tidyverse.Rmd

# The Tidyverse {#tidyverse}

```{r setup, include=FALSE}
source("etc/common.R")
```

## Questions
```{r questions, child="questions/tidyverse.md"}
```

## Learning Objectives
```{r objectives, child="objectives/tidyverse.md"}
```

There is no point in becoming fluent in Enochian if you do not then call forth a Dweller Beneath at the time of the new moon.
Similarly,
there is no point learning a language designed for data manipulation if you do not then bend data to your will.
This chapter, therefore, looks at how to do the things that R was summoned---er, designed---to do.

## How do I read data?

We begin by looking at the file `results/infant_hiv.csv`,
a tidied version of data on the percentage of infants born to women with HIV
who received an HIV test themselves within two months of birth.
The original data comes from the UNICEF site at <https://data.unicef.org/resources/dataset/hiv-aids-statistical-tables/>,
and this file contains:

```text
country,year,estimate,hi,lo
AFG,2009,NA,NA,NA
AFG,2010,NA,NA,NA
...
AFG,2017,NA,NA,NA
AGO,2009,NA,NA,NA
AGO,2010,0.03,0.04,0.02
AGO,2011,0.05,0.07,0.04
AGO,2012,0.06,0.08,0.05
...
ZWE,2016,0.71,0.88,0.62
ZWE,2017,0.65,0.81,0.57
```

The actual file has many more rows (and no ellipses).
It uses `NA` to show missing data rather than (for example) `-`, a space, or a blank,
and its values are interpreted as follows:

| Header | Datatype | Description |
|---|---|---|
| country | char | ISO3 country code of country reporting data |
| year | integer | year CE for which data reported |
| estimate | double/NA | estimated percentage of measurement |
| hi | double/NA | high end of range |
| lo | double/NA | low end of range |

We can load this data in Python like this:

```{python python-read-csv, output.lines=10}
import pandas as pd

infant_hiv = pd.read_csv('results/infant_hiv.csv')
print(infant_hiv)
```

The equivalent in R is to load the [tidyverse](glossary.html#tidyverse) collection of libraries
and then call the `read_csv` function.
We will go through this in stages, since each produces output.

```{r library-fail, eval=FALSE}
library(tidyverse)
```
```
Error in library(tidyverse) : there is no package called 'tidyverse'
```

Ah.
We must install this, which we only need to do once per machine:

```{r library-install, eval=FALSE}
install.packages("tidyverse")
```

We then load the library once per program:

```{r library-succeed, eval=FALSE}
library(tidyverse)
```
```
── Attaching packages ─────────────────────────────────────── tidyverse 1.2.1 ──
✔ ggplot2 3.1.0     ✔ purrr   0.3.0
✔ tibble  2.0.1     ✔ dplyr   0.7.8
✔ tidyr   0.8.2     ✔ stringr 1.4.0
✔ readr   1.1.1     ✔ forcats 0.3.0
── Conflicts ────────────────────────────────────────── tidyverse_conflicts() ──
✖ dplyr::filter() masks stats::filter()
✖ dplyr::lag()    masks stats::lag()
Warning messages:
1: package ‘tibble’ was built under R version 3.5.2
2: package ‘purrr’ was built under R version 3.5.2
3: package ‘stringr’ was built under R version 3.5.2
```

Note that to install, we give `install.packages` a string,
but to use,
we simply give the name of the library we want.

Asking for the tidyverse gives us eight libraries (or [packages](glossary.html#package)).
One of those, dplyr, defines two functions that mask standard functions in R with the same names.
This is deliberate, and if we need the originals, we can get them with their
[fully-qualified names](glossary.html#fully-qualified-name)
`stats::filter` and `stats::lag`.

Once we have the tidyverse loaded,
reading the file looks remarkably like reading the file:

```{r r-read-csv}
infant_hiv <- read_csv('results/infant_hiv.csv')
```

R's `read_csv` tells us more about what it has done than Pandas does.
In particular, it guesses the data types of columns based on the first thousand values
and then tells us what types it has inferred.
(In a better universe,
people would habitually use the first *two* rows of their spreadsheets for name *and units*,
but we do not live there.)

