Journal of Archaeological Science: Reports 4 (2015) 310–319

Contents lists available at ScienceDirect

Journal of Archaeological Science: Reports

journal homepage: http://ees.elsevier.com/jasrep

Color in historical ceramic typologies: A test case in statistical analysis of

replicable measurements
John M. Chenoweth a,⁎, Alan Farahani b
a
University of Michigan-Dearborn, Dept. of Behavioral Sciences, 4012 CASL, 4901 Evergreen Rd., Dearborn, MI 48170, USA
b
University of California-Los Angeles, Cotsen Institute of Archaeology, 308 Charles E Young Drive W, Los Angeles, CA 90024, USA

a r t i c l e i n f o a b s t r a c t

Article history: Three types of white-bodied, non-vitreous earthenwares distinguished primarily (though not exclusively) by
Received 6 October 2014 color and commonly known as creamware, pearlware, and whiteware are some of the most frequently encoun-
Received in revised form 21 September 2015 tered artifacts on eighteenth and nineteenth century archeological sites. However, problems exist both in the def-
Accepted 23 September 2015
inition of these types and the interpretation of their meanings. This project has applied reflectance
Available online 2 October 2015
spectrophotometry to collect replicable, highly-precise color data on a large sample of this material in order to
Keywords:
address the nature of the color differentiation and the reality of the types themselves. Linear Discriminant Anal-
Ceramics ysis (LDA) was used to assess the degree to which the visual intuitions of the archeologists making ware identi-
Color fications could be predicted by the repeatable spectrophotometer values. We do not suggest a method of proving
Historical archeology identifications “correct” or “incorrect” but analyze the nature of this attribute of the typology itself. This analysis
Reflectance spectrophotometry quantifies the level of uncertainty inherent in the types as they are currently used, addresses a longstanding ques-
Linear Discriminant Analysis (LDA) tion about illumination during identification, and discusses one possible concrete implication of these results.
Refined earthenwares This article also discusses color measurement in general, its use in archeology, and evaluations of the Munsell
BLUE mean ceramic dating (BLUE MCD)
system.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction
This project applies the replicable measurement of color to the eval-
uation of three of the most well-known ceramic types in historical ar-
cheology: the non-vitreous, white-bodied earthenwares distinguished
primarily by color and commonly known as creamware, pearlware,
and whiteware. Almost ubiquitous on sites connected to worldwide
trade routes from the mid-eighteenth to the mid-nineteenth centuries,
these three wares are some of the most useful, most discussed, and pos-
sibly some of the most controversial in archeological analysis (Majewski
and O'Brien, 1987; Miller, 1987; Miller, 1993; Miller and Hunter, 2001).
This project has sought to clarify some of these controversies
through the application of replicable, highly precise color observations
taken using a reflectance spectrophotometer. Multivariate statistical
analyses were used to evaluate the resulting data set. Rather than sug-
gesting that archeologists use this equipment to classify individual
sherds or correct visual intuitions, the goal of this work has been to as-
sess the nature of color as an attribute of these types. In effect, we assess
the extent to which replicable, independent measurements conform to
a priori groups made by researchers. This question has real implications
⁎ Corresponding author.
E-mail addresses: jmchenow@umich.edu (J.M. Chenoweth), alanfarahani@ucla.edu
(A. Farahani).
for some of the most commonly applied techniques of historical arche-
ology, such as mean ceramic dating.
In broader terms, the use of instruments designed to make precise,
replicable observations of many kinds has been growing in archeology
thanks to new instrumentation and the decreasing cost of technology.
Better observations in archeological data can enable more effective
communication and comparisons, though always ultimately to address
broad anthropological questions. That is, while archeology is and always
will be an interpretive endeavor, we can ensure that our interpretations
are based on understandings, definitions, and observations on which we
can agree or upon which our best data converge.
2. Creamware, pearlware, and whiteware: problems and questions
Ivor Noël Hume, with whom most discussions of historical artifacts
begin, suggests that the “most important development of the eigh-
teenth century” in the context of ceramics was the appearance of a
“thin, hard-firing, pale-yellow or cream-colored earthenware” that is
“now universally known as creamware” about 1762 (Noël Hume,
1969: 124). Creamware was largely supplanted by pearlware in the
1780s and 1790s, and their bodies are nearly identical but the latter
has a small amount of cobalt in the glaze, giving a whiter, bluer, or
greener appearance (Barker, 1991:167, Majewski and O'Brien, 1987:
118). Large amounts of pearlwares were produced in England and
exported from about 1775 through about 1830, when they were

http://dx.doi.org/10.1016/j.jasrep.2015.09.015
2352-409X/© 2015 Elsevier Ltd. All rights reserved.
 J.M. Chenoweth, A. Farahani / Journal of Archaeological Science: Reports 4 (2015) 310–319
overtaken in popularity by whitewares, which have a more truly white
appearance. This was a gradual process, involving the slow whitening of
pearlware, as whiteware too sought to imitate porcelain which was
growing whiter itself (Miller, 1993).
Because the history of the development of these wares is documented
relatively precisely and they represent horizons, largely supplanting one
another sequentially (although their production did overlap) these
wares have been of substantial interest for dating, especially in tech-
niques such as mean ceramic dating (South, 1977). Though made in
England, these wares are connected to the rise of industrialism and
world trade, and so can also be used to study these important develop-
ments, being found worldwide.
2.1. Differentiation of the types
Despite all the interest in these types, the distinctions between them
have been difficult to define. Noël Hume identifies creamware's glaze as
being “yellow or green in the crevices” where the glaze pools (Noël
Hume, 1969: 130) and Price terms it “deep green or yellow-green”
(Price, 1979: 10). Pearlware's glaze is “blue in crevices of footrings and
around handles” such that it “can readily be distinguished from late
creamware” (Noël Hume, 1969: 130). Lofstrum, Tordorf and George,
however, suggest that the glaze of pearlware is “uniformly greenish”
where evenly-spread (Lofstrum et al., 1982: 6) and Price calls it “an
overall blue or blue-green cast” (Price, 1979: 14). Whitewares are char-
acterized by the body color, since they may also have slight bluing in
crevices, but appear to be completely colorless in body (Miller, 1993;
Price, 1979: 13). However, these relatively simple distinctions become
difficult to implement in practice. Miller (1980: 2) points out that per-
sonal opinions about the extent of the “bluing,” for instance, coupled
with the inherent difficulty of describing and communicating color
leads different analysts to make different determinations.
