21.1.1.1 Example

To explore this in R, we will return to the commute data set we created from BARI’s Census Indicators very early on. We will visualize the relationship of commuting by automobile and median household income. Moreover, we will color each point by the rate of Bachelor’s degree attainment. As educational attainment tends to increase with income, we should expect the coloring to somewhat follow household income.

Create the visualization following the below steps.

• Use the object commute
• Plot med_house_income against (on the y axis) by_auto
• Color the plot by bach
• Add the appropriate geometry

(p <- ggplot(commute, aes(by_auto, med_house_income, color = bach)) +
  geom_point())

Wonderful! As anticipated, the color of the points are darkest at the bottom where median household income is the lowest. And as income increases, so does the brightness and saturation of our color scale. If we return to the grammar of graphics, we can take further control of the scales. In this case, the color of the scales. One of the ways we can change the color scale is to add the layer scale_color_gradient() to manually choose what colors to use in our scale. This provides a lot of flexibility to you as a information designer. Think of how you can use the color to represent what it is thatyou are visualizing. Or, how you can use colors to adhere to a color palette of your own or an organization you are working with.

scale_color_gradient() allows us to provide the colors for low values and high values. In general, you should choose a darker color to represent lower values and brighter colors for higher values. Below we add the layer and provide color codes.

You can find color codes via google by searching “color picker.” Or, you can use any number of free color palette generators online. I am personally a fan of coolors.co.

Below we’ve changed the plot to transition from a dark red to a white as values increase.

p +
  scale_color_gradient(low = "#360002", high = "white")

Try modifying the above plot by picking two colors that you think do a good job of visualizing a range of values.

21.1.1.2 Diverging colors

There are many a time when our variable of interest has a middle value and illustrating the center as well as the deviation from that center is important. To visually represent this we use what is called a diverging color scale. Diverging color scales are characterized by a middle color from which both ends of the spectrum originate. The center is to represent some middle value. If we contextualize this as z-scores and the color palette below, the center would be 0 and any negative scores would trend towards red. Whereas if they trended positive they would become more blue.

One thing to be wary of is using a divergent color scale when it is not necessary. This is an easy trap to fall into since they’re pretty cool. Remember, only use diverging color palettes if there is an existing middle value.

21.1.1.3 Example

Take our ecometrics again. Say we are interested in what the annual deviation is from the sample mean—the average for all years—of each ecometric. This is the perfect use case for a diverging color scale. This will require a bit of computational creativity. So lets work through this.

Let’s think about each of the measures we need to calculate. We need to:

• Find the number of counts for each year by ecometric.
• Find the average count for all years for each ecometric.
• Identify the deviation from the mean.

We will work through this sequentially. First and foremost we need to calculate the total number of crime reports by ecometric by year. The dataset has more than one observation per year per ecometric. We can see this by running a quick count.

count(ecometrics, measure, year)
#> # A tibble: 16 x 3
#>    measure           year     n
#>    <chr>            <dbl> <int>
#>  1 guns              2015     5
#>  2 guns              2016     5
#>  3 guns              2017     5
#>  4 guns              2018     5
#>  5 private conflict  2015     4
#>  6 private conflict  2016     4
#>  7 private conflict  2017     4
#>  8 private conflict  2018     4
#>  9 social disorder   2015     3
#> 10 social disorder   2016     3
#> 11 social disorder   2017     3
#> 12 social disorder   2018     3
#> 13 violence          2015     5
#> 14 violence          2016     5
#> 15 violence          2017     5
#> 16 violence          2018     5

