2287637
TRANSCRIPT
-
7/25/2019 2287637
1/11
Taylor & Francis, Ltd. and American Statistical Association are collaborating with JSTOR to digitize, preserve and extendaccess to Journal of the American Statistical Association.
http://www.jstor.org
Benjamin Peirce and the Howland WillAuthor(s): Paul Meier and Sandy ZabellSource: Journal of the American Statistical Association, Vol. 75, No. 371 (Sep., 1980), pp. 497-506
Published by: on behalf of theTaylor & Francis, Ltd. American Statistical AssociationStable URL: http://www.jstor.org/stable/2287637Accessed: 03-08-2015 15:08 UTC
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of contentin a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.For more information about JSTOR, please contact [email protected].
This content downloaded from 200.12.134.94 on Mon, 03 Aug 2015 15:08:46 UTCAll use subject to JSTOR Terms and Conditions
http://www.jstor.org/http://www.jstor.org/action/showPublisher?publisherCode=taylorfrancishttp://www.jstor.org/action/showPublisher?publisherCode=astatahttp://www.jstor.org/stable/2287637http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/stable/2287637http://www.jstor.org/action/showPublisher?publisherCode=astatahttp://www.jstor.org/action/showPublisher?publisherCode=taylorfrancishttp://www.jstor.org/ -
7/25/2019 2287637
2/11
Benamin
Peirce
nd
the
HowlandWil
PAUL
MEIER nd
SANDY ZABELL*
The Howlandwill ase is possibly he earliest nstance nAmerican
law of the use of probabilisticnd statistical vidence. dentifying
30 downstrokes
n
thesignaturefSylviaAnnHowland,
Benjamin
Peirce ttempted o showthat a contested ignature n a will had
been
tracedfrom nother nd genuine ignature. e argued hat
their greement
n
all 30 downstrokes as improbablen the ex-
treme under a binomialmodel. Peirce supportedhis model by
providing graphical est of goodness f fit.We give a critique
of Peirce's model and discussthe use and abuse of the product
rule formultiplyingrobabilitiesf ndependent
vents.
KEY
WORDS: Charles S. Peirce; 19th-century athematical
statistics; aw and statistics; andwritingnalysis; ndependence;
Product ule.
1. INTRODUCTION
On
July , 1865, Sylvia Ann Howland of New Bedford,
Massachusetts, died, leaving an estate of $2,025,000,
and
a single heir-her niece Hetty H. Robinson,
who
had lived with her for ome years. Two milliondollars s
a
respectable sum, even by today's inflated tandards,
and
in
1865 it was certainly prize worth ontesting.
The Howland will, dated September 1, 1863,
and a
codicil of November 18, 1864, lefthalf the estate to
a
number of individuals and institutions. t provided
that the residuewas to be held
in
trustfor he benefit
f
Hetty Robinson and, at the time
of
Hetty's death, to
be
distributed o
the
lineal descendents f Hetty's paternal
grandfather,Gideon Howland, Sr. Hetty's father had
died
a month arlier, eaving her $910,000 outright
nd
the income from
trust of $5,000,000. AlthoughHetty
Robinson was thus already a wealthy woman,
she
claimed
a
right to inheritoutright
he
entire
estate
of
her
aunt, providing
s
evidence forthis claim an earlier
will,dated January11, 1862,which eft he entire state
to
Hetty with nstructions hat
no
later
will was
to
be
honored.
Family feuds are relevanthere,
as is
a
claimed
agree-
ment
between Hetty
and her
aunt to leave mutual
wills
excludingHetty's
father.
For
our
purposes,
t
is
enough
to observe that Hetty Robinsonhad an arguable claim
*
Paul Meier
s
ProfessorfStatistics, he University
f
Chicago,
Chicago,
L
60637,
and
Sandy
Zabell
is
Associate
Professor f
Mathematics,
Northwestern
niversity, vanston,
IL
60201.
Support
for this
research
was
provided
n
part by
NSF Grant
SOC
76-80389.
The
authors re grateful
o
a
number
f friends
and
colleagues
with
whom
hey
discussed
spects
of
this
case-
P.
Diaconis, S. Fienberg, . Langbein,
P.
McCullagh,
J.
Tukey,
and H.
Zeisel, among
others.