We can now look at what `read_csv` has produced.

```{r show-tibble}
infant_hiv
```

This is a [tibble](glossary.html#tibble),
which is the tidyverse's enhanced version of R's `data.frame`.
It organizes data into named columns,
each having one value for each row.

## How do I inspect data?

We often have a quick look at the content of a table to remind ourselves what it contains.
Pandas does this using methods whose names are borrowed from the Unix shell's `head` and `tail` commands:

```{python py-show-head}
print(infant_hiv.head())
```
```{python py-show-tail}
print(infant_hiv.tail())
```

R has similarly-named functions (not methods):

```{r r-show-head}
head(infant_hiv)
```
```{r r-show-tail}
tail(infant_hiv)
```

Let's have a closer look at that last command's output:

```{r r-tail-tibble, paged.print=FALSE}
tail(infant_hiv)
```

Note that the row numbers printed by `tail` are [relative](glossary.html#relative-row-number) to the output,
not [absolute](glossary.html#absolute-row-number) to the table.
This is different from Pandas,
which retains the original row numbers.
What about overall information?

```{python data-info}
print(infant_hiv.info())
```

```{r data-summary}
summary(infant_hiv)
```

Your display of R's summary may or may not wrap,
depending on how large a screen the older acolytes have allowed you.

## How do I index rows and columns?

A Pandas DataFrame is a collection of series (also called columns),
each containing the values of a single observed variable:

```{python py-string-subscript, output.lines=NA}
print(infant_hiv['estimate'])
```

We would get exactly the same output in Python with `infant_hiv.estimate`,
i.e.,
with an attribute name rather than a string subscript.
The same tricks work in R:

```{r r-string-subscript}
infant_hiv['estimate']
```

However, R's `infant_hiv$estimate` provides all the data:

```{r r-dollar-subscript, output.lines=NA}
infant_hiv$estimate
```

Again, note that the boxed number on the left is the start index of that row.

What about single values?
Remembering to count from zero from Python and as humans do for R,
we have:

```{python py-individual-element}
print(infant_hiv.estimate[11])
```
```{r r-individual-element}
infant_hiv$estimate[12]
```

Ah---everything in R is a vector,
so we get a vector of one value as an output rather than a single value.

```{python py-len-individual-element, error=TRUE}
print(len(infant_hiv.estimate[11]))
```
```{r r-len-individual-element}
length(infant_hiv$estimate[12])
```

And yes, ranges work:

```{python py-range-estimate}
print(infant_hiv.estimate[5:15])
```
```{r r-range-estimate}
infant_hiv$estimate[6:15]
```

Note that the upper bound is the same, because it's inclusive in R and exclusive in Python.
Note also that neither library prevents us from selecting a range of data that spans logical groups such as countries,
which is why selecting by row number is usually a sign of innocence, insouciance, or desperation.

We can select by column number as well.
Pandas uses the rather clumsy `object.iloc[rows, columns]`, with the usual `:` shortcut for "entire range":

```{python py-iloc, output.lines=NA}
print(infant_hiv.iloc[:, 0])
```

Since this is a column, it can be indexed:

```{python py-iloc-indexed}
print(infant_hiv.iloc[:, 0][0])
```

In R, a single index is interpreted as the column index:

```{r single-index-is-col}
infant_hiv[1]
```

But notice that the output is not a vector, but another tibble (i.e., a table with N rows and one column).
This means that adding another index does column-wise indexing on that tibble:

```{r double-index-of-tibble}
infant_hiv[1][1]
```

How then are we to get the first mention of Afghanistan?
The answer is to use [double square brackets](glossary.html#double-square-brackets) to strip away one level of structure:

```{r double-square-on-tibble, output.lines=NA}
infant_hiv[[1]]
```

This is now a plain old vector, so it can be indexed with [single square brackets](glossary.html#single-square-brackets):

```{r double-square-then-single}
infant_hiv[[1]][1]
```

But that too is a vector, so it can, of course, be indexed as well (for some value of "of course"):

```{r double-single-single}
infant_hiv[[1]][1][1]
```

Thus,
`data[1][[1]]` produces a tibble,
then selects the first column vector from it,
so it still gives us a vector.
*This is not madness.*
It is merely...differently sane.