Other attributes of ceramics are, of course, used in their classifica-
tion, including decoration and form. Price underlines the importance
of considering the color of underglaze decoration as well when defining
typologies (Price, 1979: 15). Miller distinguishes between “pearlware”
and a variant known as “China Glaze” in that while both have a blue
tint to the glaze, the latter has decoration and forms in imitation of Chi-
nese patterns (Miller, 1993). Ceramicist David Barker describes how
some elements of decoration could be used to identify work of particu-
lar potters, although he also suggests problems with this procedure due
to molds and patterns being shared among manufacturers (Barker,
1991: 169).
In this study we examine only the question of body or glaze color but
do not mean to argue that it is the only factor used in classification. We
focus on this attribute because in many archeological applications body
color is used alone as it is often not possible to examine other attributes,
since not every sherd will be decorated or identifiable as to form. In
some distinctions, glaze color is the only distinguishing attribute; for in-
stance, David Barker writes that “apart from the blue tint of the glaze,
there is virtually nothing to distinguish [pearlware] from contemporary
creamwares” (Barker, 1991: 167). For these reasons, we feel that inves-
tigating the attribute of color more or less in isolation is informative,
even if it is not the only attribute relevant in practice.
2.2. Ware-based types and etic classifications
Miller (1980) and Majewski and O'Brien (1987) critique ware-based
schemes such as the division of creamware, pearlware, and whiteware,
because they omit the importance of decoration and because they are
“unwieldy,” since the differences they describe are difficult to define
consistently. The latter note that “critical terms either remain undefined
or are applied inconsistently or incorrectly, thus making it difficult to
use the data from such a study for comparative purposes” (Majewski
and O'Brien, 1987: 105). It seems likely that color and its communica-
tion is one of the critical terms Majewski and O'Brien consider here.
311
The problematic result is a series of typologies which are mutually-
unintelligible, and a host of publications which use the same terms in
different ways.
This project addresses the issue of consistency through instrumenta-
tion (discussed below), but the debate over these types is deeper than
simply the ability to communicate color. Miller and Hunter (2001)
have pointed out that the terms used by archeologists do not corre-
spond to any group that would have been meaningful to contemporary
potters, suggesting that merchants and customers distinguished be-
tween ceramic types more by decoration than what archeologists
term “ware.” Miller argues that “there is no way of knowing if the
archeological definition of pearlware is the same as that of 19th century
potters and merchants” (Miller, 1980: 2). There are, in effect, two paral-
lel typologies (Miller, 1993).
This discussion returns to the more than half-century old debate be-
tween James A. Ford and Albert Spaulding about archeological typolo-
gies in general. Spaulding (1953) saw types as being the result of
decisions made by people in the past: their choice of temper in ceramic
production was one “mode” while their choice of rim type was another.
Collectively (normatively) a particular group would agree on a particu-
lar series of choices among modes, and therefore create similar pottery.
That is, types were the outgrowth of culture, and would themselves
have had meaning to past peoples. Ford (1954) held the opposite
view: that types were imposed in the present by the archeologist, and
that past peoples would not have necessarily recognized our types as
important. While the debate did not remain a focus during the 1960s
rise of processual theory, it is nonetheless possible to hold that both
kinds of types exist: emic ones, having had meaning to past peoples,
and etic ones, imposed from the present by the analyst for particular
ends.
For certain questions, such as those of economic scaling, it is entirely
appropriate to focus on the emic typologies of potters and customers,
whose ideas of different groups of ceramics and their prices directly re-
late to buying habits. But in the analysis of archeologically-recovered
materials, there is sometimes a mismatch between what is needed to
assign a piece to an emic type and what information is available. For in-
stance, Miller discusses the difference between what past potters called
“China Glaze” and “Pearl White” as being one of decoration and form
(Miller, 1993), but for undecorated fragments there would be no way
to make a determination. Emic understandings can even run contrary
to traditional archeological (etic) classifications: “We need to keep the
intent of the potter in mind. If it was to produce a whiteware, then the
vessel should not be classified as pearlware because of a small amount
of cobalt used to achieve a white appearance” (Miller, 1993). We
agree that the intent of the potter is an important area of inquiry, but
it will not always be clear from archeologically-recovered materials.
Certainly it would be pessimistic to assume that because archeologists
are not always able to access such emic typologies, that we therefore
can say nothing about what is found based on its material properties.
Though etic types based on material observations—such as color
—cannot always be connected to contemporary (emic) terms, they
still speak to past practices and are worthy of consideration. The
fact that contemporaries may not have been aware of some of
these material distinctions does not make them meaningless. Trace
element analysis is a thoroughly modern, etic analysis but it provides
substantial new information relevant to reconstructing past lives
and is valid even if those making artifacts were unaware of these el-
ements' existence. Moreover, any disjunctions between etic and
emic offer substantial ground for interpretative understanding, for
if we recognize a physical distinction that was ignored in the past,
or determine that a difference considered important emicly was in
fact not clearly distinguished, then we are understanding something
about culture. But in order to recognize these distinctions, we need
thorough studies of both the emic and etic. The present study should
be classed among the latter, but should not be interpreted as a gen-
eral argument against using the former.
312 J.M. Chenoweth, A. Farahani / Journal of Archaeological Science: Reports 4 (2015) 310–319
3. The measurement of color
While color is one of the most obvious attributes of artifacts—indeed,
of any object—it is also one of the more difficult to measure and commu-
nicate. This section provides only the barest of introduction to this field;
see (Berger-Schunn, 1994; Billmeyer and Saltzman, 1981; Hunt, 1991;
Hunter and Harold, 1987) for more. While color is usually understood
as being a result of the wavelengths of the light, the perception of
color is more complex, involving a three-part relation of object, light,
and observer.
Most visible light is composed of many different wavelengths in dif-
ferent proportions. Spectrophotometry is the analysis of light, in this
case visible reflected light, through the creation of spectra describing
the reflectance percentage of light at different wavelengths. The use of
an instrument to make these observations systematizes two of the
three parts of color perception—the lightsource (the “illuminant”) and
observation—making measurements dependent only on the object
and thus replicable.