We need to tidy this up and ensure that each row is only one observation—in this case one ecometric per year—with the total count. The logic to accomplish this is to first group_by() both measure and year and then sum the n values. As such:

ecometrics %>% 
  group_by(measure, year) %>% 
  summarise(n = sum(n))
#> `summarise()` regrouping output by 'measure' (override with `.groups` argument)
#> # A tibble: 16 x 3
#> # Groups:   measure [4]
#>    measure           year     n
#>    <chr>            <dbl> <dbl>
#>  1 guns              2015  3146
#>  2 guns              2016  3111
#>  3 guns              2017  3190
#>  4 guns              2018  2585
#>  5 private conflict  2015  8063
#>  6 private conflict  2016  7807
#>  7 private conflict  2017  7410
#>  8 private conflict  2018  6592
#>  9 social disorder   2015  5043
#> 10 social disorder   2016  4707
#> 11 social disorder   2017  4876
#> 12 social disorder   2018  4168
#> 13 violence          2015 18621
#> 14 violence          2016 18080
#> 15 violence          2017 17172
#> 16 violence          2018 16630

One of the dplyr quirks is that after you summarise(), one level of grouping is removed. This is because we have already performed our aggregate measure and the last level of grouping is now unit of analysis (a row). Since we are currently grouped at the measure level and each row represent one year we are already ready to calculate the average n value by group. Rather than using summarise() as we are used to doing with summary statistics, we will use mutate() and create a column called avg with the average n value for each group. This is because we want to be able to perform a column-wise operation to subtract the average from each row’s n value.

I continue by creating three new columns. The first is avg which is set to the mean of n by group. Since the resulting value of mean(n) is a single value, each observation in the group gets that value. Second, I create a new column called deviation which subtracts the new column we created from the n column. The lastly I created a new column to contain the name of the ecometric as well as the year. This is done to ensure that in the visualization each row can be plotted.

annual_metrics <- ecometrics %>% 
  # group by measure and year
  group_by(measure, year) %>% 
  # find the total n for each measure and year
  # summarise loses the last level of grouping
  summarise(n = sum(n)) %>% 
  mutate(
    # calculate average `n` by ecometric
    avg = mean(n),
    # calculate the deviation from the mean 
    deviation = n - avg,
    # creating a new column that combines the name of the ecometric and the year
    name = glue::glue("{measure}: {year}"),
  ) 

I snuck in a new package right there. glue is a package that let’s us incorporate R expressions with plain text. Whatever is inside of the brackets { will be run as an R expression. So since I referenced the columns measure and year as "{measure}: {year}" each value of name will look something like social disorder: 2017.

Now that we have a data frame with all of the values we want to plot, let’s go ahead and do that! We will use a lollipop chart here to do so.

Remember that geom_lollipop() comes from the ggalt package. We can use the function without loading the whole package by referencing the pkgname::function_name().

Within the chart we map devation to the color aesthetic.

(p <- ggplot(annual_metrics, aes(y = deviation, x = name, color = deviation)) + 
  ggalt::geom_lollipop(size = 2) + 
  coord_flip())
#> Registered S3 methods overwritten by 'ggalt':
#>   method                  from   
#>   grid.draw.absoluteGrob  ggplot2
#>   grobHeight.absoluteGrob ggplot2
#>   grobWidth.absoluteGrob  ggplot2
#>   grobX.absoluteGrob      ggplot2
#>   grobY.absoluteGrob      ggplot2

The above graphic doesn’t take into account the middle value of 0. Because of this, we need to tell ggplot2 that this is a diverging color scale. When we were working with a normal continuous color scale we used scale_color_gradient(). Instead, since we have a middle value, to create the diverging color gradient we use scale_color_gradient2(). This adds two more arguments mid and midpoint. The former is the color of that middle value. The second is what that middle value maps to—defaults to 0, which is fine for our example.

Now if we add the below colors—a red, yellow-ish beige, and a blue—we will have a diverging color scale on the plot!

p + 
  scale_color_gradient2(low = "#A12106", mid = "#B2CEB7", high = "#004C76")

Information design is a seemingly endless field of which we only touched on a very small amount. The R community has put a lot of work into enabling the use of color for visualization purposes. The above images of color palettes were created with the help of the wonderful packages prismatic and paletteer

To explore color more check out the packages prismatic and paletteer by Emil Hvitfeld.