snht: Standard Normal Homogeneity Test
In snht: Standard Normal Homogeneity Test

Description Usage Arguments Details Value Author(s) References See Also Examples

This function performs a standard normal homogeneity test on the data supplied. This test searches the data for potential changepoints.

1 2	snht(data, period, robust = F, time = NULL, scaled = TRUE, rmSeasonalPeriod = Inf, ...)

`data`	The data to be analyzed for changepoints.
`period`	The SNHT works by calculating the mean of the data on the previous period observations and the following period observations. Thus, this argument controls the window size for the test statistics.
`robust`	Flag indicating whether or not robust estimators should be used. If T, then Huber's robust estimator for the mean and variance will be used (see ?MASS::huber).
`time`	Numeric vector specifying times for the observations. If not supplied, it is assumed that each observation occurs on one time period. If supplied, then the algorithm will create a new dataset with the same number of observations for each time unit by adding missing values.
`scaled`	In the Haimberger paper, a typo is reported in the test statistic. The denominator ought to be s^2 instead of s, as the distribution of the test statistic then becomes chi-squared assuming errors are normal. If scaled = TRUE, the corrected version is used. In this case, the test statistics can be compared to a chi-squared. However, if scaled = FALSE, there is no clear distribution to compare with.
`rmSeasonalPeriod`	This algorithm will overestimate the standard error (and hence incorrectly estimate the test statistic) if there is strong seasonality in the data. By setting rmSeasonalPeriod to some value, a GAM model will be built to capture that seasonality. Once it is estimated, it is removed and the SNHT statistic is computed on the resulting dataset. Setting this argument to Inf prevents any modeling.
`...`	Other parameters, see ?robustSNHT, ?robustSNHTunequal.

The SNHT works by calculating the mean of the data on the previous period and on the following period. The test statistic at each observation is then computed as described in Haimberger (2007). Essentially, though, it just compares the means of these two periods and normalizes by the standard deviation.

Note: if there are not enough observations both before and after the current observation, no test is performed.

Large values of the test statistic suggests the presence of a changepoint. Haimberger (see references) suggests values larger than 100 should be considered changepoints. However, this does not apply if scaled = TRUE.

Observations which are less than period away from the start or end of the dataset do not have valid SNHT statistics. Thus, the statistic for these observations is returned as NA.

Returns a data.frame, with columns score, leftMean, and rightMean. Statistic is the SNHT test statistic described above, and leftMean (rightMean) are the means to the left (right) of the current observation.

Additionally, if time is supplied, then time is returned on the output data.frame. Note that new (missing) observations were introduced to the dataset to ensure the same number of observations occur per day.

Josh Browning (jbrownin@mines.edu), Carina Schneider (carina.schneider@uzh.ch)

L. Haimberger. Homogenization of radiosonde temperature time series using innovation statistics. Journal of Climate, 20(7): 1377-1403, 2007.

huber

Other snht functions: robustSNHTunequal, robustSNHT

data = rnorm(1000)
brk = sample(1000, size=1)
data[1:brk] = data[1:brk]-2
out = snht( data, period=50, robust=FALSE )
summary(out)

data = rnorm(1000)
time = 1:1000 + rnorm(1000)
brk = sample(1000, size=1)
data[1:brk] = data[1:brk]-2
out = snht( data, period=50, time=time, robust=FALSE )
summary(out)