J Forensic Sci, January 2013, Vol. 58, No.

1
doi: 10.1111/j.1556-4029.2012.02293.x
Available online at: onlinelibrary.wiley.com

PAPER
ODONTOLOGY

H. David Sheets,1 Ph.D.; Peter J. Bush,2 B.S.; and Mary A. Bush,2 D.D.S.

Patterns of Variation and Match Rates of the

Anterior Biting Dentition: Characteristics of a
Database of 3D-Scanned Dentitions

ABSTRACT: An understanding of the variability of the anterior human dentition is essential in bitemark analysis. A collection of 1099 3D

laser scans of paired maxillary and mandibular arches were studied using geometric morphometric methods. Analyses were performed without
scale (shape only) and with scale (shape and size). Specimens differing by no more than experimentally obtained measurement error were
counted as matches, or as indistinguishable. A total of 487 maxillary (396 size preserved), 131 mandibular (83 size preserved), and one paired
dentition (two size preserved) matches were found. Principal component analysis and partial least squares revealed interpretable patterns of variation and covariation in dental shape, principally dominated by variation in dental arch width. The sensitivity of match rate to assumed degree
of measurement error was also determined showing rapid increases in match rate as measurement error increased. In conclusion, the concept of
dental uniqueness with regard to bitemark analysis should be approached with caution.

KEYWORDS: forensic science, forensic odontology, bitemarks, dental uniqueness, geometric morphometric analysis, three-dimensional
analysis

There have been a number of concerns raised about the discipline of bitemark analysis (15). It rests on the dual assumptions
that the arrangement of the teeth that produce a bitemark is
unique and that skin records sufficient detail to support statements of certainty about the relationship of the anterior dentition
to a given bitemark (1).
The entire issue of uniqueness as a testable concept in forensic
science is contentious (68). In forensic odontology, the definition of uniqueness can vary depending on whether it pertains to
victim identification or bitemark analysis. In victim identification, the combinations of restored (five surfaces possible), unrestored, and missing patterns of 32 teeth, coupled with root
morphology and trabecular bone configuration variation, combine to provide a powerful set of evidence for identification.
Typically in a bitemark, only the six anterior teeth of the maxillary and mandibular dentition leave an impression in skin, and
thus, many of the descriptors used in victim ID are not applicable (9). When the evidence is restricted to the biting surfaces of
the anterior dentition, there is a tremendous reduction in the
amount of information available. Even if all the characteristics
of the biting dentition are preserved in a bitemark, there is still
far less basis for comparison than in a case where X-rays and
other evidence of dental treatment of the entire dentition is available. Simple extension of methodology and logic of the parameters and evidence used in victim identification to study
1
Department of Physics, Canisius College, 2001 Main Street, Buffalo, NY
14208.
2
Laboratory for Forensic Odontology Research, School of Dental Medicine, SUNY at Buffalo, B1 Squire Hall, South. Campus, Buffalo, NY 14214.
Received 9 Sept. 2011; and in revised form 20 Nov. 2011; accepted 4
Dec. 2011.

60

uniqueness with respect to bitemarks is thus inappropriate, given

the drastic difference in the amount of evidence available.
There have been few studies examining the claim of dental
uniqueness with regard to bitemark analysis based on systematic
collections of specimens. Many of these have lacked a formal
statistical approach or have had a very limited sample size
(10,11). Other studies have explored the idea of dental similarity
rather than uniqueness and have defined an approach to describe
a cutoff point in which dentitions may be mathematically indistinguishable from one another (i.e., sufficiently similar) (1216).
The guidelines on bitemark analysis currently listed on the
website of the American Board of Forensic Odontology (ABFO)
(17) indicate that a bitemark analyst should state conclusions in
a case as belonging to one of the following categories: The
Biter, The Probable Biter, Not Excluded as the Biter, Excluded
as the Biter, or Inconclusive with all opinions stated to a reasonable degree of dental certainty. The first of these categories
does imply a strong identification of a particular biting dentition
to a particular bitemark.
One substantial concern raised when examining the ABFO criterion is whether or not there is a chance that two different individuals could produce a bitemark so similar that they could both
be placed into the first category: The Biter. Clearly, if there were
two individuals within this category in a given investigation,
there is a substantial potential for a miscarriage of justice. In
some sense, this is not a question of uniqueness per se, but of
the ability to effectively identify measurable differences between
two very similar individuals.
The criteria and terminology chosen for this project are that
two dentitions are said to match if the differences between
them are no larger than the differences which might occur if one
of the dentitions was measured repeatedly by the same operator,
2012 American Academy of Forensic Sciences

