You are on page 1of 22

Society for American Archaeology

Spatial Analysis: Multiple Procedures in Pattern Recognition Studies


Author(s): Harold J. Hietala and Dominique E. Stevens
Source: American Antiquity, Vol. 42, No. 4 (Oct., 1977), pp. 539-559
Published by: Society for American Archaeology
Stable URL: http://www.jstor.org/stable/278927
Accessed: 03/09/2009 14:16
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=sam.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the
scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that
promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.

Society for American Archaeology is collaborating with JSTOR to digitize, preserve and extend access to
American Antiquity.

http://www.jstor.org

SPATIAL ANALYSIS: MULTIPLEPROCEDURES


IN PATTERN RECOGNITIONSTUDIES

HAROLD J. ItETALA
DOMINIQUE E. STEVENS
The analysis of the spatial distributions of artifacts on archaeological floors has potential for the recognition
and interpretation of cultural patterns represented by prehistoric remains. This study emphasizes the use of
multiple procedures for the separate problems of pattern detection, estimation, and prediction. Aggregative and
segregative definitions are proposed for three artifact class associational modes. Definitions are also proposed for
general and special pattern prediction problems. An application to the Mousterian site of Rosh Ein Mor in Israel,
shows that multiple procedures can be extremely advantageous in the interpretation of general complex patterns.

The analysis of the spatial distributionsof artifacts on archaeologicalfloors has potential for
the recognition and interpretationof cultural patterns representedby prehistoricremains. The
primary source of analytical procedures has been the literature in ecology. Thompson (1955,
1956, 1958), Kershaw (1961), Greig-Smith(1952, 1961, 1964), Morisita (1962), and Pielou
(1959, 1960, 1969) have heavily influenced such archaeologicalapplicationsas Dacey (1963,
1973) and Whallon(1972, 1973, 1974). Although these applicationsare innovative,they should,
nevertheless,be judged relative to their potentials for identifying, measuring,and predictingthe
strength of artifact patternsin space. This paper is restrictedto excavationgridunit applications,
although the ideas are extendable, at least in principle, to sites where each artifact is mapped in
place. No attempt is made to assert that the proceduressuggestedin this paperprovidea panacea
for spatial analysis problems, but rather that many alternate procedures may be potentially
applicable.
The suggested procedures fall into three conceptual areas of methodological investigation,
correspondingto the problems of pattern detection, pattern estimation, and pattern prediction.
While some of the methods are not carriedout to their logical limits, the directionof the approach
is clear. The arguments and examples are written for applied archaeologists,as well as for
theoreticalspecialists.
Methodology

Currentlyaccepted spatial analysis methodologiesin archaeologyfocus on patterndetection as


a prerequisitefor associationalestimationwhile ignoringpatternconsistency and prediction.
PATTERN DETECTION

Some investigators,e.g., Whallon(1973) and Dacey (1973), have stated that the detection of
nonrandomspatialpatternsfor individualartifactclassesis a prerequisitefor studyingartifactclass
associations in space. Although the idea is philosophically sound, the essential unsoundnessof
potential applicationsmay be seen through a theoreticalexample. Specifically,the variance/mean
ratio statistic, based upon the Poisson distributionutilized by Dacey (1973), ignores the overall
spatial provenience of the data base. Tlle insensitivityof this statistic to nonrandompatternsis
exemplified through the example spatially portrayedin Fig. 1. The variance/meanratio statistic
equals 0.76 in both cases, A and B. Under the hypothesis of random patterning,this statistic is
distributedas a chi-squarevariate dividedby its degreesof freedom. The d.f.=15 and thus, in both
cases, the hypothesis of randompatterningis accepted with a type 1 errorvalue as largeas 0.75. It
is clear, however, that the failure to detect patterning in Case A is due to ignoringthe spatial
provenienceof the data base. Haberman(1974) has recently introduceda statisticalprocedurethat
may help with this problem.
Theoretically,it is possible for a group of people to "simultaneouslydrop"two distinct artifact
types regularlyand systematically so that each type appearsto be randomly distributed,but so
539

[Vol. 42,

AMERICAN ANTIQUITY

540

No. 4,

1977

that the two types are positivelyassociatedin space. This associationwould not be detected under
the presumption that two artifact classes cannot be spatially associated if they appear to be
randomlydistributed.

______

longitude

lonqiitude

15

14

12

13

12

11

12

10

14

15

11

10

10

1 12

ci

0o

.Th 11

10

4)1

L9

8
Case

i-4:
4J

1 131

Case

11

Fig. 1. Theoretical sample distributions for the pattern detection problem.

PATTERN ESTIMATION

The procedurediscussedby Whallon(1973) requiresthe detection of nonrandompatternsand


determinationof optimalblock size as preconditionsfor measuringcoassociationalpatterning.The
common block size requirementis inadequateto deal with the situation in which there are vast
differentialsbetween the frequenciesof artifact classes. More importantis the situation of major
differences in cluster unit sizes and shapes for the individualartifact classes. Whallon(1972)
briefly consideredthis point.
The first pitfall may be noted by consideringtwo artifactclassesthat are so differentin relative
frequenciesthat a largerblock size would be requiredfor the lesserfrequencyclass simply because
fewer artifacts are available for block size determination,while a smaller block size would be
potentiallyindicatedfor the largerfrequencyclassbecauseof its greaterartifactdensity.
The second pitfall is illustrated by considering two artifact classes that have many cluster
regions of different sizes and shapes (Fig.2). Although these two hypothetical distributions
indicate some segregation, there is no reason why they could not strongly overlap. In such a
situation,the optimal block sizes for classesA and B would not be compatible.
Apart from considerations of nonrandomness and common block size, there are some
fundamentalproblems connected with determiningmeasuresof association for the situations of
(mutual) aggregation or (mutual) segregation of artifact classes in space. The predominant
weaknesswith such measurements,in ecological and archaeologicalliterature,has been the failure
to distinguishand exploit in a well defined mannerdifferingassociational"strengths."Threelevels
of associational strength, which may prove useful in the general interpretationof differential
spatial distributions for distinct artifact classes, are defined in this paper. The following ideal
definitions of uniform, strong,and weak associationsare proposed.The weak definitionsallow for
marginal determinations of association, whereas the stronger forms require a large degree of
homogeneityor heterogeneityin the artifactclass cell probabilities.
In the specific definitions that follow, R denotes any arbitraryregion included within the
excavatedportion of the site, while R1, R2, . . ., Rk, collectively denote a specific partitionof the
excavated site into a mutually exclusive but exhaustive set of regions. The letters A, B, and C
denote distinct artifact classes, and the notation PC(Ri) denotes the theoreticalprobabilitythat a

YX

541

PATTERN RECOGNITION STUDIES

Hietala and Stevens]

X X
X

YXX

YY

XX

X X X .y
X X .jX_lLli_

XY

Y X ,_Y
yr

Fig. 2. Theoretical sample distribution profiles for the common "block size"
problem. The x entries represent high frequency cells for artifact class A while the y
entries represent the high frequency cells for artifact class B.

member of artifact class C is contained in the ith region, Ri. The standard definition of a randomly
distributed artifact class is utilized here: the probability that a member of the artifact class is
contained in R equals the area of R divided by the area of the excavated portion of the site. The
end of a definition is denoted by double periods.
Definition 1: Uniform Aggregation Two artifact classes, A and B, are said to be uniformly aggregating if,
for every region R,
PA(R) = PB(R) ..

