--- title: "Data Science for Economists" # subtitle: "
" subtitle: "Lecture 5: Data cleaning & wrangling: (1) Tidyverse" author: "Grant McDermott" date: "University of Oregon | [EC 607](https://github.com/uo-ec607)" #"`r format(Sys.time(), '%d %B %Y')`" output: xaringan::moon_reader: css: [default, metropolis, metropolis-fonts] lib_dir: libs nature: highlightStyle: github highlightLines: true highlightSpans: true countIncrementalSlides: false --- name: toc ```{css, echo=FALSE} @media print { .has-continuation { display: block !important; } } ``` ```{r setup, include=FALSE} options(htmltools.dir.version = FALSE) library(knitr) opts_chunk$set( fig.align="center", #fig.width=6, fig.height=4.5, # out.width="748px", #out.length="520.75px", dpi=300, #fig.path='Figs/', cache=T#, echo=F, warning=F, message=F ) ``` # Table of contents 1. [Prologue](#prologue) 2. [Tidyverse basics](#basics) 3. [Data wrangling with dplyr](#dplyr) - [filter](#filter) - [arrange](#arrange) - [select](#select) - [mutate](#mutate) - [summarise](#summarise) - [joins](#joins) 4. [Data tidying with tidyr](#tidyr) - [pivot_longer](#pivot_longer) / [pivot_wider](#pivot_wider) - [separate](#separate) - [unite](#unite) 5. [Summary](#summary) --- class: inverse, center, middle name: prologue # Prologue
--- # What is "tidy" data? ### Resources: - [Vignette](https://cran.r-project.org/web/packages/tidyr/vignettes/tidy-data.html) (from the **tidyr** package) - [Original paper](https://vita.had.co.nz/papers/tidy-data.pdf) (Hadley Wickham, 2014 JSS) --
### Key points: 1. Each variable forms a column. 2. Each observation forms a row. 3. Each type of observational unit forms a table. --
Basically, tidy data is more likely to be [long (i.e. narrow) format](https://en.wikipedia.org/wiki/Wide_and_narrow_data) than wide format. --- # Checklist ### R packages you'll need today ☑ [**tidyverse**](https://www.tidyverse.org/) ☑ [**nycflights13**](hhttps://github.com/hadley/nycflights13) --
I'll hold off loading these libraries for now. But you can install/update them both with the following command. ```r install.packages(c('tidyverse', 'nycflights13'), repos = 'https://cran.rstudio.com', dependencies = TRUE) ``` **Tip:** If you're on Linux, then I _strongly_ recommend installing the pre-compiled binary versions of these packages from [RSPM](https://packagemanager.rstudio.com/client/#/repos/1/overview) instead of CRAN. The exact repo mirror varies by distro (see the link). But on Ubuntu 20.04, for example, you'd use: ```r install.packages(c('tidyverse', 'nycflights13'), repos = 'https://packagemanager.rstudio.com/all/__linux__/focal/latest', dependencies = TRUE) ``` --- class: inverse, center, middle name: basics # Tidyverse basics
--- # Tidyverse vs. base R Much digital ink has been spilled over the "tidyverse vs. base R" debate. -- I won't delve into this debate here, because I think the answer is [clear](http://varianceexplained.org/r/teach-tidyverse/): We should teach the tidyverse first (or, at least, early). - The documentation and community support are outstanding. - Having a consistent philosophy and syntax makes it easier to learn. - Provides a convenient "front-end" to big data tools that we'll use later in the course. - For data cleaning, wrangling, and plotting, the tidyverse really is a no-brainer.1 .footnote[ 1 I'm also a huge fan of [**data.table**](http://r-datatable.com/). This package will be the subject of our next lecture. ] -- **But**... this certainly shouldn't put you off learning base R alternatives. - Base R is extremely flexible and powerful (and stable). - There are some things that you'll have to venture outside of the tidyverse for. - A combination of tidyverse and base R is often the best solution to a problem. - Excellent base R data manipulation tutorials: [here](https://www.rspatial.org/intr/index.html) and [here](https://github.com/matloff/fasteR). --- # Tidyverse vs. base R (cont.) One point of convenience is that there is often a direct correspondence