You are on page 1of 16

Journal of Archaeological Science 1984, 11,40%420

Integrating Pattern and Context in Spatial


Archaeology
Jan F. Simek

Traditional approaches to intrasite spatial analysis in archaeology have concentrated on


identifying associations among classes of artifacts over a site surface. This focus has
tended to ignore the possible effects of contextual constraints on inter-class relation-
ships, for example the gravity effects of hearth features on object deposition.
Consequently, sets of co-occurring artifact classes defined in space are usually ascribed
behavioural significance as tool kits, even though they may represent unrelated
objects simply discarded into the same spatial location. This paper attempts to
illustrate this problem through contextual integration of artifact and feature
distributions. Quantitative methods are employed to carry out contextual spatial
analysis of artifact distributions from the Upper Palaeolithic rockshelter site Le
Flageolet I (Dordogne, France). Results indicate that associations among artifact
classes in space might be constrained by feature locations at Le Flageolet I and that
contextual interpretations are warrantedzfor the observed inter-class relations.

Keywords: SPATIAL ANALYSIS, SPATIAL CONTEXT, LIVING FLOORS,


ACTIVITY AREAS, FEATURES, ARTIFACT ANALYSIS, SIMULATION, K-
MEANS CLUSTER ANALYSIS, REGRESSION, FRANCE, UPPER
PALAEOLITHIC

Introduction
The goals of conventional intrasite spatial analysis in archaeology have been the
identification of activities carried out at a site in the past, the areas where different
activities occurred, and the tool kit used to perform each activity. Practically, this
would seem to be a simple problem. Clusters of objects in space must be defined, the
objects composing each cluster classified and counted, and tool kits identified when
distinct assemblage variants are recognized through interclass association studies.
However, recent ethnoarchaeological work indicates that this simple, presumed relation
between archaeological location and prehistoric activity may not always hold (e.g.
Binford, 1978; Yellen, 1977). To account for possible deviation from the presumed
model, natural or prehistoric behavioural transformation processes that obscured
the expected direct relation are often invoked (Schitfer, 1976: 11-17, et pas&r).
Allowing for such transformation processes, the basic model is maintained: all other
things being equal (i.e. unmodified), spatial relations among artifact classes reflect past
human activities in their nature and location.

Department of Anthropology, University of Tennessee, Knoxville, Tennessee,


U.S.A.
405
0305-4403/84/050405+ 16 rSO3.00/0 0 1984 Academic Press Inc. (London) Limited
406 J. F. SIMEK

In this paper, I show that conventional spatial analysis based on the Activity
Model, and emphasizing object distributions, has ignored some basic spatial processes
that may cause association among artifact classes. For example, features may act as
centers of gravity for artifact distributions, and spatial associations among artifact
classes may simply be statistical effects of feature-artifact class relations. A hearth area,
serving as an activity focus for a prehistoric camp, may be the locus of deposition
for the products of tool manufacture, cooking debris, and clothing production. The
various material products of these activities were not necessarily used together but were
deposited into the same location because the activity area is contextual, i.e. in
proximity to the hearth (cf. Yellen 1977).
The formation ofartifact distributions clearly cannot be simply characterized in terms
of activities and transformations. Social relations among site occupants, duration of
occupation, the nature of site topography, and such factors as local climate, all may
condition the deposition of materials and their consequent spatial organization (e.g.
Yellen, 1977; Binford, 1978, 1981, 1983; Gould, 1980). The context of material
deposition is every bit as important (perhaps more so) than the objects involved. It
therefore requires analytic treatment.
In this paper, I explore the problem of context in archaeological spatial analysis.
To examine what is involved in contextual integration, the effects of features on the
spatial distributions of artifacts will be analyzed in order to show how contextual
analysis might be pursued in archaeology. It should be pointed out now that these
are only two among many potential contextual elements that might be considered,
e.g. topography, architecture, ecological parameters (Clarke, 1977: 9).