(1)

A nearest neighbor analogue occurs if every artifact in class A has as its nearest neighbor, among
all members of classes A and B together, a member of class B, and vice versa. Two identical
distributions of artifacts would therefore imply uniform aggregation.
Uniform aggregation can be tested for a specific partition through the multidimensional contingency table techniques discussed by Hietala (1973), Fienberg (19702 and Goodman (1969). Different anthropological applications have been provided by Muller and Mayhall (1971), Reid (1974),
and Clark (1976).
Definition 2: Strong Aggregation
Two artifact classes, A and B, are said to be strongly aggregating for a specific partition if the probabilities
Rk-, are perfectly concordant. That is, for all i
for artifact class containment in the regions, R1, R2,
and j,
pA(Ri) > PA(Rj)

if and only if
PB(Ri) > PB(Rj)..

(2)

This form of strong aggregation can be measured by Goodman and Kruskal's gamma statistic
(1963:322 ff) or Kendall's tau statistic (1955).
It should be noted that even if strong aggregation occurs for every possible partition, uniform
aggregation is not implied. In this sense, strong aggregation is weaker than uniform aggregation

542

AMERICAN ANTIQUITY

[Vol. 42,

since, for a specific partition, uniform aggregation requires homogeneity


probabilities, whereas strong aggregation only requires concordant probabilities.

No. 4,

1977

of the region

Definition 3: Weak Aggregation


Two artifact classes, A and B, are said to be weakly aggregating for a specific partition if the median
probabilities for the two classes are simultaneously smaller, or larger, than their respective region
PA(Rk)) and similarly for PB(.50), then for every
probabilities, that is, if PA(G50) = median (PA(Rl),...,
i,

PA(Ri) > PA(-50)


if and only if
PB(Ri) > PBO O) . .

(3)

This form of weak aggregation can be investigated by median split associational procedures
suggested by Pielou (1969).
Definition 4: Weak Segregation
Weak segregation of two artifact classes, A and B, is defined by substituting relation (3) by
PA(Ri) > PA(.S0)
if and only if
PB(Ri) 6 PB O) ..

(4)

Weak segregation can be investigated by the same procedures utilized for weak aggregation.
Definition 5: Strong Segregation
Strong segregation of two artifact classes, A and B, is defined by substituting relation (2) by
PA(Ri) > PA(Rj)

if and only if
PB(Ri) < PB(Rj),

.5

Strong segregation can be measured by the same procedures utilized for strong aggregation.
Definition 6: Uniform Segregation
Two artifact classes A and B, are said to be uniformly segregating for a specific partition if, for every i,
PA(Ri) > 0
if and only if
PB(Ri) = 0 ..

(6)

A nearest neighbor analogue occurs if every artifact in class A has as its nearest neighbor, among
all members of classes A and B together, a member of Class A, and similarly for artifacts in Class B.
For a particular partition, uniform segregation is generally obvious or can be approximated by the
measure of strong segregation.
It should be stated that a posteriori definitions of the partition can induce uniform segregation.
For this reason, prior grid unit definitions are useful, if not necessary.
Figure 3 summarizes the proposed statistical techniques for measuring and testing associations
between artifact class distributions.
Ideally, for nonrandomly distributed artifact classes, the hypothesis of uniform association
implies the hypotheses of strong or weak association. Similarly, strong association implies weak
association. The reverse is not true. Schematically, this is shown in Fig. 4 with the continuous
arrow directions indicating potential causality.

Hietala and Stevens]

PATTERN RECOGNITION STUDIES

543

Kind of Association

Aggregation

Segregation

Multidimensional
Chi-square
statistic

Uniform
Uni
form

Strong

Tau

Weak

or

Median

Gamnma Statistic

Split

Statistics

Fig. 3. Summary of proposed associational estimation techniques.

Unif,0orm

Uniform

Aggregation
Cl)

Seqregation

Strong
Aggregation

Q-4

Strong
Segregation

C)
rd

Weak

Aggregation

Weak

<-

> Segregation

Fig. 4. Causality relations between the proposed definitions of association.

Independence,or indeterminancy,is a plausibleresult of the lack of segregationor aggregation.


This might be considered when the aggregationor segregationhypotheses are not acceptable.
Dacey (1973:320) gives a formaldefinition for the lack of spatialassociationbetween two artifact
classes that suffers the same operationalvaguenessas our criterion. The problem is not with the
concept but ratherwith its data operationalization.
The above definitions can be utilized to measurethe degree of associationalstrengthbetween
artifact classesfor a givenpartition.To judge the robustness(relativeto differinggridunit sizes) of
these associations,one should investigatethe resultsacrossvaryingpartitions.
PATTERN CONSISTENCY AND PREDICTION

If patterns can be described for a particulararchaeologicaloccurrence (Bishop and Clark


1967:894), then it may be reasonableto hypothesize the same patterns for other archaeological

544

AMERICAN ANTIQUITY

[Vol. 42,

No. 4,

1977

occurrences that fit into similar temporal, environmental, industrial, and functional frameworks. A
general case of pattern consistency corresponds to an hypothesis of locationally imprecise but
similar patterns of aggregation and segregation. This situation could apply to sites representative of
repeated occupations on a large but stable landsurface. A stronger case of pattern consistency
corresponds to an hypothesis of identical patterning in the same relative localities. This hypothesis
is more specific and could be applied to stratified sites. Should this latter hypothesis be supported,
considerable interpretive weight may be given to continuous, rather than repeated, occupation
when artifact accumulation is consistent through depth. There are intermediate situations, but the
possibilities are too many to consider in this introductory paper.
The above two cases of consistent patterns are defined as follows:
Definition 7: Consistent Generic Patterns
Two or more artifact classes are said to be consistently patterning in the generic sense across k distinct
archaeological occurrences if the same artifact class associations are simultaneously demonstrated for the
considered occurrences. .
Definition 8: Consistent Specific Patterns
Two or more artifact classes are said to be consistently patterning in the specific sense across k distinct
archaeological occurrences if the artifact classes show consistent generic patterning in the same relative
locations .

These definitions can be thought of as predictive hypotheses. For the strongest mode of
positive association, the hypotheses can be tested by multidimensional contingency table statistics.
The weakest mode of association can be modeled and tested to give support to the interpretation
of the stronger patterns.
All of the ideas discussed in this section are illustrated in the applications to the Mousterian site
of Rosh Ein Mor.