Spectra, while perhaps the most accurate way to record the color of
an object, are difficult to work with, and hard to relate to human percep-
tion. These can be converted, with some caveats, to “tristimulus” values:
three-part measurements which approximate the amounts of primary
colors which can be mixed to achieve a certain color (Billmeyer and
Saltzman, 1981: 38). Tristimulus values are similar to the Munsell hue,
chroma, and value measurements that are more familiar to
archeologists but easier to use for statistical purposes. One of the most
common systems for representing color in tristimulus values is known
as the “CIE L*a*b*” system, which was recommended by the Commission
Internationale de l'Eclairage (International Commission on Illumination)
in 1976; it is generally abbreviated CIELAB. In this system, the “light-
ness” of a color is given by “L*” such that a perfect white has the value
of 100 and a perfect black zero (Hunt, 1991: 72–3), “a*” represents the
position of a color between red and green, while “b*” indicates the posi-
tion between blue and yellow (Hunter and Harold, 1987: 199–200).
This system is used here, since calculating differences between two
colors in the CIELAB colorspace is the same as doing three-
dimensional geometry.
4. Color measurement in archeology
4.1. Munsell
Munsell values need no introduction to most archeologists as they are
widely in use to record soil colors in the field. The Munsell system de-
scribes color based on hue (“that quality of color which we describe by
the words red, yellow, green, blue, etc.”) value (lightness or darkness),
and chroma (difference from a gray of the same value) (Billmeyer and
Saltzman, 1981: 28). Some studies have extended Munsell to other
areas of archeological analysis, such as preserved fabric (Paul, 1990),
and several have also applied it to ceramics. Most notably Prudence Rice's
seminal work, Pottery Analysis, uses Munsell to characterize ceramic color
(Rice, 1987: 339–43) and Anna Shepard suggested it for producing “reli-
able comparisons” in her classic Ceramics for the Archaeologist (Shepard,
1956: 107, 113). Nance reports using Munsell to classify precolumbian ce-
ramics by paste color, and applies this to his analysis of shifts in ceramic
color over time and inter-level mixing (Nance, 1976), while Hammond
(1964) used Munsell to assess Nabataean pottery and outlines the
Munsell values for particular types. In historical archeology, in addition
to Lofstrum et al. (1982), discussed above, which cites Munsell colors
for both bodies and decorations, Sussman used Munsell values in describ-
ing the blue decoration (not the blueish body) of pearlware, suggesting a
change from a dark blue in the eighteenth century to a more “purple-
toned” blue by the 1820s (Sussman, 1977: 108).
While it may be true that the “Munsell system is the most widely
used approach by archeologists to analyze color in artifacts” (Giardino
et al., 1998: 478), many authors have also pointed to substantial
problems with the system and its use. Considering the importance of
color to ceramic classification in general and the necessity that such
classifications be repeatable, Giardino and colleagues suggest that dif-
ferences in skill, lighting, and viewing geometry often result in inconsis-
tent color measurements (Giardino et al., 1998: 477–8). Bishop and
colleagues also point to variability in observation conditions as well as
the problem of precision (capturing the many shades which fall be-
tween Munsell chips) and reproducibility (Bishop et al., 1988: 328).
Hammond asserted that the use of Munsell on ceramics “is decidedly
an inferior method” due to variability in human perception
(Hammond, 1964: 259). Hammond actually suggests the use of a spec-
trophotometer but laments that one was not available for his study
(1964: 265).
The most often cited problem with Munsell is inconsistency. Frankel
had different people identify the same sherds repeatedly in different
conditions, showing that not only do color determinations differ be-
tween people, but also that determinations made between one day
and the next by the same person may vary, in the end suggesting that
Munsell notations provide a “false sense of accuracy” (Frankel, 1980:
36). The reliability of human-generated color-values was tested more
recently by Giardino and colleagues, who compared human-generated
color values of 52 Mississippian-period ceramics to those produced by
a spectroradiometer (similar to a spectrophotometer) and a low-cost
color sensor. They concluded that, “there can be considerable variation
in the perception of an object's color among observers” especially in
the determination of hue (Giardino et al., 1998: 480–1), possibly ac-
counting for the difference of opinion noted above in section 2.1 as to
whether pearlware is “blueish” or “greenish.”
It should be noted that none of the scholars cited above who have
used Munsell are unaware of its problems, and most make allowances
for these issues in the particular contexts of their own work. This
paper is not intended as a wholesale critique of these studies, and several
include suggestions to correct or account for some of these issues.
Gerharz and colleagues note that the surface textures and the homoge-
neity of samples complicate color determination judgments and suggest
the use of a “field of color,” comprised of several Munsell color chips, in-
stead of the choice of a single Munsell color as a partial corrective
(Gerharz et al., 1988: 91). Stanco and colleagues also note the problem
of surface textures along with that of patina and body shape, all of
which alter color perception of ceramics (Stanco et al., 2011: 338–9).
They propose a system for determining Munsell color through digital
images which they suggest will counter some of this inconsistency.
4.2. Mechanical measurement: spectrophotometers and colorimeters
Surprisingly, considering their relative accessibility and precision,
spectrophotometers and colorimeters (a similar device producing only
tristimulus values, often at a lower cost) have been little used in arche-
ology. Thompson and Jakes (2002) used a spectrophotometer as part of
a study of textile dyeing, exploring colorants suggested by archeological
and ethnohistoric evidence, and Strudwick (1991) used a colorimeter to
record colors as part of the documentation of Egyptian tomb paintings
in The Valley of the Nobles.
Two earlier studies have employed color measurement equipment
to analyze archeological ceramics, although both focus only on lower-
fired, pre-historic wares. In an exploratory study, Giardino and col-
leagues analyzed 52 sherds of Mississippian ceramics and compared
the results to human-generated color identifications. They also sug-
gested that there are “exciting prospects for expanding this type of anal-
ysis into the nonvisible (i.e., infrared, thermal, and ultraviolet) portions
of the electromagnetic spectrum” (Giardino et al., 1998: 478). Bishop
and colleagues used a Minolta chromameter to assess firing variations
through color on Hopi yellow-firing ceramics as part of an effort to cre-
ate more comprehensive analytical groups (Bishop et al., 1988: 328).
Because the examined samples had already been determined to have
similar iron content through compositional analysis, the variations in
 J.M. Chenoweth, A. Farahani / Journal of Archaeological Science: Reports 4 (2015) 310–319
color were thought to result from firing, and the recorded colors were
used to argue that ancient potters had a greater degree of control over
ceramic appearance through firing than previously thought.
