assertr tutorial

for v1.0.2

In data analysis workflows that depend on un-sanitized data sets from external sources, it’s very common that errors in data bring an analysis to a screeching halt. Oftentimes, these errors occur late in the analysis and provide no clear indication of which datum caused the error.

On occasion, the error resulting from bad data won’t even appear to be a data error at all. Still worse, errors in data will pass through analysis without error, remain undetected, and produce inaccurate results.

The solution to the problem is to provide as much information as you can about how you expect the data to look up front so that any deviation from this expectation can be dealt with immediately. This is what the assertr package tries to make dead simple.

Essentially, assertr provides a suite of functions designed to verify assumptions about data early in an analysis pipeline. This package needn't be used with the magrittr/dplyr piping mechanism but the examples in this vignette will use them to enhance clarity.

Installation

Stable version from CRAN

install.packages("assertr")

Development version from GitHub

if (!require("devtools")) install.packages("devtools")
devtools::install_github("ropenscilabs/assertr")
library("assertr")

Usage

concrete data errors

Let’s say, for example, that the R’s built-in car dataset, mtcars, was not built-in but rather procured from an external source that was known for making errors in data entry or coding.

In particular, the mtcars dataset looks like this:

head(mtcars)
#>                    mpg cyl disp  hp drat    wt  qsec vs am gear carb
#> Mazda RX4         21.0   6  160 110 3.90 2.620 16.46  0  1    4    4
#> Mazda RX4 Wag     21.0   6  160 110 3.90 2.875 17.02  0  1    4    4
#> Datsun 710        22.8   4  108  93 3.85 2.320 18.61  1  1    4    1
#> Hornet 4 Drive    21.4   6  258 110 3.08 3.215 19.44  1  0    3    1
#> Hornet Sportabout 18.7   8  360 175 3.15 3.440 17.02  0  0    3    2
#> Valiant           18.1   6  225 105 2.76 3.460 20.22  1  0    3    1

But let's pretend that the data we got accidentally negated the 5th mpg value:

our.data <- mtcars
our.data$mpg[5] <- our.data$mpg[5] * -1
our.data[4:6,]
#>                     mpg cyl disp  hp drat    wt  qsec vs am gear carb
#> Hornet 4 Drive     21.4   6  258 110 3.08 3.215 19.44  1  0    3    1
#> Hornet Sportabout -18.7   8  360 175 3.15 3.440 17.02  0  0    3    2
#> Valiant            18.1   6  225 105 2.76 3.460 20.22  1  0    3    1

Whoops!

If we wanted to find the average miles per gallon for each number of engine cylinders, we might do so like this:

library(dplyr)

our.data %>%
  group_by(cyl) %>%
  summarise(avg.mpg=mean(mpg))
#> # A tibble: 3 × 2
#>     cyl  avg.mpg
#>   <dbl>    <dbl>
#> 1     4 26.66364
#> 2     6 19.74286
#> 3     8 12.42857

This indicates that the average miles per gallon for a 8 cylinder car is a lowly 12.43. However, in the correct dataset it's really just over 15. Data errors like that are extremely easy to miss because it doesn't cause an error, and the results look reasonable.

enter assertr

To combat this, we might want to use assertr's verify function to make sure that mpg is a positive number:

library(assertr)

our.data %>%
  verify(mpg >= 0) %>%
  group_by(cyl) %>%
  summarise(avg.mpg=mean(mpg))
#> Error in verify(., mpg >= 0): verification failed! (1 failure)

If we had done this, we would have caught this data error.

The verify function takes a data frame (its first argument is provided by the %>% operator), and a logical (boolean) expression. Then, verify evaluates that expression using the scope of the provided data frame. If any of the logical values of the expression's result are FALSE, verify will raise an error that terminates any further processing of the pipeline.

We could have also written this assertion using assertr's assert function...

our.data %>%
  assert(within_bounds(0,Inf), mpg) %>%
  group_by(cyl) %>%
  summarise(avg.mpg=mean(mpg))
#> Error: 
#> Vector 'mpg' violates assertion 'within_bounds' 1 time (value [-18.7] at index 5)

The assert function takes a data frame, a predicate function, and an arbitrary number of columns to apply the predicate function to. The predicate function (a function that returns a logical/boolean value) is then applied to every element of the columns selected, and will raise an error when if it finds violations.

