## packages
library(tidyverse)
library(socsci)
library(coolorrr)
## data viz settings
set_theme()
set_palette(
qualitative = c("steelblue", "gray", "red3", "purple4", "gold2"),
sequential = c("white", "steelblue"),
diverging = c("red3", "white", "steelblue"),
binary = c("steelblue", "red3")
)
## data
url <- "http://filesforprogress.org/datasets/wthh/DFP_WTHH_release.csv"
Data <- read_csv(url)
## recodes
Data <- Data |>
mutate(
## education
educ_cat = frcode(
educ == 1 ~ "No HS",
educ == 2 ~ "HS",
educ == 3 ~ "Some College",
educ == 4 ~ "2-year Degree",
educ == 5 ~ "4-year Degree",
educ == 6 ~ "Postgrad"
),
## race
race_new = frcode(
race == 1 ~ "White",
race == 2 ~ "Black",
race == 3 ~ "Latino/a",
race == 4 ~ "Asian",
race == 5 ~ "Native American",
race == 8 ~ "Middle Eastern",
race == 6 ~ "Mixed Race",
TRUE ~ "Other"
),
## party id
pid3_new = frcode(
pid3 == 1 ~ "Democrat",
!(pid3 %in% 1:2) ~ "Independent/Other",
pid3 == 2 ~ "Republican"
),
## gender
gender_id = ifelse(
gender == 1, "Male", "Female"
),
## age
age = 2018 - birthyr,
## issues
across(
c(ICE, BAIL_item, WELTEST, PUBLICINT, GREENJOB, POLFEE),
~ frcode(
.x == 5 ~ "Strongly oppose",
.x == 4 ~ "Oppose",
.x == 3 ~ "Neither",
.x == 2 ~ "Support",
.x == 1 ~ "Strongly support"
)
)
)16 Ideas for Visualizing Survey Data
16.1 Goals
- Provide examples for how to do demographic summaries of survey samples
- Provide examples for how to summarize attitudes in a survey
In this last set of lecture notes, my hope is to arm you with plenty of examples of data visualizations with surveys. I’ve tried to include as much as possible, but I can’t show you everything. Walk through the below examples and tweak them for your own purposes as needed.
Here’s the packages and data (along with recodes) that you’ll need to walk through the below examples:
16.2 Demographic Summaries
There are many ways you could summarize the demographic profile of a sample. How complex these summaries are depends on how many demographic factors you want to consider at once. If you want to just look at one at a time, a simple barchart or histogram will do. For example, here’s a barchart of how the sample breaks down by race:
Data |>
ct(race_new) |>
ggplot() +
aes(
x = reorder(race_new, -pct),
y = pct
) +
geom_col(
fill = "gray"
) +
labs(
x = NULL,
y = NULL,
title = "Breakdown of the survey sample by race",
caption = "Source: What the Hell Happened? (Data for Progress)"
) +
scale_y_continuous(
labels = scales::percent
) +
theme(
axis.text.x = element_text(angle = 45, hjust = 1)
)
Here’s another example looking at the breakdown by respondent age:
ggplot(Data) +
aes(x = age) +
geom_histogram(
fill = "gray",
color = "black"
) +
scale_x_continuous(
n.breaks = 10
) +
labs(
x = NULL,
y = NULL,
title = "Distribution of survey respondents by age",
caption = "Source: What the Hell Happened? (Data for Progress)"
)
When we want to consider at least two demographic factors at once, things can get a little more complex. There are some tools that can help, however. One of those is xheat() from the {socsci} package. It plots a heat map based on the distribution of observations across two different discrete categories. Here’s an example of it in action with race and gender:
Data |>
xheat(gender_id, race_new) +
## ggpal works with xheat()
ggpal(
type = "sequential",
aes = "fill"
) +
labs(
x = NULL,
y = NULL,
title = "Demographic breakdown of the sample by race and gender",
caption = "Source: What the Hell Happened? (Data for Progress)"
)
It’s important to note that xheat() shows the breakdown of what’s on the y-axis by the categories on the x-axis. That means if we switched around race_new and gender_id in xheat(), we’d get different results. Whereas the previous figure showed the breakdown of race by gender, the below shows the breakdown of gender by race:
