1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
|
TO DO
------
o Generalizing 'wgt.himedian': We'd want a C API on which R builds.
There are pure R implementations:
- 'weighted.median()' in limma
and I have generalized it ---> inst/ex-funs.R
- more general code (different 'tie' strategies; weighted *quantile*s)
in /u/maechler/R/MM/STATISTICS/robust/weighted-median.R
- The 'Hmisc' package has wtd.quantile()
o covOGK():
The argument name 'weight.fn' is pretty ugly and the default function
name 'hard.rejection()' is just awful (we need a globally available
function as 'role model'.
- Could allow 'n.iter = 0' to simply compute Cov()_{ij} = rcov(X_i, X_j)
o scaleTau2(): Also do a cheap finite-sample correction [MM] !
[done partly; but undocumented, since bound to change]
--- rrcov etc ---
o rrcov.control() __ NEEDS name change ! ______
probably use mcd.control() and lts.control()
or forget about *control() completely?
since there are only a few in each ??????/
o tolellipse() --> renamed to tolEllipsePlot()
- maybe use cluster::ellipsoidPoints()
- allow other percentiles than just 97.5%
- maybe *return* something
o plot(mcd. ) [ R/covPlot.R ] : should show the call
Default for 'ask' should be smarter: depend on
prod(par("mfrow")) < #{plots} (which depends on 'classic' and p=2)
o ltsReg(): has undocumented '$resid'
in addition to '$residuals' and '$raw.residuals';
drop it or document it !
--- glmrob --
o predict.glmrob() __NEEDED__ and probably also residuals.glmrob()
should allow a 'type = *' specification like the *.glm() methods.
o glmrob(*, weights.on.x = "robCov") currently uses MASS::cov.rob(), i.e.
"MVE" and Andreas had a comment that "mcd" is worse.
But still, I (MM) strongly believe we should use covMcd() instead.
HOWEVER: Need something better when 'X' has (binary!) factors!
"hat" +- works, but needs more work
o psi.*(...., rho = FALSE/TRUE) functions from Andreas
should be replaced by using the new psi_func objects
--> ./R/psi-funs-AR.R & ./man/pkg-internal.Rd
o glmrob() needs a bit more tests in ./tests/
[also consider those from man/glmrob.Rd]
take those from Martin's old 'robGLM1' package (need more!)
o --> first test already shows that Martin's tests for "huberC == Inf"
were *not* yet moved from robGLM1 to glmrob()...
(in other words: glmrob() should work
o also, ni = 0 does not work quite as it should ( ./tests/binom-ni-small.R )
o obj $ df ... maybe should be defined -- for "glm" methods to be
applicable
o summary.glmrob() should be better documented;
we should decide if the current return value is fine.
o Eva's code (and MM's) also computed & returned the "asymptotic efficiency"!
o anova.glmrob(): More modularization, allowing to provide own 'test' function.
Test if Huber's C are different. Need theory to compare different C's and
same model (which includes classical vs robust).
o drop1() would be nice
--- nlrob ---
o nlrob() needs tests in ./tests/ -- you can take some from man/nlrob.Rd
o summary.nlrob() is currently a "no-op" -- printing it should summarize
robustness weights!
------
o Add data sets from the MMY-book -- mostly done {do we have *all* ?}
--- lmrob --- --- --- --- ---
o more tests in tests/
o fully implement and test the multivariate case (y = matrix with > 1 col.)
o src/lmrob.c :
Things to fix:
- does many many vector and vector-matrix things itself
instead of using BLAS and LAPACK
- does median() , MAD() instead of using R's sort() routines
----
o Alternative version of covOGK() for correlation-only
using's Huber's correlation formula which ensures [-1,1] range
--> ~/R/MM/Pkg-ex/robustbase/robcorgroesser1.R
and ~/R/MM/STATISTICS/robust/pairwise-new.R
o package 'riv' (author @ epfl.ch!) has 'slc()' ~= cov.S(.) -- in pure R code
doesn't Valentin have a version too?
otherwise: test this, ask author for "donation" to robustbase
o adjOutlyingness() :
typo-bug is corrected; and I have made it more pretty.
Still a bit problematic when denominator = 0
Currently leave away all the c/0 = Inf and 0/0 = NaN values.
MM: Maybe, it's the fact that the coef = 1.5 should really depend on
the sample size n and will be too large for small n (??)
--> should ask Mia and maybe Guy Brys
|