Principles of the Analysis and Variable Relations


Principles
To formalize the analysis of contextual relations between artifacts~ and features, two
key principles guide the choice of appropriate methods and the defimtron of variables to
be examined. The first principle concerns the direction of constraint or effect in the
feature-artifact relationship. Contextual analysis here will involve comparisons between
aggregate descriptions of artifact distributions and individual feature locations. This
approach is taken because I assume (following Clarke, 1977) that artifacts (defined as
portable objects) may be positioned in space relative to features but features are not
positioned relative to artifacts. I will employ the mean distance of artifacts within a
class to a given feature as such an aggregate measure.
The second principle of analysis is practical and concerns the identification of
meaningful relations among objects and features. One way to detect a significant
relation is through comparison with a non-significant one. In other words, the
observed aggregate distance between an artifact distribution and a feature might be
compared to a random one. A similar approach is often used in spatial analysis of
artifact class associations (e.g. Dacey, 1973; Whallon. 1973) and has a strong
theoretical foundation generally in quantitative spatial analysis (cf. Kershaw,
1964; Pielou, 1977). However, the problem of contextual relation is somewhat different
than that in previously used random tests (but see Hodder & Okell, 1978). I want to
determine if one distribution is independent of another, regardless of whether&- not
each is random on a global scale. In the present study, I explore simple statistical
approaches to analyzing such relations.
These principles seem reasonable when the distributions under study represent a
single occupation. Complications arise when several or many occupations compose
the distributions and cannot be separated. For example, hunters occupying a site that
had seen recent use by others might avoid refuse piles from the previous occupations.
PATTERN AND CONTEXT IN SPATIAL ARCHAEOLOGY 407

However, the problem of distinguishing meaningful spatial distributions for study is


an ubiquitous one. Unless and until either consistent use of site space over time
(redundant organization) or occupation coherence can be demonstrated, the meaning
of any kind of intrasite spatial analysis will remain unclear.

Relation
At least three relationships are involved in an examination of feature-artifact effects
(Clarke, op cit.). The first relation, among classes of artifacts, is that normally of
interest to archaeologists. Typically, spatial analysis seeks to define systematic
relationships-associations-among artifact classes. The second relationship concerns
features. Before assessing relations between artifact classes and features, patterns
present in the feature distribution must be defined. This is because constraints placed
on artifact distributions by features might reflect inter-feature configurations. The
third relation, pivotal in contextual integration, concerns the effects of features on
associations among artifacts. In other words, are the associations among artifact
classes in the distribution dependent on the locations of features? When artifact
classes are associated in space and the same classes are related to the same features,
the interclass association may be determined by context. In short, it cannot be decided
if the artifact classes are associated with each other or if their association is simply a
contextual effect.
Thus, a three-stage analytic sequence is required, even for the basic task of
identifying meaningful (i.e. independent) associations among artifact classes, (1) the
spatial relations among artifacts (described by technological or functional classes, for
example) must be assessed for spatial association, (2) relations among features have to
be characterized, and (3) patterns of artifact class association must be examined in
terms of feature distributions.

Expectations
What kinds of spatial relations between artifacts and features might be anticipated,
and which relations constitute contextual constraint? Four basic kinds of relation can be
defined. Models for these are illustrated in Figure 1. (Each relation is shown with a
schematic spatial map and two model tables: matrices of anticipated interclass
associations and of expected artifact class-feature relationships). Figure l(a) shows a
total lack of association within artifact classes and between artifacts and features.
Artifact classes are spatially independent of each other. Contextual constraint on
interclass associations is not an issue here.
A second relationship comprises two kinds of spatial patterning, representing spatial
mirror images [Figure l(b)]. The methods developed below cannot, at present,
distinguish between them quantitatively. However, the proposed techniques can
determine when such patterning is present, and simple visual inspection of plan maps
should easily discern which situation holds. In both patterns, artifact classes are
associated in space. The first pattern is that assumed in most intrasite spatial analyses.
Here associations exist among artifact classes. However, the pattern of artifact
association is independent of feature locations and can be interpreted without
reference to context. The second pattern illustrated in Figure l(b) involves interclass
associations spatially related to all features. A pattern like this implies not only that
artifact class relations are contextually constrained but also that all features serve as
context for artifact deposition in the same way. In this case, associations among
artifact classes must be interpreted in terms of context.
A third pattern involves lack of association among artifact classes and spatial
association among artifacts and features [Figure l(c)]. Here, the spatial locations of
408 J. F. SIMEK

2&3 2&3

123 ABC 123 ABC


l/ 000 l/ 000
20 / 000 2+/ 000
300/ 000 3++/ 000

la) (b)

262 3&3 G3 0
3

123 ABC 123 ABC


1 / +oo 1 / +oo
201 o+o 2+/ +oo
300/ oo+ 300/ 000
k3 (4

Figure 1. Four relations between hearth features and artifacts. In each


illustration, a schematic plan map is presented. Numbers represent artifact
classes,and letters represent features. Beneath each map, two association matrices
are shown: the expected interclass pattern (left) and the expected class-to-feature
pattern (right).

certain artifact classes are constrained by certain feature contexts, but these are non-
overlapping. There is no problem with having to consider associations among artifact
classes in contextual terms, since there are no such associations. However, contextual
constraint may be present for individual artifact categories.
A fourth pattern is where specific artifact class associations are contextually
constrained by certain features while other associations do not exist or are independent
of feature locations [Figure l(d)]. A key characteristic of this situation is that at least
some associations among artifact classes are isomorphic with artifact-feature associ-
ations. Thus, those artifact associations must be interpreted in contextual terms-
The preceding discussion has centered on spatial patterning relevant to defining
contextual effects of features on artifacts and on certain possible forms such constraint
might take. The case study which follows represents a first attempt at integrating
features and artifacts in contextual terms. As such, it is intended as illustrative rather
than definitive. The data used for the study come from Couche V at Le Flageolet
I, an Upper Palaeolithic rockshelter site in southwestern France.
PATTERN AND CONTEXT IN SPATIAL ARCHAEOLOGY 409

