <- unique(c(df$start.date, df$end.date))

ggplot(df, aes(x = mid, y = wtchange)) +
  geom_rect(aes(xmin = start.date, xmax = end.date, ymin = 0, ymax = wtchange), color = "black") +
  geom_errorbar(aes(ymin = wtchange - se, ymax = wtchange + se), width = 1) +
  scale_x_date(breaks = dates)

---

And slightly less ink demanding with geom_step :

# need to add an end date to the last period
df2 <- tail(df, 1)
df2$start.date <- df2$end.date
df2 <- rbind(df, df2)

# mid-points
df$mid <- df$start.date + as.numeric(df$end.date - df$start.date) / 2

ggplot() +
  geom_step(data = df2, aes(x = start.date, y = wtchange)) +
  geom_errorbar(data = df, aes(x = mid, ymin = wtchange - se, ymax = wtchange + se), width = 1) +
  scale_x_date(breaks = dates) +
  ylim(0, 16) +
  theme_bw()

---

On the " difftime issue":

Values of class Date can be represented internally as fractional days (see ?Date and ?Ops.Date ; try: Sys.Date() ; Sys.Date() + 0.5 ; Sys.Date() + 0.5 + 0.5 ). However, when adding a difftime object to a Date , the difftime object is rounded the nearest whole day (see x argument in ?Ops.Date ).

Let's check the calculations using your start date 2015-04-15 and end date 2015-04-30 :

mid <- (as.Date("2015-04-30") - as.Date("2015-04-15")) / 2
mid
# Time difference of 7.5 days

str(mid)
# Class 'difftime'  atomic [1:1] 7.5
# ..- attr(*, "units")= chr "days"

# calculate the midpoint using the difftime object
as.Date("2015-04-15") + mid
# [1] "2015-04-23"

# calculating midpoint using numeric object yields another date...
as.Date("2015-04-15") + as.numeric(mid)
# [1] "2015-04-22"

# But is "2015-04-15" above in fact fractional, i.e. "2015-04-22 point 5"?
# Let's try and add 0.5
as.Date("2015-04-15") + as.numeric(mid) + 0.5
# [1] "2015-04-23"
# Yes.

Thus, we use the numeric period, instead of the difftime period.