(0) Housekeeping

To save your work in RStudio, you need to put your data-analysis code in a text file. In the File menu, choose “New File > R Script.” The new script should include everything necessary to produce the computational outputs you want. It shouldn’t usually include install_github or anything like that. It should start with the libraries you need, like library(dataculture). Save the script with a name like lab7.R in a place you can find it.

To run a script you have open in RStudio, click the “Source” button. Plots—and information printed to the console—are technically “side effects.” Any graphic you make will be lost when you quit R. To save the last-plotted graphic, use the ggsave function. If I had a script like this:

library(tidyverse)
library(dataculture)
library(babynames)
babynames |>
    filter(name == "Andrew", sex == "M") |>
    ggplot(aes(year, prop)) + geom_line() + geom_point()

ggsave("lab7-andrewline.pdf")

Running this with the “Source” button, I would find a new PDF file lab7-andrewline.pdf saved in the default location for that RStudio session (usually your Home folder).

There are more elaborate ways to manage code and data files, but that will do for the moment. You can, of course, also use ggsave at the regular R prompt, or you can use the “Export” button at the top of the RStudio plot window. But then you won’t have saved the code to produce the plot.

For attendance today, submit one R script that creates at least one plot from the lab on Canvas.

(1) Let’s get cooking

The recipes data frame includes recipe data gathered from digitized editions of a selection of 74 American cookbooks published between 1803 and 1922. I have added some notes on the data in help(recipes) (KNOW THY DATA).

The ethnicgroup column indicates whether the recipe had been assigned to an ethnic group by the editors of the digital edition (based on whether the cookbook labeled it as such). Try:

recipes |> count(ethnicgroup, sort=T)

Consider: what does “ethnic” mean here? What is the opposite of “ethnic”?

The count function creates a new data frame recording how many times each value of ethnicgroup occurs in recipes (the counts are in a new column, n). sort=T means to give the results in descending order of frequency. The same result can be obtained more verbosely with

recipes |> group_by(ethnicgroup) |>
    summarize(n=n()) |>
    arrange(desc(n))

n() is a function which gives the number of rows in a group. arrange(desc(n)) means to arrange rows in descending order of n (the column).

Scrolling to the top, you will see that the most frequent value is NA. Remember that NA is a special value indicating missing data. The following lines compute a tally of NA recipes by year:

ethnic <- recipes |>
    group_by(date) |>
    summarize(nonethnic=sum(is.na(ethnicgroup)),
              ethnic=sum(!is.na(ethnicgroup))) |>
    mutate(ethnic_prop=ethnic / (ethnic + nonethnic))

1. Follow the pattern from lab 6 to produce a line chart, with date on the x axis and ethnic_prop on the y axis.

   R for Data Science, chap. 3, begins with the following simple pattern for plotting

   ggplot(data = <DATA>) + 
     <GEOM_FUNCTION>(mapping = aes(<MAPPINGS>))

   Notice carefully that there is no pipe |> after ggplot(...). Instead, there is a + followed by the function call. It might occur to you wonder what is being added to what with + here. There is no numerical addition here, just some goofy computerese.¹ Nonetheless, when you are using ggplot you chain together the functions that build and modify plots with +.

   This is not exactly the lab 6 pattern. Reproduce the line chart following this pattern exactly.

2. Is this a good visualization? What are its limitations?

3. Let’s try a different “geom.” Replace geom_line with geom_point. Now try geom_col. Make sure you can explain to each other exactly how the very same data can be used to draw three different pictures. (Hint: What is a “mapping”?)

Now make a really ugly graphic with:

ggplot(data=ethnic, mapping=aes(x=date, y=ethnic_prop)) +
    geom_point() +
    geom_line() +
    geom_col()

Notice that the aes mapping has been moved inside ggplot(). This mapping is then “passed along” to each of the successive geom layers. Try combinations of two geoms (combined with +).

I have supplied a wrangled version of recipes, ingreds, which will make it a little easier to explore questions about ingredients. Here is code to set up for a slightly more interesting graphic by calculating a basic statistic about tomatoes:

tomatoes <- ingreds |> group_by(date, recipe_id) |>
    summarize(tom=any(ingredients == "tomato")) |>
    summarize(prop=sum(tom) / n())

In the first summarize, we calculate, for each recipe, whether any of its ingredients is tomato. In the second summarize, we calculate, for each year, how many recipes have a tomato ingredient. (Taking the sum() of a logical variable counts up how many TRUE elements it has, so we are counting how many rows have a tom value of TRUE.)

