28  Lab: Axes manipulation

One way to create potentially misleading visualisations is by manipulating the axes of a plot. Here we illustrate these using one of the FiveThirtyEight data sets, which are available here.

28.1 Data wrangling

We are going to use polls from the 2020 USA presidential election. As before, we load and examine the data.

import pandas as pd 
import seaborn as sns
import altair as alt 

df_polls = pd.read_csv('data/presidential_poll_averages_2020.csv')
df_polls.head()
cycle state modeldate candidate_name pct_estimate pct_trend_adjusted
0 2020 Wyoming 11/3/2020 Joseph R. Biden Jr. 30.81486 30.82599
1 2020 Wisconsin 11/3/2020 Joseph R. Biden Jr. 52.12642 52.09584
2 2020 West Virginia 11/3/2020 Joseph R. Biden Jr. 33.49125 33.51517
3 2020 Washington 11/3/2020 Joseph R. Biden Jr. 59.34201 59.39408
4 2020 Virginia 11/3/2020 Joseph R. Biden Jr. 53.74120 53.72101

For our analysis, we are going to pick estimates from 11/3/2020 for the swing states of Florida, Texas, Arizona, Michigan, Minnesota and Pennsylvania.

df_nov = df_polls[
    (df_polls.modeldate == '11/3/2020')
]

df_nov = df_nov[
    (df_nov.candidate_name == 'Joseph R. Biden Jr.') |
    (df_nov.candidate_name == 'Donald Trump')
]

df_swing = df_nov[
    (df_nov['state'] == 'Florida') |
    (df_nov['state'] == 'Texas' ) |
    (df_nov['state'] == 'Arizona' ) |
    (df_nov['state'] == 'Michigan' ) |
    (df_nov['state'] == 'Minnesota' ) |
    (df_nov['state'] == 'Pennsylvania' ) 
]

df_swing
cycle state modeldate candidate_name pct_estimate pct_trend_adjusted
7 2020 Texas 11/3/2020 Joseph R. Biden Jr. 47.46643 47.44781
12 2020 Pennsylvania 11/3/2020 Joseph R. Biden Jr. 50.22000 50.20422
30 2020 Minnesota 11/3/2020 Joseph R. Biden Jr. 51.86992 51.84517
31 2020 Michigan 11/3/2020 Joseph R. Biden Jr. 51.17806 51.15482
46 2020 Florida 11/3/2020 Joseph R. Biden Jr. 49.09162 49.08035
53 2020 Arizona 11/3/2020 Joseph R. Biden Jr. 48.72237 48.70539
63 2020 Texas 11/3/2020 Donald Trump 48.57118 48.58794
68 2020 Pennsylvania 11/3/2020 Donald Trump 45.57216 45.55034
86 2020 Minnesota 11/3/2020 Donald Trump 42.63638 42.66826
87 2020 Michigan 11/3/2020 Donald Trump 43.20577 43.23326
102 2020 Florida 11/3/2020 Donald Trump 46.68101 46.61909
109 2020 Arizona 11/3/2020 Donald Trump 46.11074 46.10181

28.2 Default barplot

We can look at the relative performance of the candidates within each state using a nested bar plot.

ax = sns.barplot(
    data = df_swing, 
    x = 'state', 
    y = 'pct_estimate', 
    hue = 'candidate_name')

28.3 Altering the axes

Altering the axis increases the distance between the bars. Some might say that is misleading.

ax = sns.barplot(
    data = df_swing, 
    x = 'state', 
    y = 'pct_estimate', 
    hue = 'candidate_name')

ax.set(ylim=(41, 52))

What do you think?

How about if we instead put the data on the full 0 to 100 scale?

ax = sns.barplot(
    data = df_swing, 
    x = 'state', 
    y = 'pct_estimate', 
    hue = 'candidate_name')

ax.set(ylim=(0, 100))

We can do the same thing in Altair.

alt.Chart(df_swing).mark_bar().encode(
    x='candidate_name',
    y='pct_estimate',
    color='candidate_name',
    column = alt.Column('state:O', spacing = 5, header = alt.Header(labelOrient = "bottom")),
)

Note the need for the alt column. What happens if you do not provide an alt column?

Passing the domain option to the scale of the Y axis allows us to choose the y axis range.

alt.Chart(df_swing).mark_bar(clip=True).encode(
    x='candidate_name',
    y=alt.Y('pct_estimate', scale=alt.Scale(domain=[42,53])),
    color='candidate_name',
    column = alt.Column('state:O', spacing = 5, header = alt.Header(labelOrient = "bottom")),
)

28.4 Altering the proportions

We can even be a bit tricky and stretch out the difference.

alt.Chart(df_swing).mark_bar(clip=True).encode(
    x='candidate_name',
    y=alt.Y('pct_estimate', scale=alt.Scale(domain=[42,53])),
    color='candidate_name',
    column = alt.Column('state:O', spacing = 5, header = alt.Header(labelOrient = "bottom")),
).properties(
    width=20,
    height=600
)

28.5 Default line plot

It is not just bar plot that you can have fun with. Line plots are another interesting example.

