Why do we care?

NAs are squirrely when it comes to computation.¹

• NA+1 = NA
• NA-1 = NA
• NA*1 = NA
• NA/1 = NA
• NA^1 = NA
• NA^0 = 1
• NA**2 = NA

More likely than not, leaving them will throw errors.

What can we do?

We have two alternatives: remove the rows with NAs or impute values. Because the NAs are only 1.9% of the data, I’m going to avoid all the skull sweat needed for good imputation (substituting some summary statistic for the NA). An additional consideration is that the variables with the NA are a bit dodgy to begin with for reasons not germane to the subject. They also turned out not to contribute to the modeling.

How do we remove the NAs?

The dplyr filter function takes a logical argument and !isNA() provides one.

mydata_no_na <- mydata %>% filter(!is.na(VAR32))

A little sleight of hand was involved. I guessed that all of the variables with NA would be missing in each observation, so that by filtering on one, I’d hit them all. Sure enough, anyNA() confirmed my guess. If it hadn’t, of course, I would have needed to either run multiple filters or expand to

mydata_no_na <- mydata %>% filter(!is.na(VAR32) | !is.na(VAR33) ...)

Dependent variables of interest are all zero valued for some observations

These data were assembled for a purpose – to determine the influence of a group of 16 variables (different types of medical trauma) on 2 variables considered separately (different types of internal injuries to the same organ). Observations in which none of the 16 variables were coded as present (1) shed no light on the problem.

How do we identify them?

Recall that each of the 16 variables is either valued as 1 or 0. If the sum of the variables for an observation is zero, we don’t need the observation for the main purposes of the model project. There are several possible ways to sum variables rowwise. I chose a little linear algebra jujitsu.

Linear algebra!!! Relax, I don’t know much about this subject either. But I know a piece that’s simple and well suited to this problem.

Starting with mydata_no_na, the first step is to create an object for the 16 variables

    interest_vars <- c("VAR14", "VAR15", "VAR16", "VAR17", "VAR18", "VAR19", "VAR20", "VAR21", "VAR22", "VAR23", "VAR24", "VAR25", "VAR26", "VAR27", "VAR28", "VAR29")

and to use it to subset the NAs removed data

    mydata_vars <- mydata_no_na %>% dplyr::select(interest_vars)

Here comes the trick: It’s easy to convert a tibble into a matrix.

    m <- as.matrix(mydata_vars)

Why do we want a matrix? A matrix is an object that you can perform row operations on easily.

    rowtots <- rowSums(m)

We now have the sum of the 16 variables for each observation.

The return trip is equally easy.

    vcode_count <- enframe(rowtots) %>% dplyr::select(value) %>% dplyr::rename(V_COUNT = value)
    mydata_w_row_count <- as_tibble(cbind(mydata_no_na, vcode_count))
    mydata_singlets <- mydata_with_row_count %>% filter(V_COUNT == 1)

In short, we’ve added a row total field, V_COUNT and filtered for sums of 1.

The results are dramatic. We started with 22,471 observations, subtracted 419 NA, and then isolated the observations with one (only one) of the 16 variables and ended up with 9,114 observations.