Flexible Options for Handling Missing Values.

naflex: Flexible options for handling missing values

Overview

The naflex R package provides additional flexibility for handling missing values in summary functions beyond the existing options (na.rm = TRUE/FALSE) available in base R.

Most summary functions in base R e.g. mean, provide the two extreme options for handling missing values:

1. calculate the summary ignoring all missing values (na.rm = TRUE), or
2. require no missing values for the summary to be calculated (na.rm = FALSE)

In many cases, something in between these two extremes is often more appropriate. For example, you may wish to give a summary statistic if less than 5% of values are missing.

naflex provides helper functions to facilitate this flexibility. It allows for omitting missing values conditionally, using four types of checks:

1. a maximum proportion of missing values allowed
2. a maximum number of missing values allowed
3. a maximum number of consecutive missing values allowed, and
4. a minimum number of non-missing values required.

Motivation

The motivating application for producing this package was the calculation of Climate Normals: Long term averages of surface meteorological measurements e.g. total rainfall and mean temperature that provide benchmark information about the climate at specific locations. The World Meteorological Organization (WMO) Guidelines on the Calculation of Climate Normals¹ provides recommendations to standardise these calculations across countries, including handling of missing values.

For example, it recommendations that a monthly mean value calculated from daily values should only be calculated when there are no more than 10 missing values in the month and no more than 4 days of consecutive missing values. Adhering to such rules using base R requires doing further calculations and increasing the complexity and length of code. The aim of naflex is to make it easier to apply such rules routinely and efficiently as part of calculations.

Installation

Install the current release from CRAN:

install.packages("naflex")

Or install the latest development version from GitHub:

# install.packages("devtools")
devtools::install_github("dannyparsons/naflex")

Usage

The main function in naflex is na_omit_if.

When wrapped around a vector in a summary function, na_omit_if ensures that the summary value is calculated when the checks pass, and returns NA if not. The example below shows how to calculate the mean, conditionally on the proportion of missing values.

library(naflex)

x <- c(1, 3, NA, NA, 3, 2, NA, 5, 8, 7)

# Calculate if 30% or less missing values
mean(na_omit_if(x, prop = 0.3))
#> [1] 4.142857
# Calculate if 20% or less missing values
mean(na_omit_if(x, prop = 0.2))
#> [1] NA

Four types of checks are available:

1. prop: the maximum proportion (0 to 1) of missing values allowed
2. n: the maximum number of missing values allowed
3. consec: the maximum number of consecutive missing values allowed, and
4. n_non: the minimum number of non-missing values required.

If multiple checks are specified, all checks must pass for missing values to be removed. For example, although there are less than 4 missing values in x, there are two consecutive missing values, hence the consec = 1 check fails in the example below the result is NA.

# Calculate if 4 or less missing values and 1 or less consecutive missing values 
mean(na_omit_if(x, n = 4, consec = 1))
#> [1] NA

The use of %>% (“pipe”) from magrittr can be used to make the code look clearer and more familiar. The beginning of the line is now the same as standard R and it moves na_omit_if after x which then appears more like an option within the function, like na.rm, which is how you might think about na_omit_if conceptually in this case.

require(magrittr)
#> Loading required package: magrittr

sum(x %>% na_omit_if(prop = 0.25))
#> [1] NA

Note that you should not use na_omit_if with na.rm = TRUE in the summary function since this will always remove missing values so the checks are essentially ignored.

How naflex works & more details

na_omit_if works by removing the missing values from x if the checks pass, and leaves x unmodified otherwise.

# Missing values removed
na_omit_if(x, n = 4)
#> [1] 1 3 3 2 5 8 7
#> attr(,"na.action")
#> [1] 3 4 7
#> attr(,"class")
#> [1] "omit"

na_omit_if can be thought of like an extension of stats::na.omit and if missing values are removed, an na.action attribute and omit class are added for consistency with stats::na.omit.

# Missing values not removed, x is unmodified
na_omit_if(x, n = 2)
#>  [1]  1  3 NA NA  3  2 NA  5  8  7

A further set of four na_omit_if_* functions are provided for doing the same thing but restricted to a single check e.g. na_omit_if_n(x, 2).

na_check has the same parameters as na_omit_if but returns a logical indicating whether the checks pass. It is used internally in na_omit_if and may also be a useful helper function.

if (na_check(x, n = 4, consec = 1)) "NA checks pass" else "NA checks fail"
#> [1] "NA checks fail"

A set a four na_check_* functions are also provided for doing the same thing restricted to a single check e.g. na_check_prop(x, 0.2)

Finally, naflex provides a set of helper functions for calculating missing value properties used in these checks.

na_prop(x)
#> [1] 0.3
na_n(x)
#> [1] 3
na_consec(x)
#> [1] 2
na_non_na(x)
#> [1] 7

Compared to base R

In base R, this functionality can often be achieved using a combination of ifelse, is.na, rle and the option na.rm = TRUE.naflex aims to simplify, shorten and standardise this process for users.

For example, the equivalent of:

mean(na_omit_if(x, n = 4, prop = 0.2))
#> [1] NA

in base R is:

ifelse(sum(is.na(x)) <= 4 && mean(is.na(x)) <= 0.2, mean(x, na.rm = TRUE), NA)
#> [1] NA

The check for longest sequence of consecutive missing values is more complex and requires clever use of the rle function. For example,

mean(na_omit_if(x, consec = 5))
#> [1] 4.142857

is equivalent to:

r <- rle(is.na(x))
m <- r$lengths[r$values]
ifelse(max(m) <= 5, mean(x, na.rm = TRUE), NA)
#> [1] 4.142857

References

¹WMO Guidelines on the Calculation of Climate Normals↩