Transcript Robustness

Robust Statistics
Osnat Goren-Peyser
25.3.08
Outline
1.
2.
3.
4.
5.
6.
Introduction
Motivation
Measuring robustness
M estimators
Order statistics approaches
Summary and conclusions
1. Introduction
Problem definition




Let x   x1 , x2 ,..., xn  be a set of iid variables
with distribution F, sorted in increasing order
so that x1  x 2  ...  x n
x m  is the value of the m’th order statistics
An estimator ˆ  ˆ  x1 , x2 ,..., xn  is a
function of the observations.
We are looking for estimators such that
  ˆ
Asymptotic values of estimator



ˆ F 
Define: ˆ  ˆ  F  such that ˆn 



p
where ˆ  F  is the asymptotic value of
estimator ˆn at F.
Estimator ˆn is consistent for θ if: ˆ  F   
We say that ˆn is asymptotically normal with
parameters θ,V(θ) if n ˆn   
 N  0,V   
d


Efficiency

An unbiased estimator is efficient if

I   var ˆ  1
Where I   is Fisher information

An unbiased estimator is asymptotic
efficient if

lim I   var ˆ  1
n 
Relative efficiency


For fixed underlying distribution, assume two
unbiased estimators ˆ1 and ˆ2 of  .
we say ˆ is more efficient than ˆ2 if
1

 
 
var ˆ1  var ˆ2
The relative efficiency (RE) of ˆ1 estimator with
respect to ˆ2 is defined as the ratio of their variances.


 
RE ˆ2 ;ˆ1  var ˆ1
 
var ˆ2
Asymptotic Relative efficiency


The asymptotic relative efficiency (ARE)
is the limit of the RE as the sample size n→∞
For two estimators ˆ 1 ,ˆ 2 which are each
consistent for θ and also asymptotically
normal [1]:
var ˆ 1
ARE ˆ 2 ;ˆ 1  lim
n 
var ˆ 2


 
 
The location model



xi=μ+ui ; i=1,2,…,n
Where μ is the unknown location parameter,
ui are the errors, and xi are the observations.
The errors ui‘s are i.i.d random variables each
with the same distribution function F0.
The observations xi‘s are i.i.d random
variables with common distribution function:
F(x) = F0(x- μ)
Normality




Classical statistical methods rely on the assumption
that F is exactly known
2
The assumption that F  N  , 
is a normal
distribution is commonly used
But normality often happens approximately  robust
methods
Approximately normality





The majority of observations are normally distributed
some observations follow a different pattern (not normal) or
no pattern at all.
Suggested model: a mixture model
A mixture model





Formalizing the idea of F being approximate normal
Assume that a proportion 1-ε of the observations is
generated by the normal model, while a proportion ε
is generated by an unknown model.
The mixture model: F=(1- ε)G+ εH
F is a contamination “neighborhood” of G and also
called the gross error model
F is called a normal mixture model when both G and
H are normal
2. Motivation
Outliers



Outlier is an atypical observation
that is well separated from the bulk
of the data.
Statistics derived from data sets
that include outliers will often be
misleading
Even a single outlier can have a
large distorting influence on a
classical statistical methods
Without the outliers
100
90
With the outliers
180
2
1
160
80
140
70
120
60
100
50
80
40
60
30
outliers
40
20
20
10
0
-4
-2
0
values
2
4
0
-15
Estimators not sensitive to outliers
are said to be robust
-10
-5
values
0
5
Mean and standard deviation

The sample mean is defined by
1 n
x   xi
n i 1



A classical estimation for the location (center) of the data
2
2
For N  ,  , the sample mean is unbiased with N  ,  n



The sample standard deviation (SD) is defied by
1 n
2
s
x

x


 i
n  1 i 1

A classical estimation for the dispersion of the data
How much influence a single outlier can have on these
classical estimators?

