We have this unfortunate problem that we want to create a sankey diagram, but most of the existing packages are not feature complete enough. ggsankey is a popular solution, however, the sankey diagrams it produces leave much of the space unused when working with a large dataset and there are no customization options to change this. Then, there is the networkD3, which is actually just an R wrapper for the D3.js JavaScript library. Again, it inserts a lot of white whitespace and there are only some basic customization options.
Fortunately, I came across ggalluvial, which makes much better use of space. Technically, that is because it makes different assumptions about the data. An alluvial plot tracks changes of attributes across subjects over time, whereas a sankey diagram is designed to track track quantities. Because of this underlying assumption that there is a change of state between a and b, not a flow from a to b, we retain more control. For example, we can to some degree control attributes such as color even after we have already gone from a to b, for subsections of the cohort in b. The underlying data structure of a sankey diagram does not typically track observations across different stages, so there is no means of controlling attributes such as color once the train has left the station.
However, the inner workings are still a mystery to me, so let’s run some tests to see how this whole thing works.
Basic example
As you can see, the number of variables we need is quite high. x and y are quite straightforward.
xis simply a ordinal variable which denotes the axes.stratumthen denotes that value of the observations on that particular axis.alluviumallows for tracking observations across the axes.yallows us to set a count.
alluvial <- tribble(
~x, ~stratum, ~alluvium, ~y, ~label, ~fill, ~flow,
"first", "high", 1, 35, "high", "red", "yellow",
"second", "high", 1, 35, "high", "red", "yellow",
"first", "low", 2, 5, "low", "blue", "pink",
"second", "low", 2, 10, "low", "blue", "pink"
)
alluvial %>%
ggplot(aes(x = x, stratum = stratum, alluvium = alluvium, fill = fill, y = y)) +
geom_flow() +
geom_stratum() +
geom_text(aes(label = label), stat = "stratum")
The geoms do not respect the colors we selected for fill, so we have to go in
and fix that manually with scale_fill_manual. Still, this allow us to set
the color for flows and strata manually and independently.
Reordering flows
One awkward attribute of this data structure is that the factor order of the
factor variable stratum controls the order of variables across axes, so
management of the factor levels is important and awkward.
alluvial %>%
mutate(stratum = factor(stratum, c("low", "high"))) %>%
ggplot(aes(x = x, stratum = stratum, alluvium = alluvium, fill = fill, y = y)) +
geom_flow(aes(fill = flow), color = "black") +
geom_stratum() +
geom_text(aes(label = label), stat = "stratum")
Joining observations
We can merge different observations and then have them split off again.
Importantly, this impairs our ability to set colors. Notice how medium is
colored in in red, even though there should be blue in there, too. So the much
better option is to use stratum directly to control fill.
alluvial_merge <- tribble(
~x, ~stratum, ~alluvium, ~y, ~label, ~fill, ~flow,
"first", "high", 1, 35, "high", "red", "yellow",
"second", "medium", 1, 35, "medium", "red", "yellow",
"third", "high", 1, 35, "high", "red", "yellow",
"first", "low", 2, 10, "low", "blue", "pink",
"second", "medium", 2, 10, "medium", "red", "pink",
"third", "low", 2, 10, "low", "blue", "pink"
)
alluvial_merge %>%
mutate(stratum = factor(stratum, c("low", "high"))) %>%
ggplot(aes(x = x, stratum = stratum, alluvium = alluvium, fill = fill, y = y)) +
geom_flow(aes(fill = flow), color = "black") +
geom_stratum() +
geom_text(aes(label = label), stat = "stratum") +
scale_fill_manual(values = c("blue" = "blue", "red" = "red",
"pink" = "pink", "yellow" = "yellow"))
Joined observations are treated as the same object if they share start and end point after joining.
alluvial_distinct <- tribble(
~x, ~stratum, ~alluvium, ~attribute,
"first", "high", 1, "long",
"second", "medium", 1, "long",
"third", "medium", 1, "long",
"first", "low", 2, "short",
"second", "medium", 2, "short",
"third", "medium", 2, "short"
)
alluvial_distinct %>%
ggplot(aes(x = x, stratum = stratum, alluvium = alluvium, fill = stratum)) +
geom_flow(color = "black") +
geom_stratum() +
geom_text(aes(label = stratum), stat = "stratum")
That is a problem when we want to show in geom_flow any variable other than
the variables to the left and the right. We can use the fill argument to do
so, but we will need to use the after_stat function to make sure it’s
colored in correctly.
alluvial_distinct %>%
ggplot(aes(x = x, stratum = stratum, alluvium = alluvium, fill = attribute)) +
geom_flow(color = "black") +
geom_stratum(aes(fill = after_stat(stratum))) +
geom_text(aes(label = stratum), stat = "stratum")
Reordering
We need to figure out what determines the order of the inflow and outflow into the same stratum. Can we manipulate this somehow without changing the stratum?
stratum_control <- tribble(
~x, ~stratum, ~alluvium, ~attribute, ~label, ~trick,
"first", "high", 1, "high", "high", "high",
"second", "wide", 1, "high", "wide", "wide-low",
"third", "wide", 1, "high", "wide", "wide-low",
"first", "low", 2, "low", "low", "wide",
"second", "wide", 2, "low", "wide", "wide",
"third", "wide", 2, "low", "wide", "wide",
)
stratum_control %>%
mutate(stratum = factor(stratum, c("low", "medium", "high"))) %>%
mutate(attribute = factor(attribute, c("low", "medium", "high"))) %>%
mutate(trick = factor(trick, c("wide-low", "wide"))) %>%
ggplot(aes(x = x, stratum = stratum, alluvium = alluvium, fill = attribute,
label = label)) +
geom_flow(aes(stratum = trick), color = "black") +
geom_stratum(aes(fill = after_stat(label))) +
geom_text(stat = "stratum")
Fabulous, with this trick we can control flows independently from the axes.