Problems noted with this equipment generally center around its
cost. This remains an issue, with the best equipment costing several
thousand dollars. However, Giardano and colleagues' study just men-
tioned also evaluated an inexpensive, easily available color sensor and
found “good agreement between the reflectance or color measured by
the two instruments over the entire visible spectrum” including in the
analysis of ceramics (Giardino et al., 1998: 480). This opens the possibil-
ity of using less-expensive equipment for measurements which are still
far more replicable than the human eye. In addition, as discussed below,
the present study does not suggest that it is possible or even desirable
for every archeologically-recovered piece to be measured with this
equipment. At the moment, the potential for this type of analysis lies
in targeted studies aimed at clarifying specific questions and describing
the nature of larger classes of material.
5. Materials and methods
To this end, this project utilized a Konica Minolta CM-700d spectro-
photometer (loaned for the purpose by Konica Minolta Sensing US) to
take color readings of several hundred archeologically-recovered ce-
ramic sherds. Ceramics for the project were from the collections of Inde-
pendence National Historic Park in Philadelphia and were recovered
from late 18th to mid-19th century archeological contexts as part of
the National Constitution Center excavations in 2000–2003. All mea-
surements were taken on glazed portions of undecorated sherds or
areas of decorated pieces not immediately adjacent to decorations. Al-
though areas where glaze has pooled are often used in identification,
these are not present on most sherds and the goal of this study was to
evaluate the color of pieces overall, so areas of pooled glaze such as
footrings and drips were avoided as well. Sherds severely discolored
or weathered from their time in the ground or entirely covered in dec-
oration were not analyzed.
Measurements were taken originally from 699 sherds, but as the
shape of a ceramic surface may affect the readings collected by
refracting reflected light (although our results suggest that the differ-
ence is negligible) the analyses described below use only readings
from the slightly-convex sides of sherds, probably the exteriors of ves-
sels. The final data set includes 477 different pieces of ceramic, 143 of
which had been previously cataloged by project archeologists as
whiteware, 137 as pearlware, and 197 as creamware. Those doing the
original cataloging were either National Park Service staff or material
culture specialists employed by URS Corporation or Kise, Straw, and
Kolodner, contract firms hired by the Park Service. All contract firm per-
sonnel were vetted by NPS employees for relevant expertise and their
catalog entries were supervised and checked by the Park Service. All
Park Service employees relevant to this work held at least an MA degree
with a specialization in historical archeology.
Although it has long been suggested that color for these vessels may
vary across a single piece, within the confines of a two- or three-square-
inch sherd, repeated measurements produced remarkably consistent
readings. For instance measuring different parts of the same side of
one sherd of Creamware three times produced readings with a variance
of 0.28, and five measurements of another had a variance of 0.43, far
lower than the variance for the types overall. Five repeated measure-
ments on the same area of one sherd produced a variance of just
0.009. Issues with illuminants will be discussed below, but all results
considered here (other than where specifically indicated) are calculated
with CIE Standard Illuminant D65, which approximates natural
daylight.
Spectra were converted into tristimulus values using Konica
Minolta's SpectraMagix NX software, so that each reading consisted of
three numbers which could be analyzed using three-dimensional ge-
ometry in the CIELAB colorspace. The readings for each ware can be
313
given a centroid, a center-point to the three-dimensional cloud they
compose, and a variance, the average Euclidian distance from that
point; these are given in Table 1. All other statistical calculations were
made in R 3.1.1 using the “lda” function found in the package “MASS,”
and all visualizations were made using the package “ggplot2.” The R
code used to iteratively repeat Linear Discriminant Analysis as well as
many of the specifications and outputs of the model are provided as
supplemental material.
5.1. Linear discriminant analyses
To understand the nature of these types, data for these 477 sherds
were analyzed using multiple iterations of Linear Discriminant Analysis
(LDA), with the groups generated by prior archeologists as the categor-
ical outcome variable (“ware”), and the 477 L*a*b* values as the predic-
tors. LDA is a multivariate statistical technique used to maximize the
separation of two or more a priori (or “given”) categorical groups
using continuous multivariate predictors through the maximization of
within and between-group variances (Izenman, 2008: 237–238). The
linear combinations of the predictor variables are called discriminant
functions, and the relative contributions of these standardized variables,
or coefficients, are used to calculate discriminant functions (Kovarovic
et al., 2011: 3008). These discriminant functions can then be used to
predict group membership based on the existing data, along with the
posterior probabilities of group membership, and these results can be
compared to the original classification in a classificatory table (or “con-
fusion matrix”). With this method it is possible to assess whether the
variables of three-dimensional color readings can be used to achieve
separation between a priori groups, that is, groups whose membership
has already been independently assigned—in this case the groups of
creamware, pearlware, and whiteware assigned by archeologists.
A low proportion of overlap between assigned versus predicted
group membership might suggest that there are no distinguishable
groups for those variables or that the variables characterize other
groups, if any. In the present study, a low proportion of overlap would
suggest that the typology under consideration might not have indepen-
dent empirical foundation using color as the main attribute, and that the
within and between-group variation is too high for the ceramic group-
ings to be of analytic use: in other words, the mechanically-measured
color of these sherds would have little to do with their classification
based on archeologists' visual intuitions. A high proportion of overlap
would suggest that the chosen variables are the most parsimonious to
separate the a priori groupings, and thus that the data represent highly
identifiable (i.e. separable) groups that are a function of their visible
color. This does not imply, however, that this method can assess wheth-
er those making the identifications were “correct.” Instead, it provides a
way to determine whether certain variables can provide a consistent
classification for these types, i.e. the extent to which the observations
of color (L*a*b* values) can be classified into these ware groups.
LDA has wide application as a data classification tool both within ar-
cheology (Baxter, 1994; Kovarovic et al., 2011) and in other disciplines
(Hammer and Harper, 2006: 96–99, Legendre and Legendre, 2012: 616–
631), and its use in archeology has risen considerably since 2002
(Kovarovic et al., 2011: 3007). LDA makes several assumptions
concerning the underlying distributions of the variables, including
Table 1
Centroids and variances for each analyzed type.
Centroid L* Centroid a* Centroid b* Variance
Creamware 197 89.44 −0.84 13.31 2.49
Pearlware 137 86.12 −1.40 9.48 3.43
Whiteware 143 90.03 −0.31 8.99 2.13
All Sherds 477 88.66 −0.84 10.91 3.67
314 J.M. Chenoweth, A. Farahani / Journal of Archaeological Science: Reports 4 (2015) 310–319
multivariate normality, homogeneous within-group covariance matri-
ces, identical and independently distributed variables, and little to no
collinearity between variables. LDA also generally performs better
when the number of observations exceeds the number of groups and
variables.