TheData Base:RoshEin Mor; A MiddlePaleolithicSite in Israel


Rosh Ein Mor is an open-air Mousterian site located in the western part of the Divshon Plain, in
the Avdat/Aqev area of the Central Negev, Israel. The site was excavated between 1970 and 1972
by a Southern Methodist University team under the direction of A. Marks (Marks 1971; Marks, et
al. 1971; Marks and Crew 1972). It is on the western edge of a deep box canyon, the Nahal Mor, at
an elevation of 495 m and approximately 70 m above the spring of Ein Avdat. The extent of the
site has been estimated at ca. 1,200 m2, of which 45 m2 were excavated in the center portion,
which exhibited a flat surface. The four percent sample is indicated topographically in Fig. 5a.
Horizontal units of 1 m2 (Fig. Sb) followed 5 cm ground-contour levels. Artifacts were mapped in
place by horizontal position within these levels, numbered from top to bottom (Crew 1976).
The sediments in which the artifacts are preserved are primarily finegrained, filling a very
shallow, elongated bedrock depression. The deposits vary from 55 cm to 1 m in thickness,
depending on the depth of bedrock. These sediments, averaging 80 cm in thickness, are remarkably
uniform. The vast majority of the artifacts within them were lying horizontally, and were fresh,
normally displaying sharp, uncrushed edges. The sediments are interpreted (Goldberg 1976) as
being primarily of alluvial origin, but sufficiently far from the main channel so that only silts, and
not coarse gravels were deposited. Low-energy run-off from the surrounding low-relief slopes
carried some fine-grained sediments and limestone fragments down onto the site. It is suggested,
however, that the artifacts have been essentially undisturbed since the time they were originally
abandoned.
All artifacts belong within the early Levantine Mousterian (Marks and Crew 1972; Crew 1976)
and are dated to greater than 50,000 years B.P., during a time when the Negev had a
Mediterranean environment (Horowitz 1976; Munday 1976). The typology of the artifact classes
followed Bordes, and overall typological and technological similarity was noted, both across and
between levels (Crew 1976). Thus, continuity may be seen in the aggradation of both the
geological and archaeological deposits.

Hietala and Stevens]

PATTERN RECOGNITION STUDIES

545

105,0-

104 5 -104.0- --103.5

--

102.5--

102.0
101.0

~~~~~~~~~~~~~~~~4

124)20.

97.0

Fig. 5. a Topographic map of the Rosh Ein Mor area; the polygon indicates the area of excavation; b the 45
square meter unit of the initial excavation; c the 1 X3 square meter units utilized in the initial analysis.

The artifact classes chosen for this analysis are endscrapers,sidescrapers,burins, Levallois
flakes, Levalloisblades, Levalloispoints, notches, denticulates,and borers.The percentageof tools
within these classes, relative to all tool classes, is approximately92%for all levels of the site. The
artifact class percentagesare also remarkablyconsistent between levels. For the levels analyzedin

AMERICAN ANTIQUITY

546

[Vol. 42,

No. 4,

1977

this paper, the maximum percentagedifference occurs for the Levalloispoint class, which varies
from a minimum of 18.0%in level 8 to a maximumof 25.7% in level 4 (see Fig. 12c). Levels 1
through 3 were excluded to eliminate possible historic surface disturbanceeffects, and levels 10
and below were excluded to eliminate the influence of bedrock,which first appearedin Level 10.
The limited faunal remainsat Rosh Ein Morconsisted of onager(Equushemionus)and ostrich
(Struthio camelus) egg shell fragments(Tchernov 1976). The latter were scatteredthroughoutthe
site, while the few onager fragmentswere quite restrictedin location (these are representedby
counts of fragmentsover 2.5 cm. in Fig. 6).

4
22
17
16
1 31
5 17
25
3 418

VP

EN

3
1
3

2 24
22
1

4
2
3

2 23
22
1
1 15

5
5
8

32
4
13

14
20
19

3
6

3
9

3 5
54
5 4

8
11.

3 11
5 11
4 2

6
3

14

73
83
56

10
8
6

55
5
54

2
4
1

2
2
6

1
2
3

3
5
3

8
2

7
5
12

12

~~~~1~

2
Os

VF

19
11
15

VB

11
7
12

7
6
6

5
57
6

4
2

2
3

32
53
2

6
49
11
4 68

13
9
18

311.

1
1 15
12
4

2
1
1

2
1
9

53
6
54

1
3
1

5
6
6

2
3
2

2
3
1

10
8 12
11 14 17
9
5 14

3
2

3
62
31

9
6

15
13
10

10 16
11 11
7
4

8
4
5

5
4
6

2
1
1

3
3
2

3
2

3
3
2

2
7
3

13
3
6

12

1
5
2

13
3
32

9
16
26

9
5
7

8691

2
4

2
13

1
2

8
5
3

6
3
2

4
6

3 47
2
3 11
53
7

8
11
4

8
45
52
47

8
8
12

24

2
1
4

11 21 11
15
8 16
8 17
13

13
3 4
11
5 19
14 14 12

16
4
2

4
3
4

13
33
33

16
11
11

4
2

8
4
1

9
25
53
11
13
64
9
17 10
5 12

3
4

16
7
8

12
3
1
4
4 25

4
2
5

1
14

5310
5 12
6614

14
12
8

11
8

13 13
11 14
13 16

7
46
46

1
1

16
13
314

11

18
6
11

14
5
9

4
2
4

23
16
43

21
5
2 17
14
6

2
29

3
3

1
2
2

5
3
5

1
1

18
12
13

15 17 17 18
9
6
9 17
15 17 25
9

8 17 16
7
14 23 14 14
17 19 23 14

12 14 13
6 14 25
9 14 19

17
813
85
10 14
6 15 13

9
711
14
8512
11 16 10

11 12 17
6
14
912
8
7 21 15 10

4
1

2 14
16
1 12

3
2
3

2 12
3 44
3
13

4
4

Fig. 6. Raw data frequencies by level.

2
1
213

5
8

1
2
1

3
3

2
3
3 32
33
4

21

1
1

2
4

10
11
9

3
4
3

2
3
4

Hietala and Stevens]

PATTERN RECOGNITION STUDIES

547

StatisticalConsiderationsand Example Calculations


Because of many statistically small artifact class frequencies in each 1 m2 grid unit, it was
necessaryto collapse these cells into largerquadrats.This was complicatedby the violation of total
rectangularformat; thus, one collapsingprocedurewasjudiciously performedas shown in Fig. 5c.
In this way, larger sample sizes were obtained for each quadrat, with the advantagesof equal
quadratsize and the possibility of both verticaland horizontal differentiation.The raw data, by
level, for these 12 quadratsare shown in Fig. 6. Another procedureinvolvedthe collapsingof the
45 m2 units (Fig. 5b) throughlevels, yieldingthe data shown in Fig. 7.

45

Square

meter

units

12

1x3

44

32

30

42

41

16

12

16

24

23

24

97

30

39

15

12

14

11

16

25

17

84

52

46

38

10

13

23

26

24

units

45

VP

16

VP13

29

Square

meter

units

12

13

13
10

11

61
13
6

26

11

58

34

31

73

12

17

11

14

11

14

B
3
5
11

4
3

31

22

38

11

12

18

30

18

26

10

12

20

11

36

16

21
2

28

4
1

33

28

61 G

41
19

33

33

29

29

21

31

22

25

32

34

23

41

29

30

31

29

38

33

58

24

42

31

20
27

34

22

21

19

116

15

20

VF

1
3

OS32

311 VF
30 26 33
15
22

units

OS

41
4

25-3371
28 26

119 18 25

15

1x3

78
63

92
89

90

78
83

L2

93

97

67

79

9 1816LVB

10

3 11

23

20

14

26

20

23

16

23

22

19

11

30

10

13

10

60

62

60

33

20

21

22

20

25

15

13

16

26

10

11

14

11

16

18

67

44

35

45

10

20

23

114

20

20

27

31

21

17

18

23

20

10

11

15

54

79

61

36

12

1913

11G

EN

16

101
13

26

10

16

13

52

18

22

39

14

11

14

11

12

11

16

11

39

19

32

31

12

21

11

15

16

18

15

42

28

30

49

EN

R
1

1
4

11
1

12

17
1

4
1

~4

2
1

7
4

10

5
7
10
3

D
1

7
2

J7 16 13
2

4
4

6
8

15

211
18I16

15

15

2
D

11

13

25

14

21

17

13

131417

3- 96.

