reversed. You can see that a negative relationship exists between length of gestation and dream sleep and that dream sleep tends to be higher for mammals that are missing a gestation score. The number of observations missing values on both variables at the same time is printed in blue at the intersection of both margins (bottom left).

The VIM package has many graphs that can help you understand the role of missing data in a dataset and is well worth exploring. There are functions to produce scatter plots, box plots, histograms, scatter plot matrices, parallel plots, rug plots, and bubble plots that incorporate information about missing values.

18.3.3Using correlations to explore missing values

Before moving on, there’s one more approach worth noting. You can replace the data in a dataset with indicator variables, coded 1 for missing and 0 for present. The resulting matrix is sometimes called a shadow matrix. Correlating these indicator variables with each other and with the original (observed) variables can help you to see which variables tend to be missing together, as well as relationships between a variable’s “missingness” and the values of the other variables.

Consider the following code:

x <- as.data.frame(abs(is.na(sleep)))

The elements of data frame x are 1 if the corresponding element of sleep is missing and 0 otherwise. You can see this by viewing the first few rows of each:

>head(x, n=5)

BodyWgt BrainWgt NonD Dream Sleep Span Gest Pred Exp Danger

The statement

y <- x[which(apply(x,2,sum)>0)]

extracts the variables that have some (but not all) missing values, and

cor(y)

gives you the correlations among these indicator variables:

Here, you can see that Dream and NonD tend to be missing together (r = 0.91). To a lesser extent, Sleep and NonD tend to be missing together (r = 0.49) and Sleep and Dream tend to be missing together (r = 0.20).

Finally, you can look at the relationship between missing values in a variable and the observed values on other variables:

In this correlation matrix, the rows are observed variables, and the columns are indicator variables representing missingness. You can ignore the warning message and NA values in the correlation matrix; they’re artifacts of our approach.