The Data Base : Le Flageolet I Couche V


Le Flageolet and the varied materials contained in its sediments have been the subject
of a series of reports in both French and English and do not require detailed
presentation here (e.g. Rigaud, 1969, 1978, 1982; Laville et al., 1980; Simek &
Leslie, 1983; Simek, 1984). Couche V is a rich archaeological layer present over c. 32m2
in the shelter (Figure 2). The level is very thin (usually 5-8 cm thick and never
exceeding 20 cm) and stratigraphically homogeneous. A single occupation is inferred
for several reasons. First, no microstratigraphic breaks in the thin layer were detected
during excavation, and no evidence for postdepositional alteration of material positions

Figure 2. Planmap ofartifacts scattered over the Le Flageolet I Couche V surface.


410 J. F. SIMEK

is present. Second, recent (yet incomplete) refitting of lithic materials has yielded few
links, but these are generally long-distance joins, spanning the entire Couche V
surface and thickness. Third, as will be discussed shortly, only a single hearth
as normally conceived was uncovered, and there is at present no evidence for reuse
of this feature. Le Flageolet I Couche V seems to represent a single or redundant
occupation layer.
In all, four features, each involving the use of fire in some way, were uncovered
on the Couche V surface. These differ in terms of form, content, and size. Around and
within the features, 506 retouched chert artifacts were scattered (over half of them
burins) along with thousands of by-products from lithic reduction. Bone refuse was also
abundant on the Couche V surface but is not considered here.

Features
The spatial distribution of the Couche V features will be considered later, but their
formal characteristics are as follows. Feature 1 is the most extensive, with a diameter of
nearly 1.5 m. It is lined with rocks comprising both igneous cobbles, mostly of quartzite
(probably derived from the Dordogne River below) and limestone plaquettes weathered
from the shelter walls. Feature 1 fill includes tiny carbonized bone fragments mixed
with the sediment matrix, larger burnt and unburnt bone fragments, and lithic debris
including retouched tools (some burnt after retouching). Following Leroi-Gourhan
(1972) this feature is interpreted as a focaZ hearth for Couche V because of its size,
richness, and content diversity.
Feature 2 is a concentration of burnt limestone rocks c. 0.5m in diameter. Large
quantities of tiny burnt bone fragments are incorporated in the feature matrix. Feature
2 lacks other artifactual material. Its interpretation is problematic and is here
descriptively designated a burnt limestone platform.
Feature 3 is similar in size to Feature 2. However, the contents are dilferent. Both burnt
limestone and cobbles are present in Feature 3; limestone plaquettes are the more
numerous component. The feature fill includes quantities of bone and lithic debris, but
little of this is burnt. Just to the west of the feature, a large rock and a second somewhat
smaller stone serve to bound the deposit. Again, interpretation is problematic. Feature 3
is herein designated a filled rock circle.
Feature 4 is the smallest in Couche V (c. 0.3 m in diameter). In content, this feature is
the opposite of Feature 3, comprising cobbles and plaquettes with quartzite cobbles the
dominant element. Feature 4 fill includes few artifacts, and these are mostly bone
fragments. No burnt or carbonized material is present in the sediment matrix. Feature
4 is here described as an empty rock circle.

Artifacts
Artifacts recovered from Couche V are numerous and diverse. The over 500 retouched
pieces classified into the traditional tool typology compose a classic Upper
Perigordian assemblage (Rigaud, 1982). The diverse Upper Palaeolithic tool types
represented in the assemblage (Sonneville-Bordes & Perrot, 1954) are here subsumed
under general classes that stress the location of retouch on the tool blank. These general
categories (for example, endscrapers and burins) probably represent broad functional
units since they are (1) ubiquitous in Upper Palaeolithic assemblages through time, (2)
are distinct in their modes of production, and (3) may contain chronological variants
within a class but not between the classes. Six categories of retouched stone tool are
defined for the studies that follow. Table 1 presents the classes, the traditional type
numbers incorporated in each class. and the number of specimens collected from
Couche V.
PATTERN AND CONTEXT IN SPATIAL ARCHAEOLOGY 411

Table I. Artifact classes used for spatial analysis of Couche V. Listed are the
traditional type numbers included in each class, the number of artifacts in a class,
and the percentage of the total assemblage made up of each (N = 506). Repeated
type numbers are compound tools; these were treated as several looIs according to
the classes represented on a piece

percentage of total
Class Types N assemblage

Endscrapers 1-18 43 8.5


Burins 18, 19,22,3&46 256 50.8
Truncations 19,5%60 28 5.5
Retouched 61-63 43 8.5
Notched 72-14 29 5.1
Bladelets 77,84-95 36 7.1

Having described two data sets present in Le Flageolet I Couche V, contextual analysis
can proceed. As was outlined above, spatial relations among artifacts will be analyzed
first, then relations among the four features will be examined. Finally, relations between
the two element sets will be investigated.