4. Create a plot showing the proportion of tomato-recipes from over time as points tomatoes, with date on the x axis and prop on the y axis.

   1. Would using lines or bars be an improvement? Try both options, using geom_line and geom_col.
   2. Replace geom_line with geom_line(color="red"). Is this an aesthetic mapping of the data?
   3. Discuss hypotheses that might explain this historical trend. (You don’t have to try to verify these hypotheses today.)

(2) Let’s make a (square) pie

So far, in lab 6 and today, we have been making pictures of two columns of data at a time: we have usually had time on the x axis and something like name- or ingredient-frequency on the y axis. Let’s make things more complicated and look at just one variable.² The recipes data contains a column region. How many recipes belong to each region? Sure, we could count them with the count function above, but today is visualization day:

ggplot(recipes) +
    geom_bar(aes(region))

1. Remember the template for ggplots above. What are the mappings here?

   1. What is on the x axis? What determines the order? Is this a good choice?

   2. What is on the y axis? Where in the recipes data are these values?

   3. Try—and decide if you prefer—

      ggplot(recipes) +
          geom_bar(aes(y=region))

2. Produce the same chart, using geom_col instead of geom_bar. Hint: it doesn’t start ggplot(recipes). What’s the difference between geom_col and geom_bar?

3. Now make a bar plot showing counts of ethnicgroup. Is this a useful visualization? How could it be improved?

recipes also has a column classifying each recipe according to what sort of dish it is, recipe_class. Let’s try to cross-tabulate regions and recipe types visually:

ggplot(recipes) +
    geom_bar(aes(x=region, fill=recipe_class))

4. Consider once again the aesthetic mappings. What data is on the x axis and y axis? What about the color of the bars? (This is called the fill by ggplot because color is reserved for the edges of bars—try color instead of fill.) How did ggplot choose what colors to use?

   1. Can you tell from this plot whether one kind of recipe is more common to one region or another? Why not?
   2. Without doing it in R, can you imagine a better way to show the same information on the plot?

Now try

ggplot(recipes) +
    geom_bar(aes(x=region, fill=recipe_class), position="dodge")

5. Think carefully: what is the rule determining where on the x axis each bar is placed?

   1. Does this make patterns any clearer?
   2. What about position="stack"?

These position settings are called “position adjustments” in the grammar of graphics, because they change how graphical elements that would otherwise lie on top of each other are positioned. They are often used to squeeze extra information onto plots. That’s sometimes helpful but often creates problems.

We can address one perceptual problem by throwing away information. Try:

ggplot(recipes) +
    geom_bar(aes(x=region, fill=recipe_class), position="fill")

6. Look carefully at the y axis. What happened when you added position="fill"?

   1. Now discuss again the interpretation of the figure: which patterns become more visible? Are there still perceptual difficulties?

I won’t tell you how to make a pie chart. Pie charts are bad and wrong (because the eye is not very adept at comparing the relative size of angles).

So far we have only changed geoms. The next important piece of the puzzle is what the grammar of graphics calls the scale, that is, the method by which data values are mapped to graphical properties. Let’s revisit the colors in the previous plot.

Try:

ggplot(recipes) +
    geom_bar(aes(x=region, fill=recipe_class), position="fill") +
    scale_fill_hue()

7. Nothing new here, but try other colors by replacing scale_fill_hue() with: scale_fill_viridis_d() and scale_fill_grey() and scale_fill_brewer(type="qual"). Choose your favorite. But ask yourself whether we could do even better with this data.

None of this gets us very far into the substantive questions you could ask of the recipes data, but let us turn to another dataset.

(3) That’s just noise

The gssmusic data frame includes responses about music tastes from a 1993 national sample of US adults, collected as part of the General Social Survey, together with some demographic and other information.³

The pertinent interview question went like this:

I’m going to read you a list of some types of music. Can you tell me which of the statements on this card comes closest to your feeling about each type of music (HAND CARD CA TO RESPONDENT.) Let’s start with big band music. Do you like it very much, like it, have mixed feelings, dislike it, dislike it very much, or is this a type of music that you don’t know much about?