Internally, the assert function uses dplyr's select function to extract the columns to test the predicate function on. This allows for complex assertions. Let's say we wanted to make sure that all values in the dataset are greater than zero (except mpg):

our.data %>%
  assert(within_bounds(0,Inf, include.lower=FALSE), -mpg) %>%
  group_by(cyl) %>%
  summarise(avg.mpg=mean(mpg))
#> Error: 
#> Vector 'vs' violates assertion 'within_bounds' 18 times (e.g. [0] at index 1)
#> Vector 'am' violates assertion 'within_bounds' 19 times (e.g. [0] at index 4)

verify vs. assert

The first noticable difference between verify and assert is that verify takes an expression, and assert takes a predicate and columns to apply it to. This might make the verify function look more elegant--but there's an important drawback. verify has to evaluate the entire expression first, and then check if there were any violations. Because of this, verify can't tell you the offending datum.

One important drawback to assert, and a consequence of its application of the predicate to columns, is that assert can't confirm assertions about the data structure itself. For example, let's say we were reading a dataset from disk that we know has more than 100 observations; we could write a check of that assumption like this:

dat <- read.csv("a-data-file.csv")
dat %>%
  verify(nrow(.) > 100) %>%
  ....

This is a powerful advantage over assert... but assert has one more advantage of its own that we heretofore ignored.

assertr's predicates

assertr's predicates, both built-in and custom, make assert very powerful. The three predicates that are built in to assertr are

  • not_na - that checks if an element is not NA
  • within_bounds - that returns a predicate function that checks if a numeric value falls within the bounds supplied, and
  • in_set - that returns a predicate function that checks if an element is a member of the set supplied.

We've already seen within_bounds in action... let's use the in_set function to make sure that there are only 0s and 1s (automatic and manual, respectively) values in the am column...

our.data %>%
  assert(in_set(0,1), am) %>%
  ...

If we were reading a dataset that contained a column representing boroughs of New York City (named BORO), we can verify that there are no mis-spelled or otherwise unexpected boroughs like so...

boroughs <- c("Bronx", "Manhattan", "Queens", "Brooklyn", "Staten Island")

read.csv("a-dataset.csv") %>%
  assert(in_set(boroughs), BORO) %>%
  ...

Rad!

custom predicates

A convenient feature of assertr is that it makes the construction of custom predicate functions easy.

In order to make a custom predicate, you only have to specify cases where the predicate should return FALSE. Let's say that a dataset has an ID column (named ID) that we want to check is not an empty string. We can create a predicate like this:

not.empty.p <- function(x) if(x=="") return(FALSE)

and apply it like this:

read.csv("another-dataset.csv") %>%
  assert(not.empty.p, ID) %>%
  ...

Let's say that the ID column is always a 7-digit number. We can confirm that all the IDs are 7-digits by defining the following predicate:

seven.digit.p <- function(x) nchar(x)==7

A powerful consequence of this easy creation of predicates is that the assert function lends itself to use with lambda predicates (unnamed predicates that are only used once). The check above might be better written as

read.csv("another-dataset.csv") %>%
  assert(function(x) nchar(x)==7, ID) %>%
  ...

Neat-o!

enter insist and predicate 'generators'

Very often, there is a need to dynamically determine the predicate function to be used based on the vector being checked.

For example, to check to see if every element of a vector is within n standard deviations of the mean, you need to create a within_bounds predicate after dynamically determining the bounds by reading and computing on the vector itself.

To this end, the assert function is no good; it just applies a raw predicate to a vector. We need a function like assert that will apply predicate generators to vectors, return predicates, and then perform assert-like functionality by checking each element of the vectors with its respective custom predicate. This is precisely what insist does.

This is all much simpler than it may sound. Hopefully, the examples will clear up any confusion.

The primary use case for insist is in conjunction with the within_n_sds or within_n_mads predicate generator.

Suppose we wanted to check that every mpg value in the mtcars data set was within 3 standard deviations of the mean before finding the average miles per gallon for each number of engine cylinders. We could write something like this:

mtcars %>%
  insist(within_n_sds(3), mpg) %>%
  group_by(cyl) %>%
  summarise(avg.mpg=mean(mpg))
#> # A tibble: 3 × 2
#>     cyl  avg.mpg
#>   <dbl>    <dbl>
#> 1     4 26.66364
#> 2     6 19.74286
#> 3     8 15.10000

Notice what happens when we drop that z-score to 2 stardard deviations from the mean

mtcars %>%
  insist(within_n_sds(2), mpg) %>%
  group_by(cyl) %>%
  summarise(avg.mpg=mean(mpg))
#> Error: 
#> Vector 'mpg' violates assertion 'within_n_sds' 2 times (e.g. [32.4] at index 18)

Execution of the pipeline was halted. But now we know exactly which data point (and column) violated the predicate that within_n_sds(3)(mtcars$mpg) returned.

Now that's an efficient car!