Fig. 7. Raw data collapsed through levels 4-9.

PATTERN DETECTION

The pattern detection calculations in Fig. 8a are based on the variance/mean(v/m) ratio
discussed by Dacey (1973). The tabled entries are the v/m ratios and the sample sizes (n) based
upon the grid unit definition in Fig. 5c. For example, the data for endscrapers (G) in level 4

consist of the 12 frequency values in the first column and third row of Fig. 6. These frequencies
have a mean of m = 2.3333 and a variance of v = 1.5152 yielding the ratio 0.65 as shown in Fig.

8a. Under the null hypothesis of randomness,the v/m ratio possessesa chi-squareover its degrees
of freedom distribution. The hypothesis of randomnessis rejected at the .05 level if the ratio is
significant,i.e., if it exceeds the criticalvalue 1.79. The significantratiosare starred.
WEAK ASSOCIATION

For two artifact distributionsdefined over the same grid system, typical median split tables
look like those in Fig. 9, where k and n are close to N/2, x designates the diagonal with the
smallest frequencies, and N = the number of cells in the grid system. The marginal entry n (or k) is
close to N/2 since approximately this number of artifact class A (or B) frequencies are smaller than
their median frequency. The x entry in the left median split table is the number of cells for which
the A and B artifact frequencies are simultaneously smaller than their respective median
frequencies. If this value is small, relative to N/4, weak segregation between the artifact classes

may be indicated. Similarly, a small x value in the right median split table may indicate weak
aggregation. If x is close to N/4, in either table, no association may be indicated.

AMERICAN

548

VP
VB
G
EN
P
OS
VF
VB
R
D

b.

level

level

level

level

x2

X2

3.42*
1.34
1.09
1.12

120
34
24
71
2
4
152.
91
22
28

6.05*
1.79
1.27
1.71

120
25
22
67
6
4
178
92
31
39

6.58*
1.01
.65
1.32

1.29
.71
1.18
.96

132
29
28
51
4
0
145
87
32
28

.38
1.86
.58
.73

.78
3.23*
2.50*
.62

1.89*
1.48
2.55*
1.29
-

1.95*
2.13*
1.75
.64

102
38
24
63
3
7
174
117
23
29

c.

tests
association
VP, B, G, EN group.
Ihypothesis
12
I
T
t
d.f.
-value
X value

4
5
6
7
8
9

28.7
30.5
38.4
40.1
57.8
34.5

33
33
33
33
33
33

.68
.59
.24
.19
.01
.41

all

230.0

| 198

.06

level

Proportional
for
by level
level

[Vol.

42,

No.

all

1977

4,

measures.

and Patchiness

V/m ratios

a.

ANTIQUITY

X
4.98*
1.95*
1.45
1.77

1. 57
1.17
.75
1.31

levels

level

n_

3.55*
3.34*
1.33
3.06*

118
61
44
60
7
46
156
142
19
29

1.39
1.39
1.40
1.22
1.97
6.09
1.08
1.13
1.43
1.30

798
229
174
400
28

106
42
32
88
6
35
186
126
14
29

1.94*
1.61
1.43
1.54

96
991
655
141
182

association
proportional
Pairwise
an
test
p-values
for testing
all

across

of generic
levels.

VP
VP
B
G
EN

-03
.39

consistency
EN

G
.17

_19

.12
.20

__

Fig. 8. Pattern detection calculations.

A
i+

- |x |

or

--

~~~~n ~

- --l

---

Fig. 9. Median split tables.

For large values of N, as in Fig. 5b, a median split table can be analyzedby the chi-squaretest
for independence.The null hypothesis is "no association,"and weak associationwill be consistent
with a large X2 value. In both the aggregativeand segregativecases, the p-value for the null
hypothesis will be small. In this paper, the symbolic interpretations(summarizedin Fig. 12b) of
the p-valuesfor the 45 m2 gridsystem are:
Interpretation
++
+

I
--

p-value
<.05

Associational Indicator
aggregation

.05< p<.10
p>.I0

aggregation
either

.05<p<.10
<.05

segregation
segregation

PATTERN RECOGNITION STUDIES

Hietala and Stevens]

549

For a small number of cells, as in Fig. Sc, one can utilize the Hypergeometric distribution
discussed and tabulated in Lieberman and Owen (1961). Fisher's exact test, though applicable,
should not be used if the tabulated significance tables are too restrictive. Applications of
Hypergeometric distribution calculations have not previously been introduced in archaeological
spatial analysis.
In any median split table, the smallest value x may be chosen so that N>n>k>x. The
probability that a theoretical x is smaller than or equal to the observed value under the hypothesis
of no association is calculated. In the aggregative case (small x on the off diagonal and the
segregative case (small x on the main diagonal), this probability is the p-value for the hypothesis of
"no association." Due to a restricted range of p-value possibilities, the interpretations of the
calculated probabilities (summarized in Fig. 10) for the data in Fig. 6 are:
Interpretation

p-value

Associational Indicator

< .15

aggregation

> .15
< .15

either
segregation

Example calculations illustrate this procedure. Consider the median split tables of Fig. 11. For
the left table, the probability that the (-,-) cell is smaller than or equal to 1, calculated from the
Hypergeometric distribution (Lieberman and Owen 1961:36), is .040043, indicating an
interpretation of weak segregation. This table corresponds to the endscraper (G) and Levallois
flake (VF) distributions for level 8 in Fig. 6. The corresponding symbolic entry in Fig. 10 may be
noted to be "-". For the right table (N=12, n=6, k=5, x=1), the Hypergeometric probability
calculation yields a p-value of .121212 for a + interpretation.
STRONG ASSOCIATION
One measure of strong association is Kendall's tau statistic. The associated coefficient is
discussed in detail by Kendall (1955) who gives tables for testing the null hypothesis of
"independence." The alternatives, in spatial analysis situations, correspond to strong concordance
(aggregation) and strong discordance (segregation).
Some considerations should be noted in the utilization of Kendall's tau statistic in spatial
analysis. In particular, if numerous cells are simultaneously low in relative density, or empty (a
possibility when the grid system is extended beyond the reasonable site limits), concordances will
be overly abundant, and inflated tau values (on the aggregative side) will result. One way to
counter this possibility is to use Goodman and Kruskal's gamma statistic (Goodman and Kruskal
1963:322 ff, Mueller, Schuessler and Kostner 1970:218-222), which eliminates the tied cell
observation effect and only considers the observed sample concordances and discordances. This
statistic automatically rejects simultaneously empty cells as irrelevant to associational estimation.
At Rosh Ein Mor, the data are consistent in overall density across the site. However, a reduction in
the effect of tied cell observations is nevertheless important. A coefficient which reduces this
effect is Kendall's tau-b. The Kendall's tau-b p-values, with interpretations defined as in the case
for chi-square median split p-values, were obtained from the SPSS NONPAR CORR program. The
symbolic interpretations are given in Figs. 10 and 12b.
UNIFORM AGGREGATION
The uniform aggregation statistic is a direct result of testing for the existence of equal region
probabilities for two or more artifact classes. Of the three modes of association, this is the only
mode that allows for a direct test of the hypothesis of aggregation. The two weaker modes utilize a
null hypothesis of aggregation. The two weaker modes utilize a null hypothesis of "no
association," whereas the null hypothesis in this case is homogeneity of the class region
probabilities. The alternative hypothesis is the lack of uniform aggregation which, however, can
take many possible forms. A useful set of statistics, in this regard, is based upon the log linear

[Vol. 42, No. 4, 1977

AMERICAN ANTIQUITY

550

Ln

~H

+~~~~~~

~+H}+H

+I+

i-4H

Hi

H
HH

Cd0

~ ~ ~ ~ ~ ~~~~~~~~~~

vH

LrHH
H

41+

Ha%
co

IH

Ln

co

Hn

1+

f f
H

HH

4:+

41
4

II

HH

H i

+
H'

+441

+ +

iH

H0

~10

H
H

41

41

41+

.~?41+

HCA

+ 4

B
0~~~~~~~~~~~~~0

iH

U_______

0)

0~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~410

4I

to441

12~~~~~~HH

____~~~~~~~.10

.~~+41+

0%

l
04

Qc)

_____~~~~~~~~~~~~~~~~~~~~~~~

BCL)

0)4w

4141+

4141+

~+

______