Example 1 – the flour example



Consider the following 24 determinations of the copper content
in wholemeal flour (in parts per million), sorted in ascending
order [6]
2.20,2.20,2.40,2.40,2.50,2.70,2.80,2.90,
3.03,3.03,3.10,3.37,3.40,3.40,3.40,3.50,
3.60,3.70,3.70,3.70,3.70,3.77,5.28,28.95
The value 28.95 considered as an outlier.
Two cases:


Case A - Taking into account the whole data
Case B - Deleting the suspicious outlier
Example 1 – PDFs
Case A
Case B
x  4.28, s  5.30
x  3.21, s  0.69
Case B - Deleting the outlier
Case A - Using the whole data for estimation
0.18
0.18
data
sample mean
0.14
0.14
0.12
0.12
0.1
0.08
0.1
0.08
0.06
0.06
0.04
0.04
0.02
0.02
0
0
5
mean
10
15
Observation value
20
25
data
sample mean
0.16
Probability
Probability
0.16
30
outlier
0
2
2.5
3
3.5
4
Observation value
mean
4.5
5
5.5
Example 1 – arising question


Question: How much influence a single outlier can
have on sample mean and sample SD?
Assuming the outlier value 28.95 is replaced by an
arbitrary value varying from −∞ to +∞:




The value of the sample mean changes from −∞ to +∞.
The value of the sample SD changes from −∞ to +∞.
Conclusion: A single outlier has an unbounded
influence on these two classical estimators!
This is related to sensitivity curve and influence
function, as we will see later.
Handling outliers approaches

Detect and remove outliers from the
data set




Manual screening
The normal Q-Q plot
The “three-sigma edit” rule
Robust estimators!
Manual screening
Why screen the data and remove outliers is not sufficient?

Users do not always screen the data.

Outliers are not always errors!



Outliers may be correct, and very important for seeing the whole
picture including extreme cases.
It can be very difficult to spot outliers in multivariate or highly
structured data.
It is a subjective decision


Without any unified criterion: Different users  different results
It is difficult to determine the statistical behavior of the complete
procedure
The Q-Q normal plot



Manual screening tool
for an underlying
Normal distribution
A quantile-quantile plot
of the sample quantiles
of X versus theoretical
quantiles from a normal
distribution.
If the distribution of X is
normal, the plot will be
close to linear.
The “three-sigma edit” rule


Outlier detection tool for an underlying Normal distribution
Define the ratio between xi distance to the sample mean and
the sample SD:
ti 



xi  x
s
The “three-sigma edit rule”: Observations with |ti|>3 are
deemed as suspicious
Example 1 - The largest observation in the flour data has
ti=4.65, and so is suspicious
Disadvantages:


In a very small samples the rule is ineffective
Masking: When there are several outliers, their effects may interact in such
a way that some or all of them remain unnoticed
Example 2 – Velocity of light

Consider the following 20 determinations of the time
(in microseconds) needed for light to travel a
distance of 7442 m [6].
28,26,33,24,34,-44,27,16,40,-2
29, 22,24,21,25,30, 23,29,31,19


The actual times are the table values × 0.001 + 24.8.
The values -2 and -44 suspicious as outliers.
Example 2 – QQ plot
QQ Plot of Sample Data versus Standard Normal
24.84
24.83
Quantiles of Input Sample
24.82
24.81
24.8
24.79
Outlier -2
24.78
24.77
24.76
24.75
-2
Outlier -44
-1.5
-1
-0.5
0
0.5
Standard Normal Quantiles
1
1.5
2
Example 2 – Masking

Results:
ti  xi  2  1.35 , ti  xi  44   3.73

Based on the three-sigma edit rule:


the value of |ti | for the observation −2 does not indicate
that it is an outlier.
the value −44 “masks” the value −2.
Detect and remove outliers


There are many other methods for detecting outliers.
Deleting an outlier poses a number of problems:


