svyboxplot results change when using different categories versus subsets - r

Question

I am new to the survey package and have a mystery problem. I have made data weights using anesrake package and then created a survey design.
I have a problem when using svyboxplot and a grouping variable:
It draws similar boxplots for each grouping category, which is not true
When I studied the problem subsetting each of my category (15 of them)
the values are different for each area / different boxplots for each area.
Can anyone help me? I am desperate!
Here`s sample to test
library(tidyverse)
col <- tibble(
name = c("seura 1", "seura 2", "seura 3", "seura 4", "seura 5", "seura 6", "seura 7", "seura 8", "seura 9"
, "seura 10", "seura 11", "seura 12"),
riistakeskus = c("Keski-Suomi","Keski-Suomi","Keski-Suomi","Keski-Suomi","Keski-Suomi","Satakunta","Satakunta",
"Satakunta","Uusimaa", "Uusimaa","Uusimaa","Uusimaa"),
hirvi_sarvisuositus = c(1,4,5,3,7,5,3,4,6,5,8,9),
weights = c(1.1461438,1.1461438,1.1461438,1.1461438,1.1461438,0.5107815,0.5107815,0.5107815,2.0461937,
2.0461937,2.0461937,2.0461937)
)
library(survey)
my_des1 <- svydesign(data = col, weights = ~weights, ids = ~1)
b <- svyboxplot(hirvi_sarvisuositus~factor(riistakeskus), my_des1, all.outliers = F, ylim = c(0,10))
svyboxplot(hirvi_sarvisuositus~1, subset(my_des1, riistakeskus == "Keski-Suomi"), ylim = c(0,10))
svyboxplot(hirvi_sarvisuositus~1, subset(my_des1, riistakeskus == "Satakunta"), ylim = c(0,10))
svyboxplot(hirvi_sarvisuositus~1, subset(my_des1, riistakeskus == "Uusimaa"), ylim = c(0,10))

Answer 1

I had the same problem and would like to add to Anthonys answer, but I cannot comment yet.
There is an error in survey:::svyboxplot.default as Anthony indicates but it does not seem to have anything to do with data points. If you use keep.var = FALSE with FUN=svyquantile it does return the overall quantiles instead of the group specific quantiles.
Compare
svyby(~hirvi_sarvisuositus, ~riistakeskus, my_des1, svyquantile, ci = FALSE,
keep.var = FALSE, quantiles = c(0, 0.25, 0.5, 0.75, 1),
na.rm = TRUE)
with
svyquantile(~hirvi_sarvisuositus, my_des1,
quantiles = c(0, 0.25, 0.5, 0.75, 1),
na.rm = TRUE)
Note that svyquantile can not compute the SE for some quantiles.
If you use keep.var=TRUE instead and try to extract the CIs, you get quantiles by group.
svyby(~hirvi_sarvisuositus, ~riistakeskus, my_des1, svyquantile,
quantiles = c(0, 0.25, 0.5, 0.75, 1), ci=TRUE, na.rm = TRUE,
keep.var = TRUE, vartype = "ci")
However, you can't change the svyquantile function options when calling svyboxplot. This needs to be fixed in the package. You could built your boxplots yourself instead. A simple base R solution:
q <- svyby(~hirvi_sarvisuositus, ~riistakeskus, my_des1, svyquantile,
quantiles = c(0, 0.25, 0.5, 0.75, 1), na.rm = TRUE, ci=TRUE,
keep.var = TRUE,
vartype = "ci")
boxstats <- q[,2:6]
bxp(list(stats=t(as.matrix(boxstats)),
n = c(100,100,100),
names = rownames(boxstats)))
To prevent whiskers inside the box, you can change qrule to use a different way to calculate quantiles (e.g. qrule="hf7" for the quantile() default).
An alternative solution would be to use a weighted boxplot from ggplot2:
library(ggplot2)
ggplot(data=col, aes(y=hirvi_sarvisuositus, x=factor(riistakeskus), weight=weights)) +
geom_boxplot()
Please note that ggplot2 uses a slightly different estimation of the hinges, see help(geom_boxplot), which influences the results for low N.

Answer 2

great reproducible example, thank you! this result especially looks silly
svyboxplot(hirvi_sarvisuositus~riistakeskus,my_des1,ylim=c(0,10))
i think this largely happens because svyquantile just needs more data points to get reasonable estimates..
if you look at the code inside survey:::svyboxplot.default you can find the line that produces all of the same quantile results
svyby(~hirvi_sarvisuositus, ~riistakeskus, my_des1, svyquantile, ci = FALSE,
keep.var = FALSE, quantiles = c(0, 0.25, 0.5, 0.75, 1),
na.rm = TRUE)
not sure if this is really a bug that the survey package author would want to fix.. perhaps consider using the ?bxp function if your use case has that small of a dataset?