~~~~~~~~~~1

41H

+___

____

4141+

-C

~~ ~ ~ ~ ~

0)

4-)

Lr)

0
H

0tC'
0

0~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Hietala and Stevens]

PATTERN RECOGNITION STUDIES

-l

551

12

12

Fig. 11. A sample median split table.

a.

Tests
of equivalent
for
distributions

spatial
each level

collapsing
scheme

(1)

x
N.A.

VP

(2)

55.4VP
55
.46

df
p

Co-associational
for levels
4-9
Row 1-Kendalls

b.

VP B
+

N.A.

G EN VF VB R

1 |2
1
2
1

(;
G

EN

3< 4 1
3 4
34

| 1 2
1 23
1 23

2
2
2

3
3
3

1
1
1

x
df

30.4
30 . 4
15

N.A.

4|

df
p

4
4
4

2
x
df
p

1|70
N.A.

70.0
55
.09

x
df
p

N.A.

lv'F

2
x
df

N.A'.'

VB

1
2
3

1
2
3

1
2
3

1 14
2 5
3 6

4
5
6

4
5
6

2
x
df
p

- 4
4
5
6

2
x2
df
p

Test

(1):

uncollapsed

Test

(2):

collapsed

++

++

a ++++

--

++

---++

++

+
++
++

-,.

I+

I1I+11+1

I+

I+

++

I-

19.0
15
.22
c.

++

VBR
15
.70

EN

.01

4
4

G|
i

I
+

++
++,
*++

++

tests
by 45m
together.
T; Row 2-Median

76.7
55
.04

N.A.

76.9
55
.03

N.A.

N.A.

29.8
25
.23

26.2
25
.41

N.A.

Artifact

by

level

B
5.3
6.3

20.6
20.0
18.0
19.6

4.4
6.4
6.6
9.7

6.2
5.9

VF
[28.6
30.3
31-5
29.9

VB
16.1
17.2
15.9
19.0

R
5.7
3.8
5.0
3.5

D
4.7
7.1
5.3

n
696
651
660
736

29.6
26.0

J 18.6
21.2

2.0
2.7

J 4.8
1 5.7

777

4
w 5
a
6
w 7
8
9

4
5
6
7
8
9

percentages

VP
25.7
20.7

L
L

data
data

Fig. 12. P-values for performed tests.

G
4.6
4.3
3.8
4.9

EN
9.2
1 11.7
11.7
11.0
14.0
9.0

784

units
Splits.

552

AMERICAN ANTIQUITY

[Vol. 42,

No. 4,

1977

model for contingency tables (Goodman 1968, 1969, 1970, 1971; Fienberg 1970; Muller and
Mayhall 1971; Hietala 1973; Haberman 1974; Reid 1974; Clark 1976). To discuss some
possibilities,a system of notation must be introduced.
A three-way contingency table is consideredwhere factor A correspondsto I artifact classes,
factor B correspondsto the J units on the x-axis in the grid unit system, and factor C corresponds
to the K units on the y-axis. This IXJXK contingency table can be fitted by a three dimensional
log linear model. A fourth factor correspondingto an additionaldimension,such as the levels of a
site, can be introduced to give a 4-way contingency table. Using the referencedcontingency table
analysis papers,a series of 5 testable models can be identified for the three dimensionalcase. The
hypotheses and their formalinterpretationsare:
Ho= HABC: No 3-way interaction between the x-axis, y-axis and artifact class factors exist.
HI= HAB,ABC: No 3-way interaction and no artifact by x-axis interaction exist. This is formally
equivalent to stating that A and B are conditionally independent given C.
H '= HAC,ABC: No 3-way interaction and no artifact by y-axis interaction exist. This is formally
equivalent to A and C being conditionally independent given B.
H= HAB,AC,ABC: No 3-way interaction and no pairwise interactions with artifact class exist. Formally,
A is independent of (B,C).
Hs= HB,C,AB,AC,BC,ABC: No interactions and no spatial dimension effects exist.

Fig. 13 portraysthe hierarchicalnatureof these hypotheses.


H

H
AB, BC

HAB,A

Ha
,ABC
IAC

ABC

B C,AB ,AC BC,ABC


Fig. 13. Hierarchical hypotheses for investigating proportional association.

The rejection of Ho states there are complex spatial relationshipsdifferentiatingthe artifact

classes. Presuming the acceptance of Ho, the rejection of H, (or H1 ) indicates that there exists an

x-axis (or y-axis) interaction affecting the distribution of the artifacts in space. Presuming
acceptance of the above hypotheses, the acceptance of H2 indicates that the artifact classes are
proportionallyassociated (i.e., uniformly aggregating).Finally, the acceptance of H3 states that
the artifact classes are mutually randomly distributed. The goal in uniform aggregational
measurementconsists of finding those classes for which Ho, H1, H0, and H2 are accepted, while
H3 is rejected.
The uniformaggregationhypothesishas the drawbackthat it is impliedwhen the artifactclasses
are mutually randomly distributed.However,the tests of strongand weak aggregationwill tend to
not confirm this implication. Thus, investigationof H3 as well as the other modes of aggregation
should ascertainif uniformaggregationis a logical possibility.
To test the log linear hypotheses, the TRICHIprogramdeveloped at SMU was utilized. This
programwill test all but the last hypothesis, H3. The outcome of a test for hypothesis H2 was

Hietala and Stevens]

