import polars as pl
import polars.selectors as cs
from plotnine import *
from datetime import datetimeIntroduction
This document is a submission for Posit’s 2025 Plotnine Contest.
Purpose
This page includes attempted one-to-one recreations of the visualisations in “Net favourability gap between Conservatives and Labour smallest in nearly four years” in Python and plotnine, originally published by Dylan Difford, a Junior Data Journalist at YouGov, on August 19, 2025. It shows how many of YouGov’s effective data visualisations can be recreated using the plotnine Python library.
YouGov describes itself as:
…an international online research data and analytics technology group. Our mission is to offer unparalleled insight into what the world thinks.
At the core of our platform is an ever-growing source of connected consumer data that has developed daily over our 20 years of operation. We call it living data. All of our products and services draw upon this detailed understanding of our 29+ million registered panel members to deliver accurate, actionable consumer insights.
While YouGov collects and vends data relating to a myriad of topics, there are many core data visualisations which crop up time and time again in their online articles. The article chosen for recreation shows some nice examples of typical YouGov visualisations, including line and bar charts. If you would like to recreate YouGov articles yourself, the site makes it particularly easy; all data can be obtained using the “Get the data” buttons at the bottom of each plot in a YouGov web article.
On this page, some of the original YouGov plots are found in collapsible callout boxes for easy reference, though I would recommend reading the original YouGov article to get a feel for what we’re trying to recreate.
About Me
I am a senior consultant and data analyst working for an Environmental Consultancy in South Oxfordshire in the United Kingdom.
My work typically involves writing dynamic reports, creating effective data visualisations, authoring Shiny web apps, writing and maintaining R packages, facilitating training workshops, and otherwise writing code to do interesting things with data.
I’m a collaborator on the {openair} project and the lead developer on the {openairmaps} R package.
My main programming language is R and I’ve rarely used Python in anger, so consider this a warning that there will be some references to the R programming language throughout. If I’ve made some terrible Python faux-pas anywhere in this document, please let me know!
Data Visualisation
Set-Up
We’re going to start by loading some packages:
We’ll do all of our data manipulation in
polars- it’s fast and has an appealing syntax for an R user.To help a bit, we’ll also import
polars.selectors- it’s ourtidyselectequivalent.Naturally, we’ll import
plotninefor our plots.Finally, we’ll import
datetimeas we’ll be working with dates and will need to construct and manipulate them.
We’re also going to define a function from the off. It takes a load of dates. The first time it sees a new year, it’ll create a month-year label. The next time it sees the same year, it’ll just create a month label. This works near enough the same as scales::label_date_short() in R, which I don’t believe is implemented in plotnine!
def month_labels(dates):
labels = []
seen_years = set()
for d in dates:
if d.year not in seen_years:
labels.append(d.strftime("\n%b\n %Y")) # e.g. "Feb 2025"
seen_years.add(d.year)
else:
labels.append(d.strftime("\n%b")) # e.g. "Mar"
return labelsAs previously noted, all data were obtained using the “Get the data” buttons at the bottom of each plot in the YouGov web article. Their default names are just random numbers and letters, so I’ve taken the liberty to rename them to be something more evocative.
Leader Favourability Trends (Line Charts)
YougovSee original plot
The first class of charts we’ll recreate are the favourability trends. The original article has four - one for each of Keir Starmer, Nigel Farage, Kemi Badenoch and Ed Davies. The ‘Unfavourable’ trend is given in red, the ‘Favourable’ trend in purple, and some key milestones in vertical lines. The lines also get direct labels at their ends in lieu of a proper plot legend.
As we’re making four very similar charts, I’m defining a function. First, it reads and manipulates our input data:
The first chunk of this function uses the
fileargument to read a specific CSV and do some processing; the first column gets a name, the two favourable/unfavourable columns get stacked on top of one another, the date is formatted as a date, and the overall dataset is filtered to be from 2024. This last step is to match what is displayed in the YouGov graph; interestingly you get a lot more data in the download file than what is visualised.With this newly processed data, we create a separate polars DataFrame which just includes the final value for each of ‘favourable’ and ‘unfavourable’. We need this for the ‘direct labels’ at the end of the line charts.
