library(plotor)
library(dplyr)
library(datasets)
library(tidyr)
library(stats)
library(broom)
library(forcats)
library(ggplot2)
plotor
produces Odds-Ratio plots from a given logistic
regression model, as produced using the general linear model
(glm
) package.
plotor
can be installed via GitHub using the
devtools
package:
You can also install the latest released version from Cran with:
Let’s start by exploring the likelihood of surviving the Titanic disaster based on passenger economic status (class), sex, and age group.
Get and prepare data
from the datasets
package.
df <- datasets::Titanic |>
as_tibble() |>
# convert counts to observations
filter(n > 0) |>
uncount(weights = n) |>
# convert categorical variables to factors.
# we specify an order for levels in Class and Survival, otherwise ordering
# in descending order of frequency
mutate(
Class = Class |>
fct(levels = c('1st', '2nd', '3rd', 'Crew')),
Sex = Sex |>
fct_infreq(),
Age = Age |>
fct_infreq(),
Survived = Survived |>
fct(levels = c('No', 'Yes'))
)
# preview the data
df |> glimpse()
#> Rows: 2,201
#> Columns: 4
#> $ Class <fct> 3rd, 3rd, 3rd, 3rd, 3rd, 3rd, 3rd, 3rd, 3rd, 3rd, 3rd, 3rd, 3…
#> $ Sex <fct> Male, Male, Male, Male, Male, Male, Male, Male, Male, Male, M…
#> $ Age <fct> Child, Child, Child, Child, Child, Child, Child, Child, Child…
#> $ Survived <fct> No, No, No, No, No, No, No, No, No, No, No, No, No, No, No, N…
We now have a tibble of data containing four columns:
Survived
- our outcome variable describing whether
the passenger survived Yes
or died
No
,
Class
- the passenger class, either
1st
, 2nd
, 3rd
or
Crew
,
Sex
- the gender of the passenger, either
Male
or Female
,
Age
- whether the passenger was an
Adult
or Child
.
We next conduct a logistic regression of survival (as a binary
factor: ‘yes’ and ‘no’) against the characteristics of passenger class,
sex and age group. For this we use the Generalised Linear Model function
(glm
) from the stats
package, specifying:
the family as ‘binomial’, and
the formula as survival being a function of Class
,
Sex
and Age
.
# conduct a logistic regression of survival against the other variables
lr <- glm(
data = df,
family = 'binomial',
formula = Survived ~ Class + Sex + Age
)
# preview the model as a tidy table
lr |>
tidy() |>
glimpse()
#> Rows: 6
#> Columns: 5
#> $ term <chr> "(Intercept)", "Class2nd", "Class3rd", "ClassCrew", "SexFema…
#> $ estimate <dbl> -0.3762229, -1.0180950, -1.7777622, -0.8576762, 2.4200603, 1…
#> $ std.error <dbl> 0.1361769, 0.1959969, 0.1715657, 0.1573387, 0.1404093, 0.244…
#> $ statistic <dbl> -2.762751, -5.194443, -10.361993, -5.451147, 17.235750, 4.35…
#> $ p.value <dbl> 5.731642e-03, 2.053331e-07, 3.691891e-25, 5.004592e-08, 1.43…
Finally, we can plot the Odds Ratio of survival using the
plot_or
function.
This plot makes it clear that:
Children were 2.89 times more likely to survive than Adults,
Passengers in 2nd
, 3rd
class as well as
Crew
were all less likely to survive than those in
1st
class,
Women were 11.25 times more likely to survive than men.
The primary components of an Odds-Ratio plot are dots, whiskers and the line of no effect.
The dot represents the point estimate for the Odds-Ratio, which indicates how much more likely the event is than the comparator event.
The width of the whiskers represents the 95% Confidence Interval, a range of values the point estimate (the dot) is likely to fall within if the study were to be repeated, with a probability of 95%.
The line of no effect is set at a value of 1. Events whose confidence intervals touch or cross this line are considered to show no difference in likelihood than the comparator event.
The size of the dots is proportional to the number of observations.
In the above, the size of the Adult
square is much larger
than the Child
square, because there were 20 times more
adults on the ship than children. This feature can help contextualise
the findings from OR plots.
To increase the size of the font you can extend the returned plot
using the theme
function from ggplot2
. Here we set the base size of all
fonts in the plot to size 16.
Odds Ratio (OR) plots produced by plotor
are displayed
using a log10 x-axis.
By default ten breaks are shown, however, this can be altered by
extending the scale_x_log10
function from ggplot2
. Here we provide a manual list of
breaks to use:
There are three types of colours used for the dots and whiskers in the OR plot, depending on their category.
Significant
refers to dots where their results
indicate a significant finding because their 95% confidence intervals do
not touch or cross the value 1 - the line of no effect.
Comparator
refers to the level of a factor in the
model against which the Odds Ratios are calculated.
Not significant
refers to dots where their results
do not indicate a significant finding because their confidence intervals
touch or cross the line of no effect.