Spatial Relations in the Element Set Artifacts


The assessment of spatial structure in artifact distributions for Couche V has two steps.
First, patterning in the distribution map of objects must be examined to detect any
spatial concentrations or accumulations of material, i.e. clusters. This has been termed
the pattern recognition phase of spatial analysis by a number of workers (e.g. Hietala
& Stevens, 1977; Kintigh & Ammerman, 1982; Simek & Larick, 1983; Simek, 1984).
Second, accumulations must be examined for significant regularities in content
representing artifact classes that systematically co-occur over space. This has been
termed the pattern association phase of spatial analysis (Ammerman et al., 1983;
Simek, 1984).
The technique used here for pattern recognition was proposed by Kintigh &
Ammerman (1982) and illustrated with an archaeological case study by Simek &
Larick (1983; see also Simek, 1984). Detailed descriptions of the method are provided
in those papers and need not be repeated here. A k-means clustering algorithm provides
the basic tool for pattern recognition, and a series of descriptive statistics can be used
in heuristic fashion to identify and map appropriate pattern solutions for any distribu-
tion: the log(%SSE), cluster centroids, and the RI&S. When plotted against the cluster
solution scale provided by k-means analysis, the log(%SSE) allows optimal clustering
solutions to be identified. Centroids locate the centers of clusters composing a solution
as x and y coordinates. A measure of the spread of clusters around their centroids is
provided by the RMS (root mean square) of the distance among cluster members and
the center point.
Cartesian coordinates serve as the raw data for the k-means analysis of Couche V
artifact distributions. A single run was carried out combining all six data classes into a
single point scatter. Thus, clusters identified by this analysis represent spatial accumula-
tions of objects regardless of their classification. Figure 3 shows the log(%SSE) plot for
this k-means run when from one to 15 clusters are considered. Also plotted on this
Figure are log(%SSE) values derived from k-means analysis of the same data after x and
y coordinates are randomized. As discussed by Kintigh & Ammerman (1982) the
log(%SSE) decreases more rapidly for observed data than for randomized data if the
observed distribution is clustered. Figure 3 shows that the Couche V materials have a
clustered pattern.
412 J. F. SIMEK

0.25+. . .
234567 8 3 IO II 12 13 14 5
No. of clusters

Figure 3. Log(%SSE) values plotted against the number of clusters produced by


k-means analysis of Couche V tool classes. The dashed line plots the random
distribution; the solid line plots observed values for the Couche V distribution.
The arrow points to an inflection in the plot for actual data at 9 clusters,
indicating an optimal clustering configuration at that level.

Inflections in the log(%SSE) curve indicate optimal clustering solutions for the
distributions under analysis (ibid. 45). For Couche V, several optimal solutions are
indicated, e.g. at two, nine and 11 clusters. One optimal solution, a nine cluster
configuration, will be used in this analysis of contextual effects. Other optimal solutions
might also be employed. Using the centroids and RMS measures produced with the nine
cluster solution, a map of the resulting pattern can be produced (Figure 4) and the
number of artifacts of each class can be counted for each cluster (Table 2).
The nine cluster k-means solution provides one set of spatial units for association
studies. The co-occurrence of artifact classes within the cluster pattern is assessed using
Kendalls tuu-B statistic, a nonparametric measure of correlation (e.g. Blalock, 1972:
41g-421). T&B is computed from the data in Table 2 by converting raw counts for each
artifact class in a cluster to their within-class rank values. For example, Cluster 1 has the
highest within-class count for both burins (N = 47) and truncated pieces (N = 6); both
classesreceive rank values of 1 in Cluster 1. Ranks are assigned by class, since this over-
comes varying absolute abundances among artifact classes (Simek & Larick, 1983). To
preclude variable autocorrelation produced by conversion to other measures, like
percentages, ranking is carried out on the raw counts (cf. Speth & Johnson, 1976).
Tau-B is particularly useful in situations, such as the present one, where there are tied
class ranks among clusters (Blalock, lot cit.). Correlations are computed between
artifact classes because it is interclass association specifically that is of concern here.
Table 3 shows the matrix of correlation statistics obtained among the six Couche V
artifact classes. Only two positive correlation coefficients are significant at the 0.05 level,
with the highest value between notched and miscellaneous retouched pieces. Burins and
notched pieces are also correlated, but their coefficient is of lesser magnitude than for
PATTERN AND CONTEXT IN SPATIAL ARCHAEOLOGY 413

4cxl 300 200 100


SagWi dlmenslon

Figure 4. Map of nine clusters generated by k-means analysis of Couche V tool


classes. Numbers identify the clusters as in Table 2. Circle show RMS radii for
clusters.