5.2. Model training
One of the more important recent critiques of the use of LDA in ar-
cheology is by Kovarovic et al. (2011), who, among other issues, point
to the necessity of cross-validation within a data set, in order to prevent
model over-fitting (2011:3009). Model over-fitting occurs when a par-
ticular model cannot generalize to novel or unseen data because the
original model fits the specific characteristics of the training data set
too closely. They note that there are two popular methods for
preventing model over-fit. The first is the “hold-one-out” test in which
a variable portion of the data set is reserved for training and the remain-
der is used for prediction. “Training” is a process of calculating the dis-
criminant functions based on a sample of the data and applying the
results to the independent, withheld data to test the reliability of the
functions. The second is the “leave-one-out” or “jackknife” approach
where one observation is left out, the discriminant functions are then
re-calculated, and the process is repeated for every observation. The
present study employs both methods in order to address the issue of
model over-fitting.
The “confusion matrices” (the comparison of the archeologist-classi-
fied to LDA-predicted wares) presented below (Tables 2 and 3) are
based on the jackknife method. In this case, the jackknife method is pre-
ferred as it is “a more conservative approach with lower but more real-
istic success rates” (Kovarovic et al., 2011: 3009). The plots of the first
two discriminants presented below (Figs. 1–4) are based on the “hold-
one-out” training method of LDA. In this case, 10% of sherds (or about
47 sherds) were withheld iteratively without replacement from a ran-
dom sample of observations and used as training data. The 10% thresh-
old was chosen as it is conservative and attempts to predict the majority
of the data set classifications from a small subset of it. The remaining
data (90%) were then used to test the accuracy of the LDA predictions
versus archeologists' predictions, resulting in 10 independent predic-
tions on randomly drawn subsets of the same data. To ensure the reli-
ability of these model accuracy assessments, the random observations
chosen for training were permuted 1000 times for each of the 10-
prediction averages (see supplemental data). Both the jackknife method
and the “hold-one-out” method produce very similar results, reinforc-
ing the reliability of accuracy estimates between archeologists' ware
classification and the classification of the model.
6. Results
The collected data were used to address two main questions: the im-
portance of the type of illumination used when making the distinctions,
a point made central many discussions of of Munsell, and the nature of
the groups themselves (the extent to which L*a*b* measurements of
color correspond to the groups of creamware, pearlware, and
whiteware ceramics).
Table 2
Confusion matrix using illuminant D65 based on the jackknife method.
Predicted by LDA
Cream Pearl White % Mismatch
Assigned by Archeologist Cream 177 1 19 10.2%
Pearl 26 96 15 30.1%
White 7 3 133 6.9%
Overall classification error 15.3%.
Table 3
Confusion matrix using illuminant A.
Predicted by LDA
Cream Pearl White % Mismatch
Assigned by Archeologist Cream 177 1 19 10.2%
Pearl 27 95 15 30.7%
White 9 2 132 7.7%
Overall classification error 15.3%.
6.1. The reliability of types
Fig. 1 is a plot of the first two discriminant functions and shows that
L*a*b* values can be successfully used to separate each of the three ware
groups and predict a priori group membership. In both (Fig. 1A) and
(Fig. 1B), the letters “C,” “P,” and “W” represent Creamware, Pearlware,
and Whiteware. In (Fig. 1A) the letters represent the predictions of the
LDA model, while in (Fig. 1B) the letters represent the sherds as origi-
nally classed by archeologists. The biplot of the two discriminant func-
tions illustrates two separate conclusions: first, that there is
discrimination of groups by instrumentally-measured color. Second,
that there is overlap between these groups, especially at the edges of
the group centroid means (Fig. 3). The discriminant functions derived
from the “hold-one-out” training method closest to the permutation av-
erage (see supplementary material) classified 83% of the wares to the
ware specified by the archeologist using the L*a*b* values as predictor.
That is to say, archeologists' visual intuitions were coincident with the
groupings based on the spectrophotometer readings about 83% of the
time. If these archeologists' classifications had not been consistent
with L*a*b* values, a much smaller percentage of wares would have
been predicted. In Fig. 2, the letters C, P, and W represent sherds
where the LDA and archeologists agreed on a classification, while the re-
maining numbers in bold show sherds predicted by LDA to be of one
ware based on mechanical measurement but assigned by archeologists
to another group.
While the group centroids are separable, the multivariate 95% confi-
dence ellipses, shown in Fig. 3, illustrate considerable overlap between
groups. As a whole, ware groups are difficult to separate due to a large
degree of internal variance, and as can be seen in Fig. 4, which shows nu-
merous discrepancies between archeologists' classification and LDA
prediction at the outer edges of overlap between the groups. The extent
to which these outliers are classification cataloging errors versus being
the result of actual variation present in ware groups may be one contrib-
uting factor here, and is indicated in cases where a particular sherd lies
well within the 95% confidence ellipse of one type only but has been
classified by archeologists to a different type. In this study, a visual in-
spection of these figures suggests that perhaps a dozen sherds might
be considered cataloging errors of the type which could be eliminated
by spectrophotometric classification, if such analysis was practical.
However what would be called cataloging errors are not the cause of
most of the disagreements between LDA and archeologists' classifica-
tions shown here. It is clear from these data that some sherds cannot
be classified as falling clearly into one or another of these groups. That
is, the colors of some of these sherds fall between groups, blurring the
groups at the edges (Fig. 4). It is important to note that this is not neces-
sarily a “mistake” by the original catalogers, and that these pieces have
not necessarily been “mistyped.” Rather, some sherds may not be clearly
assignable to one group or another on the basis of color. While other at-
tributes can sometimes be used to make determinations, as discussed
above, the identification of some undecorated sherds without a priori
information about dating or manufacturer and which fall between
these LDA groupings is fundamentally ambiguous. The ambiguity of
these types has long been recognized by ceramicists and has led to the
calls to abandon type-based assessment of these materials discussed
above. In counterpoint, our results also suggest that types based on
 J.M. Chenoweth, A. Farahani / Journal of Archaeological Science: Reports 4 (2015) 310–319 315

Fig. 1. Plot of first two discriminant functions of the “hold-one-out” training model closest to the permutation average, where the symbols in (A) represent the ware classification predicted
by the LDA model, and (B) where the symbols represent archeologists' ware classification. The x and y axes represent the first two linear discriminants.