Affects the distribution theory  Underestimating data
variability
Depends on the user’s subjective decisions  difficult to
determine the statistical behavior of the complete
procedure.
Robust estimators provide automatic ways of detecting,
and removing outliers
Example 1 – Comparing the sample
median to the sample mean
Case C - Testing the median estimator
data
Case A
Case A
Case B
Case B
0.16
0.14
Probability
0.12
sample mean
sample median
sample mean
sample median
0.1
• Case A: med = 3.3850
• Case B: med = 3.3700
• The sample median fits the bulk
of the data in both cases
• The value of the sample
median does not change from
−∞ to +∞ as was the case
for the sample mean.
0.08
0.06
0.04
0.02
0
2
2.5
3
3.5
4
Observation value
4.5
5
5.5
Median A
Mean B
Mean A
Median B
• The sample median is a good
robust alternative to the
sample mean.
Robust alternative to mean


Sample median is a very old method for estimating the “middle”
of the data.
The sample median is defined for some integer m by
if n is odd , n  2m  1
 x m 

Med  x    x m  x m 1
 
 
if n is even, n  2m

2





For large n and F  N  ,  2 , the sample median is
approximately N   ,  2 2n 
At normal distribution:
ARE(median;mean)=2/π≈64%
Single outlier affection





The sample mean can be upset completely by a
single outlier
The sample median is little affected by a single
outlier
The median is resistant to gross errors whereas the
mean is not
The median will tolerate up to 50% gross errors
before it can be made arbitrarily large
Breakdown point:


Median: 50%
Mean: 0%
Mean & median –
robustness vs. efficiency

For mixture mode:
ε
1.4
1.68
1.8
1.8
4
1
1.57
1.75
1.7
2.5
1.84
5
1
1.57
2.2
1.7
3.4
1.86
6
1
1.57
2.75
1.71
4.5
1.87

2
2n 1      
10
1
1.57
5.95
1.72
10.9
1.9
20
1
1.57
20.9
1.73
40.9
1.92
nvar(med)
The sample median variance is
approximately
The gain in robustness due to using the median is paid for
by a loss in efficiency when F is very close to normal
nvar(med)
1.57
nvar(mean)
1
nvar(mean)
n

3
The sample mean variance is
1      2
0.1
τ
nvar(med)


0.05
nvar(mean)

F  1    N   ,1   N  , 2
0
So why not always use the
sample median?



If the data do not contain outliers, the sample median
has statistical performance which is poorer than that
of the classical sample mean
Robust estimation goal: “The best of both worlds”
We shall develop estimators which combine the low
variance of the mean at the normal, with the
robustness of the median under contamination.
3. Measuring robustness
Analysis tools



Sensitivity curve (SC)
Influence function (IF)
Breakdown point (BP)
Sensitivity curve


SC measures the effect of different locations of
an outlier on the sample
The sensitivity curve of an estimator ̂ for the
samples x1 , x2 ,..., xn is:
SC  x0   ˆ  x1 , x2 ,..., xn , x0   ˆ  x1 , x2 ,..., xn 

where x0 is the location of a single outlier
Bounded SC(x0)  high robustness!
SC of mean & median
-3
5
4
4
3
3
2
2
1
1
SC

n=200
F=N(0,1)
SC

5
-3
Sample mean
x 10
0
-1
-2
-2
-3
-3
-4
-4
-5
0
outlier
5
10
Sample median
0
-1
-5
-10
x 10
-5
-10
-5
0
outlier
5
10
Standardized sensitivity curve

The standardized sensitivity curve is
defined by
SCn  x0  
ˆ  x1 , x2 ,..., xn , x0   ˆ  x1 , x2 ,..., xn 
1  n  1
What happens if we add one more observation
to a very large sample?
Influence function



The influence function of an estimator̂ (Hampel, 1974)
is an asymptotic version of its sensitivity curve.
It is an approximation to the behavior of θ∞ when the
sample contains a small fraction ε of identical outliers.
It is defined as
ˆ  1    F   x   ˆ  F 
IFˆ  x0 , F   lim
 0