After the predicate generator, insist takes an arbitrary number of columns just like assert using the syntax of dplyr's select function. If you wanted to check that everything in mtcars is within 10 standard deviations of the mean (of each column vector), you can do so like this:

mtcars %>%
  insist(within_n_sds(10), mpg:carb) %>%
  group_by(cyl) %>%
  summarise(avg.mpg=mean(mpg))
#> # A tibble: 3 × 2
#>     cyl  avg.mpg
#>   <dbl>    <dbl>
#> 1     4 26.66364
#> 2     6 19.74286
#> 3     8 15.10000

Aces!

I chose to use within_n_sds in this example because people are familiar z-scores. However, for most practical purposes, the related predicate generator within_n_mads is more useful.

The problem with within_n_sds is the mean and standard deviation are so heavily influenced by outliers, their very presence will compromise attempts to identify them using these statistics. In contrast with within_n_sds, within_n_mads uses the robust statistics, median and median absolute deviation, to identify potentially erroneous data points.

For example, the vector <7.4, 7.1, 7.2, 72.1> almost certainly has an erroneous data point, but within_n_sds(2) will fail to detect it.

example.vector <- c(7.4, 7.1, 7.2, 72.1)
within_n_sds(2)(example.vector)(example.vector)
#> [1] TRUE TRUE TRUE TRUE

whereas within_n_mads will detect it at even lower levels of power....

example.vector <- c(7.4, 7.1, 7.2, 72.1)
within_n_mads(2)(example.vector)(example.vector)
#> [1]  TRUE  TRUE  TRUE FALSE
within_n_mads(1)(example.vector)(example.vector)
#> [1]  TRUE  TRUE  TRUE FALSE

Tubular!

row-wise assertions and row reduction functions

As cool as it's been so far, this still isn't enough to consitute a complete grammar of data integrity checking. To see why, check out the following small example data set:

example.data <- data.frame(x=c(8, 9, 6, 5, 9, 5, 6, 7,
                             8, 9, 6, 5, 5, 6, 7),
                         y=c(82, 91, 61, 49, 40, 49, 57,
                             74, 78, 90, 61, 49, 51, 62, 68))
(example.data)
#>    x  y
#> 1  8 82
#> 2  9 91
#> 3  6 61
#> 4  5 49
#> 5  9 40
#> 6  5 49
#> 7  6 57
#> 8  7 74
#> 9  8 78
#> 10 9 90
#> 11 6 61
#> 12 5 49
#> 13 5 51
#> 14 6 62
#> 15 7 68

Can you spot the brazen outlier? You're certainly not going to find it by checking the distribution of each column! All elements from both columns are within 2 standard deviations of their respective means.

Unless you have a really good eye, the only way you're going to catch this mistake is by plotting the data set.

plot(example.data$x, example.data$y, xlab="", ylab="")

plot of chunk unnamed-chunk-24

Ok, so all the ys are roughly 10 times the xs except the outlying data point.

The problem having to plot data sets to catch anomalies is that it is really hard to visualize 4-dimensions at once, and it is near impossible with high-dimensional data.

There's no way of catching this anomaly by looking at each individual column separately; the only way to catch it is to view each row as a complete observation and compare it to the rest.

To this end, assertr provides two functions that take a data frame, and reduce each row into a single value. We'll call them row reduction functions.

The first one we'll look at is called maha_dist. It computes the average mahalanobis distance (kind of like multivariate z-scoring for outlier detection) of each row from the whole data set. The big idea is that in the resultant vector, big/distant values are potential anomalous entries. Let's look at the distribution of mahalanobis distances for this data set...

maha_dist(example.data)
#>  [1]  1.28106379  3.10992407  0.25081851  1.35993969 12.81898913
#>  [6]  1.35993969  0.26181283  0.47714597  0.87804987  2.95741956
#> [11]  0.25081851  1.35993969  1.29208587  0.28235776  0.05969507

maha_dist(example.data) %>% hist(main="", xlab="")

plot of chunk unnamed-chunk-25

There's no question here as to whether there's an anomalous entry! But how do you check for this sort of thing using assertr constructs?

Well, maha_dist will typically be used with the insist_rows function. insist_rows takes a data frame, a row reduction function, a predicate-generating function, and an arbitrary number of columns to apply the predicate function to. The row reduction function (maha_dist in this case) is applied to the data frame, and returns a value for each row. The predicate-generating function is then applied to the vector returned from the row reduction function and the resultant predicate is applied to each element of that vector. It will raise an error if it finds any violations.

As always, this undoubtedly sounds far more confusing than it really is. Here's an example of it in use

example.data %>%
  insist_rows(maha_dist, within_n_mads(3), everything())
#> Error: Data frame row reduction violates predicate 'within_n_mads' 1 time (at row number 5)

Check that out! To be clear, this function is running the supplied data frame through the maha_dist function which returns a value for each row corresponding to its mahalanobis distance. (The whole data frame is used because we used the everything() selection function.) Then, within_n_mads(3) computes on that vector and returns a bounds checking predicate. The bounds checking predicate checks to see that all mahalanobis distances are within 3 median absolute deviations of each other. They are not, and the pipeline errors out.