The interview continued with the same question about 17 other music genres. These columns of gssmusic give each respondent’s rating:

gssmusic_doc |> filter(id %in% c(gssmusic_musicvars)) |>
    select(id, text) |>
    mutate(text=str_extract(text, "\\. [^.]*$")) |>
    mutate(text=str_sub(text, start=3)) |>
    knitr::kable()

Let’s begin by visualizing a few individual item scores.⁴

1. The country column gives responses about country and western music. Make a bar chart of these responses.⁵

   1. Pay attention to the warning message. Given the formulation of the question above, what will be the consequences of ignoring it?

   2. Make similar bar charts for other genres that interest you—can you find the genres with differently shaped bar charts?

People in different age groups have different tastes. Let’s explore this a little:

2. Start by making a scatterplot with age on the x axis and the response to the question about rap music on the y axis. What can you see, and what is hard to see?

   1. Look at gssmusic |> count(age). What further problem does this suggest?
   2. One solution to the problem in part (a) is to add random noise to each plotted point (think carefully about the nature of the aesthetic mapping this entails). There is another position adjustment, position="jitter", that adds random noise to each point. Try it here and see what you can see.

Another solution is to look at age groups. The following incantation produces a new column where respondents are grouped in ten-year intervals:

music_age <- gssmusic |>
    mutate(age_group=cut(age, breaks=18 + 10 * (0:7))) |>
    select(age_group, all_of(gssmusic_musicvars))

3. Now make a plot with bars for how each age group rates rap music. Hint: you did this already with the recipes data.

I have included a set of other variables to explore in addition to age. sex codes for sex; educ is years of education (goes up to 20); degree codes for degrees (see code values on the GSS website); realinc shows the respondent’s reported household income category in 1986 dollars. I included the GSS race variable, but KNOW THY DATA: there are only three values in the data, “white,” “black,” and “other.”

(4) Go forth and multiply

Since the 1990s, a great deal of sociological research has explored the degree to which people express preferences for a variety of different kinds of music (or other kinds of cultural goods). I have created two additional columns that summarize the 18 music-genre questions by showing how many the respondent “liked” or “liked very much” (music_likes) and how many they “disliked” or “disliked very much” (music_dislikes).

We could repeat the bivariate analyses above. But since we’ve had gender and education on the brain, let’s look at those variables. Try:

ggplot(gssmusic) +
    geom_bar(aes(music_likes)) +
    facet_grid(cols=vars(degree))

1. Explain what each sub-plot shows, and why there are six of them. What do you observe about the relation between educational credentials and music tastes here?

1. One can make a “smooth” version of these charts using a so-called kernel density estimate, which is a fancy way of averaging nearby bars together. Try geom_density and consider the advantages and disadvantages of this display. (Hint: look at the axes.)

   2. Try the same multiple plot but with facet_grid(rows=vars(degree)). Better or worse? Why?

Though it is stretching the data, we can try a further cross-tabulation:

ggplot(gssmusic) +
    geom_density(aes(music_likes)) +
    facet_grid(cols=vars(degree), rows=vars(sex))

2. Now begin to explore musical dislikes. Can you make a small-multiples plot which reveals whether particular demographic categories are more or less disposed to express intolerance for one kind of music or another?

3. Return to one of the “dodged” bar charts above and reformulate it as a faceted small-multiples plot.

Lab 6 had an example of the other faceting option, facet_wrap. Try

ggplot(gssmusic) +
    geom_density(aes(music_likes)) +
    facet_wrap(vars(degree), nrow=2)

(5) Not just decorative: axes

There is much more to learn about fine-tuning the appearance of plots. You have probably heard before about the importance of choosing a good scale for the axes. The range of data displayed on each axis is adjusted with coord_cartesian:

ggplot(gssmusic) +
    geom_point(aes(educ, realinc)) +
    coord_cartesian(xlim=c(0, 12), ylim=c(0, 50000))

Finally, and perhaps most importantly of all, label axes and title your plots! This can be done with the labs layer:

ggplot(gssmusic) +
    geom_point(aes(educ, realinc, color=sex), position="jitter") +
    coord_cartesian(xlim=c(0, 13), ylim=c(0, 50000)) +
    labs(x="years of education",
         y="reported income (1986$)",
         title="Income vs. education",
         color="Respondent sex",
         subtitle="This is DEFINITELY NOT THE WHOLE STORY",
         caption="Data from the 1993 General Social Survey, graphic by YOU")

1. Try to make the most deceptive possible version of an income vs. education plot.