On the subject of the direct labels; the very last method applied in the first chunk of our function uses the
extra_separgument. As we’re using direct labels, if the ends of the lines are very close together these can potentially overlap one another. This argument lets us add a little extra spacing between the values.I happen to know that ‘unfavourable’ is always higher for these four politicians (and therefore wants that extra spacing adding) - if it was more variable I’d have to write something clever to find the higher of the two values and add to that one, but in this case we can hard-code it.
There’s also a world in which
extra_sepisn’t needed at all and it could auto-detect an appropriate gap to add if the two values are too close.We could also play around with a package like adjustText (inspired by R’s ggrepel) which can automatically reduce overlaps
All that being said, here I’d like finer control over the output so we’re doing it by hand! Note that the adjusted values are saved as their own variable as we still need the actual values to plot the lines and label with.
Then we construct our plot. The syntax is very very similar to R’s ggplot2, by design. There are some steps here that are worth further mention, however:
The direct labels on the YouGov plots have a bold title (Favourable/Unfavourable) and a normal weight value label beneath them. To my knowledge, this requires two
geom_text()calls - one withfontweight="bold"and the other without. I usednudge_yandnudge_xto ensure the labels are correctly placed relative to one another and the end of the line - the bold title is directly to the right, and the normal weight value below that. In R, we might use a package like marquee to do this in one step.A few things about axis scales:
The x- and y-axis scales can take a
range()for thebreaksargument as well as a list, which is convenient. One thing that totally caught me out as an R user wasrange(0, 70, 10)resolves to[0, 10, 20, ..., 50, 60]not[0, 10, 20, ..., 60, 70]- hence the use ofrange(0, 71, 10). In R,seq(0, 70, 10)would include70.For the x-axis we use the
start_dateargument to specify a specific start date for that axis, as they vary between politicians.The
expandargument is also used to surpress the automatic paddingplotnine(andggplot2) adds to axis ranges -[0,0]removes all of this, though the x-axis does retain some so that the direct labels are still readable.We also use our
month_labelsfunction inlabels. This works becauselabelscan take any callable that recieves the axis breaks and then does whatever arbitrary thing you’d like to them, as long as it returns a label for plotting.
Various
theme()options are added. Notably the legend is removed withlegend_position, andplot_title_positionis set to"plot". This second argument is quite subtle, but it makes the title sit on top of the y-axis labels rather than be aligned with the actual plotting area.
Have a look in each tab to see the plots.
def plot_individual_favourability(who, file, start_date, extra_sep=0):
# read data into python
plot_data = (
pl.read_csv(file)
# rename the first column to be 'date'
.rename({"X.1": "date"})
# pivot the table longer to stack the rankings
.unpivot(["Favourable", "Unfavourable"], index="date")
# coerce date to an actual date
.with_columns(pl.col("date").str.to_datetime("%d/%m/%Y"))
# to match yougov article - filter for 2024
.filter(pl.col("date").dt.year() >= 2024)
# if the fields are very close together, might need to nudge them further apart
.with_columns(
pl.when(pl.col("variable") == "Unfavourable")
.then(pl.col("value") + extra_sep)
.otherwise(pl.col("value") - extra_sep)
.alias("text_value")
)
)
# get the final values in the data - needed for direct labels
max_dates = plot_data.sort("date").group_by("variable").tail(1)
# construct plot
plot = (
ggplot(plot_data, aes(x="date", y="value", color="variable"))
# geometries
+ geom_vline(
xintercept=datetime(2024, 7, 4), linetype="dotted", color="#A6A6A6"
)
+ geom_line()
# need two texts here - one bold for the variable, one normal for the value
+ geom_text(
mapping=aes(label="variable", y="text_value"),
data=max_dates,
show_legend=False,
ha="left",
nudge_x=10,
fontweight="bold",
size=8,
)
+ geom_text(
mapping=aes(label="value", y="text_value"),
data=max_dates,
show_legend=False,
ha="left",
nudge_x=10,
nudge_y=-3,
size=8,
)
# scales
+ scale_color_manual(
values={"Favourable": "#9f29ff", "Unfavourable": "#e42119"}
)
+ scale_y_continuous(limits=[0, 70], breaks=range(0, 71, 10), expand=[0, 0])
+ scale_x_datetime(
labels=month_labels,
date_breaks="3 month",
limits=[start_date, datetime(2025, 12, 1)],
expand=[0, 0.1],
)
# themes
+ theme_minimal()
+ theme(
axis_text_y=element_text(color="#A6A6A6"),
axis_text_x=element_text(color="#7B7B7B"),
panel_grid=element_line(color="#E8E8E8"),
panel_grid_major_x=element_blank(),
panel_grid_minor_x=element_blank(),
panel_grid_minor_y=element_blank(),
axis_line_x=element_line(color="#3e3f41"),
legend_position="none",
plot_title=element_text(face="bold", size=22 / 1.5),
plot_subtitle=element_text(face="light", size=12 / 1.5),
plot_title_position="plot",
)
# labels & annotations
+ labs(
x="",
y="",
title=who + " favourability tracker",
subtitle="Do you have a favourable or unfavourable opinion of the following? ["
+ who
+ "] %\n",
)
+ annotate(
geom="text",
x=datetime(2024, 7, 10),
y=2.5,
label="General election",
size=8,
ha="left",
color="#3e3f41",
fontstyle="italic",
)
)
return plotKeir Starmer is our current Prime Minister and the leader of the nominally centre-left Labour Party. Since the election his favourability rating has been decreasing, although it has remained relatively stable in the recent months.
