Refuse a categorical column where a numeric 0/1 treatment is required

Checks that a column meant as a binary treatment or outcome indicator is numeric 0/1 and not a factor or character column. Coercing a factor with as.numeric() returns the level indices (1, 2, ...) rather than the data, and a mis-ordered level silently flips every observation; this guard stops before that happens. Call it at the top of any function that needs a 0/1 column.

Usage

guard_binary(x, col)

Arguments

x

The column to check.

col

Character string: the column's name, used only in the error message.

Value

TRUE, invisibly, when the column is a 0/1 indicator; otherwise an error.

Details

The checks are: x is an atomic vector; it is not a factor or character vector; its distinct non-missing values are all 0 or 1. Logical vectors pass, because TRUE and FALSE compare equal to 1 and 0. Missing values are ignored. Each failure is an error that names the column.

See also

guard_recode to build an explicit indicator.

Examples

guard_binary(c(0, 1, 1, 0, NA), "treated")
guard_binary(c(TRUE, FALSE), "arrested")

# a factor is refused rather than coerced
try(guard_binary(factor(c("control", "treated")), "treated"))
#> Error : Column ‘treated’ is categorical (‘control’, ‘treated’...). Refusing to coerce: as.numeric() on a factor returns level INDICES (1, 2, ...), not your data, and a mis-ordered level silently relabels every observation. Encode explicitly first, e.g. guard_recode() + as.integer(x == "treated_label").

# so is a numeric column with other values
try(guard_binary(c(0, 1, 2), "treated"))
#> Error : Column ‘treated’ must be binary 0/1 (saw: 0, 1, 2).