Table 2. Matrix of artifact class frequencies for Couche V nine cluster K-means
solution

Cluster number

Class 1 2 3 4 5 6 7 8 9

Endscrapers 3 5 I 5 3 2 9 2 1
Burins 41 26 38 45 22 20 14 24 20
Truncations 6 14 2 3 4 3 2 3
Retouched 4 6 15 14 5 9 2
Notches 2 5 6 6 I o 3 4 2
Bladelets 5 4 0 4 12 2 3 3 3

Total 61 41 62 61 42 32 37 44 31

notched and retouched pieces. This correlation table is analogous to the model tables for
interclass relations illustrated in Figure 1.
In summary, nine clusters or concentrations of material scattered over the Couche V
surface compose one optimal pattern solution. Within this pattern of material
accumulations, three artifact classes have significant spatial associations. Notched
pieces and miscellaneous retouched pieces have the strongest correlation, but notches
and burins are also associated within the cluster pattern.
Results like these are the usual products of intrasite spatial analysis in archaeology.
414 J. F. SIMEK

Table 3. Matrix of Kendalls tau-B sratistics for Couche V arrijacr class


associations compukd from Table 2. Sig@ican[ correlations (P~0.05) are
underlined. Sign$cancc tesfs are two-tailed

ChSS Endscrapers Burins Truncations Retouched Notches Bladelets

Endscrapen:
Burim - o-1494
Truncations -0.0318 0~0000
Retouched 0.0606 0.0870 -0.2772 -
Notches 0.3638 04348 - 0.4004 m
Bladelets -0.2188 0.2689 -0.2540 - 0.3638 PO.1213

The artifact pattern has been characterized and co-occurring sets of artifact classes have
been identified. But are these sets tool kits in the sense of independent associations
among object classes? Or might the observed associations relate to contextual
constraints imposed by features? To answer these questions, we must go beyond this
normally final stage of analysis.

Spatial Structure in the Element Set Features


Any overall patterning in feature distributions may affect the relations between
features and artifacts. Thus, patterns among the former are investigated before the two
data sets are compared. Relations among features are examined using euclidian
distances among their center points. The matrix of inter-feature distances is shown in
Table 4 (mean = 2.8544, S.D. = 1.4828). The mean distance between nearest neighbors
is 1.8069 m. Examining the feature distribution (Figure 5) it appears that a linear pattern
may characterize this configuration. The features seem to run from the shelter wall in the
northeast towards the shelter opening in the southwest, with the Focal Hearth occupying
the position nearest the wall. A regression of x coordinates against their corresponding
y values at each feature center confirms this impression. The following regression line is
defined :
y = 576.2+ 1.7720
The fit of data points to this line has an Y coefficent of 0.965 (P <O.OOl). The regression
line is plotted over the feature distribution on Figure 5 to compare the observed and
abstracted linear patterns for the features.
For Couche V, some artifact classes are associated over space while some are not, and
the pattern of features is strongly linear. Thus we might expect a complex relational
pattern between features and the artifact classes already shown to associate over space.
It is the examination of this relationship that constitutes contextual integration.

Spatial Relations Among Artifacts and Features


For the most part, procedures for assessing the nature and degree of relations between
element sets have remained visual in archaeology. The inspection technique, where
spatial analysis of artifacts is performed and results compared to feature positions
without using quantitative methods, has produced some important results (e.g.
Leroi-Gouhan & Brezillon, 1966, 1972; Movius 1966; Speiss, 1979; Binford, 1983).
However, post hoc intuitive correlation between spatial patterns of artifacts and
features rarely effects the interpretation of artifact class associations. Typical interpreta-
tions rely on the Activity Model : associations among artifact classes are tool kits that
may have been used together near the feature (cf. Binford, 1983: 147-148). Because of
PATTERN AND CONTEXT IN SPATIAL ARCHAEOLOGY 415

Table 4. Matrix of interfeature distances calculatedfrom center offeature to center


offeature. All distances are expressed in meters

Feature 1 2 3 4

1. (Focal hearth)
2. (Burnt limestone platform) 2.5821 -
3. (Filled rock circle) 3.7011 1.1327
4. (Empty rock circle) 52839 2.7209 1.7059 ~