PATTERN RECOGNITION STUDIES

553

given the interpretation ++ when its p-value exceeded .30, an interpretation + when its p-value was
between .10 and .30, an indeterminate interpretation (I) when its value was between .05 and .10,
and a - interpretation when its value was smaller than .05. It must be noted that the last outcome
does not imply segregation but that the hypothesis of uniform aggregation is not supported. The
TRICHI program does not handle tables for which zero marginal frequencies exist. Thus, some of
the tests could not be performed. This program is available through the SMU Department of
Statistics.
In the applications to Rosh Ein Mor, the aggregative hypothesis, H2, was tested for multiple
artifact groups. The symbolic interpretations of these results are given in Figs. 9b and 10 for the
data base shown in Fig. 6.
PATTERN CONSISTENCY
Some forms of pattern consistency can be related to uniform aggregation. While the general
setting requires a four dimensional model, an equivalent three dimensional model can be
constructed. Consider the model where the factors A, P, and L correspond respectively to the
artifact classes, the quadrat units as a partition of the horizontal space, and the levels (or sites).
If the artifact classes under consideration are uniformly aggregating in each level, then the
hypothesis H2, defined above, must be true for each level. That is, H4: A,P are conditionally
independent given L is the formal hypothesis for testing generic consistency
within the uniform aggregation mode. If each artifact class exhibits a single spatial pattern for all
levels, then H5: P, L are conditionally independent given A, is true. If the hypotheses H4 and H5
are simultaneously true, then the artifact classes must be uniformly aggregating in the same relative
locations in each level. This is formally equivalent to H6: (A,L) are independent of P, which is the
hypothesis of specific consistency for the uniform aggregation mode of association.
The tests for these hypotheses are based upon chi-square distributions. The p-values for the
performed tests are given in Figs. 8b, 8c, and 12a.
The weaker modes of association are discussed in the applications to give support to the
stronger hypotheses.
Applications
Rosh Ein Mor gave rise to many associational regularities and irregularities between artifact
classes when investigated by separate levels and, thus, can be utilized to illustrate spatial analysis
concepts.
DETECTION OF PATTERNING
The detection of nonrandom patterning of an artifact class in space has been based upon the
variance/mean ratio or a variant of such, as noted earlier. The block choice systems introduced by
Morisita (1962) and Grieg-Smith (1961), statistically motivated by Thompson (1955), and
introduced into archaeology by Whallon (1973), as well as the variance/mean ratio (Dacey 1973),
are special cases of a general system of variance analysis for a sequence of specified block sizes.
The block choice systems are primarily oriented toward the determination of regional definitions
which allow for optimal associational detection between two or more artifact classes. The optimal
block size decision, if one exists, is based upoIn the detection of significantly nonrandom artifact
class distributions relative to specified block sizes. unfortunately, the testing procedures are based
on the null hypothesis of randomness. There are practical problems in nonrandom patterning
detection when based upon the variance/mean ratio (or its affiliated block size procedures), if
considered as an admissability criterion for associational determination. It is instructive to show
what might have happened had such a criterion been adhered to in this study.
Fig. 8a summarizes the variance/mean ratios for the data base in Fig. 6. The starred values
indicate the ratios statistically inconsistent with the randomness hypothesis at a .05 type 1 error
value. Only one artifact class, Levallois points (VP), has strong indicators of nonrandomness across

554

AMERICAN ANTIQUITY

[Vol. 42,

No. 4,

1977

all six levels. The opposit extreme is seem in the denticulates (D) that always support the
randomness hypothesis. Since only 15 of the 48 randomness tests indicate a nonrandom
alternative, the number of combinations amendable to study under the variance/meanratio
criterion is extremely limited. By separate levels, only Levels 7 and 9 possess as many as four
nonrandom indicators. These two levels, on the other hand, possess only two nonrandom
indicatorsin common, viz., the Levallois point (VP) and Levalloisflake (VF) classes.In general,if
studies were restricted to only those classes with nonrandom indicators, Levallois points (VP)
would have been compared to Levallois blades (VB), and sidescrapers(R) in Level 6; Levallois
blades (VB), Levallois flakes (VF), and endscrapers(G) in Level 7; burins (B) in Level 8; and
burins (B), notches (EN) and Levallois flakes (VF) in Level 9. No comparisonswould have been
performedin Levels 4 and 5. Basedupon this criterion,associationalinvestigationwould not have
been implemented across the six levels. Furthermore,152 of the 168 pairwisecomparisonswould
not have been considered.
ESTIMATIONOF PATTERNING
Fig. 10 summarizesthe resultsof the pairwiseassociationaltests (uniform, strong,and weak) by
level. The lower left hand side of this matrix portraysa subjectiveimpressionacrosslevels of the
site.
It may be noted that Levalloispoints (VP), burins (B), endscrapers(G), and notches (EN) are
seemingly aggregatingin all levels. Of the 105 tests between these classesacrossthe six levels, only
three have negative indicators. These three all involve uniform aggregationtests for which a
negative entry does not imply segregation.Thus, the four classes seem to exhibit strong aggregation and are likely to be uniformly aggregating.On the other hand, the Levalloisflake (VF) class
has most of its conclusive indicators directed toward segregationfrom the Levalloispoint (VP),
burin(B), endscraper(G), and notch (EN) classes.
The other comparisonsdo not yield as strongsubjectiveimpressions.It is difficult to ascertain,
for example, if the Levallois point (VP) and Levallois blade (VB) classes are consistently
associating. Also, the indicators of uniform aggregationbetween the notch (EN) and denticulate
(D) classes are a likely result of the near random patternsindicated for both. Thus, their overall
interrelationshipappearsto be indeterminate.
In the remainingcases (including the VB-VF and VB-D comparisons)aggregativeor segregative
trendsfor the indicatorsets are not clear.
From the comparisonsin Fig. 10 it seems reasonableto posit that, for the site as a whole, the
Levallois point (VP), burin (B), endscraper(G) and notch (EN) classes are strongly aggregating,
while Levallois flakes (VF) are segregatingfrom this complex. Denticulates may be spatially
"independent"of the aggregativegroup whereassidescrapers(R) appearto be "independent"of
all other classes.
Before the spatial distributions are discussed further, the impact of block size should be
considered. As was noted in the methodology section, the common block size requirementis
inadequate in situations of large frequency differences between artifact classes and major
differencesin the sizes and shapesof individualclass clusters.The formersituationis illustratedin
Fig. 12c. The latter may be derived from Figs. 6 and 7. This derivationis exemplified in the
"theoretical patterns" column of Fig. 14, which is discussed later. For example, the burin (B),
Levalloisflake (VF), Levalloisblade (VB), sidescraper(R), and bone (OS) classesare differentially
distributedin space. A largerblock size would, therefore,inhibit patterndiscovery.Smallerblock
sizes, on the other hand, would have obfuscatedsome of the previouslyidentified patterns.
CONSISTENCYOF PATTERNING
The previousinvestigationsindicate that artifact associationsare fairlyconsistent acrosslevels.
It seems reasonableto ascertainthe nature of these pattern regularitiesin space and time. Space
regularitiesshould lead to similar artifact class associations when the grid unit definitions are
varied. Time regularitiesshould be reflected by similarassociationalpatterns from level to level,

Hietala and Stevens]