> **Subsetting data frames:**
>
> When we are working with data frames (including tibbles),
> subsetting with a single vector selects columns, not rows,
> because data frames are stored as lists of columns.
> This means that `df[1:2]` selects two columns from `df`.
> However, in `df[2:3, 1:2]`, the first index selects rows, while the second selects columns.

## How do I calculate basic statistics?

What is the average estimate?
We start by grabbing that column for convenience:

```{python py-average-index}
estimates = infant_hiv.estimate
print(len(estimates))
```
```{python py-estimates-mean}
print(estimates.mean())
```

This translates almost directly to R:

```{r r-average-index}
estimates <- infant_hiv$estimate
length(estimates)
```
```{r r-estimates-mean}
mean(estimates)
```

The void is always there, waiting for us...
Let's fix this in R first by telling `mean` to drop NAs:

```{r r-remove-na}
mean(estimates, na.rm = TRUE)
```

And then try to get the statistically correct behavior in Pandas:

```{python py-remove-na}
print(estimates.mean(skipna=False))
```

Many functions in R use `na.rm` to control whether `NA`s are removed or not.
(Remember, the `.` character is just another part of the name)
R's default behavior is to leave `NA`s in, and then to include them in [aggregate](glossary.html#aggregation) computations.
Python's is to get rid of missing values early and work with what's left,
which makes translating code from one language to the next much more interesting than it might otherwise be.
But other than that, the statistics works the same way in Python:

```{python py-min-max-std}
print("min", estimates.min())
print("max", estimates.max())
print("std", estimates.std())
```

Here are the equivalent computations in R:

```{r r-min-max-std}
print(glue("min {min(estimates, na.rm = TRUE)}"))
print(glue("max {max(estimates, na.rm = TRUE)}"))
print(glue("sd {sd(estimates, na.rm = TRUE)}"))
```

A good use of aggregation is to check the quality of the data.
For example,
we can ask if there are any records where some of the estimate, the low value, or the high value are missing,
but not all of them:

```{python py-check-null}
print((infant_hiv.hi.isnull() != infant_hiv.lo.isnull()).any())
```
```{r r-check-null}
any(is.na(infant_hiv$hi) != is.na(infant_hiv$lo))
```

## How do I filter data?

By "[filtering](glossary.html#filter)", we mean "selecting records by value".
As discussed in Chapter \@ref(basics),
the simplest approach is to use a vector of logical values to keep only the values corresponding to `TRUE`.
In Python, this is:

```{python py-simple-filter}
maximal = estimates[estimates >= 0.95]
print(len(maximal))
```

And in R:

```{r r-simple-filter}
maximal <- estimates[estimates >= 0.95]
length(maximal)
```

The difference is unexpected.
Let's have a closer look at the result in Python:

```{python py-maximal, output.lines=NA}
print(maximal)
```

And in R:

```{r r-maximal, output.lines=NA}
maximal
```

It appears that R has kept the unknown values in order to highlight just how little we know.
More precisely,
wherever there was an `NA` in the original data
there is an `NA` in the logical vector
and hence an `NA` in the final vector.
Let us then turn to `which` to get a vector of indices at which a vector contains `TRUE`.
This function does not return indices for `FALSE` or `NA`:

```{r r-which}
which(estimates >= 0.95)
```

And as a quick check:

```{r r-length-which}
length(which(estimates >= 0.95))
```

So now we can index our vector with the result of the `which`:

```{r r-maximal-which}
maximal <- estimates[which(estimates >= 0.95)]
maximal
```

But should we do this?
Those `NA`s are important information,
and should not be discarded so blithely.
What we should *really* be doing is using the tools the tidyverse provides
rather than clever indexing tricks.
These behave consistently across a wide scale of problems
and encourage use of patterns that make it easier for others to understand our programs.