400 300 200 loo


Sagittal dimension
Figure 5. Couche V feature plan. Shaded areas show feature locations and size.
Line indicates regression through feature centers.

this presumptive bias, the technical problem of assessing contextual effects between
element sets has been ignored.
As was stated earlier, one possible approach to examining associations between
element sets involves comparing observed relations to random ones. In such an
approach, a relevant random pattern is required. Monte Carlo simulation can be used
to model random distributions for such comparison (Aldenderfer, 1981). Monte Carlo
is, in a practical sense, the simplest simulation procedure, but it has had little application
in archaeology despite its obvious utility in certain problem situations (see Kintigh,
1984). The goal of the simulation is to determine, based on series of sampling trials from
a random probability model, the parameters of a random distribution of specific form.
The Monte Carlo method is preferred here because, as will be seen, it allows aggregate
measures for random populations to be generated that are directly comparable to the
data at hand and controls for the possible stochastic production of random but atypical
distributions.
416 J. F. SIMEK

For my purposes, an added advantage of the Monte Carlo technique is that it


provides control over variance in aggregate measures when populations of different sizes
are considered. Regardless of the number of artifacts involved in computing mean
artifact to feature distances, the average will be a parametric value. However, even
though variance about the mean is also parametric, it typically will be greater for a series
of small point scatters than for large ones. That this is true will be demonstrated below.
Thus, many simulations must be carried out to accurately determine the expected
variance for random scatters of different numbers of points.

The procedure
The Monte Carlo simulations used here involved the generation of 50 random point
distributions for each artifact class size (e.g. 50 scatters of 43 points, corresponding to the
number of endscrapers), within a simulation frame analogous to the Couche .V surface.
Fur each random pattern, the mean distance is calculated between all points and
stipulated locations corresponding to the centers of the four features. The average mean
for all 50 runs, along with the S.D. among means, is then determined. These means-of-
means are the basic data for comparisons with observed artifact-to-feature distances.
Means-of-means are useful values because they are parametric and normally
distributed (Blalock, 1972: 180-184). These characteristics allow simple and robust
statistical tests to be used in comparing the random patterns with the observed relations
between artifacts and features.
I will use the Students t-test to test the hypothesis that the random and observed mean
distance values are the same. In these tests, the variance around both means must be
known. As was stated above, the variance within random patterns of different numbers
of points decreases as the number of points increases. Figure 6 shows the S.D. of values
around the simulated mean for each of the four feature locations in relation to the
number of points in the scatter. In all four cases, clear, dependent relations are evident.
Correlations between standard deviations and sample size in all cases exceed -0.96
(P<0401).

L 2

0 50 100 150 200 250 1i0


No. of objects in o class

Figure 6. Standard deviations around simulated means-of-means plotted with


size of point scatter for each of four feature locations.
PATTERN AND CONTEXT IN SPATIAL ARCHAEOLOGY 417

Table 5. Students t-values and associated degrees of freedom resulting from


comparisons between simulated random distribution means and observed mean
distances from artifacts to features. Signifcant t-values are underlined P ~0.05).
Signl$cance tests are two-tailed

Feature 1 Feature 2 Feature 3 Feature 4

Class t d.f. t d.f t d.f t d.f

Endscrapers 0.37 42 0.65 42 0.30 42 0.84 42


Burins * 256 -0.18 256 0.78 256 1.64 256
Truncations -090 27 -0.33 27 -0.27 27 0.77 27
Retouched -351 42 933 42 - 1.56 42 2 42
Notches -2-65 28 -381 28 ~ 1.79 28 1.32 28
Bladelets -0.62 35 - 0.62 35 -0.93 35 -0.33 35

Table6. SortedandsimpliJiedmatrixofrelationsamong features andartifact classes


in Couche V

Feature

Class 1 2 3 4

Bladelets I I
Endscrapers i ::
Truncations :: i
Burins + :: ::
Notches + + :: I
Retouched + + / -

+, Indicates that a class concentrates on a feature; -, indictes segregation


between a class and a feature; /, indicates that artifacts in a class are randomly
scattered around a feature.

Acceptance of the null hypothesis in the r-tests indicates no association between an


artifact class and a particular feature. Rejection of the null hypothesis means that the
observed distribution is not randomly scattered around the feature. A nonrandom
relation can have two forms. If the observed point pattern is concentrated on a feature (is
associated with it) then the r-value will be significant and negative. If the artifact
distribution avoids the feature (is segregated from it) the r-statistic will be significant
and positive. For this reason, two-tailed t-tests are performed in each comparison.