Data |> # flip race and gender around:
xheat(race_new, gender_id) +
ggpal(
type = "sequential",
aes = "fill"
) +
labs(
x = NULL,
y = NULL,
title = "Demographic breakdown of the sample by race and gender",
caption = "Source: What the Hell Happened? (Data for Progress)"
) +
theme(
axis.text.x = element_text(angle = 45, hjust = 1)
)
If you want to instead show the distribution across all categories rather than by the one on the x-axis, you’ll need to do a little more work. Here’s an example using the same variables above to show the distribution of survey respondents by race and gender without grouping the summary by gender or race:
Data |>
group_by(race_new, gender_id) |>
count() |>
ungroup() |>
mutate(
pct = n / sum(n)
) |>
ggplot() +
aes(
x = gender_id,
y = race_new,
fill = pct,
label = scales::percent(round(pct, 3))
) +
geom_tile(
color = "black"
) +
geom_text() +
theme(
legend.position = ""
) +
ggpal(
type = "sequential",
aes = "fill"
) +
labs(
x = NULL,
y = NULL,
title = "Demographic breakdown of the sample by race\nand gender",
caption = "Source: What the Hell Happened? (Data for Progress)"
)
Not all pairs of variables you want to consider will be discrete. Sometimes you may want to show how a continuous variable breaks down by a discrete category. That’s easy enough to do with a faceted histogram. Here’s an example using age (continuous) and gender (discrete):
ggplot(Data) +
aes(x = age) +
geom_histogram(
color = "black",
fill = "gray"
) +
facet_wrap(~ gender_id) +
scale_x_continuous(
n.breaks = 10
) +
labs(
x = NULL,
y = NULL,
title = "Breakdown of respondent age by gender",
caption = "Source: What the Hell Happened? (Data for Progress)"
)
If we want to move beyond two variables at once, we have options, too. Here’s an extended version of the above figure that adds partisanship into the mix. Instead of using facet_wrap() as in the above example, note the use of facet_grid().
ggplot(Data) +
aes(x = age) +
geom_histogram(
color = "black",
fill = "gray"
) +
## use facet_grid() instead of facet_wrap()
facet_grid(gender_id ~ pid3_new) +
scale_x_continuous(
n.breaks = 10
) +
labs(
x = NULL,
y = NULL,
title = "Breakdown of respondent age by gender and\npartisanship",
caption = "Source: What the Hell Happened? (Data for Progress)"
)
We can do something similar when all three variables are discrete, too. Let’s replace age with race and do a column plot instead:
Data |>
group_by(gender_id, pid3_new) |>
ct(race_new) |>
ggplot() +
aes(
x = pct,
y = race_new
) +
geom_col(
fill = "gray"
) +
facet_grid(gender_id ~ pid3_new) +
scale_x_continuous(
labels = scales::percent
) +
labs(
x = NULL,
y = NULL,
title = "Breakdown of respondent race by partisanship\nand gender",
caption = "Source: What the Hell Happened? (Data for Progress)"
)
If we don’t like the above approach (it is a little hard to draw comparisons), we instead could use a stacked barchart:
Data |>
group_by(gender_id, pid3_new) |>
ct(race_new, cum = T) |> # computes cumulative pct
ggplot() +
aes(
x = pct,
y = pid3_new,
fill = race_new
) +
geom_col() +
facet_wrap(~ gender_id) +
scale_x_continuous(
labels = scales::percent
) +
labs(
x = NULL,
y = NULL,
title = "Breakdown of respondent race by partisanship\nand gender",
caption = "Source: What the Hell Happened? (Data for Progress)",
fill = NULL
)
There’s nothing to stop you from going further and showing the breakdown of the data using four (or more) variables. However, for each new variable you want to throw in the mix, there are diminishing returns to interpretability. In most cases, three variables in one data viz is more than enough to tell a compelling story. Go beyond three at your own risk.
