movAv {berryFunctions}

Moving average

Description

Weighted moving average (running mean) with overlapping windows

Usage

movAv(dat, width = 7, weights = rep(1, width), quiet = FALSE)

Arguments

dat

Vector with regularly spaced data

width

Odd integer specifying window width. DEFAULT: 7

weights

Vector with weights. Sum is normalized to 1. DEFAULT: rep(1,width)

quiet

Logical: suppress allNA message and even width warning? DEFAULT: FALSE

Details

Width has to be odd, so there is a defined middle point of each window. Even inputs will be changed with a warning (unless quiet=TRUE).
Weights doesn't have to be symmetrical, but is always mapped to the middle of each window!
If there are NAs in the window, the corresponding weight is distributed evenly to the other weights.

Value

Vector of the same length as the original input. Padded with NAs at width/2 margin elements

Note

You can specify just one of weights or width.

Author(s)

Berry Boessenkool, [e-mail], ca 2012

See Also

movAvLines, filter, decompose, smooth, loess, rollapply (no overlapping!)

Examples

# general usage -------------------------------------------------------------
set.seed(29); a <- runif(40, 5,50)
data.frame(a, movAv(a))

# final and commencing NAs are kept, middle ones are filled:
a[c(1:10, 18:26, 32:40)] <- NA
data.frame(a, movAv(a))

set.seed(29); a <- runif(60, 5,50)
plot(a, type="o", pch=16, las=1)
lines(movAv(a), col=2, lwd=3) # shows trends, signal in the noise
lines(movAv(a,3), col=4, lwd=3)
lines(movAv(a,15), col=3, lwd=3) # degree of smoothing depends on window width

# Weights:
plot(a, type="o", pch=16, las=1)
lines(movAv(a), col=2, lwd=3) # uniform weight within running window
# Triangular weights react stronger to extrema:
lines(movAv(a, weights=c(1,2,4,6,4,2,1)), col=4, lwd=3)

plot(c(Nile), type="l")
lines(movAv(Nile,20), col=4, lwd=4)
lines(movAv(Nile,21), col=3) # even widths are changed to a higher value


# smoothing intenstiy -------------------------------------------------------
plot(1871:1970, Nile, type="l", col=8)
movAvLines(1871:1970, Nile, lwd=3)

for(i in 1:30*2-1)
  {
  plot(a, type="o", pch=16, las=1, main=paste("moving average, width =", i))
  lines(movAv(a, i), col=2, lwd=4)
  }
# How to lie with moving averages: compare width 29 with 49 - the "trend"
# appears to be in opposite direction! (OK, this is random data anyways).

b <- rep(a, each=10)+runif(600, -10, 20)
plot(b, type="l")
lines(movAv(b), col=2, lwd=4)
lines(movAv(b, 35), col=4, lwd=4)
lines(movAv(b, 101), col=5, lwd=4)  # choose width according to scale!


# Deviance from running mean can identify outlier:
nile <- c(Nile)
op <- par(mfrow=c(3,1), mar=c(1,3,2.5,0), cex.main=1, las=1)
plot(nile, type="l", main=c("original Nile data",""), xlab="", xaxt="n")
lines(movAv(nile,5), lwd=2, col=2)
title(main=c("", "5-element running mean (moving average)"), col.main=2)
box("figure")
plot(nile-movAv(nile,5), type="o", pch=16, col=4,
      main="difference  ( original data - moving average )", xlab="", xaxt="n")
abline(h=0)
box("figure")
par(mar=c(3,3,1,0))
hist(nile-movAv(nile,5), breaks=25, xlim=c(-500,500), col=4, main="Deviances")
abline(v=0, lwd=5) # the deviances are pretty symmetric.
# If this were shifted more strongly to the left, we could say:
# movav(5) overestimates minima more than it underestimates maxima
# This would happen if low values peak away further and more shortly
par(op)