Special
thanks re due
to David
Wallace
for
xtensive iscussion
oncerning
he inference
roblem
that
arose, nd Stephen tigler orhelpful orrespondence
n
the
historical ackground.he staff
f
the
New Bedford
ublic
Library
were
especially ccommodating
n
locating
nd
arranging
orthe
copying
f
unique source documents nd exhibits,
nd
to
them
the authors
xpress heir pecialgratitude.
that the earlierwill should be recognized.But this the
Executor,
Thomas Mandell, declined to do. Not only
did he claim
that the later will governed,but he
also
contended
hat two of the three ignatures n the
earlier
will were tracings from the
third, thus providingan
independent ground for
invalidating that will (see
Appendix 1).
Hetty Robinson
retained
distinguished ounsel and
sued.'
Thomas Mandel retained equally distinguished
counsel and defended.
Between them, both sides called
as
witnesses ome
of
the most
prominent cademicians
of
the day. Oliver
Wendell
Holmes, Sr.,
Parkman
Pro-
fessor fAnatomy nd Physiologyn theMedical School
of
Harvard
University,
xamined
the
contested signa-
tures
under
a
microscope,
s did Louis
Agassiz, and
testified for Robinson) that
he could find no
evidence
of
pencil
marks
uch
as
mighthave
appeared
in
tracing.
(The
cross-examination f ProfessorHolmes was
both
brief nd
painful;
ee
Appendix
2.) Among
those
testify-
ing for he Executor were
Benjamin Peirce, Professor f
Mathematics at Harvard, and his
son,
Charles
Sanders
Peirce, who
was later to do importantwork in several
branches
of
philosophy.
The Peirces were
among the
first o contribute o the development f mathematical
statistics
n
the
United States;
a
recent urvey of their
work n thisarea is given n Stigler 1978, pp.
244-251).
Professor Peirce
undertook
to
demonstrate by
sta-
tistical
means
that the disputed ignatures n the second
page of the
will
were indeed forgeries.
is
method
was
to
contrast he similarities etween
one of the
disputed
signatures
referred
o
in
the trial record
as
signature
#10) and
the unquestioned original (referredto
as
signature# 1) with the lesser
degree
of
similarity
o be
found
n
pairing
42
other
signaturespenned by Sylvia
Ann
Howland
in
her
later years
on
other
documents.
(See Figures
A
and
B.)
It
is
this analysis
with
which
we shall be concerned.
The Peirceanalysis, s weshallsee, sat once ngenious
and in
some
respects naive. It
leads
to
a
distributional
problem, discussed
in
Appendix
4,
which is
nontrivial
and has some
nterest
n
its
own
right.
he data
available
1December, 865. Testimonywas taken n 1867, decision
ren-
dered October 1868 (Robinson . Mandell, 20 Fed.
Cas. 1027).
A copyofthe trial ranscripts available n theNew Bedford ublic
Library nd the testimonyn the text and Appendixes and 3
is
extracted rom hat source.
? Journal
f the
American
Statistical
Association
September
1980,Volume
75, Number
371
Applications
Section
497
This content downloaded from 200.12.134.94 on Mon, 03 Aug 2015 15:08:46 UTCAll use subject to JSTOR Terms and Conditions
http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp -
7/25/2019 2287637
3/11
498
Journal
f the American
tatistical
ssociation,
eptember
980
A.
The
Unquestioned
riginal
ignature
no.
1)
and
the
Two
Disputed
Signatures
nos.
10
and
15)
in
the record
is
sufficient o allow
some
interesting
analyses,
but
tantalizing
or
ts failure o include
details
of
major
relevance for the
problem
at issue.
However,
as weproceed, tis welltokeep nmind hat no organized
theory
of statistical nference
xisted at the
time,
and
that,
for
example,
casual
imputations
of
independence
weremade almost
utomatically y
those
eeking
o
bring
considerationsof
probability
to bear on
problems
of
human
affairs.
We
should
not
be
surprised,
herefore,
o
find
uch
mputations
s a
matter
f
course
n
the
present
problem.
2.
THE
PROBLEM
Let
us
start
with
the
evidence
as
given,
making
assumptions s
needed, nd returnater to judge which
seem
plausible and
which
not. The
evidence
consisted
of
photographic
opies
of 42
signatures,
n
addition
to
the
original
#1)
and one
of the
purported
opies
(#
10).