For our simple line plot, we will need the poll data for a single state.

df_texas = df_polls[
    df_polls['state'] == 'Texas'
]

df_texas_bt = df_texas[
    (df_texas['candidate_name'] == 'Donald Trump') |
    (df_texas['candidate_name'] == 'Joseph R. Biden Jr.')
]

df_texas_bt.head()
cycle state modeldate candidate_name pct_estimate pct_trend_adjusted
7 2020 Texas 11/3/2020 Joseph R. Biden Jr. 47.46643 47.44781
63 2020 Texas 11/3/2020 Donald Trump 48.57118 48.58794
231 2020 Texas 11/2/2020 Joseph R. Biden Jr. 47.46643 47.44781
287 2020 Texas 11/2/2020 Donald Trump 48.57118 48.58794
455 2020 Texas 11/1/2020 Joseph R. Biden Jr. 47.45590 47.43400

The modeldate column is a string (object) and not date time. So we need to change that: we will create a new datetime column called date.

df_texas_bt['date'] = pd.to_datetime(df_texas_bt.loc[:,'modeldate'], format='%m/%d/%Y').copy()

df_texas_bt.dtypes
cycle                          int64
state                         object
modeldate                     object
candidate_name                object
pct_estimate                 float64
pct_trend_adjusted           float64
date                  datetime64[ns]
dtype: object

Create our line plot.

alt.Chart(df_texas_bt).mark_line().encode(
    y=alt.Y('pct_estimate', scale=alt.Scale(domain=[42,53])),
    x='date',
    color='candidate_name')

Sometimes multiple axis are used for each line, or in a combined line and bar plot.

The example here uses a dataframe with a column for each line. Our data does not have that.

df_texas_bt
our_df = df_texas_bt[['candidate_name', 'pct_estimate', 'date']]
our_df
candidate_name pct_estimate date
7 Joseph R. Biden Jr. 47.46643 2020-11-03
63 Donald Trump 48.57118 2020-11-03
231 Joseph R. Biden Jr. 47.46643 2020-11-02
287 Donald Trump 48.57118 2020-11-02
455 Joseph R. Biden Jr. 47.45590 2020-11-01
... ... ... ...
28931 Donald Trump 49.09724 2020-02-29
28963 Joseph R. Biden Jr. 45.30901 2020-02-28
28995 Donald Trump 49.09676 2020-02-28
29027 Joseph R. Biden Jr. 45.30089 2020-02-27
29058 Donald Trump 49.07925 2020-02-27

502 rows × 3 columns

Pivot table allows us to reshape our dataframe.

our_df = pd.pivot_table(our_df, index=['date'], columns = 'candidate_name')
our_df.columns = our_df.columns.to_series().str.join('_')
our_df.head()
pct_estimate_Donald Trump pct_estimate_Joseph R. Biden Jr.
date
2020-02-27 49.07925 45.30089
2020-02-28 49.09676 45.30901
2020-02-29 49.09724 45.30896
2020-03-01 49.09724 45.30895
2020-03-02 48.91861 45.37694

Date here is the dataframe index. We want it to be a column.

our_df['date1'] = our_df.index
our_df.columns = ['Trump', 'Biden', 'date1']
our_df.head()
Trump Biden date1
date
2020-02-27 49.07925 45.30089 2020-02-27
2020-02-28 49.09676 45.30901 2020-02-28
2020-02-29 49.09724 45.30896 2020-02-29
2020-03-01 49.09724 45.30895 2020-03-01
2020-03-02 48.91861 45.37694 2020-03-02

Creating our new plot, to fool all those people who expect Trump to win in Texas.

base = alt.Chart(our_df).encode(
        alt.X('date1')
)

line_A = base.mark_line(color='#5276A7').encode(
    alt.Y('Trump', axis=alt.Axis(titleColor='#5276A7'), scale=alt.Scale(domain=[42,53]))
)

line_B = base.mark_line(color='#F18727').encode(
    alt.Y('Biden', axis=alt.Axis(titleColor='#F18727'), scale=alt.Scale(domain=[35,53]))
)

alt.layer(line_A, line_B).resolve_scale(y='independent')

Did you see what I did there?

Of course, mixed axis plots are rarely purely line plots. Instead they can be mixes of different axis. For these and other plotting mistakes, the economist has a nice article here. You may want to try some of these plots with this data set or the world indicators dataset from a few weeks ago.