between a tidyverse command and its base R equivalent. These generally follow a `tidyverse::snake_case` vs `base::period.case` rule. E.g. Compare: | tidyverse | base | |---|---| | `?readr::read_csv` | `?utils::read.csv` | | `?dplyr::if_else` | `?base::ifelse` | | `?tibble::tibble` | `?base::data.frame` | Etcetera. If you call up the above examples, you'll see that the tidyverse alternative typically offers some enhancements or other useful options (and sometimes restrictions) over its base counterpart. -- **Remember:** There are (almost) always multiple ways to achieve a single goal in R. --- # Tidyverse packages Let's load the tidyverse meta-package and check the output. ```{r tverse, cache = FALSE} library(tidyverse) ``` -- We see that we have actually loaded a number of packages (which could also be loaded individually): **ggplot2**, **tibble**, **dplyr**, etc. - We can also see information about the package versions and some [namespace conflicts](https://raw.githack.com/uo-ec607/lectures/master/04-rlang/04-rlang.html#59). --- # Tidyverse packages (cont.) The tidyverse actually comes with a lot more packages than those that are just loaded automatically.1 ```{r tverse_pkgs} tidyverse_packages() ``` We'll use several of these additional packages during the remainder of this course. - E.g. The **lubridate** package for working with dates and the **rvest** package for webscraping. - However, bear in mind that these packages will have to be loaded separately. .footnote[ 1 It also includes a *lot* of dependencies upon installation. This is a matter of some [controversy](http://www.tinyverse.org/). ] --- # Tidyverse packages (cont.) I hope to cover most of the tidyverse packages over the length of this course. Today, however, I'm only really going to focus on two packages: 1. [**dplyr**](https://dplyr.tidyverse.org/) 2. [**tidyr**](https://tidyr.tidyverse.org/) These are the workhorse packages for cleaning and wrangling data. They are thus the ones that you will likely make the most use of (alongside **ggplot2**, which we already met back in Lecture 1). - Data cleaning and wrangling occupies an inordinate amount of time, no matter where you are in your research career. --- # An aside on pipes: %>% We already learned about pipes in our [lecture](https://raw.githack.com/uo-ec607/lectures/master/03-shell/03-shell.html#91) on the bash shell. The tidyverse loads its own pipe operator, denoted `%>%`. I want to reiterate how cool pipes are, and how using them can dramatically improve the experience of reading and writing code. Compare: ```{r, eval = F} ## These next two lines of code do exactly the same thing. mpg %>% filter(manufacturer=="audi") %>% group_by(model) %>% summarise(hwy_mean = mean(hwy)) summarise(group_by(filter(mpg, manufacturer=="audi"), model), hwy_mean = mean(hwy)) ``` -- The first line reads from left to right, exactly how I thought of the operations in my head. - Take this object (`mpg`), do this (`filter`), then do this (`group_by`), etc. The second line totally inverts this logical order (the final operation comes first!) - Who wants to read things inside out? --- # An aside on pipes: %>% (cont.) The piped version of the code is even more readable if we write it over several lines. Here it is again and, this time, I'll run it for good measure so you can see the output: ```{r pipe} mpg %>% filter(manufacturer=="audi") %>% group_by(model) %>% summarise(hwy_mean = mean(hwy)) ``` Remember: Using vertical space costs nothing and makes for much more readable/writeable code than cramming things horizontally. -- PS — The pipe is originally from the [**magrittr**](https://magrittr.tidyverse.org/) package ([geddit?](https://en.wikipedia.org/wiki/The_Treachery_of_Images)), which can do some other cool things if you're inclined to explore. --- name: nativepipe # A further aside on the base R pipe: |> The magrittr pipe has proven so successful and popular, that the R core team [recently announced](https://t.co/HVECPENQ5C?amp=1) a "native" pipe