SHEETS ET AL.

the repeated measurement error (10,1216). If the difference

between the two is within the expected range of error in the
measurements involved, then one cannot claim that there is any
meaningful, measurable difference between the two biting dentitions (although it may be possible that differences exist in the
rest of the dentition, the posterior teeth, not recorded in a
bitemark).
The pragmatic definition used here of the term match does
not concur perfectly with the terminology used in the ABFO
guidelines, but it provides a working definition of what a match
is, specifically that there are no measurable differences between
the two sets of teeth given the measurement methods available. If
there are no individualizing characteristics detectable between
dentitions at some measurement resolution, then there seems little
possibility of assigning one and only one individual to the singular
category of The Biter. The limitations of measurement would suggest that an analysis of a bitemark could result in the identification
of more than one individual as The Biter.
The goals of this study were first to expand an existing database of 497 paired three-dimensional (3D) scans by over 600
additional specimens, to observe the impact of increased sample
size on match rate (14). In the earlier study of 3D-scanned data,
the comparison methods used did not take size into account, but
rather removed all size differences. In this study, both size-preserving and size-removing methods were used to determine the
impact of including size information on the match rate (13,16).
Second, the increase in match rate according to loss of measurement resolution was examined in order to understand the
sensitivity to increase in measurement error. There is a second
factor in bitemark analysis, that of transfer of the dental pattern
to skin. Distortion in the transfer is inevitable, such that precise
replication of dental shape will not be attained in the tissue (13).
This factor has the effect of reducing resolution and increasing
measurement error, potentially allowing many more dental
matches to occur. This study explores that concept by demonstrating the number of matches that occur when measurement
error threshold is increased.
Third, we examined several measures of size variation in the
data set, both the familiar arch width and the centroid size measure common in geometric morphometrics (18,19). Correlation
structures between the two measures, and between the maxillary
and mandibular dentition were examined.
Last, a range of statistical tools were used in conjunction with
landmark methods to produce a summary of the variation present
in this large data set, as an approach to understanding shape variation present in the population of dentitions. The same types of
methods were used in studying patterns of variation in bitemarks
and in 2D dentitions to allow systematic comparisons of patterns
appearing in both types of data (13). A database of this size
allows a more robust understanding of the types of variation that
are possible in the dentition and subsequently bitemarks.
Materials and Methods
All necessary Human Subject Institutional Review Board protocols were completed for this project and exemption was
granted. A collection of 1106 paired sets of 3D-scanned maxillary and mandibular dentitions were obtained from a commercial
dental laboratory, which produced these scans for the production
of occlusal guards (night guards). The 3D virtual models that
made up this data set were produced by laser scanning dental
stone models that had been mailed to the commercial laboratory.
The nominal spatial resolution of the laser scanner was 100