316 J.M. Chenoweth, A. Farahani / Journal of Archaeological Science: Reports 4 (2015) 310–319

Fig. 2. Distribution of discrepancies between LDA and analysts categorizations, where C(reamware), P(earlware), and W(hiteware) represent the concordance between LDA and a priori
ware assignation, and the remaining numbers represent, via the key, cases where the archeologist's classification differed from the predicted (by LDA) ware.

glaze tinting are not complete fantasies either; real patterning does
exist when the color of these ceramics is assessed with replicable
measurements.
The data collected here allow us to put a number on this ambiguity.
Using the L*a*b* values as predictor, the discriminant functions of the
more conservative “leave-one-out” jackknife method classified 85% of
the wares, or 406 out of 477, to the ware specified by the archeologist.
That is to say, archeologists' visual intuitions can be explained as coinci-
dent with more repeatable methods about 85% of the time. In theory,
approximately 15% of the ceramics known as cream, pearl, and
whitewares should equally fall into this “blurry” category, which prob-
ably includes a small number of potential cataloging errors as well as
sherds whose color falls between two or more groups.
Interestingly, this blurriness does not apply to all types equally. As
shown in Table 2, only 7% of whiteware and 10% of creamware, respec-
tively, were classified differently by LDA and the archeological techni-
cians, showing that archeologists' identification and L*a*b* values are
tightly matched. More problematic is that 30% of pearlware was not pre-
dicted by LDA as belonging to archeologists' initial ware classification.
The variability is probably based on the fact that the colored glaze that
is key to defining pearlware is created by the inclusion in the glaze of co-
balt, which could be added in any amount, resulting in a greater or lesser
degree of “bluing.” Some different categorizations might be a result of
other factors besides glaze tinting being involved in identification of
types by archeologists, particularly for pearlware. For instance, Miller
argues that blue-tinted glazes should be distinguished into at least

Fig. 3. Ellipses of the confidence level of each group at the 95% level based on a multivariate t-distribution — the symbols represent the predicted associations based on the LDA model.
 J.M. Chenoweth, A. Farahani / Journal of Archaeological Science: Reports 4 (2015) 310–319
Fig. 4. Ellipses of the confidence level of each predicted group at the 95% level based on a multivariate t-distribution — the symbols represent the discrepancies between the LDA model and
analysts' prior categorization.
three (emic, potters') categories: pearlware, “China glaze” imitating
Chinese forms and decorations, and even whitewares if the tint is min-
imal, the piece late, and the intention of the potter is to make a whiter
ware not to imitate porcelain (Miller, 1993). However if this is so, it
calls into question the overlap of these two attributes in the typology,
color and decoration; this remains to be assessed in another study.
In addition, we suggest that the influence of other attributes on
archeologists' categorizations, and thus possibly on disjunctions be-
tween the L*a*b* values and archeologists classifications, is likely to be
minimal for two reasons. The first is that this study used primarily
(though not exclusively) undecorated individual sherds rather than
complete vessels, where glaze color was the only attribute available
for classification between the three types considered here. The ques-
tions of time and potters' intent were not available to either the cata-
logers or the LDA. Second is that when decoration or form were
considered by archeologists, China Glaze was distinguished only as a
variant of pearlware by the cataloging protocol in use by NPS. Though
incommensurate with emic potters' typologies, this makes sense from
an archeological perspective, as undecorated sherds could be consid-
ered part of either type. Therefore the classifications by LDA and
archeologists considered here should be largely commensurate, and
the results given above indicative of color variation.
6.2. Metamerism and illuminants
Color scientists often remind us that the phenomenon of color is
made up of three parts: light, object, and observer. Differences in per-
ceived color can be related to any of these. “Metamerism” occurs when
changes in illuminant (light source) or observer make two objects
which appeared to have the same color no longer look the same. This oc-
curs because color spectra are more complex than either tristimulus
values such as L*a*b* numbers or human perception: “even though
two objects may have different spectral reflectance curves, they match
if their CIE tristimulus values are the same for a given illuminant and ob-
server” (Billmeyer and Saltzman, 1981: 53). So two objects, say a ceramic
sherd and a Munsell chip, may appear to be identical under one light and
quite different under another.
As a result, lighting conditions for color comparisons are given sub-
stantial discussion in color science works (Hunter and Harold, 1987)
and have been noted by archeologists concerned with color analysis
(Bishop et al., 1988: 328, Giardino et al., 1998: 477–8, Shepard, 1956:
111). This project was in a position to analyze whether or not the
317
illuminant used in making observations has a significant effect on the
results of those observations in the case of these particular ceramic
types. The “leave-one-out” jackknife method was used for cross-
validation on the same 477 measurements using a different illuminant
than the D65 one used in the above discussions: CIE Illuminant “A”
which approximates incandescent artificial light. Again, the model pre-
dicted the ware to which a given piece would be classified given its
L*a*b* values, this time a set generated using illuminant A, and com-
pared these to the actual classifications made by archeologists. The re-
sults are given in Table 3, and show a net change of only three sherds
of 477 (0.6%) classified differently as a result of the different illuminant.
Only 6 individual sherds in total were classified differently depending
on the illuminant (1.2%). This is substantially smaller than the overall
classification error of 15% inherent in these materials established in
the last section. This suggests that illuminant may have far less of a
role in the identification of these particular types than thought. Put sim-
ply, the additional precision of uniform lighting is outweighed by the in-
herent uncertainty of these particular types.
7. Implications and conclusions
It is quite clear that “wares are not static entities” (Majewski and
O'Brien, 1987: 106) and that the differences between creamware,
pearlware, and whiteware are the “result of an evolution of one type
out of another” (Miller, 1980: 2). This project has clarified the extent
and nature of the variation inherent in these types as a result of this pro-
cess of change, showing that fully 85% of the time sherds were repeated-
ly classifiable based on color. As was noted above, this project has not
attempted to determine if those making the initial identifications were
“correct” but rather to assess the extent to which a consistent classifica-
tion can be made at all for these types: the extent to which observations
of color produce identifiable groups. Our results show that there is in-
herent ambiguity, but we have quantified the ambiguity of these types
and this allows its implications to be considered.
One example is in mean ceramic dating (MCD), one of the most com-
mon uses of the typology considered here. This is the calculation of an
average production age for a ceramic assemblage based on the frequen-
cy and mean production date of each type included (South, 1977). The
MCD technique is, it should be noted, a rough measurement which
has been much discussed in the field (see Chenoweth, in press).