This is probably the most powerful construct in assertr--it can find a whole lot of nasty errors that would be very difficult to check for by hand.

Part of what makes it so powerful is how flexible maha_dist is. We only used it, so far, on a data frame of numerics, but it can handle all sorts of data frames. To really see it shine, let's use it on the iris data set, that contains a categorical variable in its right-most column...

head(iris)
#>   Sepal.Length Sepal.Width Petal.Length Petal.Width Species
#> 1          5.1         3.5          1.4         0.2  setosa
#> 2          4.9         3.0          1.4         0.2  setosa
#> 3          4.7         3.2          1.3         0.2  setosa
#> 4          4.6         3.1          1.5         0.2  setosa
#> 5          5.0         3.6          1.4         0.2  setosa
#> 6          5.4         3.9          1.7         0.4  setosa

iris %>% maha_dist %>% hist(main="", xlab="")

plot of chunk unnamed-chunk-27

Looks ok, but what happens when we accidently enter a row as a different species...

mistake <- iris
(mistake[149,5])
#> [1] virginica
#> Levels: setosa versicolor virginica
mistake[149,5] <- "setosa"

mistake %>% maha_dist %>% hist(main="", xlab="")

plot of chunk unnamed-chunk-28

mistake %>% maha_dist %>% which.max
#> [1] 149

Look at that! This mistake can easily be picked up by any reasonable bounds checker...

mistake %>% insist_rows(maha_dist, within_n_mads(7), everything())
#> Error: Data frame row reduction violates predicate 'within_n_mads' 1 time (at row number 149)

insist and insist_rows are both similar in that they both take predicate generators and not actual predicates. What makes insist_rows different is its usage of a row-reduce data frame.

assert has a row-oriented counterpart, too; it's called assert_rows. insist is to assert as insist_rows is to assert_rows.

assert_rows works the same as insist_rows, except that instead of using a predicate generator on the row-reduced data frame, it uses a regular-old predicate.

For an example of a assert_rows use case, let's say that we got a data set (another-dataset.csv) from the web and we don't want to continue processing the data set if any row contains more than two missing values (NAs). You can use the row reduction function num_row_NAs to reduce all the rows into the number of NAs they contain. Then, a simple bounds checker will suffice for ensuring that no element is higher than 2...

read.csv("another-dataset.csv") %>%
  assert_rows(num_row_NAs, within_bounds(0,2), everything()) %>%
  ...

assert_rows can be used for anomaly detection as well. A future version of assertr may contain a cosine distance row reduction function. Since all cosine distances are contrained from -1 to 1, it is easy to use a non-dynamic predicate to disallow certain values.

combining chains of assertions

Let's say that as part of an automated pipeline that grabs mtcars from an untrusted source and finds the average miles per gallon for each number of engine cylinders, we want to perform the following checks...

  • that the dataset contains more than 10 observations
  • that the column for 'miles per gallon' (mpg) is a positive number
  • that the column for 'miles per gallon' (mpg) does not contain a datum that is outside 4 standard deviations from its mean, and
  • that the am and vs columns (automatic/manual and v/straight engine, respectively) contain 0s and 1s only

This could be written thusly:

mtcars %>%
  verify(nrow(mtcars) > 10) %>%
  verify(mpg > 0) %>%
  insist(within_n_sds(4), mpg) %>%
  assert(in_set(0,1), am, vs) %>%
  group_by(cyl) %>%
  summarise(avg.mpg=mean(mpg))
#> # A tibble: 3 × 2
#>     cyl  avg.mpg
#>   <dbl>    <dbl>
#> 1     4 26.66364
#> 2     6 19.74286
#> 3     8 15.10000

Ew, there are four lines of assertions before the real fun starts. We can make look much better by abstracting out all the assertions:

check_me <- . %>%
  verify(nrow(mtcars) > 10) %>%
  verify(mpg > 0) %>%
  insist(within_n_sds(4), mpg) %>%
  assert(in_set(0,1), am, vs)

mtcars %>%
  check_me %>%
  group_by(cyl) %>%
  summarise(avg.mpg=mean(mpg))
#> # A tibble: 3 × 2
#>     cyl  avg.mpg
#>   <dbl>    <dbl>
#> 1     4 26.66364
#> 2     6 19.74286
#> 3     8 15.10000

Awesome! Now we can add an arbitrary number of assertions, as the need arises, without touching the real logic.

Citing

Tony Fischetti (2016). assertr: Assertive Programming for R Analysis Pipelines. R package version 1.0.2. https://cran.rstudio.com/package=assertr

License and bugs

Back to top

comments powered by Disqus