DENTAL MATCH RATE IN THREE DIMENSIONS

61

microns. Of these 3D models, 497 pairs were taken from an earlier data set (14) and 609 pairs were newly measured for this
study.
After initial matching studies, seven of these specimens were
identified as being repeated scans of the same individuals, which
were removed, leaving 1099 distinct individuals. This was an
interesting early result, as these blind repeats indicated the ability
of the shape comparison process to detect identical dentitions
and thus to serve as internal controls.
The dental models provided to the dental laboratory for 3D
scanning were from private practice patients from dental practices across the United States. All patient identifying information
was stripped from the data prior to any additional processing.
This was a sample of convenience and contained a wide range
of alignment patterns, from relatively straight to fairly
malaligned.
Shape Measurement
Landmark measurements (LM) were taken on the 3D-scanned
dentitions by placing 10 data points along the incisal edge of
each of the six anterior teeth, using the Landmark digitizing program (20). The dentitions were rotated in 3D space within the
software as landmark points were placed along the incisal edges
using changes in 3D perspective to verify accurate landmark
placement. This resulted in 60 total points along the incisal
edges of the anterior teeth of both the maxillary and the mandibular arches, recording a wide range of geometric information
including mesial to distal width, midpoints, angulation, incisal
edge shape of each tooth, malalignment patterns, relative tooth
heights and positions within the arch. It is emphasized that
throughout this study, the measurements refer strictly to the six
anterior teeth in the maxilla and mandible as this is the relevant
portion of the dentition with regard to bitemarks.
Procrustes and Procrustes Size-Preserving Superimposition
When comparing shapes represented as sets of landmark data,
it is first necessary to superimpose the sets of data. Much of the
work in geometric morphometrics has focused on separating the
shape of the object from the size of the object, so that the Procrustes method used in this field superimposes the data by
removing all differences attributable to size. For most forensic
work, size is regarded as a highly important aspect of the data,
and so a variant superimposition method, called Procrustes SizePreserving (Procrustes-SP) was used in this project as well as
Procrustes.
Procrustes methods and Procrustes-SP (also called size and
shape analysis) have been discussed extensively in numerous
other publications (11,1316,18,19,2123). Both methods simply
minimize the summed squared distances between corresponding
landmarks on two or more specimens by translating and rotating
the specimens to match one another as well as possible. The
Procrustes method used also requires that the centroid size of the
two specimens be scaled to one, so that the size of the two specimens also matches.
The use of a least squares criteria for the quality of this match
is a feature shared with many other statistical procedures including simple regression models for example, so that the procedure,
while somewhat complex, is not outlandish by statistical standards. The least squares approach also allows for the computation of a distance metric, a standard descriptor of how different
two shapes are, called the Procrustes distance, or the Procrustes-

62

JOURNAL OF FORENSIC SCIENCES

SP distance, depending on whether or not the size information is

preserved. These distance measures can be used either to characterize variability of measurements, to characterize a data set, or
to understand the measurement resolution of a system.
There is a wide range of methods of working with shape statistics (19,21,22), the Procrustes method used here is technically
termed a Partial Procrustes method, which is the generally preferred one of several alternative forms (21,23).
Match Criteria-Repeated Measures
Use of the match criteria that there are no measurable differences between specimens requires that we understand the variability possible in repeated measurements of a single specimen.
This variation sets the limits of useful resolution in a measurement system. The total error variance may be expressed as the
sum of contributions to the measurement error from a series of
sources: operator error in placing landmark, inter-operator error
because of differences between operators, error owing to the variability of taking a dental impression (including different materials and methods at this stage), error caused by the casting of the
impression (including variations in casting material) and variation introduced by the 3D-scanning process, including that
because of the mount of the cast in the scanning apparatus. The
total variance may thus be expressed as the summed random
contributions of each of these terms.
r2total r2operator r2among operators r2impression r2casting r2scanning
1
Not included in this model are systematic biases in any of
these terms, as such factors may be detected and removed. In
this study, all digitization was by a single operator, and thus, the
inter-operator term is zero. Multivariate analysis of variance
methods are generally used to decompose the total variance into
the contributions from each term, by systematically analyzing
repeated measurements in each category of the factors contributing to variation. As we were working with scans obtained from
a commercial laboratory intended for dental treatment, we did
not have the ability in this study to measure the contributions of
the impression, casting and scanning terms, which would have
required a series of repeating impressions, casts and scans of
several dentitions to estimate the contributions of these terms.
We do believe that these terms are relatively small relative to
the digization term (r2operator ), but it is clear that by neglecting
the impression, casting and scanning variances, we are underestimating the total variability in the measurements, so that our estimates of match rates will be an underestimate of the true match
rate obtained by considering all sources of variance.
To determine the measurement resolution (r2operator ), the scans of
three maxillary and three mandibular specimens were digitized 10
times each by the same operator, and the scatter of Procrustes distances and Procrustes-SP distances of each specimen about the
mean measurement of that specimen was determined. From these
distances, the average root mean square (RMS) distance scatter of
repeated measurement specimens about the mean was determined
(roperator). This value characterizes the measurement resolution
obtained in the study, expressed as a Procrustes or Procrustes-SP
distance. The RMS scatter is analogous to a standard deviation,
although it is not measured in the same way, nor does it have the
same statistical properties. However, experiments have empirically
indicated that 9396% of repeatedly digitized specimens lie within
twice the RMS scatter level.