Under certain circumstances and with certain assumptions, this sug-
gests a possible mean occupation date for a site. In other cases,
318 J.M. Chenoweth, A. Farahani / Journal of Archaeological Science: Reports 4 (2015) 310–319
Table 4
Mean ceramic date calculations (totals based on percentages do not total precisely due to rounding to whole sherds, since fractional sherds are not meaningful units).
Archeologist's classifications LDA classifications
Creamware (1762–1820) 197 210
Pearlware (1775–1830) 137 100
Whiteware (1820–1900) 143 167
Total 477 477
MCD 1815.7 1818.1
BLUE MCD 1807.3 1808.9
mismatch between MCD and known occupation dates suggests some-
thing important about human behavior (Turnbaugh and Turnbaugh,
1977). Table 4 shows MCD calculations on the sample used in this
study derived from archeologist's classifications and from the assign-
ments made by LDA based on the spectrophotometric data. The dates
are taken from South (1977) but that for pearlware is updated to reflect
more recent work by Miller (1987) and Barker (1991).
We also incorporate a newer MCD technique based on finding the
best linear unbiased estimator, producing what is called the “BLUE
MCD.” This method was developed in archeological analysis by Fraser
Neiman and Karen Smith (Neiman and Smith, 2005) and has been
used to counteract the problem of skewing due to uneven manufactur-
ing ranges in ceramic types by Galle (2010, 2011) and Arendt (2013).
However, there is an error in the published version of the equation in
all of these sources. The correct version of the equation is given below.
 2
t
∑i¼1 mi pi si1=6
¼ ness, porosity, and chemical composition: these are all areas for further
MCDBLUE  2 :
t
∑i¼1 pi si1=6
It should be noted that, as far as we can determine, there are no
problems with other aspects of the data or calculations in the cited stud-
ies, but simply that the equation was misprinted. The technique works
by incorporating into the equation the overall range of the dates of pro-
duction, so that those types with extremely long manufacturing spans
such as tin-enameled wares (often called “delft”) play less of a factor
than more tightly-dated materials.
The LDA model presented above was similar to the classifica-
tions of archeologists 85% of the time, but about 15%—mainly
pearlware—disagreed. For this sample, based on the separately-
measured ambiguities of each type (creamware 10.2%, pearlware
30.1% and whiteware 6.9%) and the archeologists' original identifi-
cations, we can expect about 404 sherds to be confidently identi-
fied, while about 73 will fall somewhere in between groups. If all
of these un-certain sherds were considered to be creamware, the
chronologically earliest type, the MCD would be about 1813 and
the BLUE MCD 1805, while if these were all counted as whiteware,
the latest type, the BLUE MCD would be closer to 1813 and the
traditional MCD 1823..
This difference is minor, but it is not suggested here that the spectro-
photometer data reveal a “correct” classification, and therefore that one
of these dates is correct and the other an error. Rather, this study has
provided a way to independently assess the ambiguity inherent in the
attribute of color, and this MCD data suggests the implications of that
ambiguity. In other words, the actual MCD for this assemblage could
fall anywhere between these dates not because of errors in identifica-
tion but because of the ambiguity of these types themselves.
Of course, this ceramic dating analysis is merely illustrative and not
generalizable until it is considered in the context of actual test cases,
since it is based on an artificially selected assemblage. For many studies,
the potential variation in MCD or other analyses caused by this issue
may be insignificant. Other studies will want to consider the possibility
that their classifications of these materials have this ambiguity, which in
turn produces variability in MCD and other calculations based on this
Confident arch. IDs Questionable arch. IDs Possible extreme old Possible extreme new
177 20 (10%) 250 177
95 42 (30%) 95 95
132 11 (7%) 132 205
404 73 477 477
1812.88 1823.46
1805.1 1813.2
typology. This result may also be another reason to move away from
the traditional calculation of MCDs and rely instead on a range of
dates, as has been occasionally suggested (e.g. Bartovics, 1981), or con-
duct multiple versions of the MCD analysis and report and compare all
these results as a group (Chenoweth, in press).
Beyond the implications for studies of creamware, pearlware, and
whiteware in particular, this project has aimed to contribute to a
broader discussion of how archeological data are collected and analyzed
in the twenty-first century, particularly for historical archeology. This
analysis, ultimately, raises as many questions as it answers, as all scien-
tific endeavors should. The question of how color varies across a single
complete piece, how much it varies over time and the precise chronol-
ogy of these changes (as with the lessening of the “creaminess” of
creamware or the whitening of pearlware), the relationship of body
color measured here with decoration and form, the relationship be-
tween emic and etic typologies, and a comparison of these results
with other materials science techniques, such as those measuring hard-
analysis raised by what we have presented here.
While archeology is and always will be an interpretive endeavor,
when addressing these questions and others we can ensure that our in-
terpretations are based on data created by understandings, definitions,
and observations which can be repeated and are clearly communicated.
Instruments manufactured to conduct repeatable observations, here of
color, can be used to create observations that are comparable across
researchers.
Acknowledgments
The authors would like to thank Jed Levin and the staff of Indepen-
dence National Historic Park for access to their collections. Thanks also
to Douglas Mooney for information about cataloging. Funding for data
collection was provided by fellow research funds from the Thinking
Matters (formerly “Introduction to the Humanities” or IHUM) program
of Stanford University. We would also like to thank Dan Foley and
Konica Minolta Sensing, US, for the loan of equipment for the analysis,
Dustin Stansbury for his statistical advice, and Hans Bernard for his
thoughts on this manuscript. Thanks also to Jillian Galle and Alan
Wiggins for discussions of the BLUE MCD technique.
References
Arendt, B., 2013. Return to Hopedale: excavations at Anniowaktook Island, Hopedale, Lab-
rador. Canadian Journal of Archaeology/Journal Canadien d'Archéologie 37, 302–330.
Barker, D., 1991. William Greatbatch: A Staffordshire Potter. Jonathan Horne Publications,
London.
Bartovics, A.F., 1981. The Archaeology of Daniels Village: An Experiment in Settlement Ar-
chaeology (Ph.D. dissertation) Department of Anthropology, Brown University, Prov-
idence, RI.
Baxter, M.J., 1994. Stepwise discriminant analysis in archaeometry: a critique. J. Archaeol.
Sci. 21, 659–666.