plot_individual_favourability(
"Keir Starmer", "assets\\data\\yougov_keir.csv", start_date=datetime(2024, 1, 1)
)
Nigel Farage is the leader of the new right-wing populist Reform UK Party. While Reform UK is a young party with very few MPs, it has had recent success in local elections and has received a lot of media attention. Nigel Farage’s favourability has been relatively stable since the general election.
plot_individual_favourability(
"Nigel Farage", "assets\\data\\yougov_nigel.csv", start_date=datetime(2024, 6, 1)
)
Kemi Badenoch is the Leader of the Opposition as well as the leader of the right-wing Conservative Party. The Conservatives (or ‘Tories’) significantly lost the last general election under the leadership of Rishi Sunak to the Labour Party after 14 years in power. Kemi Badenoch succeeded Sunak a few months later.
Kemi’s plot has an extra vertical annotation, which lets us take advantage of plotnine’s capability for “post-hoc” plot tweaking. As our function returns a plotnine object, we can add additional geom_vline()s and annotate()s to create a marker for the date at which she took over the Conservative party.
(
plot_individual_favourability(
"Kemi Badenoch",
"assets\\data\\yougov_kemi.csv",
start_date=datetime(2024, 5, 12),
)
# the function returns a plotnine object, so we can adjust it some more
+ geom_vline(xintercept=datetime(2024, 10, 31), linetype="dotted", color="#A6A6A6")
+ annotate(
geom="text",
x=datetime(2024, 11, 5),
y=1.5,
label="Becomes\nConservative leader",
size=8,
ha="left",
va="bottom",
color="#3e3f41",
fontstyle="italic",
)
)
Ed Davey is the leader of the third biggest party in the House of Commons, the more centrist Liberal Democrats. After a few years in the wilderness after an unfortunate coalition with the Tories, the Liberal Democrats performed well at the last general election, now having the largest proportion of seats in the Commons the Lib Dems have ever won.
Ed Davey’s plot is interesting in that his “Unfavourable” and “Favourable” ratings are quite similar to one another. This is why we included the extra_sep argument in the plotting function; this adds a bit of padding between the two ratings to prevent them from overlapping. As in Kemi’s plot, we can use annotate() to add the short dashed lines to help label the data lines.
(
plot_individual_favourability(
"Ed Davey",
"assets\\data\\yougov_ed.csv",
start_date=datetime(2024, 5, 12),
extra_sep=2,
)
+ annotate(
geom="segment",
x=datetime(2025, 8, 15),
xend=datetime(2025, 8, 22, 12),
y=33,
yend=35,
linetype="dotted",
color="#e42119",
)
+ annotate(
geom="segment",
x=datetime(2025, 8, 15),
xend=datetime(2025, 8, 22, 12),
y=30,
yend=28,
linetype="dotted",
color="#9f29ff",
)
)
The last plot of this type is a “net favourablity” tracker, found near the bottom of the original article. This shows that Reform UK’s net favourability, while negative, is the highest when compared to Labour and the Conservatives, who both show very similar net favourability.
This plot is broadly similar to the plots above, including the direct labels and the dotted lines seen on Ed Davey’s plot. This plot demonstrates another use of the labels argument of scales functions, which can override existing scale labels with any arbitrary strings. In this case, it is used to add the ± symbol to “0” and the + symbols to positive numbers, which are not present by default.
