![Page 1: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/1.jpg)
Robust Imputation of Missing Values in CompositionalData Using the R-Package robCompositions
Matthias Templ1,2, Peter Filzmoser1, Karel Hron3
1 Department of Statistics and Probability Theory, TU WIEN, Austria2 Department of Methodology, Statistics Austria
3 Department of Mathematical Analysis and Applications of Mathematics, Palacký
University, Olomouc, Czech Republic
Brussels, Feb. 19, 2009
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 1 / 28
![Page 2: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/2.jpg)
Content
1 Compositional Data
2 Transformation
3 Imputation Methods
4 Simulation Results
5 R-package robCompositions
6 Conclusion
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 2 / 28
![Page 3: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/3.jpg)
Compositional Data
Compositional (Closed) Data
Multivariate data that sum up to a constant (e.g. 100%):
x = (x1, . . . , xD)t , xi > 0,
D∑i=1
xi = κ
(the constant κ could be di�erent for each observation as well)
The set of all closed observations with positive values forms a simplex
sample space.
the ratios between the parts are of interest.
Key reference:
J. Aitchison. The Statistical Analysis of Compositional Data. Chapman
and Hall, London, U.K., 1986.
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 3 / 28
![Page 4: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/4.jpg)
Compositional Data
Compositional (Closed) Data
Multivariate data that sum up to a constant (e.g. 100%):
x = (x1, . . . , xD)t , xi > 0,
D∑i=1
xi = κ
(the constant κ could be di�erent for each observation as well)
The set of all closed observations with positive values forms a simplex
sample space.
the ratios between the parts are of interest.
Key reference:
J. Aitchison. The Statistical Analysis of Compositional Data. Chapman
and Hall, London, U.K., 1986.
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 3 / 28
![Page 5: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/5.jpg)
Compositional Data
Compositional Data: Example
0.50 0.55 0.60 0.65 0.70
0.10
0.15
0.20
0.25
0.30
high−quality production
low
−qu
ality
pro
duct
ion
1.0 1.5 2.0 2.5
−0.
50.
00.
51.
0
log(hq production/machine repair)
log(
low
−qu
ality
pro
duct
ion/
mac
hine
rep
air)
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 4 / 28
![Page 6: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/6.jpg)
Compositional Data
Compositional Data: Example
qu1 qu2 qu3 sum
1 52 42 6 100%2 52 44 4 100%3 47 48 5 100%
......
......
...
22 14 47 39 100%23 24 56 20 100%
question1 question2
question3
0.8
0.6
0.4
0.2
0.8
0.6
0.4
0.2
0.8
0.6
0.4
0.2
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 5 / 28
![Page 7: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/7.jpg)
Compositional Data
Compositional Data: Example
qu1 qu2 qu3 sum
1 52 42 6 100%2 52 44 4 100%3 47 48 5 100%
......
......
...
22 14 47 39 100%23 24 56 20 100%
question1 question2
question3
0.8
0.6
0.4
0.2
0.8
0.6
0.4
0.2
0.8
0.6
0.4
0.2
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 5 / 28
![Page 8: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/8.jpg)
Compositional Data
Compositional Data: Expentidures
housing foodstu� alcohol tobacco other goods
1 640 328 147 169 1962 1800 484 515 2291 9123 2085 445 725 8373 17324 616 331 126 117 1495 875 368 191 290 2756 770 364 196 242 236...
......
......
...
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 6 / 28
![Page 9: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/9.jpg)
Compositional Data
Compositional Data: Example
1
1
00
First compositional part
Sec
ond
com
posi
tiona
l par
t
1
1
00
0.1
0.2
0.5
0.6
First compositional partS
econ
d co
mpo
sitio
nal p
art
Left plot: Two-part compositional data without the constraint of constant sum. Thepoints could be varied along the lines from the origin without changing the ratio of the
compositional parts.Right plot: The points at the boundary are more distant than the central points. The
Aitchison distance accounts for this fact.
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 7 / 28
![Page 10: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/10.jpg)
Compositional Data
Aitchison Distance and the Simplex
A distance measure that is accounting for this relative scale property is the
Aitchison distance (Aitchison, 1992, Aitchison et al., 2000), de�ned for two
compositions x = (x1, . . . , xD)t and y = (y1, . . . , yD)t as
d2A(x , y) =
1
D
D−1∑i=1
D∑j=i+1
(ln
xi
xj− ln
yi
yj
)2
.
As an example, the boundary points in the previous Figure (right) have an
Aitchison distance of 0.33, whereas the central points have Aitchison
distance 0.08.
Replacing the Euclidean distance by the Aitchison distance is
necessary because the simplex sample space has a di�erent geometrical
structure than the classical Euclidean space.