would be coming to base R, denoted `|>`.1 For example: ```r mtcars |> subset(cyl==4) |> head() mtcars |> subset(cyl==4) |> d => lm(mpg ~ disp, data = d) ``` .footnote[1 That's actually a `|` followed by a `>`. The default font on these slides just makes it look extra fancy.] -- At the time of writing this native pipe is only available in the [development](https://stat.ethz.ch/R-manual/R-devel/library/base/html/pipeOp.html) version of R. (I'll show an in-class demo.) This native pipe complements some other new cool features, like support for ["lambda" functions](https://stackoverflow.com/questions/16501/what-is-a-lambda-function) in R. - So, worth watching this space. --- class: inverse, center, middle name: dplyr # dplyr
--- # Aside: dplyr 1.0.0 release Some of the **dplyr** features that we'll cover today were introduced in [version 1.0.0](https://www.tidyverse.org/blog/2020/06/dplyr-1-0-0/) of the package. - Version 1.0.0 is a big deal since it marks a stable code base for the package going forward. However, at the time of writing these slides, it had only come out very recently. - Please make sure that you are running at least **dplyr** 1.0.0 before continuing. ```{r dplyr_vers, cache=FALSE} packageVersion('dplyr') # install.packages('dplyr') ## install updated version if < 1.0.0 ``` -- *Note:* **dplyr** 1.0.0 also notifies you about grouping variables every time you do operations on or with them. YMMV, but, personally, I find these messages annoying and so prefer to [switch them off](https://twitter.com/MattCowgill/status/1278463099272491008). ```r options(dplyr.summarise.inform = FALSE) ## Add to .Rprofile to make permanent ``` --- # Key dplyr verbs There are five key dplyr verbs that you need to learn. 1. `filter`: Filter (i.e. subset) rows based on their values. 2. `arrange`: Arrange (i.e. reorder) rows based on their values. 3. `select`: Select (i.e. subset) columns by their names: 4. `mutate`: Create new columns. 5. `summarise`: Collapse multiple rows into a single summary value.1 .footnote[ 1 `summarize` with a "z" works too. R doesn't discriminate against uncivilised nations of the world. ] --
Let's practice these commands together using the `starwars` data frame that comes pre-packaged with dplyr. --- name: filter # 1) dplyr::filter We can chain multiple filter commands with the pipe (`%>%`), or just separate them within a single filter command using commas. ```{r filter1} starwars %>% filter( species == "Human", height >= 190 ) ``` --- # 1) dplyr::filter *cont.* Regular expressions work well too. ```{r filter2} starwars %>% filter(grepl("Skywalker", name)) ``` --- # 1) dplyr::filter *cont.* A very common `filter` use case is identifying (or removing) missing data cases. ```{r filter3} starwars %>% filter(is.na(height)) ``` --
To remove missing observations, simply use negation: `filter(!is.na(height))`. Try this yourself. --- name: arrange # 2) dplyr::arrange ```{r arrange1} starwars %>% arrange(birth_year) ``` -- *Note:* Arranging on a character-based column (i.e. strings) will sort alphabetically. Try this yourself by arranging according to the "name" column. --- # 2) dplyr::arrange *cont.* We can also arrange items in descending order using `arrange(desc())`. ```{r arrange2} starwars %>% arrange(desc(birth_year)) ``` --- name: select # 3) dplyr::select Use commas to select multiple columns out of a data frame. (You can also use "first:last" for consecutive columns). Deselect a column with "-". ```{r select1} starwars %>% select(name:skin_color, species, -height) ``` --- # 3) dplyr::select *cont.* You can also rename some (or all) of your selected variables in place. ```{r select2} starwars %>% select(alias=name, crib=homeworld, sex=gender) ``` -- If you just want to rename columns without subsetting them, you can use `rename`. Try this now by replacing `select(...)` in the above code chunk with `rename(...)`. --- # 3) dplyr::select *cont.* The `select(contains(PATTERN))` option provides a nice shortcut in relevant cases. ```{r select3} starwars %>% select(name, contains("color")) ``` --- # 3) dplyr::select *cont.* The `select(..., everything())` option is another useful shortcut if you only want to bring some variable(s) to the "front" of a data frame. ```{r select4} starwars %>% select(species, homeworld, everything()) %>% head(5) ``` --