# read the "main parties" dataset
main_parties = (
pl.read_csv("assets\\data\\yougov_tories_labour_reform.csv")
.rename({"X.1": "date"})
.unpivot(["Conservative", "Labour", "Reform UK"], index="date")
.with_columns(pl.col("date").str.to_datetime("%d/%m/%Y"))
.filter(pl.col("value").is_not_null())
# again, we need to nudge the data around a bit
.with_columns(
pl.when(pl.col("variable") == "Labour")
.then(pl.col("value") + 2)
.when(pl.col("variable") == "Conservative")
.then(pl.col("value") - 2)
.otherwise(pl.col("value"))
.alias("text_value")
)
)
# needed for plotting
max_dates = main_parties.sort("date").group_by("variable").tail(1)
# axis labels needed here! Needs "+" for positive values, and "+/-" for zero
numbers = [
f"+{i}" if i > 0 else f"±{abs(i)}" if i == 0 else str(i) for i in range(-60, 21, 10)
]
# construct plot
(
ggplot(main_parties, aes(x="date", y="value", color="variable"))
# geometries
+ geom_hline(yintercept=0, color="#3e3f41")
+ geom_line()
+ geom_text(
mapping=aes(label="variable", y="text_value"),
data=max_dates,
show_legend=False,
ha="left",
nudge_x=30,
nudge_y=0,
fontweight="bold",
size=8,
)
+ geom_text(
mapping=aes(label="value", y="text_value"),
data=max_dates,
show_legend=False,
ha="left",
nudge_x=30,
nudge_y=-3,
size=8,
)
# scales
+ scale_x_datetime(
date_breaks="1 year",
date_labels="%b %Y",
limits=[datetime(2020, 1, 1, 0), datetime(2026, 8, 1)],
expand=[0, 0.1],
)
+ scale_y_continuous(
limits=(-65, 20), breaks=range(-60, 21, 10), labels=numbers, expand=(0, 0)
)
+ scale_color_manual(
values={"Reform UK": "#0082c7", "Labour": "#c20800", "Conservative": "#003cab"}
)
# themes
+ theme_minimal()
+ theme(
axis_text_y=element_text(color="#A6A6A6"),
axis_text_x=element_text(color="#7B7B7B"),
panel_grid=element_line(color="#E8E8E8"),
panel_grid_major_x=element_blank(),
panel_grid_minor_x=element_blank(),
panel_grid_minor_y=element_blank(),
legend_position="none",
plot_title=element_text(face="bold", size=22 / 1.5),
plot_subtitle=element_text(face="light", size=12 / 1.5),
plot_title_position="plot",
)
# labels and annotations
+ labs(
x="",
y="",
title="Net favourability gap between Tories and Labour is at\nlowest level since September 2021, though both trail\nReform UK",
subtitle="Do you have a favourable or unfavourable opinion of the following? [Net score]\n",
)
+ annotate(
geom="segment",
x=datetime(2025, 8, 15),
xend=datetime(2025, 9, 5),
y=-37,
yend=-35,
linetype="dotted",
color="#c20800",
)
+ annotate(
geom="segment",
x=datetime(2025, 8, 15),
xend=datetime(2025, 9, 5),
y=-39,
yend=-41,
linetype="dotted",
color="#003cab",
)
)
Leader Favourability By Party (Bar Charts)
YougovSee original plot
This next collection of plots are quite appealing bar charts showing overall favourability for different national and international political figures by UK political affiliation. We might call these kinds of plots ‘small multiples’. On the face of them they look quite simple, but these plots have quite a lot of features that make recreation with plotnine (or even ggplot2) tricky:
You could create a
facet_grid(), but the “row” labels are on top of each row, not to the left or right of it.There are direct value labels, but whether they are to the right of the bar or left aligned to the whole plotting area depends on the value being presented.
When the direct value labels are on top of the bar, they need to be a colour with sufficient contrast to be read.
The plots have overall titles and, in one instance, a caption.
So, the strategy here is to actually create multiple individual plots and then assemble them together. plotnine has a basic implementation of some of the functionality of R’s patchwork, which lets you assemble plots with simple mathematical operators like +, | and /.