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 8 / 28
![Page 11: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/11.jpg)
Transformation
Logratio Transformations
Family of one-to-one transformations from the simplex to the real space
(Aitchison, 1986):
additive logratio (alr) transformation
centred logratio (clr) transformation
isometric logratio (ilr) transformation
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 9 / 28
![Page 12: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/12.jpg)
Transformation
alr Transformation
Divide all values by the j-th part:
x (j) =(x
(j)1 , . . . , x
(j)D−1
)t=
(log
x1
xj, . . . , log
xj−1xj
, logxj+1
xj, . . . , log
xD
xj
)t
The index j ∈ {1, . . . ,D} refers to the �ratioing� variable.
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 10 / 28
![Page 13: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/13.jpg)
Transformation
clr Transformation
Divide all values by the geometric mean:
y = (y1, . . . , yD)t =
logx1
D
√∏Di=1 xi
, . . . , logxD
D
√∏Di=1 xi
t
Advantage: symmetric with respect to variables, easier interpretation
Disadvantage: singularity problem, because
logx1
D
√∏Di=1 xi
+ · · ·+ logxD
D
√∏Di=1 xi
=
D∑j=1
log(xj)−1
D
D∑j=1
D∑i=1
log(xi ) = 0
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 11 / 28
![Page 14: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/14.jpg)
Transformation
ilr Transformation
Take an orthonormal basis V = (v1, . . . , vD−1) (of dimension D × D − 1)
with
v i =
√i
i + 1
(1
i, . . . ,
1
i,−1, 0, . . . , 0
)for i = 1, . . . ,D − 1,
in the hyperplane H : y1 + · · ·+ yD = 0 in IRD .
The ilr-transformed data are
z = (z1, . . . , zD−1)t = V ty .
zi are coe�cients to the chosen basis.
Advantage: no singularity problem, good geometric properties
Disadvantage: zi is not easy to interprete.
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 12 / 28
![Page 15: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/15.jpg)
Transformation
Properties of the ILR Transformation
1
1
00
First compositional part
Sec
ond
com
posi
tiona
l par
t
−2.2 −1.2 −0.2 0.8 1.8 2.8
ilr transformed original data
−2.2 −1.2 −0.2 0.8 1.8 2.8
ilr transformed scaled data
Left plot: Two-part compositional data without the constraint of constant sum (symbols◦), and projections on the line indication a constant sum of 1 (symbols +).
Right plot: In the upper part the ilr tranformed original data (with symbols ◦ are shown.The lower plot shows the ilr transformed data with constant sum constraint (symbols
+). This demonstrates that the constant sum constraint does not change the ilrtransformed data.
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 13 / 28
![Page 16: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/16.jpg)
Transformation
Special Choice of ILR Variables
If, for example, the missing values are mainly contained in the �rst
compositional part of the data, one can choose the ilr transformation as
ilr(x) = (z1, . . . , zD−1)t , zj =
√D − j
D − j + 1ln
D−j
√∏Dl=j+1 xl
xj,
with j = 1, . . . ,D − 1.
Only this choice of the balances guarantees that missing values in x1 does
not a�ect z2, . . . , zD−1.
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 14 / 28
![Page 17: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/17.jpg)
Transformation
Outliers
1
1
00
First compositional part
Sec
ond
com
posi
tiona
l par
t
−2.0 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
ilr transformed data
Left plot: Two-part compositional data consisting of three groups. While the relativeinformation of the groups with symbols ◦ and + is similar, the data points corresponding
to the open triangles contain very di�erent information.Right plot: The ilr transformed data reveal that the group with open triangles areindeed di�erent. They are potentially in�uencing non-robust statistical methods.
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 15 / 28
![Page 18: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/18.jpg)
Imputation Methods
KNN Imputation
When imputing one missing value we
use the Aitchison distance to �nd k nearest neighbors.
adjust the corresponding cells according to the overall size of the
parts.
take the median of these cells to impute the missing.
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 16 / 28
![Page 19: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/19.jpg)
Imputation Methods
Iterative Model Based Imputation
Start: knn solution.
Order the data so that the �rst variable includes the highest amount
of missing values, . . .
Untill convergence:For i in 1 : D
Apply a speci�c, well-de�ned ilr-transformationUpdate former missing values in zi by regression imputation in theilr-space; zi is chosen as the response variable.back-transformation to the original space
end inner �loop�
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 17 / 28
![Page 20: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/20.jpg)
Simulation Results
Simulated Data
x1 x2
x3
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●●●
●
●●●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
−6 −4 −2 0 2 4 6−
4−
20
24
68
z1
z 2
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 18 / 28
![Page 21: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/21.jpg)
Simulation Results
Results0.