Suppose
we
are
given
some
index of
agreement be-
tween
signatures. n the
present
case
Professor
Peirce
chose the
numberof
coincidences-in
length and
posi-
tion-among
the
30
downstrokes
n each
of
a
pair
of
signatures,
when
the two
were
uperimposed.We
might
then et
Xi,
.. .,
X42
be the
values of
that
ndex
obtained
in
matching
ach
of
the
42
uncontested
nd
uninvolved
signatures
with
#1,
and
let
Y
be
the
value
of
that
index
in
matching#10 with
#
1.
We
could
then
pose the
prob-
lem as one of udgingwhetherY could reasonably have
been drawn
at random from
he
same
population
that
gave
rise
to
the
sample
of
size 42.
Since,
n
this
case,
Y
=
30,
that
s, #
1
and
#10
coin-
cide
in
all
30
downstrokes,
nd
the
Xi
are
presumably
B. Some of
the
42
Comparison ignatures
J
,
49
od- Zn9Sy0,f Z.o{e
4
O
t.^:Fg.v..*~~~~4-
,-.fA
w
,72
?o
At KX
'
r4
tjtt..
o.
4s
r
'7
4.
4/
7/
i14-L
p0r.zzo.
jy?
2/6to
. /is
7$6./CA
r%,r
? *; ;- ie1*
;
f
,
4
0
0
.,,,ta'47'
/
LW4zEtv;
A1
~ ~ ~ ~ ~ ~ ~ ~ ~
A
.
/K.
*rt
-
7/25/2019 2287637
4/11
Meier
nd Zabell:
Benjamin
eirce
nd
the Howland
Will
499
all smaller
han 30,
we might ssign
a significance
rob-
ability
of
1/43
=
2.3%
to
this outcome.
This
is
a small
enoughprobability o
raise suspicion,
but far
from ver-
whelming.2
If we had
available the individual
Xi
values
(which
we do not)
we might
e willing o
proceed little
more parametrically.
We
can determine see
Table 1)
that
the
average
index in
random pairings
of the 42
signatures s near 6 with a standarddeviationof about
2.5.
Thus
Y is nearly 10 sample
standard
deviations
away from
X
and,
if
we believe
that the
population
sampled
is
at least approximately
normal, the
dis-
crepancy
would
be
udged
significantt
an extreme evel.)
Professor eirce
proceeded
n a different
ay and came
to
the conclusion hat
the odds against
such
a coincidence
as that
between ignatures
#
1
and #
10
were stronomical
(of
order
1020). To see how he
arrived at this
striking
result,
et
us
follow
the
order of
testimony
n
the
case.
2.1 Testimony
f Charles
anders
eirce
The son, Charles Sanders Peirce-then a memberof
the
staffof the
United States Coast
Survey-testified
on
the work he carried out under
the directionof
his
father.
Charles
was provided with
photographic opies
of
signatures
#1
and
#10 (the
third
signature
s
not
mentioned
n
the testimony)
and with copies of 42
signatures
of
Sylvia
Ann Howland on leases
and
other
documents.Following
nstructions
iven
to
him
by
his
father,
Charles undertook
o
match every possible
pair
of
the
42
uncontested
signatures
and
to
observe
the
agreement
etween
ach
pair
n
respect
o
the
number ut
of
30
identified downstrokes
n which
they
were
in
agreement.There being ( ) = 861 possible pairs, and
30 downstrokes
o
be checked for each,
this meant
a
total of
25,830 comparisons,
or
which Charles
reported
agreement
in
5,325
or
20.6
percent
of them-very
nearly
1
in 5. More particularly,
e
gave
the
distribu-
tion
of number
f
agreements
s
shown n
Table
1.
Thus,
for
xample,
15
pairs
each had twocoincidences ut
of
30,
for total
of
2
X
15
=
30 strokes
oinciding.
The
precise
details
of the
matching rocess
nd of
the
criteria for
judgment
of
agreement
were
only briefly
touched
on
in
C.S.
Peirce's
testimony:
Q.
Whatwas
the standard
..
of coincidence?
A. I
considered
hat
two
ines
coincidedwhenever
hey
coin-
cided as well as .
.
. nos. 1 and 10, in those same lines.