In dimensionless Procrustes units, the RMS scatter in this data

was 0.02 for both the maxillary and the mandibular dentition, so
the matching criteria were twice this, 0.04. As Procrustes landmarks are scaled to the centroid size, this measurement indicates
our typical total error in measuring all 60 landmarks was about
2% of the total size of the dentition.
When using Procrustes-SP methods, the RMS scatter per LM
point was roughly 0.2 mm, still roughly 2% of the centroid
size, which is consistent with a broad definition of 1 mm width
of the incisal edges. A maximum error of 0.4 mm per point
would still leave each LM on the incisal surface in all cases.
There may of course also be a lateral component, a sliding of
points along the top margin of the tooth contributing to this
RMS scatter as well. The RMS scatter summed over all 60
points (all six anterior teeth) in the mandibular dentition was
3.1 mm and 4.28 mm in the maxillary dentition, and thus, the
net error was 2% of total size. Simply expressed, placement of
LM points on the incisal edges was precise, and the placement
error was small.
The substantial difference between the two arches is because
of the increased size of the maxillary structure. The error as a
percentage of size was similar. In considering this error, it is
important to note that the RMS scatter reflects only the error in
repeatedly placing landmarks on a single specimen, not errors
appearing in retaking a dental cast, or in reproducing the 3D
scan of that cast. Future work should ideally include an estimate
of the contribution of these error sources as well.
Arch Width and Centroid Size
The typical measure of size used in geometric morphometrics
is centroid size, the square root of the summed squared distances
of landmarks from their centroid, or average (18,19). Centroid
size is effectively a measure of how scattered the points in the
system are about their average and is advantageous in many calculations in that all landmarks are treated equally, none carrying
more weight than others in the calculations. However, centroid
size is difficult to conceptualize, so arch width was also calculated, measured from the central point of opposing canine for
each specimen from the landmark coordinates. The mean, minimum, maximum, standard deviation, and coefficient of variance
of both arch width and centroid size for the maxillary and the
mandible was calculated. Also correlation of arch width with
centroid size in the maxillary and mandible was calculated, as
was the correlation between maxillary and mandible in these two
variables.
Principal Component Analysis Patterns of Variation
Principal component analysis (PCA) is a mathematical way of
displaying the strongest independent patterns of variation in a
multivariate data set (24). These methods have a long history of
use in various forms of biometry and are typically used to determine patterns of variation in a complex data set (22). When used
with shape data, the resulting principal component (PC) axes
may be displayed as patterns of shape differences in the data set.
These axes are often drawn as deformation grids for 2D data
(22,23), showing the differences interpolated across the entire
structure (18). Plots typically show the changes in shape associated with positive scores on the PCA axis, negative scores imply
the opposite or inverse patterns of changes. In 3D data, it is easiest to simply plot the landmark positions of the mean shape and
the PC axis.

SHEETS ET AL.

