strata ordering fails to respect factor levels

[corybrunson]

Originally raised in a comment on #6 . See example here.

[corybrunson]

The problem is that some strata are used at multiple axes. Inside stat_alluvium(), when to_lodes() combines factor levels to create a single column comprising all strata, these levels are unique()ed, which puts them in order of appearance; "Level 2" and "Level 3" therefore appear first, among the other strata at axis1, while the other levels at axis2 appear after them. I'm not sure what the best way is to resolve this, e.g. annotate the strata to distinguish them or keep track of their positions and restore them later.

In the meantime, @svenhalvorson, you can dodge the problem either (a) by slightly changing the names of the strata at one axis:

levels(subsank_math$`Winter Projection`) <-
  paste0(levels(subsank_math$`Winter Projection`), " ")

or (b) by assigning the same, larger set of levels to both axes:

all_levels <- apply(
  expand.grid(c(" (+)", "", " (-)"), paste("Level", 5:1)),
  1,
  function(x) paste(rev(x), collapse = "")
)
subsank_math$subgroup <-
  factor(subsank_math$subgroup, levels = all_levels)
subsank_math$`Winter Projection` <-
  factor(subsank_math$`Winter Projection`, levels = all_levels)

[corybrunson]

Current plan for to_lodes():

1. Include a parameter, maybe axis.strata, with options TRUE to paste stratum values with axis values or FALSE (default) to put combined stratum levels in order of appearance. Also allow the user to provide a vector of combined stratum levels to use in the given order?
2. If axis.strata is not set, and any stratum appears at more than one axis, then print a warning.
3. Have stat_*() functions accept axis.strata and (a) pass it to to_lodes() if data is in alluvia format or (b) print a warning if data is already in lodes format.

[corybrunson]

An attempt to resolve this issue is at the order branch. It adds a parameter relevel.strata that can be logical (ignored if FALSE) or a character vector of (a subset of) the levels in the desired order. Some examples of its use can be found in ex-alluvial-data.r (accessible via help(to_lodes)).

I'm not especially fond of this patch, since it requires passing (yet another) cumbersome parameter to every ggalluvial layer. But the problem is likely to be rare enough that it shouldn't be a problem (and it can be easily programmed). Feedback would be very welcome!

[svenhalvorson]

Thank you corybrunson! I've got some beautiful graphs out of this package!

[corybrunson]

@svenhalvorson , that's great to know! Thank you for saying so. : )

Have you tried the relevel.strata parameter (still in the order branch)? I think what might work better is to give the user a few simple options:

• NULL, the default and current behavior, to add new levels for each axis in sequence (in the order of their levels, if they are factor variables)
• "appearance" to put them in order of appearance (first along strata within axes, then along axes)
• "unique" to add a suffix (probably the axis number) to each duplicated stratum level before combining them, analogous to make.unique()

Anything more creative than that, and it's probably better for the user to handle it manually (as in the solution posted on SO). Does that sound reasonable?

[rbutleriii]

Hi @corybrunson , trying to accomplish something similar, but following along with the solutions and converting to lodes form I get errors that it isn't a recognized alluvial form, despite is_lodes_form returning true.

library(ggalluvial)
a <- data.frame(
  "Older structure" = c(
    "Class", 
    "Neighborhood", 
    "Neighborhood", 
    "Subclass", 
    "Cluster", 
    "None"
    ), 
  "Current structure" = c(
    "Neurotransmitter type", 
    "Neighborhood", 
    "Class", 
    "Subclass", 
    "Supertype", 
    "Cluster"
  ),
  N = rep(1, 6)
)
a$Older.structure <- factor(a$Older.structure, levels = unique(a$Older.structure))
a$Current.structure <- factor(a$Current.structure, levels = a$Current.structure)
ggplot(data = a, aes(axis1 = Older.structure, axis2 = Current.structure, y = N)) +
  geom_alluvium() +
  geom_stratum() +
  geom_text(stat = "stratum", aes(label = after_stat(stratum))) +
  scale_x_discrete(limits = c("Older structure", "Current structure")) +
  theme_classic() +
  ggtitle("Naming structure changes",
          "Yao et al 2021 --> Yao et al 2023")

Produces plot with second axis strata out of order:

Trying to lodes form can't get past checks

b <- to_lodes_form(a[1:2])
is_lodes_form(b, key = x, value = stratum, id = alluvium, silent = TRUE)
# [1] TRUE
ggplot(data = b, aes(x = x,  stratum = stratum, alluvium = alluvium)) +
  geom_alluvium() +
  geom_stratum() +
  geom_text(stat = "stratum", aes(label = after_stat(stratum))) +
  scale_x_discrete(limits = c("Older structure", "Current structure")) +
  theme_classic() +
  ggtitle("Naming structure changes",
          "Yao et al 2021 --> Yao et al 2023")

Error in `geom_alluvium()`:
! Problem while computing stat.
ℹ Error occurred in the 1st layer.
Caused by error in `setup_data()`:
! Data is not in a recognized alluvial form (see `help('alluvial-data')` for details).
Run `rlang::last_trace()` to see where the error occurred.

Trying in a form more like the stackoverflow version still errors at the same step

b$N <- 1
ggplot(b, aes(x = x,  stratum = stratum, alluvium = alluvium, weight = N, label = stratum)) +
  geom_flow(aes(fill = stratum)) +
  geom_stratum() +
  geom_text(stat = "stratum") +
  scale_x_discrete(limits = c("Older structure", "Current structure")) +
  theme_classic() +
  ggtitle("Naming structure changes",
          "Yao et al 2021 --> Yao et al 2023")

[corybrunson]

Hi @rbutleriii and i apologize for taking a while to respond—i've been traveling and distracted for several weeks.

I think what you need to do here is define both variables as factors with the same combined set of levels—and with the levels in the desired order. Could you give that a try and report back?

If that's not clear then let me know and i'll do an illustration using your data within a few days.