The tests
Monte Carlo simulation runs and r-tests were carried out for each of the six Couche V
artifact classes. The resulting t-values and associated degrees of freedom for all tests are
presented in Table 5. Table 6 simplifies these data in a sorted matrix showing the kind and
direction of relation between each artifact class and each feature. This table is analogous
to the model artifact-feature relation tables presented in Figure 1.
Of the six artifact classes analyzed, three have no significant relation with any feature.
These are endscrapers, bladelets, and truncated pieces. Their lack of association with
features parrallels their lack of interclass association over the surface. These classes
conform to the independently structured distributions presented in Figure la. Three
artifact classes have spatial associations with features : burins, notches, and retouched
pieces. These classes also have associations with each other. The strongest interclass
418 J. F. SIMEK

association is between retouched and notched pieces. These two classes are associated
with Features 1 and 2. Moreover, retouched pieces are segregated from Feature 4. In
light of the feature pattern, this suggests compaction of this class towards one end of the
feature line. Burins and notches are both associated with Feature 1. These two classes
are also correlated over the cluster pattern. In short, for burins, notches, and retouched
pieces, interclass associations are isomorphic with artifact class-feature associations.
This situation conforms to conditions modelled in Figure Id and represents contextual
constraint.
To summarize these analyses, the spatial associations among artifact classes present
on the Couche V surface may not be independent of feature locations. Based on previous
discussion, we cannot tell if interclass associations represent co-occurring sets of artifacts
or the association of individual classes of artifacts with certain features. Thus, activity
interpretations are unwarranted.

Discussion and Conclusions


I have attempted to show that analysis of spatial associations solely among artifact
classes, by failing to account for possible contextual effects, may detect statistically
significant sets of classes having no independent causal relationship. These sets may be
misinterpreted in terms of activities based on what is, in effect, incomplete investigations
of spatial structure. Because I have examined only one possible contextual relation,
causal interpretation cannot be offered here. Before such interpretation is warranted,
other relevant relations must be integrated into the analysis (e.g. the relation between
features and the internal rockshelter topography). However, the demonstrated overlap
between artifact class and feature associations in Couche V precludes the immediate
interpretation of co-occurring artifact sets as activity units. It also points out that global
measures, such as the ones used here, may not be sensitive enough to local spatial process
to be helpful in detecting the kinds of activity sets that archaeologists assume caused the
spatial patterns they identify.
It seems clear that the analysis only of artifact distributions on a global scale cannot be
considered sufficient research to warrant activity based interpretation. In short, the
traditional approach to spatial analysis in archaeology, which concentrates on site-wide
patterning among artifacts, needs to incorporate potential contextual effects on artifact
class associations.
Based on the preceeding analyses, we can say that Couche V from Le Flageolet I has
the following spatial structure. The four features on the site surface have a linear pattern.
Three artifact classes are associated with features and three are not. For the three
associated classes, the Focal Hearth serves as the primary context. For two of these
classes, the Burnt Limestone Platform is a second contextual feature. The only
associations among artifact classes that occur over the surface are among those classes
having feature associations. The features, then, serve as the context of artifact
association. We do not know if the artifact associations are independent or are due to the
association of individual classes with the features. Thus, global-scale tool kits cannot be
inferred despite the presence of co-occurring artifact classes. Activity sets must be sought
on a local scale, and this clearly requires a new analytic approach to the whole problem.
In conclusion, I have tried to show the importance of determining if context constrains
artifact distributions. The multi-stage strategy used here has been successful in a single
case study for defining relations within and between different data sets. Analysis was
oriented towards defining the structural characteristics of the distribution, i.e. the
principles of organization underlying observed patterns. Such an approach results in the
interpretation of spatial patterning in terms of depositional contexts rather than specific
activities. Some may consider such resolution of interpretation less interesting than
PATTERN AND CONTEXT IN SPATIAL ARCHAEOLOGY 419

