Error properties of estimates derived from imputation differ from those of regression-based
estimates because the two methods include a different mix of error components. This function
computes a partitioning of error statistics as proposed by Stage and Crookston (2007).
An object of class yai computed with method="mahalanobis".
...
Other objects of class yai for which statistics are desired. All
objects should be for the same data and variables used for the first argument.
scale
When TRUE, the errors are scaled by their respective standard deviations.
pzero
The lower tail p-value used to pick reference observations that are zero
distance from each other (used to compute rmmsd0).
plg
The upper tail p-value used to pick reference observations that are
substantially distant from each other (used to compute rmsdlg).
seeMethod
Method used to compute SEE: seeMethod="lm" uses lm
and seeMethod="gam" uses gam. In both cases, the model formula is
a simple linear combination of the X-variables.
A list that contains several data frames. The column names of each are a combination
of the name of the object used to compute the statistics and the name of the statistic. The
rownames correspond the the Y-variables from the first argument. The data frame names are as follows:
common
statistics used to compute other statistics.
name of first argument
error statistics for the first yai object.
names of ... arguments
error statistics for each of the remaining yai objects,
if any.
see
standard error of estimate for individual regressions fit for
corresponding Y-variables.
rmmsd0
root mean square difference for imputations based on method="mahalanobis"
(always based on the first argument to the function).
mlf
square root of the model lack of fit: sqrt(see^2 - (rmmsd0^2/2)).
rmsd
root mean square error.
rmsdlg
root mean square error of the observations with larger distances.
sei
standard error of imputation sqrt(rmsd^2 - (rmmsd0^2/2)).
dstc
distance component: sqrt(rmsd^2 - rmmsd0^2).
Note that unlike Stage and Crookston (2007), all statistics reported here are in the natural
units, not squared units.
Stage, A.R.; Crookston, N.L. (2007). Partitioning error components
for accuracy-assessment of near neighbor methods of imputation.
For. Sci. 53(1):62-72.
http://www.treesearch.fs.fed.us/pubs/28385
See Also
yai, TallyLake
Examples
require (yaImpute)
data(TallyLake)
diag(cov(TallyLake[,1:8])) # see col A in Table 3 in Stage and Crookston
mal=yai(x=TallyLake[,9:29],y=TallyLake[,1:8],
noTrgs=TRUE,method="mahalanobis")
msn=yai(x=TallyLake[,9:29],y=TallyLake[,1:8],
noTrgs=TRUE,method="msn")
# variable "see" for "mal" matches col B (when squared and scaled)
# other columns don't match exactly as Stage and Crookston used different
# software to compute values
errorStats(mal,msn)