PCA is a method of organizing data, rather than a statistical

test (24). It arranges the data into conveniently interpretable variables (PCA scores) along each PC axis. A large score indicates
a large difference from the mean form, along the pattern of
shape difference depicted by the deformation grid plot, or 3D
landmark plot of the change in shape along the axis. A negative
score indicates a set of changes opposite to those depicted by
the deformation grid. Each PC axis explains some percentage of
the total shape variation in the data set, ordered by decreasing
percentages of the total variance.
Partial Least Squares Patterns of Covariation
The partial least squares (PLS) method is a way of studying
the covariation of two blocks of multivariate data measured on
the same individuals (25). The method employs a singular value
decomposition of a covariance matrix calculated between the
two blocks of data (26). The result is a set of PLS axes which
explain or express covariation in much the same way the PC
axes express variation (2730). The PLS method results in
paired PLS axes, one for each block. In this study, the two
blocks are the 60 landmarks measured on the incisal edges of
the maxillary and mandibular dentitions of the individuals in the
collection.
Much like PCA, specimens have scores along each PLS axis,
and the scores along each axis have maximum possible covariation between the two blocks. In other words, the PLS scores in
one block are the best possible predictors of the scores (and thus
the shapes) in the other block. This approach is very similar to a
regression model, except that regression models typically assume
that one variable is predicted by the other. PLS methods explain
covariation, not causal, predictive relationships as appear in
regression models.
As PLS axes are ordered by the covariance explained, methods have been developed to use permutation methods to determine how many of the PLS axis are statistically meaningful
(25). In some cases, the axes may be statistically meaningful
(i.e., nonrandom), but explain so little covariation as to be meaningless.
Results
An initial determination of matches indicated five matches
based on Procrustes and seven matches based on Procrustes-SP.
Careful examination of the data indicated that seven of the specimens in these matches were among repeated 3D scans of the
same individuals, so these repeated scans were removed prior to
the final analysis, leaving 1099 specimens in the data set. Interestingly, the 3D scans of one particular individual were taken a
year apart and were still within twice the repeated measurements
RMS scatter range, indicating that in this case at least, the differences because of the intervening time, retaking of the cast and
rescanning of that cast were still within the measurement error.
This is some indication that the digitization process is the largest
contribution to the total variance and that the matching procedure does work on real data.
The matching rates based on Procrustes and Procrustes-SP
superimpositions were determined, and a substantial number of
matches were found for the maxillary dentition as compared to
the mandible. This rate further decreased when matches were
searched between the maxillary and mandibular sets (Table 1).
There was one remaining match in both maxilla and mandible of
two different specimens under the Procrustes protocol. This was

DENTAL MATCH RATE IN THREE DIMENSIONS

63

TABLE 1Match rates in the maxillary and mandibular dentitions, and in

both combined, based on Procrustes and Procrustes-SP superimpositions.
The matching criterion used was twice the RMS scatter of repeatedly measured specimens.
Procrustes
Number
of Matches
Maxilla
Mandible
Both

Number of
Individuals

Procrustes-SP
Number
of Matches

Number of
Individuals

1691
487
763
129
131
75
1
2
2
(603, 351 total comparisons, 1099 individuals)

396
83
4

RMS, root mean square; Procrustes-SP, Procrustes Size-Preserving.

the same pair seen in an earlier study using the same data set
(14). Under the Procrustes-SP procedure, there were two
matched pairs of different specimens, for a total of four individuals with a match. The two pairs under Procrustes-SP were not
the same individuals who matched under Procrustes.
The exploration of the dependence of match rate on measurement error is summarized in Table 2. Clearly, the rates of
matches in this large population increases rapidly as measurement resolution or repeatability decreases, as represented by the
increased RMS errors used to calculate this table.
The characteristics of the observed ranges of the measurement
of arch width and the centroid size are shown in Table 3. A plot
of the relationship between the maxillary and mandibular arch
width is shown in Fig. 1. The correlations between arch widths
and centroid sizes are listed in Table 4.
Patterns of Variation
The PCA, based on Procrustes superimposed data, reveals the
predominant patterns of variation in the anterior dentition of both
the maxillary and the mandibular arches. Plots of the measured
landmarks for the first six PC axes are shown in Figs 2 and 3.
The figures show the data in two distinct views: along the incisal
edges of the dentition (occlusal view) and in frontal view. These
views are shown in Fig. 4 with representative 3D scans. The
data in Figs 2 and 3 for the PC axes are shown as floating gray
crosses. The average dentition over the entire set is shown as a
series of black dots connected by solid lines along the incisal
edges. The crosses indicate the pattern of differences for specimens with positive scores along the axis; specimens with negative scores would have the reversed pattern. The PC axes are
independent axes that attempt to summarize variation; any real
specimen could have a complex mixture of two or more of these
patterns of variation.
The patterns shown in Figs 2 and 3 are readily interpretable,
representing easily explained patterns of variation in the data as
presented in Table 5. It should be noted that the data shown in
Figs 2 and 3 are 3D in nature and that rotation in 3D space permits a better understanding of the shape variance.
Partial Least Squares: Decomposition of Covariance
The PLS method displays patterns of covariation of two sets
of measurements, much in the way PCA displays patterns of variation. The PLS results reported here are thus describing patterns
of correlated or covarying departures of individuals from the
mean shape of both the maxilla and the mandible. Permutation
tests of the singular values and of the correlations between PLS
axes scores indicated that the first 11 PLS axes explained more