 ˆ
    1    F   0   0

0
Where  x is the point-mass at x0 and  stands for “limit
from the right”
0
IF main uses

ˆ  1    F   x


is the asymptotic value of the
estimate when the underlying distribution is F and a
fraction ε of outliers is equal to x0.
0
IF has two main uses:

Assessing the relative influence of individual observation
towards the value of estimate


Unbounded IF  less robustness
Allowing a simple heuristic assessment of the asymptotic
variance of an estimate
IF as limit version of SC

The SC is a finite sample version of IF

If ε is small



ˆ  1    F   x   ˆ  F    IFˆ  x0 , F 

0

bias  ˆ 1    F   x0  ˆ  F    IFˆ  x0 , F 
for large n: SCn  x0   IFˆ  x0 , F 
where ε=1/(n+1)
Breakdown Point




BP is the proportion of arbitrarily large
observations an estimator can handle before
giving an arbitrarily large result
Maximum possible BP = 50%
High BP  more robustness!
As seen before:


Mean: 0%
Median: 50%
Summary



SC measures the effect of different
outliers on estimation
IF is the asymptotic behavior of SC
The IF and the BP consider extreme
situations in the study of contamination.


IF deals with “infinitesimal” values of ε
BP deals with the largest ε an estimate can
tolerate.
4. M estimators
Maximum likelihood of μ


Consider the location model, and assume that F0
has density f 0
The likelihood function is:
n
L  x1 , x2 ,..., xn ;     f 0  xi   
i 1

The maximum likelihood estimate (MLE) of μ is
ˆ  arg max L  x1 , x2 ,..., xn ;  

M estimators of location (μ)


MLE-like estimators: generalizing ML estimators
If we have density f 0 , which is everywhere positive
The MLE would solve:
n
(*)

Let
  
ˆ  arg min    xi   

   log f 0
i 1
, if this exists, then
n
(**)
  x  ˆ   0
i 1



i
M estimator can almost equivalently described by ρ or Ψ
If ρ is everywhere differentiable and ψ is monotonic, then the
forms (*) and (**) are equivalent [6]
If ψ continuous and increasing, the solution is unique [6]
Special cases
The sample mean
The sample median

  x   x2 2


  x  x

n

1
i 1
i 1
1

  x   sign  x   0
1

 I  x  0  I  x  0
x0
x0
x0
n
n
  xi  ˆ   0  ˆ  n  xi
  x  x

 sign  x  ˆ   0 
i
i 1
n
  I  x  ˆ   0  I  x  ˆ   0  0
i 1
i
i
 #  xi  ˆ   #  xi  ˆ   0  ˆ  Med  x 
Special cases: ρ and ψ
Squared errors
Absolute errors
4
2
2
1
0
-4
0
-4
-2
0
2
x
Squared errors
4
4
1
2
0.5
(x)
(x)
Sample
mean
(x)
3
(x)
6
0
-2
-4
-4
-2
0
2
x
Absolute errors
4
0
-0.5
-2
0
x
2
4
-1
-4
-2
0
x
2
4
Sample
median
Asymptotic behavior of
location M estimators

For a given distribution F, assume ρ is differentiable and ψ is
increasing, and define 0  0  F  as the solution of
EF  x  0   0

For large n,
ˆ 
 0
p
and the distribution of estimator
is approximately N  0 , n  with

If
̂
(***)

  EF   x 
0
is uniquely defined then is consistent at F [3]
2

̂
 EF0   x 
2
Desirable properties

M estimators are robust to large proportions of outliers


The IF is proportional to ψ


Ψ function may be chosen to bound the influence of outliers
and achieve high efficiency
M estimators are asymptotically normal