Berger-Schunn, A., 1994. Practical Color Measurement: A Primer for the Beginner, a Re-
minder for the Expert, Wiley, New York.
Billmeyer Jr., F.W., Saltzman, M., 1981. Principles of Color Technology. Second ed. John
Wiley & Sons, New York.
Bishop, R.L., Canouts, V., Atley, S.P.D., Qöyawayma, A., Aikins, C.W., 1988. The formation of
ceramic analytical groups: Hopi Pottery Production and Exchange, a. C. 1300–1600.
Journal of Field Archaeology 15, 317–337.
 J.M. Chenoweth, A. Farahani / Journal of Archaeological Science: Reports 4 (2015) 310–319
Chenoweth, J.M., 2015. Collecting, Updating, and Building on the Classics, in: Chenoweth.
In: M., J. (Ed.), The Historical Archaeology Laboratory Handbook. Society for Historical
Archaeology, Germantown, MD. in press.
Ford, J.A., 1954. On the concept of types. Am. Anthropol. 56.
Frankel, D., 1980. Munsell colour notation in ceramic description: an experiment. Aust.
Archaeol. 33-37.
Galle, J.E., 2010. Costly signaling and gendered social strategies among slaves in the eigh-
teenth-century chesapeake: an archaeological perspective. Am. Antiq. 75, 19–43.
Galle, J.E., 2011. Assessing the Impacts of Time, Agricultural Cycles, and Demography on
the Consumer Activities of Enslaved Men and Women in Eighteenth-Century
Jamaica and Virginia. In: Delle, J.A., Hauser, M., Armstrong, D.V. (Eds.), Out of Many,
One People: The Historical Archaeology of Colonial Jamaica. The University of Ala-
bama Press, Tuscaloosa, pp. 211–242.
Gerharz, R.R., Lantermann, R., Spennemann, D.R., 1988. Munsell color charts: a necessity
for archaeologists? Aust. J. Hist. Archaeol. 6, 88–95.
Giardino, M., Miller, R., Kuzio, R., Muirhead, D., 1998. Analysis of ceramic color by spectral
reflectance. Am. Antiq. 63, 477–483.
Hammer, Ø., Harper, D.A.T., 2006. Paleontological Data Analysis. Blackwell Pub, Malden,
MA.
Hammond, P.C., 1964. The physical nature of nabataean pottery. Am. J. Archaeol. 68,
259–268.
Hunt, R.W.G., 1991. Measuring Colour, 2nd ed. ed., E. Horwood, New York.
Hunter, R.S., Harold, R.W., 1987. The Measurement of Appearance. Second Edition. John
Wiley and Sons, New York.
Izenman, A.J., 2008. Modern Multivariate Statistical Techniques: Regression, Classifica-
tion, and Manifold Learning. Springer, New York.
Kovarovic, K., Aiello, L.C., Cardini, A., Lockwood, C.A., 2011. Discriminant function analyses
in archaeology: are classification rates too good to be true? J. Archaeol. Sci. 38,
3006–3018.
Legendre, P., Legendre, L., 2012. Numerical Ecology. Elsevier Science, Burlington.
Lofstrum, E., Tordoff, J.P., George, D.C., 1982. The seriation of historic earthenwares in the
Midwest, 1780–1870. Minnesota Archaeologist 41, 3–29.
Majewski, T., O'Brien, M.J., 1987. The use and misuse of nineteenth-century English and
American ceramics in archaeological analysis. Advances in Archaeological Method
and Theory 11, 97–209.
Miller, G.L., 1980. Classification and economic scaling of 19th-century ceramics. Hist.
Archaeol. 14, 1–40.
319
Miller, G.L., 1987. Origins of Josiah Wedgwood's “pearlware". Northeast Historical Archae-
ology 16, 83–95.
Miller, G.L., 1993. Thoughts Towards A User's Guide to Ceramic Assemblages, Part IV:
Some Thoughts on Classification of White Earthenwares, Council for Northeast His-
torical Archaeology Newsletter. 26, pp. 4–7.
Miller, G.L., Hunter, R., 2001. How creamware Got the blues: the origins of China glaze and
pearlware. Ceramics in America 2001, 135–161.
Nance, C.R., 1976. Artifact attribute covariation as the product of inter-level site mixing.
Midcont. J. Archaeol. 1, 229–235.
Neiman, F.D., Smith, K., 2005. How Can Bayesian Smoothing and Correspondence Analysis
Help Decipher the Occupational Histories of Late-eighteenth Century Slave Quarters
at Monticello? Paper presented at the Society for American Archaeology conference,
Salt Lake City, UT. [Available online in poster form: http://www.monticello.org/sites/
default/files/media/temp/Bayesian%20Smoothing%20and%20Correspondence%
20Analysis%20Poster_1.pdf].
Noël Hume, I., 1969. A Guide to Artifacts of Colonial America. Knopf, New York.
Paul, A., 1990. The use of color in paracas necropolis fabrics: what does it reveal about the
organization of dyeing, designing, and society? Natl. Geogr. Res. 6, 7–21.
Price, C.R., 1979. 19th Century Ceramics in the Eastern Ozark Border Region. Center for Ar-
chaeological Research, Southwest Missouri State University, Springfield, MO.
Rice, P.M., 1987. Pottery Analysis. University of Chicago Press, Chicago.
Shepard, A.O., 1956. Ceramics for the Archaeologist. Carnegie Institution of Washington,
Washington, DC.
South, S., 1977. Method and Theory in Historical Archaeology. Academic Press, New York.
Spaulding, A.C., 1953. Statistical techniques for the discovery of artifact types. Am. Antiq.
18, 305–313.
Stanco, F., Tanasi, D., Bruna, A., Maugeri, V., 2011. Automatic color detection of archaeo-
logical pottery with munsell system. Proceedings of the International Conference
on Image Analysis and Processing 16, 337–346.
Strudwick, N., 1991. An objective colour-measuring system for the recording of egyptian
tomb paintings. J. Egypt. Archaeol. 77, 43–56.
Sussman, L., 1977. Changes in pearlware dinnerware, 1780–1830. Hist. Archaeol. 11,
105–111.
Thompson, A.J., Jakes, K.A., 2002. Replication of textile dyeing with sumac and bedstraw.
Southeast. Archaeol. 21, 252–256.
Turnbaugh, W., Turnbaugh, S.P., 1977. Alternative applications of the mean ceramic date
concept for interpreting human behavior. Hist. Archaeol. 11, 90–104.