64

JOURNAL OF FORENSIC SCIENCES

TABLE 2This table shows the dependence of the number of matches and number of individuals matching in this data set, under both criteria, for the maxilla
and mandible, and both, as a function of increasing RMS scatter level, or decreased measurement resolution. The percentages are of the actual repeated
measure scatter, so 125% represents a 25% increase in measurement error (603,351 total comparisons, 1099 individuals).
Fraction of RMS Scatter
100%

Procrustes
Maxilla
Mandible
Both
Procrustes-SP
Maxilla
Mandible
Both

125%

150%

175%

Matches

Individuals

Matches

Individuals

Matches

Individuals

Matches

Individuals

1691
129
1

487
131
2

21,358
3119
526

873
500
246

85,282
18,543
6826

1007
769
579

184,407
5497
30,056

1061
925
813

763
75
2

396
83
4

9660
1658
166

826
451
144

39,854
9510
2056

1001
759
502

93,091
29,308
20,217

1059
931
773

RMS, root mean square; Procrustes-SP, Procrustes Size-Preserving.

TABLE 3Characteristics of the distribution of arch widths and centroid

sizes in this population. Dimensions are in mm. The coefficient of variation
is the standard deviation divided by the mean and allows comparison of the
variability despite differences in the mean or scale of the data.
Mandible

Maximum
Minimum
Mean
Standard deviation
Coefficient of variation

Maxilla

Arch
Width

Centroid
Size

Arch
Width

Centroid
Size

35.3
19.7
24.8
2.0
0.080

93.2
54.8
69.5
5.0
0.071

42.2
25.3
33.7
2.4
0.071

127.0
76.8
97.4
6.5
0.067

TABLE 4Correlations between size measurements (arch width and centroid size[CS]) in the maxilla and the mandible.
Variables

Correlation (R2)

Maxillary arch widthmandibular arch width

Maxillary CSmandibular CS
Maxillary arch widthmaxillary CS
Mandibular arch widthmandibular CS

0.4098
0.4525
0.8958
0.8871

between wide and narrow arches. The PLS analysis indicates this
pattern is strongly shared between the maxilla and mandible.
Discussion

FIG. 1Arch width in the maxilla versus arch width in the mandible. The
R2 value is 0.4098.

covariation than 95% of randomly permuted versions (25) of the

data set did. The first four PLS axes appear to be easily interpreted and explained a total of 81.9% of the total covariance
between the maxillary and mandibular measurements.
The PLS axes thus produced are shown in Fig. 5 and may be
interpreted in the same manner as the PCA axes (Table 6). Perhaps not surprisingly, the first PLS axes (65.4% of covariance)
describes the same pattern as seen in the first PC, a contrast

