Scribd
Upload
57 views8 pages

Peirce Sub

Peirce's criterion for rejecting outliers in datasets has been used for over 150 years. The author implements Peirce's method as an R function to automate outlier removal for large datasets generated in their lab. They compare Peirce's criterion to other outlier detection methods using several sample datasets. Peirce's criterion performs comparably to other methods while rejecting fewer observations per application. The author also explores the theoretical underpinnings and limitations of Peirce's criterion as described in the original 1852 paper.

Uploaded by

neverwolf
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
57 views8 pages

Peirce Sub

Peirce's criterion for rejecting outliers in datasets has been used for over 150 years. The author implements Peirce's method as an R function to automate outlier removal for large datasets generated in their lab. They compare Peirce's criterion to other outlier detection methods using several sample datasets. Peirce's criterion performs comparably to other methods while rejecting fewer observations per application. The author also explores the theoretical underpinnings and limitations of Peirce's criterion as described in the original 1852 paper.

Uploaded by

neverwolf
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 8

JSS

Journal of Statistical Software


MMMMMM YYYY, Volume VV, Issue II. http://www.jstatsoft.org/
Peirces Criterion for the Rejection of Non-Normal
Outliers; Dening the Range of Applicability
Christopher Dardis
Barrow Neurological Institute, Phoenix, Arizona
Abstract
Peirces criteria for the rejection of non-normal outliers has been with us for over 150
years. Here, I present an implementation of the method in R. A number of examples are
presented and I discuss its range of applicability. Finally, I give illustrations from the
early literature on the method.
Keywords: Peirce, outlier, R.
1. Introduction
Peirces criterion for the rejection of non-normality have been with us for over 150 years. It
was, in fact, the rst criterion developed for the exclusion of outliers. I became interested in
his methods during the course of some lab research, where it became clear that the techniques
we were using were producing occasional grossly erroneous results.
I was persuaded of the merits of the technique by a paper from Ross (2003), which gave a
simple but practical technique for applying the method. However given the volume of data
our lab was generating, I sought to automate the method as a function in R (Team (2012)).
I was also interested to see if the technique generalized to a broader range than that given in
the above paper, which is limited to rejecting up to 9 outliers from 60 observations.
The original paper Peirce (1852) describes a technique for rejecting doubtful observations in
the case of those arising from observing planetary motion. It is assumed that unusual depar-
tures from normality (a Gaussian distribution) of observations are the result of the observer
rather the planets themselves. I believe similar assumptions are worthwhile in generalizing
the ndings of the natural sciences.
In brief, his technique was to generate probabilities of error occurring in the system where all
N observations are retained vs. that where k are rejected. He then rejected k observations if
2 Peirce
the new system (i.e., rejecting k) is closer to normal than the old.
2. Practical application
To begin with, an illustration of the merit of his technique following, with comparisons
with a number of standard existing techniques. While multiple methods are already im-
plemented in the outliers package in R, the following are limited to removing only one value:
chisq.out.test, dixon.test, grubbs.test. There is limited literature as to how legiti-
mate it is to repeat them multiple times on a dataset.
A leading alternative for rejecting non-normal values is Chauvenets criterion Chauvenet.R.
There is a lack of consensus as to whether it is legitimate to repeatedly apply the function,
so I have provided the option loop=TRUE to do so. Repeated application tends to further
shrink sets, whereas Peirces criterion does not suer this disadvantage. I took four sample
sets; that from Ross (2003), one from the National Institute of Standards and Technology
Natrella (2012) and two that are already available with R. The latter two are cautionary tales:
TeachingDemos in regard to repeated application of Chauvenets, compositions sa.outliers for
the perils of a set with complete separation. These sets are shown in Figure 1 and the results
in Table 1. Full details in PeirceVsChauvenet.R.
Another approach to rejecting outliers has been proposed by bbalibor
TM
. Their goal is to
determine an average value for a series of submissions (referred to as libor, basically a nominal
rate of interest). They suggest that a reasonable approach in the case of 16 submissions is
to eliminate the upper and lower 4, leaving 8 and then taking the mean of these 8. (Method
explained here). I compared this to using Peirces criterion on each set of 16 observations,
then averaging those, gure 2. (This is based on part of a complete set available from google
docs). Although there is no reason a priori to assume that the submissions on which libor is
based should follow a normal distribution, it can clearly be seen that Peirces criterion gives a
result almost identical to its rival, and excludes far fewer observations per application (range
is 1-4 in this example, vs 8 each time). This is akin to saying that excluding larger numbers
of outliers tends to have little eect on the mean value for this type of data. Alternatively,
one may say that the bbalibor
TM
method is to be preferred owing to its simplicity.
Dataset No. observations Method (no. removed)
Peirce Chauvenet Chauvenet
(repeated)
Ross 10 2 2 8
NIST 90 11 3 13
TeachingDemos 100 7 6 17
compositions sa.outliers 300 71 0 0
Table 1: Performance of outlier detection methods.
3. Methods
I sought to duplicate the table and results from Ross (2003). I was able to do so by following
Journal of Statistical Software 3
Figure 1: Illustration of datasets used to test outlier methods.
Figure 2: Libor calculated traditionally and using Peirces criterion.
4 Peirce
Figure 3: Values of R from % values of k and m. Sample size N=1000.
the methods in Gould (1855), (see PeirceGould.R). However, the upper limit for N, the
number of observations in the sample, is limited by Rs representation of large numbers, which
is given by .Machine$double.xmax and is 143 on my device. A more ecient technique for
achieving the same result already exists in C, and I implemented this as Peirce.R. Both
methods rely on generating R, the ratio of the absolute error of one measurement to the
sample standard deviation :
R =
|x
i
x|