16.3 Summarizing Attitudes
We have similar data visualization choices when we want to show how an attitude either breakdowns by different demographic factors in the data or else is correlated with other attitudes. Just like with demographic factors, your data visualizations can range from simple to complex.
Here’s an example showing how attitudes with respect to six different policies breakdown. While it doesn’t show how attitudes vary by different factors in the data, it manages to summarize attitudes for all six at once:
Data |>
pivot_longer(
cols = c(ICE, BAIL_item, WELTEST, PUBLICINT, GREENJOB, POLFEE)
) |>
mutate(
name = frcode(
name == "ICE" ~ "De-fund ICE",
name == "BAIL_item" ~ "End Cash Bail",
name == "WELTEST" ~ "Drug-testing for Welfare",
name == "PUBLICINT" ~ "Public Internet Company",
name == "GREENJOB" ~ "Green Jobs for Unemployed",
name == "POLFEE" ~ "Polution Fees for Companies"
)
) |>
group_by(name) |>
ct(value, show_na = F) |>
ggplot() +
aes(x = value, y = pct, fill = name) +
geom_col(show.legend = F) +
facet_wrap(~ name) +
scale_y_continuous(
labels = scales::percent
) +
labs(
x = NULL,
y = NULL,
title = "Support for different policy proposals",
caption = "Source: What the Hell Happened? (Data for Progress)"
) +
theme(
axis.text.x = element_text(angle = 45, hjust = 1)
)
We can also convert this into a stacked version to make comparisons easier:
Data |>
pivot_longer(
cols = c(ICE, BAIL_item, WELTEST, PUBLICINT, GREENJOB, POLFEE)
) |>
mutate(
name = frcode(
name == "ICE" ~ "De-fund ICE",
name == "BAIL_item" ~ "End Cash Bail",
name == "WELTEST" ~ "Drug-testing for Welfare",
name == "PUBLICINT" ~ "Public Internet Company",
name == "GREENJOB" ~ "Green Jobs for Unemployed",
name == "POLFEE" ~ "Polution Fees for Companies"
)
) |>
group_by(name) |>
ct(value, show_na = F, cum = T) |>
ggplot() +
aes(x = pct, y = name, fill = value) +
geom_col() +
scale_x_continuous(
labels = scales::percent
) +
labs(
x = NULL,
y = NULL,
title = "Support for different policy proposals",
caption = "Source: What the Hell Happened? (Data for Progress)",
fill = NULL
) +
ggpal(
aes = "fill",
guide = guide_legend(reverse = T)
)
Summarizing attitudes that are measured using continuous scales can be done similarly, for example, with a histogram, like I’ve done below using racial resentment, hostile sexism, and fear of demographic change:
Data |>
pivot_longer(
cols = c(racial_resentment_scaled, hostile_sexism_scaled, fear_of_demographic_change_scaled)
) |>
mutate(
name = str_replace_all(name, "_", " ") |>
str_remove(" scaled")
) |>
ggplot() +
aes(x = value) +
geom_histogram(
color = "black",
fill = "gray"
) +
facet_wrap(~ name) +
scale_x_continuous(
breaks = 0:1,
labels = c("low", "high")
) +
labs(
x = NULL,
y = NULL,
title = "Distribution of attitudes toward different marginalized\ngroups",
caption = "Source: What the Hell Happened? (Data for Progress)"
) +
theme(
axis.text.x = element_text(hjust = 0:1)
)
You can also use a “ridge” plot, for continuous data. A ridge plot is similar to a density plot, but it lets you show multiple distributions at once without the need to use facet_wrap() or facet_grid().