00.
51.
01.
52.
02.
53.
03.
5
Rel
ativ
e A
itchi
son
dist
ance
s
0 5 10 15 20 25
knn (Euclidean), k=6knn (Euclidean), k=8knn (Euclidean), k=10iterative LS (no transf.)iterative LTS (no transf.)
Outlier group 1 [%]Outlier group 2 [%]
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
Rel
ativ
e A
itchi
son
dist
ance
s
0 5 10 15 20 25
knn (Aitchison), k=6knn (Aitchison), k=8knn (Aitchison), k=10iterative LS (ilr)iterative LTS (ilr)
Outlier group 1 [%]Outlier group 2 [%]
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 19 / 28
![Page 22: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/22.jpg)
R-package robCompositions
Usage
http:
//cran.r-project.org/web/packages/robCompositions/index.html
> library(robCompositions)
> help(package=robCompositions)
Description:
Package: robCompositions
Type: Package
Title: Robust Estimation for Compositional Data.
Version: 1.1
Date: 2009-01-22
Depends: utils, e1071, robustbase, compositions, car, MASS
Author: Peter Filzmoser, Karel Hron, Matthias Templ
Maintainer: Matthias Templ <[email protected]>
Description: This first version of the package includes methods for
imputation of compositional data including robust
methods and Anderson-Darling normality tests for
compositional data. The package will be enhanced with
other multivariate methods for compositional data in
near future.
License: GPL-2
LazyLoad: yes
Built: R 2.8.0; ; 2009-01-22 16:53:39; windows
Index:
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 20 / 28
![Page 23: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/23.jpg)
R-package robCompositions
Data
We use the randomly generated data as used in the previous Figure.
> head(x)
[,1] [,2] [,3]
[1,] 0.29395572 0.16181078 0.09169296
[2,] 0.24290463 0.24092547 0.16041012
[3,] NA 0.05278444 0.51727452
[4,] NA 0.09599913 0.11838661
[5,] 0.31172499 0.22095742 0.35843191
[6,] 0.02038967 0.04858723 0.55728004
> dim(x)
[1] 100 3
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 21 / 28
![Page 24: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/24.jpg)
R-package robCompositions
Data
x1 x2
x3
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
−2 −1 0 1 2
−2
02
46
8z1
z 2
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 22 / 28
![Page 25: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/25.jpg)
R-package robCompositions
Missing Values
> library(VIM)
> plot(aggr(x))
X1
X2
X3
Nu
mb
er
of M
issin
gs
05
10
15
Co
mb
ina
tio
ns
X1
X2
X3
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 23 / 28
![Page 26: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/26.jpg)
R-package robCompositions
Missing Values
> library(VIM)
> plot(aggr(x))
X1
X2
X3
Nu
mb
er
of M
issin
gs
05
10
15
Co
mb
ina
tio
ns
X1
X2
X3
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 23 / 28
![Page 27: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/27.jpg)
R-package robCompositions
Imputation with robCompositions
> xImp <- impKNNa(x, k = 6)
> class(xImp)
[1] "imp"
> methods(class = "imp")
[1] plot.imp print.imp summary.imp
> xImp
---------------------------------------
[1] "31 missing vales were imputed"
---------------------------------------
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 24 / 28
![Page 28: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/28.jpg)
R-package robCompositions
Imputation with robCompositions
> xImp <- impKNNa(x, k = 6)
> class(xImp)
[1] "imp"
> methods(class = "imp")
[1] plot.imp print.imp summary.imp
> xImp
---------------------------------------
[1] "31 missing vales were imputed"
---------------------------------------
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 24 / 28
![Page 29: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/29.jpg)
R-package robCompositions
Imputation with robCompositions
> xImp <- impKNNa(x, k = 6)
> class(xImp)
[1] "imp"
> methods(class = "imp")
[1] plot.imp print.imp summary.imp
> xImp
---------------------------------------
[1] "31 missing vales were imputed"
---------------------------------------
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 24 / 28
![Page 30: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/30.jpg)
R-package robCompositions
Imputation with robCompositions
> xImp <- impKNNa(x, k = 6)
> class(xImp)
[1] "imp"
> methods(class = "imp")
[1] plot.imp print.imp summary.imp
> xImp
---------------------------------------
[1] "31 missing vales were imputed"
---------------------------------------
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 24 / 28
![Page 31: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/31.jpg)
R-package robCompositions
Imputation with robCompositions
> xImp <- impKNNa(x, k = 6)
> class(xImp)
[1] "imp"
> methods(class = "imp")
[1] plot.imp print.imp summary.imp
> xImp
---------------------------------------