*Note:* The new `relocate` function available in dplyr 1.0.0 has brought a lot more functionality to ordering of columns. See [here](https://www.tidyverse.org/blog/2020/03/dplyr-1-0-0-select-rename-relocate/). --- name: mutate # 4) dplyr::mutate You can create new columns from scratch, or (more commonly) as transformations of existing columns. ```{r mutate1} starwars %>% select(name, birth_year) %>% mutate(dog_years = birth_year * 7) %>% mutate(comment = paste0(name, " is ", dog_years, " in dog years.")) ``` --- # 4) dplyr::mutate *cont.* *Note:* `mutate` is order aware. So you can chain multiple mutates in a single call. ```{r mutate2} starwars %>% select(name, birth_year) %>% mutate( dog_years = birth_year * 7, ## Separate with a comma comment = paste0(name, " is ", dog_years, " in dog years.") ) ``` --- # 4) dplyr::mutate *cont.* Boolean, logical and conditional operators all work well with `mutate` too. ```{r mutate3} starwars %>% select(name, height) %>% filter(name %in% c("Luke Skywalker", "Anakin Skywalker")) %>% mutate(tall1 = height > 180) %>% mutate(tall2 = ifelse(height > 180, "Tall", "Short")) ## Same effect, but can choose labels ``` --- # 4) dplyr::mutate *cont.* Lastly, combining `mutate` with the new `across` feature in dplyr 1.0.0+ allows you to easily work on a subset of variables. For example: ```{r, mutate4} starwars %>% select(name:eye_color) %>% mutate(across(where(is.character), toupper)) %>% #<< head(5) ``` --
*Note:* This workflow (i.e. combining `mutate` and `across`) supersedes the old "scoped" variants of `mutate` that you might have used previously. More details [here](https://www.tidyverse.org/blog/2020/04/dplyr-1-0-0-colwise/) and [here](https://dplyr.tidyverse.org/dev/articles/colwise.html). --- name: summarise # 5) dplyr::summarise Particularly useful in combination with the `group_by` command. ```{r summ1} starwars %>% group_by(species, gender) %>% summarise(mean_height = mean(height, na.rm = TRUE)) ``` --- # 5) dplyr::summarise *cont.* Note that including "na.rm = TRUE" (or, its alias "na.rm = T") is usually a good idea with summarise functions. Otherwise, any missing value will propogate to the summarised value too. ```{r summ2} ## Probably not what we want starwars %>% summarise(mean_height = mean(height)) ## Much better starwars %>% summarise(mean_height = mean(height, na.rm = TRUE)) ``` --- # 5) dplyr::summarise *cont.* The same `across`-based workflow that we saw with `mutate` a few slides back also works with `summarise`. For example: ```{r, summ4} starwars %>% group_by(species) %>% summarise(across(where(is.numeric), mean, na.rm=T)) %>% #<< head(5) ``` --
*Note:* Again, this functionality supersedes the old "scoped" variants of `summarise` that you used prior to dplyr 1.0.0. Details [here](https://www.tidyverse.org/blog/2020/04/dplyr-1-0-0-colwise/) and [here](https://dplyr.tidyverse.org/dev/articles/colwise.html). --- # Other dplyr goodies `group_by` and `ungroup`: For (un)grouping. - Particularly useful with the `summarise` and `mutate` commands, as we've already seen. -- `slice`: Subset rows by position rather than filtering by values. - E.g. `starwars %>% slice(c(1, 5))` -- `pull`: Extract a column from as a data frame as a vector or scalar. - E.g. `starwars %>% filter(gender=="female") %>% pull(height)` -- `count` and `distinct`: Number and isolate unique observations. - E.g. `starwars %>% count(species)`, or `starwars %>% distinct(species)` - You could also use a combination of `mutate`, `group_by`, and `n()`, e.g. `starwars %>% group_by(species) %>% mutate(num = n())`. --- # Other dplyr goodies (cont.) There are also a whole class of [window functions](https://cran.r-project.org/web/packages/dplyr/vignettes/window-functions.html) for getting leads and lags, ranking, creating cumulative aggregates, etc. - See `vignette("window-functions")`. --