When ψ is odd, bounded and monotonically increasing, the BP
is 0.5
Can be also consistent for μ
M estimators can be chosen to completely reject outliers
(called redescending M estimators) , while maintaining a
large BP and high efficiency
Disadvantages


They are in general only implicitly
defined and must be found by iterative
search
They are in general will not be scale
equivariant
Huber functions



A popular family of M estimator (Huber,1964)
This estimator is an odd nondecreasing ψ function
which minimizes the asymptotic variance among all
estimator satisfying: IF ( x)  c  
where c   2
Advantages:


Combines sample mean for small errors with sample median
for gross errors
Boundedness of ψ
Huber ρ and ψ functions

2

x

k  x   
2
2
k
x

k


if x  k
if x  k
With derivative 2 k  x  , where

 x
 k  x  
 sign  x  k
 0  x   sign  x 
if x  k , k  0
if x  k , k  0
Huber ρ and ψ functions
ρ (x)
Ψ(x)
Ricardo A. Maronna, R. Douglas Martin and V´ıctor J. Yohai, Robust Statistics: Theory and
Methods , 2006 John Wiley & Sons
Huber functions – Robustness
& efficiency tradeoff
Increasing v

Special cases:


median

Decreasing
robustness

mean
Ricardo A. Maronna, R. Douglas Martin and
V´ıctor J. Yohai, Robust Statistics: Theory and
Methods , 2006 John Wiley & Sons

K=0  sample median
K ∞  sample mean
Asymptotic variances at
normal mixture model with
G = N(0, 1) and H = N(0, 10),
The larger the asymptotic
variance, the less efficient
estimator, but the more
robust.
Efficiency come on the
expense of robustness
Redescending M estimators



Redescending M estimators have ψ functions which are nondecreasing near the origin but then decrease toward the axis far
from the origin
They usually satisfy ψ(x)=0 for all |x|≥r, where r is the
minimum rejection point.
Accept the completely outliers rejection ability, they:




Do not suffer form masking effect
Has a potential to have high BP
Their ψ functions can be chosen to redescend smoothly to zero 
information in moderately large outliers is not ignored completely
 improve efficiency!
A popular family of redescending M estimators (Tukey), called
Bisquare or biweight
Bisquare ρ and ψ functions


2 3
 
if x  k
1  1   x / k  
  x  

if x  k
1
2
With derivative 6  x  / k , where
  x
  x   x 1   
  k 
2
2

 I  x  k

Bisquare ρ and ψ functions
ρ (x)
Ψ(x)
Ricardo A. Maronna, R. Douglas Martin and V´ıctor J. Yohai, Robust Statistics: Theory and
Methods , 2006 John Wiley & Sons
Bisquare function – efficiency


ARE(bisquare;MLE)=
ARE
0.8
0.85
0.9
0.95
k
3.14
3.44
3.88
4.68
Achieved ARE close to 1
Choice of ψ and ρ



In practical, the choice of ρ and ψ
function is not critical to gaining a good
robust estimate (Huber, 1981).
Redescending and bounded ψ functions
are to be preferred
Bounded ρ functions are to be preferred
Bisquare function is a popular choice
5. Order statistics
approaches
The β-trimmed mean




Let   0, 0.5 and m     n  1 
The β-trimmed mean is defined by
nm
1
x 
x i 

n  2m i  m 1
Where [.] stands for the integer part and x i 
denotes the ith order statistic.
x is the sample mean after the m largest
and the m smallest observation have been
discard
The β-trimmed mean – cont.