## How do I write tidy code?

The six basic data transformation operations in the tidyverse are:

- `filter`: choose observations (rows) by value(s)
- `arrange`: reorder rows
- `select`: choose variables (columns) by name
- `mutate`: derive new variables from existing ones
- `group_by`: define subsets of rows for further processing
- `summarize`: combine many values to create a single new value

`filter(tibble, ...criteria...)` keeps rows that pass all of the specified criteria:

```{r filter-as-function}
filter(infant_hiv, lo > 0.5)
```

Notice that the expression is `lo > 0.5` rather than `"lo" > 0.5`.
The latter expression would return the entire table
because the string `"lo"` is greater than the number 0.5 everywhere.

But how is it that the name `lo` can be used on its own?
It is the name of a column, but there is no variable called `lo`.
The answer is that R uses [lazy evaluation](glossary.html#lazy-evaluation):
function arguments aren't evaluated until they're needed,
so the function `filter` actually gets the expression `lo > 0.5`,
which allows it to check that there's a column called `lo` and then use it appropriately.
This is much tidier than `filter(data, data$lo > 0.5)` or `filter(data, "lo > 0.5")`.

Many languages rely on lazy evaluation,
and when used circumspectly,
it allows us to produce code that is easier to read.
We will explore it further in Chapter \@ref(nse).

We can make data anlaysis code more readable by using the [pipe operator](glossary.html#pipe-operator) `%>%`:

```{r filter-in-pipe}
infant_hiv %>% filter(lo > 0.5)
```

This may not seem like much of an improvement,
but neither does a Unix pipe consisting of `cat filename.txt | head`.
What about this?

```{r filter-complex}
filter(infant_hiv, (estimate != 0.95) & (lo > 0.5) & (hi <= (lo + 0.1)))
```

It uses the vectorized "and" operator `&` twice,
and parsing the condition takes a human being at least a few seconds.
Its pipelined equivalent is:

```{r filter-complex-pipe}
infant_hiv %>% filter(estimate != 0.95) %>% filter(lo > 0.5) %>% filter(hi <= (lo + 0.1))
```

Breaking the condition into stages like this often makes reading and testing much easier,
and encourages incremental write-test-extend development.

Let's increase the band from 10% to 20%:

```{r filter-wider-band}
infant_hiv %>% filter(estimate != 0.95) %>% filter(lo > 0.5) %>% filter(hi <= (lo + 0.2))
```

and then order by `lo` in descending order,
breaking the line the way the [tidyverse style guide][tidyverse-style] recommends:

```{r filter-arrange}
infant_hiv %>%
  filter(estimate != 0.95) %>%
  filter(lo > 0.5) %>%
  filter(hi <= (lo + 0.2)) %>%
  arrange(desc(lo))
```

We can now [select](glossary.html#select) the three columns we care about:

```{r filter-arrange-select}
infant_hiv %>%
  filter(estimate != 0.95) %>%
  filter(lo > 0.5) %>%
  filter(hi <= (lo + 0.2)) %>%
  arrange(desc(lo)) %>%
  select(year, lo, hi)
```

Once again,
we are using the unquoted column names `year`, `lo`, and `hi`
and letting R's lazy evaluation take care of the details for us.