Later,
under
cross-examination,
harles
added
some
further etails:
I
made
it
a
condition that
in
shifting
one
photograph
over
another
n
order
o
make as
many
ines s
possible
oincide,-
2
One
might ry
to
improve
n this a
little
by
considering
he
probability
hat
out of 44 signatures, given
pair
will have
the
greatest
ndex
of coincidences,
hat is,
(4Y4)1
=
.0011. (Strictly
speaking,
we
are
not
given
the
number
f coincidences
etween
either
ignature
#1
or
#10
on the
one
hand,
and
any
of the 42
other ignatures
n the other, ut
presumably one
of these
even
approaches
he 30
of #1 and #10.) The appropriateness
f
such
a significanceest is an interestingssue, but one that we will
not touch
on.
1. C. S. Peirce's Table of Coincidences
Coincidences Number
of Pairs Number of
Strokes
?l0
0
2
15
30
3
97 291
4
131
524
5 147 735
6 143 858
7 99
693
8 88
704
9
55
495
10
34
340
11
17 187
12 15 180
13
20 288
30_
861 5325
that the line of writinghould not be materially hanged.
By materially
hanged, mean so much
changed hat there
couldbe
no
question hattherewas a difference
n the general
direction f he two
ignatures.
2.2 Testimony
f Benjamin eirce
C.S.
Peirce
was
followed o the stand
by his father,
who testified s to his own
examination nd comparison
of
#
1
and
#10.
The coincidences extraordinary
nd of such kind s
ir-
resistably
o suggest esign,nd especially
hetracingf
10
over
1.
There re smalldifferences
. . butthe differences
re
such
as
to
strengthen he argument fordesign.
.
.
.
The
co-
incidences
re manifestly
hose of position; while the
dif-
ferences re thoseof form.Coincidence f position s easily
effected y design, nd can be tolerably
well accomplished
by
an unskilful
and.
t
especiallybelongs
o
the
downward
stroke.
. .
. It involves a
close agreement n the average posi-
tion of the
stroke,
n
agreement
n its slant,
nd no extrava-
gant
diversity
f
curvature.
ut
coincidence
f form s ex-
ceedingly
ifficult....
In
a
long ignature. omplete oincidence
fposition
s
an
infallible
vidence
f
design.
The mathematical
iscussion
of
this
subject has never, o my knowledge,
een proposed,
but it is
not
difficult;
nd a numerical xpression pplicable
to
this
problem, he correctness
f which
would
be instantly
recognized
y all the mathematiciansf
the world, an be
readily btained.
ProfessorPeirce
went on to describe
his
son's
work
and, without ualification, sserted hatsince theoverall
proportion
of matches was very
nearly one-fifth, he
probability
of finding
30 matches
in
a
given pair
of
signatures
was
one
in 530
or
once
in
2,666
millions
of
millionsof
millions.
(Note:
There
is some error
here,
since 530
=
9.31
X 1020.)
Lest the
point
be
missed,
Peirce
added:
This
number
ar
transcends uman
experience.
o
vast an
improbability
s
practically
n
impossibility.
uch evanescent
shadows
f
probability
annot
belong
o actual ife.
They
are
unimaginablyess
han hose east
hings
hich he
aw cares
notfor.
When asked
the purpose of his son's display of
the
distribution f the number of coincidences,Peirce ex-
This content downloaded from 200.12.134.94 on Mon, 03 Aug 2015 15:08:46 UTCAll use subject to JSTOR Terms and Conditions
http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp -
7/25/2019 2287637
5/11
500 Journal
f
the Americantatistical
ssociation,eptember980
2. BenjaminPeirce's Table (and
chi-squared omputations)
Number of
Agreements Obs. Exp. (O
-
E)
(O
-
E)2IE
? l?0 9.1 -9.1 9.1
2 15 29.0 -14.0 6.7
3
97
67.6
29.4
12.8
4 131 114.1 16.9 2.5
5
147
148.3 -1.3 0.0
6 143 154.5 -11.5 0.9
7
99 132.4 -33.4 8.4
8 88 95.2 -7.2 0.5
9
55
58.2 -3.2 0.2
10 34 30.5 3.5 0.4
11 17 13.9 3.1 0.7
12
15 5.5 9.5 16.4
13
20 2.7 17.3
112.1
30J
861 861 0.0 170.7
plained that
It was orderedby
me
in
order
to
obtain
as severe
a
test
as
possible
from
bservation
or
the
criticism f
my
mathe-
maticaldeductions.