Limit cases



 the sample mean
 the sample median
Distribution of trimmed mean



β=0
β 0.5
The exact distribution is intractable
For large n the, distribution under location
model is approximately normal
BP of β % trimmed mean = β %
Example 1 – trimmed mean
All data
Delete outlier
Mean
4.28
3.2
Median
3.38
3.37
Trimmed mean 10%
3.2
3.11
Trimmed mean 25%
3.17
3.17
Median and trimmed mean are less sensitive to
outliers existence
The W-Winsorized mean

xw, 
The W-Winsorized mean is defied by
n  m 1
1

  m  1 x m 1   x i    m  1 x n  m  
n
i m 2


The m smallest observations are replaced by the
(m+1)’st smallest observation, and the m largest
observations are replaced by the (m+1)’st largest
observations
Trimmed and W-Winsorized
mean disadvantages



Uses more information from the sample than
the sample median
Unless the underlying distribution is
symmetric, they are unlikely to produce an
unbiased estimator for either the mean or the
median.
Does not have a normal distribution
L estimators


Trimmed and Winsorized mean are special cases of L
estimators
L estimators is defined as Linear combinations of order
statistics:
n
ˆ    i x i 
i 1
Where the

i ' s
are given constants
1
For β-trimmed mean:  i 
I m 1  i  n  m
n  2m
L vs. M estimators

M estimators are:




More flexible
Can be generalized straightforwardly to
multi-parameter problems
Have high BP
L estimator

Less efficient because they completely
ignore part of the data
6. Summary & conclusions
SC of location M estimators
n=20
xi~N(0,1)
Trimmed mean: α=25%
Huber: k=1.37
Bisquare: k=4.68
The effect of increasing
contamination on a sample
Replace m points by a fixed value x0=1000

biased SC  m  ˆ  x0 ,..., x0 , xm1 ,..., xn   ˆ  x1, x2 ,..., xn 
n=20
xi~N(0,1)
Trimmed mean: α=8.5%
Huber: k=1.37
Bisquare: k=4.68

IF of location M estimator


IF is proportional to ψ (Huber, 1981)
In general
  x0  ˆ  
IFˆ  x0 , F  
E   x  ˆ  
BP of location M estimator

In general



When ψ is odd, bounded and monotonically increasing, the
BP is 50%
Assume k1      , k2      are finite, then
the BP is: min  k1 , k2  k1  k2
Special cases:


Sample mean = 0%
Sample median = 50%
Comparison between different
location estimators
Estimator
BP
SC/IF/ψ
Redescending ψ
Efficiency in
mixture
model
unbounded
No
low
Median 50%
bounded
No
low
Huber 50%
bounded
No
high
Bounded at 0
Yes
high
bounded
No
Mean 0%
Bisquare 50%
x% trimmed mean x%
Conclusions



Robust statistics provides an alternative
approach to classical statistical methods.
Robust statistics seeks to provide methods
that emulate classical methods, but which are
not unduly affected by outliers or other small
departures from model assumptions.
In order to quantify the robustness of a
method, it is necessary to define some
measures of robustness
Efficiency vs. Robustness


Efficiency can be achieved by taking ψ
proportional to the derivative of the loglikelihood defined by the density of F:
ψ(x)=-c(f’/f)(x), where c is constant≠0
Robustness is achieved by choosing ψ that
is smooth and bounded to reduce the
influence of a small proportion of
observations
References
1.
2.
3.
4.
5.
6.
7.
8.
Robert G. Staudte and Simon J. Sheather, Robust estimation and
testing, Wiley 1990.
Elvezio Ronchetti, “THE HISTORICAL DEVELOPMENT OF ROBUST
STATISTICS”, ICOTS-7, 2006: Ronchetti, University of Geneva,
Switzerland, 2006.
Huber, P. (1981). Robust Statistics. New York: Wiley.
Hampel F. R., Ronchetti, E. M., Rousseeuw, P. J. and Stahel, W. A.
(1986). Robust Statistics: The Approach Based on Influence Functions.
New York: Wiley.
Tukey
Ricardo A. Maronna, R. Douglas Martin and V´ıctor J. Yohai, Robust
Statistics: Theory and Methods , 2006 John Wiley & Sons
B. D. Ripley , M.Sc. in Applied Statistics MT2004, Robust Statistics,
1992–2004.
Robust statistics From Wikipedia.