The final set of dplyr "goodies" are the family of join operations. However, these are important enough that I want to go over some concepts in a bit more depth... - We will encounter and practice these many more times as the course progresses. --- name: joins # Joins One of the mainstays of the dplyr package is merging data with the family [join operations](https://cran.r-project.org/web/packages/dplyr/vignettes/two-table.html). - `inner_join(df1, df2)` - `left_join(df1, df2)` - `right_join(df1, df2)` - `full_join(df1, df2)` - `semi_join(df1, df2)` - `anti_join(df1, df2)` (You find find it helpful to to see visual depictions of the different join operations [here](https://r4ds.had.co.nz/relational-data.html).) -- For the simple examples that I'm going to show here, we'll need some data sets that come bundled with the [**nycflights13**](http://github.com/hadley/nycflights13) package. - Load it now and then inspect these data frames in your own console. ```{r flights, echo = F} library(nycflights13) ``` ```{r, eval = F} library(nycflights13) flights planes ``` --- # Joins (cont.) Let's perform a [left join](https://stat545.com/bit001_dplyr-cheatsheet.html#left_joinsuperheroes-publishers) on the flights and planes datasets. - *Note*: I'm going subset columns after the join, but only to keep text on the slide. -- ```{r join1} left_join(flights, planes) %>% select(year, month, day, dep_time, arr_time, carrier, flight, tailnum, type, model) ``` --- # Joins (cont.) (*continued from previous slide*) Note that dplyr made a reasonable guess about which columns to join on (i.e. columns that share the same name). It also told us its choices: ``` *## Joining, by = c("year", "tailnum") ``` However, there's an obvious problem here: the variable "year" does not have a consistent meaning across our joining datasets! - In one it refers to the *year of flight*, in the other it refers to *year of construction*. -- Luckily, there's an easy way to avoid this problem. - See if you can figure it out before turning to the next slide. - Try `?dplyr::join`. --- # Joins (cont.) (*continued from previous slide*) You just need to be more explicit in your join call by using the `by = ` argument. - You can also rename any ambiguous columns to avoid confusion. ```{r join2} left_join( flights, planes %>% rename(year_built = year), ## Not necessary w/ below line, but helpful by = "tailnum" ## Be specific about the joining column ) %>% select(year, month, day, dep_time, arr_time, carrier, flight, tailnum, year_built, type, model) %>% head(3) ## Just to save vertical space on the slide ``` --- # Joins (cont.) (*continued from previous slide*) Last thing I'll mention for now; note what happens if we again specify the join column... but don't rename the ambiguous "year" column in at least one of the given data frames. ```{r join3} left_join( flights, planes, ## Not renaming "year" to "year_built" this time by = "tailnum" ) %>% select(contains("year"), month, day, dep_time, arr_time, carrier, flight, tailnum, type, model) %>% head(3) ``` -- Make sure you know what "year.x" and "year.y" are. Again, it pays to be specific. --- class: inverse, center, middle name: tidyr # tidyr
--- # Key tidyr verbs 1. `pivot_longer`: Pivot wide data into long format (i.e. "melt").1 2. `pivot_wider`: Pivot long data into wide format (i.e. "cast").2 3. `separate`: Separate (i.e. split) one column into multiple columns. 4. `unite`: Unite (i.e. combine) multiple columns into one. .footnote[ 1 Updated version of `tidyr::gather`. 2 Updated version of `tidyr::spread`. ] --