PATTERN RECOGNITION STUDIES

Theoretical

levels

Patterns

4
o<

VP
,

_-

06

__+

4
+

CS03B

[Y1F]F

+ 0

j 2

jL
_j +

6+

43

_-

71F51741-[771

j j j

52

12

6-1

0 4

NA

4heortica

13

0 3

2 3

0 5

-6

03

-++

Rg+

6 1

+33

~~~~~
21 I

4-9

+6_F71150F71][771F7_1+F771

1+

-5

04
+

+05

33

361

3I

555

5j
9 01

0
+

90

06

1t

|
6+

90-90

60
6

|0

51

Fig. 14. Theoretical patterning. Rows give observed values; columns give the theoretical pattern.

556

AMERICAN ANTIQUITY

[Vol. 42,

No. 4,

1977

which would either maintain themselves in the same local areas (specific consistency) or be
locationally different from level to level (generic consistency).
To investigate these possibilities the following four hypotheses are considered:
HI:
H2:
H3:
H4:

Spatial consistency through levels for individual artifact classes.


Associational consistency by level (generic consistency).
HI and H2 together (specific consistency).
Associational consistency for the data collapsed across levels.

The first three hypotheses allow for distinguishing generic from specific consistency, while the
fourth hypothesis allows for a robust determination of grid unit dependency given that specific
consistency is a logical possibility. For H4, this dependency may be investigated by collapsing the
data across levels using the 45 square meter grid unit system.
The components of the first hypothesis are tested by the chi-square homogeneity test. These
results are given in Fig. I 2a. In these tests, burins (B), endscrapers (G), notches (EN), sidescrapers
(R) and denticulates (D) required the additional collapsing of the adjacent cells in order to meet
the assumptions for the test statistic. The hypotheses for Levallois points (VP), endscrapers (G),
notches (EN), sidescrapers (R), and denticulates (D), are overwhelmingly accepted. The
hypotheses for burins (B), Levallois flakes (VP), and Levallois blades (VB) are strongly affected by
the upper two levels. If one concentrates only on the Levallois point (VP), burin (B), endscraper
(G), and notch (EN) classes, by summing the corresponding independent chi-square statistics, a
chi-square value of 1 16.5 with 100 degrees of freedom results yielding a p-value in excess of 0.10.
Thus, the primary aggregative cluster (Levallois points, burins, endscrapers, and notches) is in a
reasonable acceptance range for this hypothesis.
An alternate method for examining HI is a comparison of "theoretical" and observed patterns,
using a median split approach. The "theoretical" patterns for this study were subjectively derived
from the previous results in conjunction with the collapsed 1X3 m2 units shown in Fig. 7. These
are depicted in Fig. 14 where the + signs in the "theoretical patterns" column indicate where the
highest cell frequencies are expected. Most classes overwhelmingly fit their expectations. The only
contradiction, among 53 possibilities, occurs in Level 4 where the Levallois flake (VF) observation
is in near total negation of its expectation. In all other instances the results are not negative, with
46 of the 52 cases supporting the hypothesis. On a quasi-statistical basis, using Fisher's test for
combining "independent" tests of significance, the general hypothesis is accepted. This result is
indicated by the last column which collapses the data across Levels 4-9. It should be noted that the
borer (P) class fits the expectation of the aggregative group and, thus, might be considered to be
weakly aggregating with it. The sample sizes are unfortunately too small to consider higher
aggregative modes. The bone (OS) class has its own consistent scenario. This pattern does not seem
to be related to the others.
The second hypothesis, generic consistency, is tested for the artifact classes in the aggregative
group. The results are portrayed in Figs. 8b and 8c. In Fig. 8b, it is seen that the only negative
result is within Level 8 where the p-value for acceptance of the uniform aggregative hypothesis is
.01. The notch (EN) class in this level causes the disruption of the hypothesis. For all levels
combined, however, the hypothesis is accepted at a p-value in excess of .05. Fig. 8c gives the
pairwise uniform aggregational test p-values between these four artifact classes, collapsed across
levels, suggesting generic and specific consistency as logical conclusions.
From the results of testing HI and H2, it is seen that evidence exists for specific consistency of
the aggregative group for the uniform aggregation mode. Borers (P) are weakly aggregating with
this group, Levaflois flakes (VF) are strongly segregating from this defined complex, and
sidescrapers (R), as well as denticulates (D) possess their own patterns. The latter two classes are
seemingly independent of the others. Finally, Levallois blades (VB) seem to be, at best,
indeterminate.
The results of testing the fourth hypothesis, for the weak and strong association modes, are
shown in Fig. 12b. This table is structurally comparable in interpretation to Fig. 10, although it
uses the collapsed 45 m2 data given in Fig. 7. The results are consistent with the previously

Hietala and Stevens]

PATTERN RECOGNITION STUDIES

557

considered hypotheses. A major weakness occurs with respect to the notch (EN) and denticulate
(D) classes which appear to strongly aggregate. Fig. 7 indicates this tendency, which may not be
spurious, although it certainly is inconsistent across the six levels. This association might be
investigated with further data analysis and excavation.
At a more detailed level, the data obviously do not indicate totally consistent patterns of tool
class associations. In fact, there is some indication of potential shifts in these associations through
the deposits. Of particular note is that Level 4 exhibits strong aggregation between Levallois flakes
and the primary aggregative group, and in Level 9, there may be strong aggregation between
denticulates and the primary group. These patterns may be indicative of expanded activities, at
least within the excavated portion of Rosh Ein Mor. On the other hand, these apparent anomalies
may be the result of ordinary random processes common within human cultural systems and the
natural universe.
The major patterns of association, however, are remarkably consistent across levels. The results
of testing generic consistency suggest one possible Mousterian "tool kit." This consists of
endscrapers, Levallois points, burins, notched pieces and borers. Normally segregated from this
"tool kit" are Levallois flakes. While the results are ambiguous, it may be that Levallois flakes and
Levallois blades either constitute another "tool kit" or, more likely, they represent debitage.
Further work is needed to clarify this point. The spatial relationships between these Levallois
categories and the traditional forms of debitage such as primary elements, cores, flakes, etc., would
assist in this clarification.
The recognition of these possible "tool kits" may bring to mind the work of Binford and
Binford (1966) where factor analysis was used to isolate "tool kits." Unfortunately, there is little
comparability between the two methodologies. Regardless of whether their approach is
appropriate for such studies, the Binfords were only concerned with the association of artifacts in
an aspatial context, while this study has isolated apparent "tool kits" within spatial subsets of each
excavated level.
The results of testing specific consistency suggest a different view of Middle Paleolithic
occupations which have traditionally been thought of as repeated encampments. It was initially
hypothesized that the "occupations" at Rosh Ein Mor were, in fact, of this nature. That is,
generic, but not specific pattern consistency was expected between the occurrences defined by the
investigated 5 cm levels. However, not only did the artifact classes consistently aggregate and
segregate by levels, they did so maintaining the same localities regardless of depth. To test that
these results were not due to the chosen subregional units, a secondary data base was obtained by
collapsing the information from Levels 4-9 into a single data base corresponding to the original 45
square meter grid units (Fig. 5b). This secondary data base (Fig. 7) supported the original results.
Combined with the continuous aggradation of geological and archaeological deposits, these results
strongly suggest that the occupation at Rosh Ein Mor was sedentary. This phenomenon may well
exist at other Middle Paleolithic sites.