Initial inspection of the matching results revealed that there

were seven duplicate scans within the data set. This unexpected
finding served to demonstrate and validate the ability of the
matching process to identify the same individuals. The duplicates
served as a useful (if inadvertent) internal control for the study.
A principal effect of database expansion was the large number
of matches when the maxillary dentition was viewed separately
from the mandibular, a number that increased significantly from
the earlier study (14). However, the number of matches in the
combined data set did not increase when considering Procrustes
superposition. There was only one matching pair, which had
already been seen in an earlier study (14). The use of Procrustes-SP methods (13,16,21), which had not been available in the
earlier study, produced two matching pairs, of four individuals
total, which did not include the individuals matching under the
Procrustes criterion. The individuals matching under Procrustes
but not under Procrustes-SP evidently differed in size. The
matches under Procrustes-SP but not under Procrustes must have
been virtually identical in size, so that the increased variance
because of the uncertainty in size measurements overcame the
dissimilarities in shape that allowed us to distinguish them under
the Procrustes (shape only) procedure.
The effects of including the third dimension in analysis can be
seen as these are relatively low match rates, compared to those
seen in 2D data (12,15). It is difficult to make particularly strong
statements about the effectiveness of Procrustes methods relative
to Procrustes-SP based on these observed matches, owing to the
limited number of total matched pairs. Detailed 3D data did
reduce the match rate relative to that seen in 2D data, whether
scale information was included or not. The effect of including

SHEETS ET AL.

DENTAL MATCH RATE IN THREE DIMENSIONS

65

FIG. 2Principal component (PC) axes 16 (AF) of the maxillary dentition in frontal and occlusal views. The data for the PC axis variation are shown as
floating gray crosses. The average dentition over the entire set is shown as a series of black dots connected by solid lines along the incisal edges. The crosses
indicate the pattern of differences for specimens with positive scores along the axis; specimens with negative scores would have the reversed pattern. It should
be noted that the data shown in Figs 2 and 3 are 3D in nature and that rotation in 3D space permits a better understanding of the shape variance.

FIG. 3Principal component 16 (AF) for the mandibular dentition, in occlusal and frontal view.

66

JOURNAL OF FORENSIC SCIENCES

FIG. 4Views of representative 3D scans in the same orientation as the data presented in Figs 2 and 3.

TABLE 5Results of PCA of maxillary and mandibular anterior dentition.

PCA
Axis

Percentage of
Variance

Mandible: 54.0
Maxilla: 39.8

Mandible: 8.8
Maxilla: 12.6

Mandible: 6.7
Maxilla: 7.0

Mandible: 4.2
Maxilla: 5.2

Mandible: 3.9
Maxilla: 4.8

Mandible: 3.6
Maxilla: 3.8

Pattern Implied by the Axis

Arch width; positive scores indicate a relatively
wide arch, negative scores indicate a
relatively narrow arch, in both maxilla
and mandible
Positive scores indicate a labial displacement of
both central incisors, relative to a lingual
displacement of both lateral incisors, with a
slight labial shift of the canines in the
maxilla
This axis implies a pattern of leftright asymmetry
in both maxilla and mandible, although the
specific details of the asymmetry differ. In the
mandible, there is a bulge of all teeth to one
side, whereas in the mandible, all four incisors
shift in a line relative to the canines
Asymmetry in the location of the central incisors
in the mandible. One shifted lingually and the
other labially, while the adjacent lateral incisors
shift in the opposite directions. In the maxilla,
this is a different pattern, a labial shift of one
central incisor with an accompanying lingual
shift of the lateral incisor and canine
In the mandible, this axis implies opposing lateral
and lingual shifts of the central incisors, while
in the maxilla, this axis describes outward shifts
of both canines
Opposing linguallabial shifts of the canines and
lateral incisors appear in the maxillary, while
asymmetric linguallabial shifts appear in the
two lateral incisors of the maxilla, with some
changes in the orientation of one canine

PCA, principal component analysis.