To facilitate comparison,
Peirce
provided
a
table,
recapitulatinghis son's
counts,
together
with the
ex-
pected number f counts
under
the
binomial model
/30\ 1
i4 30-j
El
Yj}
=
861
(
3)(1)()
(See Table 2. The table given by Peirce omitsthe 15
cases with two or fewer
coincidences,
nd
the
20 with
more
than
12.)3
Peirce
commented:
This is
a species of comparison amiliar
o mathematicians
and such
a
coincidence
s above
would
be
regarded
s
quite
extraordinary,
nd a
sufficient
emonstrationhat the true
law
of
the case was embodied
n
the
formula.
A
graphical display
of
the comparison
was
also
pro-
vided,
which
s
reconstructed
n
Figure
C
according
to
the
descriptiongiven
in the trial record
(the
actual
display
was
omittedfrom he trial record).
3.
CRITIQUE
OF PEIRCE'SMODEL
3.1 Goodness
f Fit
How
convincing
s
Peirce's
model
and his
evidence
for it? No
formal
theory
of
goodness-of-fitesting
ex-
istedat the time,and Peirce can scarcelybe faultedfor
failing
to
provide
one.
Still,
the quality
of fit
can
rea-
sonably be questioned. Peirce's graph
shows
only
the
middle
range,
and
even
here there
appears
to
be
a
con-
sistent
rend n the
deviations.What the graph
does not
show
s the
substantial bsence of really poormatches-
a The table n the trialrecord ives41 cases as undistributed,
but it is clear from eirce's estimonymmediatelyollowinghat
35 was meant. My son had also 35 undistributedases
...
no cases of 0 or 1
agreement,
with
9 expected-and
the
considerable excess of outstandingly ood
matches-20
(with 13 or more
greements),
with ess than 3 expected.
If
we calculate the
usual
goodness-of-fithi-squared
for this
frequency
able we
get X2
=
170.7 on 12
df.
We can
improve
this a little
by choosingp
to
minimize
chi-squared,
but the minimizing p
=
.211
actually
worsens things n the middle range of the graph and
gives
a xI
=
135.3,
which
remains n
exceedingly oor
fit. (The appropriateness
f the
chi-squared
tatistic s
discussed
n
Sec.
3.2.)
If one
compares
Peirce's
graphical
test of fit with
the
chi-squared
statistic,
t
is
obvious
that
the visual
impact
of the
graphical display
is to
weight
ll
deviations
from he
expected values
equally,
while
chi-squared weights
such
deviations relative to
the
magnitude
of each
expected
value.
For
this
reason,
discrepancies
etween the
observed
and
expected
values
in
the
tails,
even
if
they
had been
plotted,
might not
have appeared
to be more
serious than
those in
the
middleofthe distribution, hile t is precisely he con-
tribution
of the
tails
to
chi-squared
that
makes the
latter ignificant.
Clearly
the most
frustratingspect
of
this
data set is
its
lack
of
detail with
regard to
the 20
good matches.
Were even
one of them to show
30
agreements,
r
even
29,
the persuasiveness
of
the mathematical
evidence
would evaporate.
Indeed,
as
shown
n
the earlier
table,
we
can
calculate
from
the data
given
that
the total
number
of agreements
n
these 20 cases was
288,
or
an
average of
14.4
for
those
signatures
with 13 or more
agreements.
hus,
although
one
or two
very highvalues
remains numerical
possibility,
t seems
likely
that few
ifany values were as high s 20.
Unfortunately
for
the
record,
Professor Peirce's
demeanor
nd
reputation
s
a
mathematician
must have
been
sufficiently
ntimidating
to
deter
any
serious
mathematical
ebuttal.He
was made
to confess lack
of
any general xpertise
n
udginghandwriting,
ut
he was
C.
Distribution
f Coincidences
FREQUE14CY
200
---
OBSERVED
FREQUENCY
EXPECTED
FREQUENCY
150-
00- S 0 1 0 5
I~~~
~NME OF GREMNT
This content downloaded from 200.12.134.94 on Mon, 03 Aug 2015 15:08:46 UTCAll use subject to JSTOR Terms and Conditions
http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp -
7/25/2019 2287637
6/11
Meier nd
Zabell:
Benjamin eirce
nd
the Howland
Will
501
not cross-examined
t all on the
numerical nd mathe-
matical
parts of
his
testimony,
et alone made to
list the
20 high values
(see Appendix3).