traditional analytic goals. I must counter that more detailed interpretation is clearly
unwarranted. It should be stressed that the actual techniques used in this study are first
attempts at solving the problem and certainly require refinement or replacement. If,
however, the research discussed here stimulates further work on contextual integration
in spatial archaeology, then it will have achieved its ultimate goal.
Acknowledgements
I would like to thank J.-Ph. Rigaud for access to the superb data from Le Flageolet
and for his help and guidance over the years. R. R. Larick ran the k-means program at
SUNY-Binghamton. A. J. Ammerman, R. C. Dunnell, and D. K. Grayson read and
commented on previous drafts of this paper. I also thank K. W. Kintigh and an
anonymous reviewer for the Journal of Archaeological Science who contributed helpful
observations on the submitted manuscript. All results are, of course, own responsibility.
Research funds were generously provided by the L. S. B. Leakey Foundation.
References
Aldenderfer, M. (1981). Computer simulation for archaeology: an introductory essay. In (J.
Sabloff, Ed.) Simulations in Archaeology. Albuquerque: University of New Mexico Press.
Ammerman, A., Kintigh, K. & Simek, J. (1983). Recent developments in the application of the
k-means approach to spatial analysis. Paper presented at the Fourth International Flint
Symposium, Brighton, England.
Binford, L. (1978). Dimensional analysis of behaviour and site structure : learning from an Eskimo
hunting stand. American Antiquity 43,33&361.
Binford, L. (1981). Behavioural archaeology and the Pompeii premise. Journal of
Anthropological Research 37, 195208.
Binford, L. (1983). In Pursuit of the Past: Decoding the Archaeological Record. New York:
Thames and Hudson.
Blalock, H. (1972). Social Statistics. New York: McGraw-Hill.
Clarke, D. (1977). Spatial information in archaeology. In (D. Clarke, Ed.) Spatial Archaeology.
London : Academic Press.
Daccy, M. (1973). Statistical tests of spatial association in the locations of tool types. American
Antiquity 38, 320-328.
Gould, R. (1980). Living Archaeology. New York: Cambridge University Press.
Hietala, H. &Stevens, D. (1977). Spatial analysis: multiple procedures in pattern recognition
studies. American Antiqutty 42, 539559.
Hodder, I. & Okell, E. (1978). An index for assessing the association between distributions of points
in archaeology. In (I. Hodder, Ed.) Simulation Studies in Archaeology. New York:
Cambridge University Press.
Kershaw, K. (1964). Quantitative and Dynamic Ecology. London: Edward Arnold.
Kintigh, K. (1984). Measuring archaeological diversity by comparison with simulated assemblages.
American Antiquity 49,4454.
Kintigh, K. & Ammerman, A. (1982). Heuristic approaches to spatial analysis in archaeology.
American Antiquity 47,31-63.
Laville, H., Rigaud, J.-Ph. & Sackett, J. (1980). Rockshelters of the Perigord: Geological
Stratigraphy and Archaeological Succession. New York : Academic Press.
Leroi-Gourhan, A. & Brezillon, M. (1966). LHabitation Magdalenienne No. 1 de Pincevent pres
de Montereau (Seine-et-Marne). Gallia Prhhistoire 9,263385.
Leroi-Gourhan, A. & Brezillon, M. (1972). Fouilles de Pincevent : essai dAnalyse ethnographique
dun habitat Magdalenien. 7eme Supplement a Gallia Prehistoire. Paris: C.N.R.S.
Movius, H. (1966). The hearths of the Upper Perigordian and Aurignacian horizons at the Abri
Pataud, Les Eyzies (Dordogne), and their possible significance. American Anthropologist 68,
296325.
Pielou, E. C. (1977). Mathematical Ecology. New York: John Wiley and Sons.
Rigaud, J.-Ph. (1969). Note preliminaire de la stratigraphie du gisement du Flageolet I
(commune de Bezenac, Dordogne). Bulletin de la Socihth Prehistorique Francaise 667375.
420 J. F. SIMEK

Rigaud, J.-Ph. (1978). The significanceof variability among lithic artifacts: a specificcasefrom
southwestern France. Journal ofAnthropological Research 34299310.
Rigaud, J.-Ph. (1982). Le Paleolithique en Perigord: Les DonnePs du Sud-Ouest Sarladais et leur
Implications. Thesede Doctorat dEtat es Sciences.Universite de Bordeaux, No. 737.
Schiffer, M. (1976). Behavioural Archaeology. New York: Academic Press.
Simek, J. (1984). A K-means Approach to the Analysis of Spatial Structure in Upper Palaeolithic
Habitation Sites: Le Flageolet Zand Pincevent 36. B. A. R. International Series.Oxford: B. A. R
(in press).
Simek, J. & Larick, R. (1983). The recognition ofmultiple spatial patterns: a casestudy from the
French Upper Palaeolithic. Journal of Archaeological Science 10, 165-l 80.
Simek, J. & Leslie, P. (1983). Partitioning Chi-square for the analysis of frequency table data : an
archaeological application. Journal of Archaeological Science 10,7%85.
Sonneville-Bordes,D. & Perrot, J. (1954). Essaidadaptation des methodes statistiques au
paleolithique superieur: premiers resultats. Bulletin de la Socikte Prehistorique Francaise
50-51.
Speiss,A. (1979). Reindeer and Caribou Hunters: an Archaeological Study. New York: Academic
Press.
Speth, J. &Johnson, G. (1976). Problems in the use of correlation for the investigation oftool kits
and activity areas.In (C. Cleland, Ed.) Cultural Change and Continuity, pp. 3557. New York:
Academic Press.
Whallon, R. (1973). Spatial analysisof occupation floors: the application ofdimensional analysis
of variance. American Antiquity 38,266278.
Yellen, J. (1977). Archaeological Approaches to the Present. New York : Academic Press.

You might also like