Once again we’ve defined a function. Some notes:
The
is_green_onlyargument exists because there’s a specific plot which only contains two categories (green voters vs all britons) which affects some of the internal plotting parameters like how much to nudge labels by and the colours to use.We use multiple
geom_text()calls - one for labels below a certian threshold that go to the right of the bar, and one for labels above that threshold that are left-aligned on the whole plot. In the former, the colour is a dark grey. For the latter, the colour needs to vary based on the political party, so it is mapped to a variable. We create a colourscalethat’s purely for the text colour. This is done manually, but one could imagine writing a function to pick the “best contrast” for the party colours (in R I’d useprismatic::best_contrast()for this; I’m sure there’ll be a Python equivalent).I’ve used
coord_flip(). Inggplot2the plot automatically detects the orientation of bars/boxplots/etc.; here I found I needed to have the continuous variable be on the “y” forgeom_col()and then flip the coordinate after-the-fact.The politician’s name is used as a subtitle, which will make sense when multiples of these plots are put together!
def plot_leader_favourability(data, who, is_green_only=False):
plot_data = data.filter(pl.col("who") == who)
# Set thresholds and colors based on plot type
if is_green_only:
text_threshold = 5
text_y_pos = 1
nudge_y = 1
fill_colors = {"All Britons": "#9f29ff", "2024 Green voters": "#31caa8"}
text_colors = {"All Britons": "white", "2024 Green voters": "#333333"}
else:
text_threshold = 30
text_y_pos = 5
nudge_y = 5
fill_colors = {
"All Britons": "#9f29ff",
"2024\nGreen": "#31caa8",
"2024\nLabour": "#c20800",
"2024\nLib Dem": "#ffba22",
"2024\nConservative": "#003cab",
"2024\nReform UK": "#06a6ee",
}
text_colors = {
"All Britons": "white",
"2024\nGreen": "#333333",
"2024\nLabour": "white",
"2024\nLib Dem": "#333333",
"2024\nConservative": "white",
"2024\nReform UK": "white",
}
plot = (
ggplot(plot_data, aes(y="value", x="what"))
+ geom_col(aes(y=100), fill="#3333330d")
+ geom_col(aes(fill="variable"))
+ geom_text(
data=plot_data.filter(pl.col("value") > text_threshold),
mapping=aes(label="value", y=text_y_pos, color="variable"),
ha="left",
size=8,
)
+ geom_text(
data=plot_data.filter(pl.col("value") <= text_threshold),
mapping=aes(label="value", y="value"),
ha="left",
size=8,
color="#333333",
nudge_y=nudge_y,
)
+ coord_flip()
+ facet_grid(cols="variable")
+ scale_y_continuous(limits=[0, 100])
+ scale_fill_manual(values=fill_colors)
+ scale_color_manual(values=text_colors)
+ theme_minimal()
+ theme(
axis_text_x=element_blank(),
panel_grid=element_blank(),
legend_position="none",
axis_text_y=element_text(ha="left"),
strip_text=element_text(ha="left"),
plot_subtitle=element_text(face="bold", size=12 / 1.5),
plot_title_position="plot",
)
+ labs(x="", y="", subtitle=who)
)
return plotThis plot compares the favourability of Keir Starmer to the fledgling left-wing “Your Party” run by Jeremy Corbyn and Zarah Sultana, both historically on the Labour hard-left.
To construct the overall assembly title I created a dummy plot that is effectively only a title and subtitle, and added this to the plot assembly. Patchwork has the plot_annotation() function that doesn’t seem to be mirrored in plotnine yet, so this was the workaround I found!
your_party_leaders = (
pl.read_csv("assets\\data\\yougov_leftwing.csv")
.rename({"X.1": "who", "X.2": "what"})
.unpivot(index=["who", "what"], on=["All Britons", cs.starts_with("2024")])
.with_columns(pl.col("variable").str.replace("<br>", "\n"))
.with_columns(
pl.col("variable").cast(pl.Categorical),
pl.col("what").cast(pl.Enum(["Unfavourable", "Favourable"])),
)
)
(
ggplot()
+ theme_void()
+ labs(
title="How do Britons' attitudes towards Corbyn and Sultana\ncompare to their views on Starmer?",
subtitle="Do you have a favourable or unfavourable opinion of the following? %",
)
+ theme(
plot_title=element_text(face="bold", size=22 / 1.5),
plot_subtitle=element_text(face="light", size=12 / 1.5),
plot_title_position="plot",
aspect_ratio=0.001,
plot_margin=0,
)
) / plot_leader_favourability(
your_party_leaders, "Jeremy Corbyn"
) / plot_leader_favourability(
your_party_leaders, "Zarah Sultana"
) / plot_leader_favourability(
your_party_leaders, "Keir Starmer"
)
This plot is much the same as the domestic leaders, but shows various international political figures - President of Ukraine Volodymyr Zelenskyy, US President Donald Trump, US Vice President JD Vance, and President of Russia Vladimir Putin.