3.2
Model Assumptions
The frequent aralysis
of the law before demonstra-
tion of mathematicshas long been noted. An extreme
view,
expressedby
Laurence
Tribe
in
his 1971 Harvard
Law
Review rticle Trial by Mathematics
(Tribe
1971)
is that laymen
are too readily mpressed nd misled
by
such
presentations
nd that they should
therefore e
excluded-at
least in jury trials.
Whatever one
thinks
of this
position
as a
generality, eirce's
model, as
well
as his data analysis,
deserves closer
examination han it
receivedduring
he trial.
One might,
for example, question the
assumption-
required
for the binomial model-that
the probability
of match houldbe the
same at each of
the 30 positions.
One can easily imagine that certain strokes-perhaps
those early
in the signature-would
have a
higher
probability
f matching han do
others.The assumption
of
equal probability
could readily
have been
checked
in the course
of C.S. Peirce's work,but
it seems never
to
have
been questioned. (Of course,
the effect f such
inequalitywould
be to reduce he
variance n the
number
of matches, and thus to
make the probability
of 30
matches ven smaller
han that quoted.)
Far
more serious
is
the implicit
ssumption of inde-
pendence.
One would expect
that agreement
n, say,
positions
1
and 3, would
make
far
more ikely
n
agree-
ment
n position
2.
The effect
f such dependence
would
be to increase he variance and thus increase the prob-
ability of 30 matches
over that quoted,
quite possibly
by
orders of
magnitude.
To the extent
that the
data
bears on this
point,
t does
indeed
suggest
he
possibility
of dependence by exhibiting
what appear to
be
far
too
many cases with over
12
matches
to agree
with the
binomial model.
A
final
critique
of Peirce's
discussion
s
the absence
of
any comment
on correlation
between signatures
made
close
in
time. It
is
at
least plausible
that signatures
ate
in
life
would differ ubstantially
from hose
made at
a
younger
ge,
and also
plausible
that
signatures
made at
a
single sittingwould be more similar than those made
days
or weeks
apart. (Similar
problems
arise
in
the
statistical
analysis of
authorship,
where both style
and
vocabulary
may depend
on the writer's
ubject
matter
and
may vary over
his
ifetime;
ee Mosteller
nd
Wallace
1964,
pp. 18-22, 195-199,
265-266.)
Since many
of
the
42
signatures
sed
in
the study
were on leases
and other
dated
documents,
this
question
could
also
have
been
examined,
ut there
s no
evidence
hat t
was considered.
In
part
these challenges o
the fairness f
Peirce's
test
can
be
met by the nonparametric
rocedure
uggested
earlier.The questions
of uniform robability f
matching
in each of the 30 positions, nd
independencebetween
them,
would then be
irrelevrant,.
The
problem of time
dependence,on the other
hand, remains n full
force.)
Thattest,
however, s noted
before, ieldsonly
strongly
suggestive
significance
robability,but nothing
at
all
as
compelling s Peirce's
calculation.
For all that one
can raise questions about
Peirce's
model
and
analysis, t remains
o consider he extent o
which
the evidence available
really contradicts
t. Al-
thoughthe conventional hi-squared est clearlyrejects
the binomial
model, t
is
not
at
all
clear that the chi-
squared
statistic should have
even
approximately he
familiar
distribution.Under
Peirce' model the
prob-
ability that a random
pair of true
signatures should
match in
k
stokes is indeed binomial,
as
claimed,
and
thus
the
expected fraction f
pairs with
k
agreements s
also binomial.
However, the (42)
=
861 pairs are
gener-
ated from nly 42
signatures nd are not
at all the same
as
861 independent rials
from binomial
population. t
is to be
expected, herefore,hat
the relative
frequencies
will
show variability
greater han
multinomial nd that
the chi-squared statistic will also exhibit greater vari-
ability. To
assess more
rigorously he possible
lack of
fit
of
Peirce's
model
we
turn
to
considerationof the
rather
arge excess of
pairs with over
12
matches (20
observed,
only 2.7 expected). The
Markov
inequality
(e.g.,
see
Billingsley1979, p. 65), of
which
the familiar
Chebyshev inequality
is
a direct
corollary, tates
that
for
ny nonnegative andom
variableZ,
one has
Pr{Z
>
t} < t-'E{Z}
.
Taking
Z
=
the number
f pairs
out
of the
861
showing
13
or
more
agreements,
nd
t
=
20, we
find
PrIZ
>
t}