pcafill: PCA-based missing-value filling

In metno/esd: Climate analysis and empirical-statistical downscaling (ESD) package for monthly and daily data

PCA-based missing-value filling

Description

Fills missing (station) values by predicting their values using multiple regression. The regression uses as input pcincipal components from PCA from the same (group of station) data, but where series with missing data have been excluded. This makes sense for (station) data where most of the variability is accounted for by a few leading modes. This method is not expected to be useful when there are many large data gaps.

Usage

pcafill(
  X,
  insertmiss = 0,
  ip = 1:4,
  mnv = 0,
  complete = FALSE,
  test = FALSE,
  verbose = FALSE
)

Arguments

X	station data (group of stations)
insertmiss	Used for testing and evaluating. Missing data are introduced to test the predictive capability
ip	Number of EOFs/PCAs to include in filling in. In many cases, it may be useful to keep this to a small set of values.
mnv	Minimum number of valid data points for any given time. Can be used to get around the problem with too many missing data
complete	Use pattern projection between PCA pattern and original data to get a complete record - otherwise a subset of times with sufficient data.
test	Extra test - debugging
verbose	Print diagnostics - debugging
N	Number of runs in Monte-Carlo simulation
max.miss	Maximum NAs to insert (insertmiss) in Monte-Carlo simulations
x	time series for calibrating regression analysis
y	PC input for regression analysis

Details

This function is handy for the downscaling of PCAs. See Benestad, R.E., D. Chen, A. Mezghani, L. Fan, K. Parding, On using principal components to represent stations in empirical-statistical downscaling, Tellus A 28326, accepted.

Value

The same as the input - station object with filled-in values

See Also

PCA, allgood

Examples

data('Tx.Norway')
X <- annual(Tx.Norway,FUN='mean',nmin=200)
ok<- apply(X,1,nv)
X <- subset(X,it=ok > 0)
Y <- pcafill(X)

plot(PCA(Y))
plot(c(coredata(Y)),c(coredata(X)))

## Monte-Carlo test with random selection of data points set to NA:
Y.test <- pcafill.test(X,max.miss=10,ip=1:3)
cor(Y.test)

metno/esd documentation built on March 9, 2024, 11:21 a.m.