international_leaders = (
pl.read_csv("assets\\data\\yougov_international.csv")
.rename({"X.1": "who", "X.2": "what"})
.unpivot(index=["who", "what"], on=["All Britons", cs.starts_with("2024")])
.with_columns(pl.col("variable").str.replace("<br>", "\n"))
.with_columns(
pl.col("variable").cast(pl.Categorical),
pl.col("what").cast(pl.Enum(["Unfavourable", "Favourable"])),
)
)
(
ggplot()
+ theme_void()
+ labs(
title="YouGov international favourability ratings, August 2025",
subtitle="Do you have a favourable or unfavourable opinion of the following? %",
)
+ theme(
plot_title=element_text(face="bold", size=22 / 1.5),
plot_subtitle=element_text(face="light", size=12 / 1.5),
plot_title_position="plot",
aspect_ratio=0.001,
plot_margin=0,
)
) / plot_leader_favourability(
international_leaders, "Volodymyr Zelenskyy (net +49)"
) / plot_leader_favourability(
international_leaders, "Donald Trump (net -61)"
) / plot_leader_favourability(
international_leaders, "JD Vance (net -55)"
) / plot_leader_favourability(
international_leaders, "Vladimir Putin (net -86)"
)
At the original time of writing of the YouGov article, a leadership election was occurring in the UK Green Party; our green, left-wing party which has historically had little representation in the House of Commons. Britons were surveyed on their opinions of various Green leadership figures (and a fictitious candidate to guage reflexive political opinions).
This plot is cut from the same cloth as the others but the overall layout is somewhat different, hence the need for the different parameters to control label placement and so on. Here we add an additional mock caption to explain the non-existent Andrew Farmer MP.
green_party_leaders = (
pl.read_csv("assets\\data\\yougov_green.csv")
.rename({"X.1": "who", "X.2": "what"})
.unpivot(index=["who", "what"], on=["All Britons", cs.starts_with("2024")])
.with_columns(
pl.col("variable").cast(pl.Categorical),
pl.col("what").cast(pl.Enum(["Don't know", "Unfavourable", "Favourable"])),
)
)
(
ggplot()
+ theme_void()
+ labs(
title="Green leadership figures largely unknown, even by\nthose who have voted for the party",
subtitle="Do you have a favourable or unfavourable opinion of the following? %",
)
+ theme(
plot_title=element_text(face="bold", size=22 / 1.5),
plot_subtitle=element_text(face="light", size=12 / 1.5),
plot_title_position="plot",
aspect_ratio=0.001,
plot_margin=0,
)
) / plot_leader_favourability(
green_party_leaders, "Carla Denyer", is_green_only=True
) / plot_leader_favourability(
green_party_leaders, "Adrian Ramsay", is_green_only=True
) / plot_leader_favourability(
green_party_leaders, "Zack Polanski", is_green_only=True
) / plot_leader_favourability(
green_party_leaders, "Ellie Chowns", is_green_only=True
) / plot_leader_favourability(
green_party_leaders, "Andrew Farmer*", is_green_only=True
) / (
ggplot()
+ theme_void()
+ labs(
caption="* Andrew Farmer is a fake politician, used to test how many respondents reflexively say they have an opinion of a\nnon-existent figure"
)
+ theme(
plot_caption=element_text(ha="left", face="light", size=12 / 1.5),
plot_caption_position="plot",
aspect_ratio=0.001,
plot_margin=0,
)
) + theme(figure_size=(7, 7))
Overall Senior Political Figure Favourability (Stacked Bars)
YougovSee original plot
Frankly, this is where the wheels start falling off!
This is a big summary stacked bar chart which shows a Favourable, Unfavourable and “Don’t Know” score for a load of different political figures, both domestic and international. On first glance, you might assume we could do what we did for the previous bar charts. However, the height of each “subplot” would be different as the numbers of figures in each category are different. Adjusting the heights of the different subplots would be difficult, so we’re just going to have to use facet function to get something approximating the original plot.
