Roll Hampel

Apply a moving-window Hampel function to a numeric vector.

roll_hampel(
  x,
  width = 1L,
  by = 1L,
  align = c("center", "left", "right"),
  na.rm = FALSE,
  min_valid = NULL
)

Arguments

x

Numeric vector.

width

Integer width of the rolling window.

by

Integer shift by which the window is moved each iteration.

align

Character position of the return value within the window. One of: "left" | "center" | "right".

na.rm

Logical specifying whether NA values should be removed before the calculations within each window.

min_valid

Integer minimum number of non-NA values that must fall within a window for that window to return a non-NA result. Supplying min_valid implies NA-tolerant counting within each window (as if na.rm = TRUE), regardless of na.rm. The default, NULL, applies no minimum.

Value

Numeric vector of the same length as x.

Details

The Hampel filter is a robust outlier detector using Median Absolute Deviation (MAD).

For every index in the incoming vector x, a value is returned that is the Hampel function of all values in x that fall within a window of width width. The score at each index compares the value at that index (x[i]) with the median and MAD of its window: abs(x[i] - median) / (1.4826 * MAD). Larger scores indicate values less consistent with their neighborhood. With align = "center" the tested value sits at the middle of its window; with align = "left" or "right" it sits at the window edge. Outlier detection (see findOutliers()) is normally done with the default align = "center".

The align parameter determines the alignment of the return value within the window. Thus:

• align = "left" [*------] will cause the returned vector to have width - 1 NA values at the right end.

• align = "center" [---*---] will cause the returned vector to have NA values at either end as needed for centered alignment.

• align = "right" [------*] will cause the returned vector to have width - 1 NA values at the left end.

For large vectors, the by parameter can be used to force the window to jump ahead by indices for the next calculation. Indices that are skipped over will be assigned NA values so that the return vector still has the same length as the incoming vector. This can dramatically speed up calculations for high resolution time series data.

Examples

x <- c(0, 0, 0, 1, 1, 2, 2, 4, 6, 9, 0, 0, 0)
roll_hampel(x, 3)
#>  [1]        NA 0.0000000 0.0000000 0.0000000 0.0000000 0.0000000 0.0000000
#>  [8] 0.0000000 0.0000000 0.6744908 0.0000000 0.0000000        NA