[1] "31 missing vales were imputed"
---------------------------------------
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 24 / 28
![Page 32: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/32.jpg)
R-package robCompositions
Imputation with robCompositions
> xImp <- impKNNa(x, k = 6)
> class(xImp)
[1] "imp"
> methods(class = "imp")
[1] plot.imp print.imp summary.imp
> xImp
---------------------------------------
[1] "31 missing vales were imputed"
---------------------------------------
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 24 / 28
![Page 33: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/33.jpg)
R-package robCompositions
Imputation with robCompositions
> names(xImp)
[1] "xOrig" "xImp" "criteria" "iter" "w" "wind" "metric"
> xImp$xImp[1, 3]
[1] 0.09169296
> xImp1 <- impCoda(x, method = "lm")
> xImp2 <- impCoda(x, method = "ltsReg")
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 25 / 28
![Page 34: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/34.jpg)
R-package robCompositions
Imputation with robCompositions
> names(xImp)
[1] "xOrig" "xImp" "criteria" "iter" "w" "wind" "metric"
> xImp$xImp[1, 3]
[1] 0.09169296
> xImp1 <- impCoda(x, method = "lm")
> xImp2 <- impCoda(x, method = "ltsReg")
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 25 / 28
![Page 35: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/35.jpg)
R-package robCompositions
Imputation with robCompositions
> names(xImp)
[1] "xOrig" "xImp" "criteria" "iter" "w" "wind" "metric"
> xImp$xImp[1, 3]
[1] 0.09169296
> xImp1 <- impCoda(x, method = "lm")
> xImp2 <- impCoda(x, method = "ltsReg")
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 25 / 28
![Page 36: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/36.jpg)
R-package robCompositions
Imputation with robCompositions
> names(xImp)
[1] "xOrig" "xImp" "criteria" "iter" "w" "wind" "metric"
> xImp$xImp[1, 3]
[1] 0.09169296
> xImp1 <- impCoda(x, method = "lm")
> xImp2 <- impCoda(x, method = "ltsReg")
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 25 / 28
![Page 37: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/37.jpg)
R-package robCompositions
Imputation with robCompositions
> names(xImp)
[1] "xOrig" "xImp" "criteria" "iter" "w" "wind" "metric"
> xImp$xImp[1, 3]
[1] 0.09169296
> xImp1 <- impCoda(x, method = "lm")
> xImp2 <- impCoda(x, method = "ltsReg")
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 25 / 28
![Page 38: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/38.jpg)
R-package robCompositions
Diagnostics
> plot(xImp2 , which = 1)
||| | || | |||| | || || | || | || ||| || | || | ||| || | || ||| || | |||| ||| || || || | | | | || || | ||| | || ||| | || ||| || ||| | ||| | | |||| || |
x1
0.0 0.2 0.4 0.6
●
●●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●●●
0.0
0.2
0.4
0.6
0.8
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●●
●
0.0
0.2
0.4
0.6
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●●
●
●
●●
●
●
●
●●
●
●●●●● | || ||| | | | || || | || ||| | || | |||| ||| ||||| || | || |||| | || ||| || || || || || | || |||| ||| ||| ||| ||||| ||| ||| | ||| | ||||||||
x2
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●●
●
●
●●
●
●
●
●●
●
●●●●●
0.0 0.2 0.4 0.6 0.8
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
● ●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●●●●
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
| | ||| ||| | | || ||| ||| ||| || | || ||| ||| ||||||| || || |||| | || | || || || || |||||| ||| | ||| ||| ||| || ||| || | ||| | |||| | |||||
x3
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 26 / 28
![Page 39: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/39.jpg)
R-package robCompositions
Diagnostics
> plot(xImp2 , which = 3)
x1 x2
x3
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●●
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 27 / 28
![Page 40: Robust Imputation of Missing Values in Compositional Data Using the R …ec.europa.eu/eurostat/documents/1001617/4398393/S6P3... · 2014-09-25 · Robust Imputation of Missing Values](https://reader033.vdocuments.mx/reader033/viewer/2022060209/5f047bea7e708231d40e329e/html5/thumbnails/40.jpg)
Conclusion
Conclusion
We tested more than 20 imputation procedures which all were
outperformed by our method (Hron, Templ, Filzmoser, 2008) for
compositional data.
Robustness is an issue. We proposed new robust imputation methods
for compositional data.
R-package robCompositions includes these methods, but other
methods are implemented as well. Diagnostic tools are available
within the package.
A lot of important issues were not mentoined in this presentation, but
they have been discussed in our NTTS-paper or in Hron, Templ,
Filzmoser (2008).
Templ, Filzmoser, Hron (TUW) Robust Imputation in CoDa Brussels, Feb. 19, 2009 28 / 28