In recent versions of ggplot2 we could use facet_wrap() with one column of facets and a combination of the space and scales arguments to create something very similar to the YouGov plot. In older version of ggplot2, we could use ggforce::facet_col() to achieve much the same thing. No similar functionality seems to exist in plotnine, so we’re going to stick with facet_grid() and live with having the subplot labels be on the right hand side of the plot.
A couple of curious bugs or limitations of plotnine emerged when trying this:
I could not seem to left-align the facet labels if there were new lines (
\n) within them.Left-aligning the y-axis labels placed the start point in different places for the different facets. This isn’t ideal and makes the plot look less cohesive, and obviously doesn’t match YouGov’s formatting.
Anyway, the Python code below shares many of the same themes as the other plots, with some new strategies needed for this plot:
There’s a lot of
pl.Enum()to ensure categories appear in the correct order; this is similar to usingfactor()in R. To order the leaders by favourability was quite round-the-houses; in R I’d have usedforcats::fct_reorder()to do it in one line.To achieve the correct value labels I used
position_stack()with thevjustargument. Very small value labels don’t appear in the data, so I filtered the dataset within thegeom_text()call. It’s not possible to use anudgeargument withposition_stack(), and adding a value to the data directly will throw offposition_stack()s calculations, so I’ve cheated and just added a space (" ") to the beginning of the label.The x scale extends to
101- this isn’t a funky Python counting thing this time, though - the YouGov plot has this ragged edge where, I imagine, values have been rounded and now don’t always add up to100.
# read the leaders dataset
leaders = (
pl.read_csv("assets\\data\\yougov_politicians.csv")
.rename({"X.1": "category", "X.2": "who"})
.unpivot(["Favourable", "Don't know", "Unfavourable"], index=["category", "who"])
.with_columns(
# need to order the variable
pl.col("variable").cast(pl.Enum(["Unfavourable", "Don't know", "Favourable"])),
# Stack categories by replacing spaces with new lines, then order to match yougov plot
pl.col("category")
.str.replace_all(" ", "\n")
.cast(
pl.Enum(
["Party\nleaders", "Other\nsenior\npoliticians", "Political\nparties"]
)
),
)
)
# get the order of the leaders based on favourability
ordered_leaders = (
leaders.filter(pl.col("variable") == "Favourable")
.sort("value")
.get_column("who")
.to_list()
)
# order the leaders column to by favourability
leaders = leaders.with_columns(pl.col("who").cast(pl.Enum(ordered_leaders)))
# construct plot
(
ggplot(leaders, aes(y="value", x="who"))
# geometries
+ geom_col(aes(fill="variable"))
+ geom_text(
data=leaders.with_columns(
pl.when(pl.col("value") <= 3)
.then(pl.lit(""))
.otherwise(" " + pl.col("value").cast(pl.String))
.alias("value_str")
),
mapping=aes(y="value", label="value_str", group="variable", color="variable"),
position=position_stack(vjust=0),
size=8,
ha="left",
show_legend=False,
)
# facet
+ facet_grid(rows="category", scales="free_y", space="free")
# scales
+ coord_flip()
+ scale_fill_manual(
values={
"Favourable": "#9f29ff",
"Don't know": "#ccd1db",
"Unfavourable": "#ff412c",
},
breaks=["Favourable", "Don't know", "Unfavourable"],
)
+ scale_color_manual(
values={
"Favourable": "white",
"Don't know": "#333333",
"Unfavourable": "white",
}
)
+ scale_y_continuous(limits=[0, 101])
# themes
+ theme_minimal()
+ theme(
axis_text_x=element_blank(),
panel_grid=element_blank(),
legend_position="top",
axis_text_y=element_text(ha="left"),
strip_text=element_text(ha="left", face="bold", rotation=0),
plot_title=element_text(face="bold", size=22 / 1.5),
plot_subtitle=element_text(face="light", size=12 / 1.5),
plot_title_position="plot",
plot_caption=element_text(ha="left", face="light", size=12 / 1.5),
plot_caption_position="plot",
panel_spacing=0.05,
figure_size=(7, 7),
)
# labels
+ labs(
x="",
y="",
fill="",
title="YouGov political favourability ratings, August 2025",
subtitle="Do you have a favourable or unfavourable opinion of the following? %",
caption="* Andrew Farmer is a fake politician, used to test how many respondents reflexively say they have an opinion of a\nnon-existent figure",
)
)
Now these limitations and shortcomings are somewhat frustrating and, just to get this as a “win”, I have also produced this final plot from the article in R’s ggplot2 below. Two points to note here, though:
plotnineis relatively new, so its no surprise it can’t do everythingggplot2can.ggplot2is old enough to drink in my country, whereasplotnineis only half-way through primary school, so its quite impressive thatplotninecan get as close as it does on its own.The
ggplot2implementation still doesn’t quite look right - the category labels aren’t aligned with the plot. You can get part of the way there with the arguably off-label move of passing a negative value tohjustintheme(strip.text), or could split the plots up andpatchworkthem back together (although you’d need to do the maths on the bar heights so they’re all consistent, and again that’s not convenient).