(1)
Where x is the sample mean. R depends on k, the number of outliers proposed to be rejected
and m the number of unknown quantities. The meaning of this latter may be destined to
remain obscure; however it appears to be something akin to degrees of freedom, i.e., the
number of independent processes that are giving rise to outliers in the data. Gould (1855)
acknowledges that the cases of m > 2 are of little practical signicance.
An illustration of the range of values of R for k = 0 100% and m = 0 100% of N is shown
in Figure 3, with details in PeirceLimits.R. For values of m = 1 we can see R dropping
below 0 at the point where k > 90% i.e., it is meaningless to try to reject more than 90%
of a given dataset. Additionally, for low values of k, increasing m does reduce R.
As an aside, I sought to generate the values in Table III in Gould (1855), giving values of
NlogQ for N observations and k proposed rejections. I followed his equation (B.) which is:
Q
k
=
k
k
(N k)
N
k
N
k
(2)
whence
N log
10
Q = N log
10
k

k
k
(N k)
N
k
N
k
(3)
However in faithfully replicating the table, I found additional adjustments were necessary
such that
Nlog
10
Q 10 + Nlog
10
Q (4)
Journal of Statistical Software 5
Figure 4: Values of NlogQ for N and k, per Gould paper.
and
Nlog
10
Q < 0.05 Nlog
10
Q 100 + Nlog
10
Q (5)
The reason for this is not entirely self-evident. An illustration of the function is shown in
Figure 4, which may be manipulated, if desired, with NlogQLimits.R:
4. Original paper
Peirce himself appears to have been aware of the diculties in interpreting his orininal paper.
I perceive that the theory of my criterion has been frequently misunderstood.
I presume this to be due in a great degree to the conciseness of the argument with
which it was published.
Peirce (1877)
However this did not prevent the methods becoming widely adopted in his own time, largely
due to the clarity of Goulds implementation. I had some diculty in replicating all of the
results from the original paper Peirce (1852). However a number of formulas are of interest.
6 Peirce
Figure 5: Probability of an error occurring, varying by mean error of system.
(These functions with corresponding plots are included in Pierce1852.R). His expression for
the probability of certain error in a system with mean error is given by:
() =
1

2
e


2
2
2
(6)
This is illustrated for a number of sample ranges of interest in Figure 5:
His next formula gives the probability of an error in a system which exceeds the required
limit, x:
(x) =
2


x
e

1
2
x
2
(7)
The reader may recognize as closely related to the complementary error function, erfc,
which is already implemented in R as NORMT3::erfc:
erfc(x) =
2


x
e
t
2
dt (8)
A comparison of both is shown in Figure 6.
His next equation for the probability of k observations exceeding the required limit x I took
to be:
P =

(x)
(x)

k
(9)
Journal of Statistical Software 7
Figure 6: Comparison of Peirces with erfc.
shows probability varying by ratio of limit of acceptable error to mean error.
whereby he derives:
P =
1

Nk
2
N
k
2
e
N+m+kx
2
2
(x)
k
(10)
However substituting arbitrary values in both lead to values of > 1. Again, the reason for
this is not entirely self-evident to me.
5. Conclusion
Peirce deserves credit as the rst to suggest a method of excluding outliers. Given the preva-
lence of normal distributions in routine observations, a revival of his methods may be timely.
I hope these illustrations will clarify the range over which his methods may be applied.
6. Acknowledgements
Knud Thomsen for sharing the method in C. Kevin Mullin for converting the method to R.
References
Gould BA (1855). On Peirces Criterion for the Rejection of Doubtful Observations, with
Tables for Facilitating its Application. Astronomical Journal, 4(83), 8187. doi:10.1086/
100480. URL http://adsabs.harvard.edu/abs/1855AJ......4...81G.
Natrella M (2012). NIST/SEMATECH e-Handbook of Statistical Methods. URL http:
//www.itl.nist.gov/div898/handbook/index.htm.
Peirce B (1852). Criterion for the Rejection of Doubtful Observations. The As-
tronomical Journal, 2(45), 161163. URL http://articles.adsabs.harvard.edu/
8 Peirce
cgi-bin/nph-iarticle_query?1852AJ......2..161P;data_type=PDF_HIGHhttp:
//adsabs.harvard.edu/full/1852AJ......2..161P.
Peirce B (1877). On Peirces Criterion. Proceedings of the American Academy of Arts and
Sciences, 13, 348351. URL http://www.jstor.org/stable/25138498.
Ross S (2003). Peirces Criterion for the Elimination of Suspect Experimental Data. Journal
of Engineering Technology, 20(2), 112. URL http://classes.engineering.wustl.edu/
2009/fall/che473/handouts/OutlierRejection.pdf.
Team RDC (2012). R: A Language and Environment for Statistical Computing. R Foun-
dation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0, URL http:
//www.R-project.org/.
Aliation:
Christopher Dardis
Department of Neurology
Barrow Neurological Institute
350 W. Thomas Road
Phoenix, AZ 85013
E-mail: [email protected]
URL: https://christopherdardis.wordpress.com/
Journal of Statistical Software http://www.jstatsoft.org/
published by the American Statistical Association http://www.amstat.org/
Volume VV, Issue II Submitted: yyyy-mm-dd
MMMMMM YYYY Accepted: yyyy-mm-dd

You might also like