Rather than selecting these three columns,
we can [select *out*](glossary.html#negative-selection) the columns we're not interested in by negating their names.
This leaves the columns that are kept in their original order,
rather than putting `lo` before `hi`,
which won't matter if we later select by name,
but *will* if we ever want to select by position:

```{r select-out}
infant_hiv %>%
  filter(estimate != 0.95) %>%
  filter(lo > 0.5) %>%
  filter(hi <= (lo + 0.2)) %>%
  arrange(desc(lo)) %>%
  select(-country, -estimate)
```

Giddy with power,
we now add a column containing the difference between the low and high values.
This can be done using either `mutate`,
which adds new columns to the end of an existing tibble,
or with `transmute`,
which creates a new tibble containing only the columns we explicitly ask for.
(There is also a function `rename` which simply renames columns.)
Since we want to keep `hi` and `lo`,
we decide to use `mutate`:

```{r mutate-new-column}
infant_hiv %>%
  filter(estimate != 0.95) %>%
  filter(lo > 0.5) %>%
  filter(hi <= (lo + 0.2)) %>%
  arrange(desc(lo)) %>%
  select(-country, -estimate) %>%
  mutate(difference = hi - lo)
```

Does the difference between high and low estimates vary by year?
To answer that question,
we use `group_by` to [group](glossary.html#group) records by value
and then `summarize` to aggregate within groups.
We might as well get rid of the `arrange` and `select` calls in our pipeline at this point,
since we're not using them,
and count how many records contributed to each aggregation using `n()`:

```{r summarize-and-count}
infant_hiv %>%
  filter(estimate != 0.95) %>%
  filter(lo > 0.5) %>%
  filter(hi <= (lo + 0.2)) %>%
  mutate(difference = hi - lo) %>%
  group_by(year) %>%
  summarize(count = n(), ave_diff = mean(year))
```

How might we do this with Pandas?
One approach is to use a single multi-part `.query` to select data
and store the result in a variable so that we can refer to the `hi` and `lo` columns twice
without repeating the filtering expression.
We then group by year and aggregate, again using strings for column names:

```{python equivalent-to-pipeline}
data = pd.read_csv('results/infant_hiv.csv')
data = data.query('(estimate != 0.95) & (lo > 0.5) & (hi <= (lo + 0.2))')
data = data.assign(difference = (data.hi - data.lo))
grouped = data.groupby('year').agg({'difference' : {'ave_diff' : 'mean', 'count' : 'count'}})
print(grouped)
```

There are other ways to tackle this problem with Pandas,
but the tidyverse approach produces code that I find more readable.

## How do I model my data?

Tidying up data can be as calming and rewarding in the same way as knitting
or rearranging the specimen jars on the shelves in your dining room-stroke-laboratory.
Eventually,
though,
people want to do some statistics.
The simplest tool for this in R is `lm`, which stands for "linear model".
Given a formula and a data set,
it calculates coefficients to fit that formula to that data:

```{r simple-formula}
lm(estimate ~ lo, data = infant_hiv)
```

This is telling us that `estimate` is more-or-less equal to `0.0421 + 1.0707 * lo`.
The `~` symbol is used to separate the left and right sides of the equation,
and as with all things tidyverse,
lazy evaluation allows us to use variable names directly.
In fact,
it lets us write much more complex formulas involving functions of multiple variables.
For example,
we can regress `estimate` against the square roots of `lo` and `hi`,
though there is no good statistical reason to do so:

```{r complex-formula}
lm(estimate ~ sqrt(lo) + sqrt(hi), data = infant_hiv)
```

One important thing to note here is the way that `+` is overloaded in formulas.
The formula `estimate ~ lo + hi` does *not* mean "regress `estimate` against the sum of `lo` and `hi`",
but rather, "regress `estimate` against the two variables `lo` and `hi`":

```{r double-regression}
lm(estimate ~ lo + hi, data = infant_hiv)
```

If we want to regress `estiate` against the average of `lo` and `hi`
(i.e., regress `estimate` against a single calculated variable instead of against two variables)
we need to create a temporary column:

```{r regress-temporary}
infant_hiv %>%
  mutate(ave_lo_hi = (lo + hi)/2) %>%
  lm(estimate ~ ave_lo_hi, data = .)
```

Here, the call to `lm` is using the variable `.` to mean "the data coming in from the previous stage of the pipeline".
Most of the functions in the tidyverse use this convention
so that data can be passed to a function that expects it in a position other than the first.