Let's practice these verbs together in class. - Side question: Which of `pivot_longer` vs `pivot_wider` produces "tidy" data? --- name: pivot_longer # 1) tidyr::pivot_longer ```{r pivot_longer1} stocks = data.frame( ## Could use "tibble" instead of "data.frame" if you prefer time = as.Date('2009-01-01') + 0:1, X = rnorm(2, 0, 1), Y = rnorm(2, 0, 2), Z = rnorm(2, 0, 4) ) stocks stocks %>% pivot_longer(-time, names_to="stock", values_to="price") ``` --- # 1) tidyr::pivot_longer *cont.* Let's quickly save the "tidy" (i.e. long) stocks data frame for use on the next slide. ```{r pivot_longer2} ## Write out the argument names this time: i.e. "names_to=" and "values_to=" tidy_stocks = stocks %>% pivot_longer(-time, names_to="stock", values_to="price") ``` --- name: pivot_wider # 2) tidyr::pivot_wider ```{r pivot_wider1, dependson=tidy_stocks} tidy_stocks %>% pivot_wider(names_from=stock, values_from=price) tidy_stocks %>% pivot_wider(names_from=time, values_from=price) ``` --
Note that the second example — which has combined different pivoting arguments — has effectively transposed the data. --- # Aside: Remembering the pivot_* syntax There's a long-running joke about no-one being able to remember Stata's "reshape" command. ([Exhibit A](https://twitter.com/helleringer143/status/1117234887902285836).) It's easy to see this happening with the `pivot_*` functions too. However, I find that I never forget the commands as long as I remember the argument order is *"names"* then *"values"*. --- name: separate # 3) tidyr::separate ```{r sep1} economists = data.frame(name = c("Adam.Smith", "Paul.Samuelson", "Milton.Friedman")) economists economists %>% separate(name, c("first_name", "last_name")) ``` --
This command is pretty smart. But to avoid ambiguity, you can also specify the separation character with `separate(..., sep=".")`. --- # 3) tidyr::separate *cont.* A related function is `separate_rows`, for splitting up cells that contain multiple fields or observations (a frustratingly common occurence with survey data). ```{r sep2} jobs = data.frame( name = c("Jack", "Jill"), occupation = c("Homemaker", "Philosopher, Philanthropist, Troublemaker") ) jobs ## Now split out Jill's various occupations into different rows jobs %>% separate_rows(occupation) ``` --- name: unite # 4) tidyr::unite ```{r unite1} gdp = data.frame( yr = rep(2016, times = 4), mnth = rep(1, times = 4), dy = 1:4, gdp = rnorm(4, mean = 100, sd = 2) ) gdp ## Combine "yr", "mnth", and "dy" into one "date" column gdp %>% unite(date, c("yr", "mnth", "dy"), sep = "-") ``` --- # 4) tidyr::unite *cont.* Note that `unite` will automatically create a character variable. You can see this better if we convert it to a tibble. ```{r unite2} gdp_u = gdp %>% unite(date, c("yr", "mnth", "dy"), sep = "-") %>% as_tibble() gdp_u ``` -- If you want to convert it to something else (e.g. date or numeric) then you will need to modify it using `mutate`. See the next slide for an example, using the [lubridate](https://lubridate.tidyverse.org/) package's super helpful date conversion functions. --- # 4) tidyr::unite *cont.* *(continued from previous slide)* ```{r unite3, message=F} library(lubridate) gdp_u %>% mutate(date = ymd(date)) ``` --- # Other tidyr goodies Use `crossing` to get the full combination of a group of variables.1 ```{r cross1} crossing(side=c("left", "right"), height=c("top", "bottom")) ``` .footnote[ 1 Base R alternative: `expand.grid`. ] -- See `?expand` and `?complete` for more specialised functions that allow you to fill in (implicit) missing data or variable combinations in existing data frames. - You'll encounter this during your next assignment. --- class: inverse, center, middle name: summary # Summary
--- # Key verbs ### dplyr 1. `filter` 2. `arrange` 3. `select` 4. `mutate` 5. `summarise` ### tidyr 1. `pivot_longer` 2. `pivot_wider` 3. `separate` 4. `unite` -- Other useful items include: pipes (`%>%`), grouping (`group_by`), joining functions (`left_join`, `inner_join`, etc.). --- class: inverse, center, middle # Next lecture: Data cleaning and wrangling: (2) data.table
```{r gen_pdf, include = FALSE, cache = FALSE, eval = TRUE} infile = list.files(pattern = '.html') pagedown::chrome_print(input = infile, timeout = 100) ```