## open {ggridges}
library(ggridges)
## make the plot
Data |>
pivot_longer(
cols = c(racial_resentment_scaled, hostile_sexism_scaled, fear_of_demographic_change_scaled)
) |>
mutate(
name = str_replace_all(name, "_", " ") |>
str_remove(" scaled")
) |>
ggplot() +
aes(x = value, y = name) +
geom_density_ridges(
aes(height = ..density..),
stat = "density",
trim = T
) +
labs(
x = NULL,
y = NULL,
title = "Distribution of attitudes toward different marginalized\ngroups",
caption = "Source: What the Hell Happened? (Data for Progress)"
)
We can also use a “raincloud” plot. It’s very similar to the above, but it adds more elements. Specifically, it blends together a density plot, a boxplot, and a dotplot/histogram, giving the appearance of rain clouds. Check it out below. Make sure to install {ggrain}.
## install.packages("ggrain")
library(ggrain)
Data |>
pivot_longer(
cols = c(racial_resentment_scaled, hostile_sexism_scaled, fear_of_demographic_change_scaled)
) |>
mutate(
name = str_replace_all(name, "_", " ") |>
str_remove(" scaled")
) |>
ggplot() +
aes(x = name, y = value, fill = name, color = name) +
geom_rain(alpha = 0.6) +
coord_flip() +
labs(
x = NULL,
y = NULL,
title = "Distribution of attitudes toward different marginalized\ngroups",
caption = "Source: What the Hell Happened? (Data for Progress)"
) +
ggpal(aes = "fill") +
ggpal(aes = "color") +
theme(
legend.position = ""
)
We can also summarize the degree of agreement on different issues, as is relevant when we want to know whether a consensus or division exists among survey respondents. While we could just look at the data and make a judgement about the level of consensus that exists among the individuals who took a survey, putting an exact number on it helps to remove some subjectivity from our evaluation. We can measure agreement in many different ways, but one simple approach is to calculate a Herfindahl Index based on survey responses. This is just the sum of the squared proportion of responses in each of the possible response categories for a given question. The output from this approach gives us a value between 0 and 1, where 1 means perfect agreement. If you’re wondering, the number returned from this approach is an estimate of the probability that two randomly selected respondents to the survey have the same attitude. The higher this probability, the more general agreement there is among the people who took the survey.
Here’s an example of how to do it using the six different policy issues mentioned in an earlier example. The below code first creates an agreement() function that calculates the probability that two randomly selected individuals in the data had the same response to a survey question. After pivoting the data and grouping by the six policy issue questions, counts and proportions of responses are calculated. The agreement() function is then applied to these counts to give us a question specific measure of agreement, which we can then plot. The data viz I produced below is quite involved, so go slowly to make sure you understand what’s going on. Rather than produce a barchart with geom_col() (that’d be easy), I made a horizontal lollipop plot and used some extra hacks to have the relevant policy question labels appear inline, above the relevant lollipop.
## a function to calculate "agreement":
agreement <- function(x) {
p <- (x / sum(x, na.rm = T))^2
sum(p, na.rm = T)
}
## start prepping the data
Data |>
pivot_longer(
cols = c(ICE, BAIL_item, WELTEST, PUBLICINT, GREENJOB, POLFEE)
) |>
mutate(
name = frcode(
name == "ICE" ~ "De-fund ICE",
name == "BAIL_item" ~ "End Cash Bail",
name == "WELTEST" ~ "Drug-testing for Welfare",
name == "PUBLICINT" ~ "Public Internet Company",
name == "GREENJOB" ~ "Green Jobs for Unemployed",
name == "POLFEE" ~ "Polution Fees for Companies"
)
) |>
group_by(name) |>
ct(value, show_na = F) |>
## calculate agreement and save the top response:
summarize(
agreement = agreement(pct),
attitudes = unique(value[pct == max(pct)])
) |>
ungroup() |>
## update the name to include the top attitude and reorder:
mutate(
name = reorder(
paste0(name, " (", attitudes, ")"), -agreement
)
) |>
ggplot() +
aes(x = agreement, y = name) +
geom_pointrange(
aes(xmin = 0, xmax = agreement)
) +
labs(
x = "Agreement",
y = NULL,
title = "Which issues do respondents most agree on?"