## How do I create a plot?

One of R's greatest strengths is the tools it gives us for seeing the hitherto unseeable.
The most popular of these tools is `ggplot2`,
which implements and extends the patterns described in @Wilk2005.
Every chart it creates has a [geometry](glossary.html#geometry) that controls how data is displayed,
and a [mapping](glossary.html#mapping) that controls how values are mapped to geometric properties.
For example,
these lines of code create a scatter plot showing the relationship between `lo` and `hi` values in the infant HIV data:

```{r basic-plot}
ggplot(infant_hiv) + geom_point(mapping = aes(x = lo, y = hi))
```

Looking more closely:

-   The function `ggplot` creates an object to represent the chart with `infant_hiv` as the underlying data.
-   `geom_point` specifies the geometry we want (points).
-   Its `mapping` argument is assigned an [aesthetic](glossary.html#aesthetic)
    that specifies `lo` is to be used as the `x` coordinate and `hi` is to be used as the `y` coordinate.
-   The elements of the chart are combined with `+` rather than `%>%` for historical reasons.

Let's create a slightly more appealing plot by dropping NAs,
making the points semi-transparent,
and colorizing them according to the value of `estimate`:

```{r plot-after-drop}
infant_hiv %>%
  drop_na() %>%
  ggplot() +
  geom_point(mapping = aes(x = lo, y = hi, color = estimate), alpha = 0.5) +
  xlim(0.0, 1.0) + ylim(0.0, 1.0)
```

We set the transparency `alpha` outside the aesthetic because its value is constant for all points:
if we set it inside `aes(...)`,
we would be telling ggplot2 to set the transparency according to the value of the data.
We specify the limits to the axes manually with `xlim` and `ylim` to ensure that ggplot2 includes the upper bounds:
we found by trial and error that without this,
all of the data would be shown,
but the upper label "1.00" would be omitted.

This plot immediately shows us that we have some outliers.
There are far more values with `hi` equal to 0.95 than it seems there ought to be,
and there are eight points running up the left margin that seem troubling as well.
Let's create a new tibble that doesn't have these:

```{r plot-remove-outliers}
infant_hiv %>%
  drop_na() %>%
  filter(hi != 0.95) %>%
  filter(!((lo < 0.10) & (hi > 0.25))) %>%
  ggplot() +
  geom_point(mapping = aes(x = lo, y = hi, color = estimate), alpha = 0.5) +
  xlim(0.0, 1.0) + ylim(0.0, 1.0)
```

We can add the fitted curve by including another geometry:

```{r plot-with-fit}
infant_hiv %>%
  drop_na() %>%
  filter(hi != 0.95) %>%
  filter(!((lo < 0.10) & (hi > 0.25))) %>%
  ggplot() +
  geom_point(mapping = aes(x = lo, y = hi, color = estimate), alpha = 0.5) +
  geom_smooth(method = lm, mapping = aes(x = lo, y = hi), color = 'red') +
  xlim(0.0, 1.0) + ylim(0.0, 1.0)
```

But wait:
why is this complaining about missing values?
Some online searches and guidance from the hill gods led to the discovery that
`geom_smooth` adds virtual points to the data for plotting purposes,
some of which lie outside the range of the actual data,
and that setting `xlim` and `ylim` then truncates these.
(Differently sane...)
The safe way to control the range of the data is to add a call to `coord_cartesian`,
which effectively zooms in on a region of interest:

```{r plot-cartesian}
infant_hiv %>%
  drop_na() %>%
  filter(hi != 0.95) %>%
  filter(!((lo < 0.10) & (hi > 0.25))) %>%
  ggplot() +
  geom_point(mapping = aes(x = lo, y = hi, color = estimate), alpha = 0.5) +
  geom_smooth(method = lm, mapping = aes(x = lo, y = hi), color = 'red') +
  coord_cartesian(xlim = c(0.0, 1.0), ylim = c(0.0, 1.0))
```

## Key Points
```{r keypoints, child="keypoints/tidyverse.md"}
```

```{r links, child="etc/links.md"}
```