Conclusionsand Recommendations
The methodology of spatial analysis can be characterized as a particular emphasis within
pattern recognition problems. Within these, there are three primary areas defined by the problems
of detection, estimation, and consistency (prediction). The pattern detection problem was shown
to be weak in application. On the other hand, pattern estimation considerations have been
deficient in that the concept of "estimation" has not been adequately studied and articulated.
Finally, pattern consistency, or prediction, has not previously been defined and exploited.
The emphasis of this paper minimized the role of the detection problem relative to estimation
and consistency. The results of this study indicate that simplistic procedures for spatial analysis,
such as used by Whallon (1973) and Dacey (1973), are potentially misleading. It may be that the
new archaeological principles, insistent upon a deductive-nomological approach, have been
partially responsible for this impasse. Undoubtedly, a balance between inductive and deductive

558

AMERICAN ANTIQUITY

[Vol. 42,

No. 4,

1977

principles is required for archaeological theory to finely mesh with archaeological data. If
deduction takes the upper hand, then pattern detection rules, for example, necessarily follow.
Alan Watts (1972:28) argues against pure deduction, among other things, as follows:
... a scanning process that observes the world bit by bit soon persuades its user that the world is a great
collection of bits, and these he calls separate things or events. We often say that you can only think of one
thing at a time. The truth is that in looking at the world bit by bit, we convince ourselves that it consists of
separate things, and so give ourselves the problem of how these things are connected and how they cause and
effect each other. The problem would never have arisen if we had been aware that it was just our way of
looking at the world which had chopped it up into separate bits, things, events, causes, and effects
[emphasis addedi.

It is emphasized, therefore, that numerous inductive-deductive approaches may be useful. A


panacea methodology does not exist for any problem and certainly not for spatial analysis.
Acknowledgments. We wish to acknowledge Professor Anthony Marks, Southern Methodist University, for his
extensive help in the promotion, review, and interpretation of the results of this preliminary study.
The data upon which this study is based were obtained through National Science Foundation grants GS-3019
and GS-28602X.
Binford, L. R. and S. R. Binford
1966 A preliminary analysis of functional variability in the Mousterian of Levallois facies. American
Anthropologist, 68, Part 2: 238-95.
Bishop, W. W. and J. D. Clark, editors
1967 Background to Evolution in Africa. The University of Chicago Press, Chicago.
Clark, G. A.
1976 More on Contingency table analysis, decision making criteria and the use of log linear models.
American Antiquity 41: 259-73.
Crew, H. L.
1976 The Mousterian site of Rosh Ein Mor. In Prehistory and Paleoenvironments in the Central Negev,
Israel, Vol. I, edited by A. E. Marks, pp. 75-112. SMU Press, Dallas.
Dacey, M. F.
1963 Order neighbor statistics for a class of random patterns in multidimensional space. Annals of the
Association of American Geographers 53: 505-15.
1973 Statistical tests of spatial association in the locations of tool types. American Antiquity 38: 320-28.
Fienberg, S.
1970 The analysis of multidimensional contingency tables. Ecology 51: 419-33.
Goldberg, P.
1976 Upper Pleistocene geology of the Avdat/Aqev area. In Prehistory and Paleoenvironments in the
Central Negev, Israel, Vol. I, edited by A. E. Marks pp. 25-56. SMU Press, Dallas.
Goodman, L.A. and W. H. Kruskall
1963 Measures of association for cross classifications III: approximate sampling theory. Journal of the
American Statistical Association, 5 8: 310-64.
Goodman, L.A.
1968 The analysis of cross-classified data: independence, quasi-independence and interactions in
contingency tables with or without missing entries. Journal of the American Statistical Association, 63:
1091-131.
1969 How to ransack social mobility tables and other kinds of cross classification tables. American Journal
of Sociology 75: 1-40.
1970 The multivariate analysis of qualitative data: interactions among multiple classifications. Journal of
the American Statistical Association 65: 226-5 6.
1971 Partitioning of X2, analysis of marginal contingency tables, and estimation of expected frequencies in
multidimensional contingency tables. Journal of the American Statistical Association 66: 33945.
Grieg-Smith, P.
1952 Ecological observations on degraded and secondary forest in Trinidad, British Indies, II: structure of
the communities. Journal of Ecology 40: 316-30.
1961 Data on pattern within plant communities. Journal of Ecology 49: 695-708.
1964 Quantitative plant ecology, 2nd edition. Plenum Press, New York.
Haberman, S. J.
1974 Log-linear models for frequency tables with ordered classifications. Biometrics 30: 589-600.
Hietala, H. J.
1973 Adaptive parametric estimation and classification for multinomial distributions with applications to
log linear models and contingency tables. Unpublished PhD dissertation. Department of Biostatistics,
U.C.L.A.

Hietala and Stevens]

PATTERN

RECOGNITION

STUDIES

559

Horowitz, A.
1976 Late quaternary paleoenvironments of prehistoric settlements in the Avdat/Aqev area. In Prehistory
and Paleoenvironments in the Central Negev, Israel, Vol. I. edited by A. E. Marks, pp. 57-68. SMU Press,
Dallas.
Kendall, M.G.
1955 Rank Correlation Methods, 2nd edition. Hafner Publishing Company, New York.
Kershaw, K. A.
1961 Association and co-variance analysis of plant communities. Journal of Ecology 49: 643-54.
Lieberman, G. J. and D. B. Owen
1961 Tables of the Hypergeometric Probability Distribution. Stanford University Press, Stanford.
Marks, A. E.
1971 Settlement patterns and intrasite variability in the central Negev, Israel. American Anthropologist 73:
1237-44.
Marks, A. E., J. Phillips, H. Crew and R. Ferring
1971 Prehistoric sites near Ein Avdat in the Negev. Israel Exploration Journal 21: 13-24.
Marks, A. E. and H. Crew
1972 Rosh Ein Mor, an open-air Mousterian site in the Central Negev, Israel. Curent Anthropology 13:
591-93.
Morisita, M.6
1962 I -index, a measure of dispersion of individuals. Research in Population Ecology, IV, pp. 1-7. Kyushu
University, Fukuoka.
Mueller, J., K. Schuessler and H. Costner
1970 Statistical Reasoning in Sociology, 2nd edition. Houghton Mifflin, Boston.
Muller, T. P. and J. T. Mayhall
1971 Analysis of contingency table data on Torus Mandibularis using a log linear model. American Journal
of PhysicalAnthropology 34: 149-53.
Munday, F. C.
19 7 6 The Avdat/Aqev Area: its habitat and geographic setting. In Prehistory and Paleoen vironmen ts in the
Central Negev, Israel, Vol. I, edited by A. E. Marks, pp. 9-25. SMU Press, Dallas.
Pielou, E. C.
1959 The Use of point-to-plant distances in the study of the pattern of plant populations. Journal of
Ecology 47: 607-13.
1960 A single mechanism to account for regular, random and aggregated populations. Journal of Ecology
48: 575-84.
1969 An Introduction to Mathematical Ecology. Wiley-Interscience, New York.
Reid, D.
1974 Some comments on typologies in archaeology and an outline of a methodology. American Antiquity
39: 216-42.
Tchernov, E.
1976 Some late quaternary faunal remains from the Avdat/Aqev area. In Prehistory and Paleoenvironments
in the Central Negev, Israel, Vol. I, edited by A. E. Marks, pp. 69-74. SMU Press, Dallas.
Thompson, H. R.
1955 Spatial point processes, with applications to ecology. Biometrika 42: 102-15.
1956 Distribution of distance to n-th. neighbor in a population of randomly distributed individuals. Ecology
37: 391-94.
1958 The statistical study of plant distribution patterns using a grid of quadrats. Australian Journal of
Botany 6: 32242.
Watts, A.
1972 On the taboo against knowing who you are. Vintage Books, New York.
Whallon, R.
1972 Spatial analysis of paleolithic occupation areas. In The explanation of culture change: models in
prehistory, edited by C. Renfrew, pp. 115-30. University of Pittsburgh Press, Pittsburgh.
1973 Spatial analysis of occupation floors, I. Application of dimensional analysis of variance. American
Antiquity 38: 266-78.
1974 Spatial analysis of occupation floors, II. Application of nearest neighbor analysis. American Antiquity
39:16-34.

You might also like