) +
facet_wrap(~ name, scales = "free_y", ncol = 1) +
scale_y_discrete(
breaks = NULL
) +
theme(
strip.text = element_text(hjust = 0, face = "bold.italic"),
strip.background = element_blank(),
panel.grid.major = element_blank()
)
We can also use this approach to measure agreement on continuous policy preferences, too. Here’s an example using the hostile sexism, racial resentment, and fear of demographic change indexes:
Data |>
pivot_longer(
cols = c(racial_resentment_scaled, hostile_sexism_scaled, fear_of_demographic_change_scaled)
) |>
mutate(
name = str_replace_all(
name, "_", " "
) |>
str_remove_all(" scaled")
) |>
group_by(name) |>
count(value) |>
drop_na() |>
## calculate agreement and save the top response:
summarize(
agreement = agreement(n),
mean = round(sum(value * n) / sum(n), 2)
) |>
ungroup() |>
## update the name to include the top attitude and reorder:
mutate(
name = reorder(
paste0(name, " (mean = ", mean, ")"), -agreement
)
) |>
ggplot() +
aes(x = agreement, y = name) +
geom_pointrange(
aes(xmin = 0, xmax = agreement)
) +
labs(
x = "Agreement",
y = NULL,
title = "Which issues do respondents most agree on?",
) +
facet_wrap(~ name, scales = "free_y", ncol = 1) +
scale_y_discrete(
breaks = NULL
) +
theme(
strip.text = element_text(hjust = 0, face = "bold.italic"),
strip.background = element_blank(),
panel.grid.major = element_blank()
)
Finally, if we’re interested in how different attitudes are correlated in the data, we have a number of options. You already are familiar with many, like scatter plots, so I want to introduce something new: a correlation plot. We’ve been talking about correlations among variables a lot so far, but we’ve really only done so in the colloquial sense. The term “correlation” also has a technical meaning in statistical analysis, specifically, as a formal measure of the degree to which two variables have a linear relationship. The estimate of this measure is called the correlation coefficient. It can take a value between -1 and 1, where 1 means the two variables in question have a perfect positive linear relationship, -1 means the two variables have a perfect negative linear relationship, and 0 means the two variables are completely orthogonal (they have no relationship).
We can calculate correlations for multiple variables at once in R using the cor() function. The below code uses this function to create a correlation matrix (a matrix that contains all the bivariate correlations for all possible combinations of variables in the data). I have it do so only for the three continuous indexes we’ve been playing around with so far by only selecting the relevant columns, then giving them to cor(). Since there are some missing values in the data, I set the option use = "pairwise.complete.obs", which is similar to na.rm = T in mean().
## make the correlation matrix
corr_mat <- Data |>
select(
racial_resentment_scaled, hostile_sexism_scaled, fear_of_demographic_change_scaled
) |>
cor(use = "pairwise.complete.obs")
## update names
colnames(corr_mat) <- str_replace_all(colnames(corr_mat), "_", " ") |>
str_remove(" scaled") -> rownames(corr_mat)
## print
corr_mat racial resentment hostile sexism
racial resentment 1.0000000 0.5884866
hostile sexism 0.5884866 1.0000000
fear of demographic change 0.5887728 0.5968202
fear of demographic change
racial resentment 0.5887728
hostile sexism 0.5968202
fear of demographic change 1.0000000Once I have the correlation matrix, I can use ggcorrplot() from the {ggcorrplot} package to make a data visualization with the above correlation matrix:
## install.packages("ggcorrplot")
library(ggcorrplot)
ggcorrplot(corr_mat,
type = "lower",
lab = T) +
theme(legend.position = "") +
labs(
title = "Correlation in attitudes toward marginalized groups"
)