TipLet’s see it in R
library(ggplot2)
readr::read_csv("assets/data/yougov_politicians.csv") |>
dplyr::rename(category = X.1, who = X.2) |>
tidyr::pivot_longer(-(1:2)) |>
dplyr::mutate(
favourable = ifelse(name == "Favourable", value, NA),
who = forcats::fct_reorder(who, favourable),
name = factor(name, c("Unfavourable", "Don't know", "Favourable")),
category = factor(category, c("Party leaders", "Other senior politicians", "Political parties"))
) |>
ggplot(aes(x = value, y = who)) +
geom_col(aes(fill = name)) +
geom_text(
aes(label = ifelse(value < 5, "", paste(" ", value)), group = name, color = name),
hjust = 0,
position = position_stack(vjust = 0),
show.legend = FALSE
) +
theme_minimal() +
theme(
axis.text.y = element_text(hjust = 0),
axis.text.x = element_blank(),
strip.text = element_text(hjust = 0, face = "bold"),
strip.clip = "off",
panel.grid = element_blank(),
legend.position = "top",
legend.justification = "left",
plot.title = element_text(size = 22, face = "bold"),
plot.subtitle = element_text(size = 12),
plot.title.position = "plot",
plot.caption = element_text(size = 10, hjust = 0, face = "italic"),
plot.caption.position = "plot"
) +
facet_wrap(vars(category), scales = "free_y", space = "free_y") +
coord_cartesian(clip = "off") +
scale_x_continuous(expand = expansion()) +
labs(
y = NULL,
x = NULL,
fill = NULL,
title = "YouGov political favourability ratings, August 2025",
subtitle = "Do you have a favourable or unfavourable opinion of the following? %",
caption = "* Andrew Farmer is a fake politician, used to test how many respondents reflexively say they have an opinion of a non-\nexistent figure"
) +
scale_fill_manual(
values = c(
"Favourable" = "#9f29ff",
"Don't know" = "#ccd1db",
"Unfavourable" = "#ff412c"
),
breaks = c("Favourable", "Don't know", "Unfavourable")
) +
scale_color_manual(
values = c(
"Favourable" = "white",
"Don't know" = "#333333",
"Unfavourable" = "white"
),
breaks = c("Favourable", "Don't know", "Unfavourable")
)
ggsave("assets/media/R_plot.png", width = 8, height = 10, dpi = 300, device = "png")
Wrap-up
Good news - that’s all of the plots! To close, I just want to give a few thoughts to summarise as an R user coming to plotnine and Python more generally:
The syntax of
plotnineis truly almost identical toggplot2, so it really is easy to move from one to the other. Yes, the odd argument is named slightly differently (e.g.,haandvato align text), but these instances are few and far between. Despite never really writing any proper Python in the past, betweenplotnineandpolarsit really wasn’t a struggle. Speaking of…polarsfeels nice to use in a waypandasnever quite did. It really does feel close todplyr- it just, by necessity, lacks the convenience brought by non-standard evaluation in R (e.g., needing to saypl.with_columns(pl.col("x"))rather than justmutate(x)).One of the best things about
ggplot2is its extensibility: the ecosystem of custom stats, geoms, facets, and scales gives you an incredible range of options for building exactly the plots you need. I’m not yet sure how extensibleplotnineis, but I imagine many of these capabilities will arrive in time.Speaking as someone who works with a lot of weather data, I’d love
coord_polar()to come toplotnineso I can make a wind rose, please! 🙂When learning a new tool, the challenge isn’t just understanding how it works - it’s also figuring out what to do with it. Recreating existing visualisations, especially those based on freely available data (like YouGov’s), is a great way to cut through that problem. You have a clear objective, real data, and a concrete end result to aim for.