The colours for these points can be changed by extending the output
using scale_colour_manual
function from ggplot2
with a named vector specifying colour
values for the three types of colours:
This data
set comes from a case-control study of oesophageal cancer in
Ile-et-Vilaine, France. In addition to the outcome variable,
Group
, identifying who is a case (developed cancer) or a
control (disease free), it contains three explanatory variables:
agegp
- the age group of each participant,
alcgp
- the alcohol consumption of each participant,
measured in grams per day,
tobgp
- the tobacco consumption of each participant,
measured in grams per day.
To look at the likelihood of a participant to develop oesophageal cancer we can perform logistic regression against these variables.
df <- datasets::esoph |>
# convert aggregated data to tidy observational data
tidyr::pivot_longer(
cols = c(ncases, ncontrols),
names_to = 'Group',
values_to = 'people'
) |>
uncount(weights = people) |>
# prepare the variables
mutate(
# convert the intervention group to a factor
Group = Group |>
case_match('ncases' ~ 'Case', 'ncontrols' ~ 'Control') |>
fct(levels = c('Control', 'Case')),
# remove the ordering from these factors so the glm model doesn't treat
# them as numeric
agegp = agegp |> factor(ordered = F),
alcgp = alcgp |> factor(ordered = F),
tobgp = tobgp |> factor(ordered = F)
)
# preview the data
df |> glimpse()
#> Rows: 975
#> Columns: 4
#> $ agegp <fct> 25-34, 25-34, 25-34, 25-34, 25-34, 25-34, 25-34, 25-34, 25-34, 2…
#> $ alcgp <fct> 0-39g/day, 0-39g/day, 0-39g/day, 0-39g/day, 0-39g/day, 0-39g/day…
#> $ tobgp <fct> 0-9g/day, 0-9g/day, 0-9g/day, 0-9g/day, 0-9g/day, 0-9g/day, 0-9g…
#> $ Group <fct> Control, Control, Control, Control, Control, Control, Control, C…
Next we carry out the logistic regression and then plot the results.
# conduct the logistic regression
lr <- glm(
data = df,
family = 'binomial',
formula = Group ~ agegp + alcgp + tobgp
)
# plot the odds ratio plot with customised title
plot_or(lr) +
labs(title = 'Likelihood of developing oesophageal cancer')
From this we can see there is a strong link between age and
likelihood of cancer. Compared with those in the 25-34
years group there is a statistically significant increased likelihood of
being in the case cohort of those in the 45-54
years group
(43 times more likely), 55-64
years group (76 times more
likely), 65-74
years group (133 times more likely), and
75+
years group (124 times more likely).
There is also a strong link between alcohol consumption and likelihood of cancer. Compared with those who consumed the least alcohol, defined as between 0 and 39 g/day, those who consumed more alcohol are more at risk of developing cancer with the greatest risk in those who consumed more than 119 g/day, putting them at 36 times more likely to develop cancer.
Tobacco use is a more nuanced picture. There was no detectable
difference in the likelihood of developing cancer for those in the first
three groups (0-9g/day
, 10-19g/day
and
20-29g/day
) - seen by the confidence intervals crossing the
line of no effect. However, there was a statistically significant
increased risk of developing cancer in those who consumed the most
tobacco, 30+g/day
, putting them at 5 times the risk.
Replacing variable names with a more descriptive label makes the
plots more accessible to those not involved in the analysis. For
example, Alcohol consumption (g/day)
is a more
user-friendly label than the name of the variable,
alcgp
.
There are some amazing packages designed to help label your data. In
the below example we use the labelled
package to label our data before analysing and plotting it.
# library to apply labels to data
library(labelled)
# create a list of variable = labels
var_labels <- list(
agegp = 'Age group',
alcgp = 'Alcohol consumption',
tobgp = 'Tobacco consumption',
Group = 'Developing oesophageal cancer'
)
# label the variables in our data
labelled::var_label(df) <- var_labels
# preview the data with labels appplied
labelled::look_for(df)
#> pos variable label col_type missing values
#> 1 agegp Age group fct 0 25-34
#> 35-44
#> 45-54
#> 55-64
#> 65-74
#> 75+
#> 2 alcgp Alcohol consumption fct 0 0-39g/day
#> 40-79
#> 80-119
#> 120+
#> 3 tobgp Tobacco consumption fct 0 0-9g/day
#> 10-19
#> 20-29
#> 30+
#> 4 Group Developing oesophageal cancer fct 0 Control
#> Case
Analyse the data using logistic regression as before and plot the result.
# conduct the logistic regression
lr <- glm(
data = df,
family = 'binomial',
formula = Group ~ agegp + alcgp + tobgp
)
# plot the odds ratio plot using variable labels
plot_or(lr)
plot_or
recognises the use of labels and uses these in
preference to variable names wherever available.
Using variable labels makes plots easier to read and more accessible, and is especially useful where you want to include the chart in reports or publications.