Note that the descriptions are stated in terms of positive scores, negative
scores simply reverse the pattern.

size in the analysis did not have as large an effect on the matching rate as might have been expected from first principles.
The total number of matches did not increase quite as much
as expected, when the sample size was more than doubled relative to the earlier study. As the number of possible pairwise
comparisons in a data set is roughly proportional to the number
of specimens squared (N(N1)/2), the expectation would have
been that the number of matches should quadruple, which was
not seen. This probably indicates that the sample size is still not
large enough for us to obtain a robust estimate of the underlying
match rate at the maximum resolution obtained for this data.
Lower resolutions of course drastically increased the match rates,
putting a premium on measurement resolution in studying the
human dentition. Estimation of optimal sample size may be elusive, as this was an empirical study and it is thought that match
rate is dependant on the population studied. Attempts to use fitted statistical models to estimate the optimal sample size are certainly possible, but difficulties do arise in finding an appropriate
distributional model when working with complex multivariate
data. In light of this, we used only empirical approaches.
This study did work with a population of convenience, and
one that may be expected to be biased toward relatively high
levels of dental care given the source of the scans. It also did
not consider the effect of missing teeth on match rate, or indeed
attempt to determine the rate of incidence of missing teeth in a
general population, both of which are clearly important in
assessing the utility of bitemark analysis.
The rate of increase in the number of matches as the matching
distances was increased (corresponding to progressive loss of
measurement resolution, or incorporation of estimates of the
other sources of measurement variation) shown in Table 2 is
informative about the relationship between Procrustes and Procrustes-SP. These results show that Procrustes-SP methods do
produce about 15% fewer matches than the Procrustes methods,

SHEETS ET AL.

DENTAL MATCH RATE IN THREE DIMENSIONS

67

FIG. 5First four partial least squares axes (AD), showing patterns of covariation in the maxillary and mandibular dentition (occlusal view with maxilla at
top).

TABLE 6Results of PLS analysis of maxillary and mandibular anterior

dentition.
PLS
Axis

Percentage of
Covariance

1
2

65.4
7.2

3
4

5.0
4.3

Pattern Implied by the Axis

A contrast between wide and narrow arches
Strong labial displacement of both central incisors of
both the upper and lower dentition
Strong leftright asymmetry pattern
Opposing tilts of the central incisors and lateral
incisors of both maxilla and mandible, and minor
amounts of rotation of the canines

PLS, partial least squares.

Note that the descriptions are stated in terms of positive scores, negative
scores simply reverse the pattern.

so that the incorporation of size information is important in this

data set, although the contribution of size is not as dominant as
one might initially expect. This indicates that the counterintuitive
increase in match rate discussed earlier was probably simply a
chance event because of the very low probability of matching.
This table also indicates that match rates rise rapidly, a 25%
increase in the measurement error resulted in a change from two
matching individuals to 246 under Procrustes and from four to
144 under Procrustes-SP. This indicates that small changes in
measurement resolution are critical to the ability to discriminate
between specimens in this data. A 25% increase in measurement
error has dramatically limited the ability to distinguish individuals from one another.
The arch widths seen in this data set appear similar to those
reported elsewhere (31,32). The correlation of the less familiar
centroid size with the commonly used arch width is relatively
high in both the maxilla and the mandible (Table 4). These
results are reported largely to help characterize this data set,
allowing comparison with data sets in other studies. Neither arch

width nor centroid size correlates well between the maxilla and
the mandible. Not surprisingly, the coefficient of variation in the
mandible was higher than in the maxilla for both arch width and
centroid size, reflecting the higher structural variance in the mandible. Centroid size also displayed less variability (based on the
coefficient of variation) than appears in the arch width. Centroid
size incorporates all 60 landmarks in the study, averaging out
localized variation, while arch width uses only two, so that tilts
or malalignments in the canines contribute additional variation to
this size measure.
Correlation of arch width or centroid size between the upper
and lower dentition was small. This effect may be because the
measurements are all along the incisal surfaces of the six anterior
teeth, which is simply not capturing enough information about
the dentition as a whole to see the expected strong size correlation. Studies comparing whole dentition patterns with those seen
here on the anterior dentition might prove quite informative.
Correlation of centroid size with the measured arch width was
much higher, indicating that both measures of size were reflecting highly similar differences in the population.
The results of the PCA and PLS analysis were readily interpretable and appeared to indicate general patterns in the population, suggesting that the sample size was large enough that
stable patterns (beyond the first and second PCA axes) were
emerging in these analyses. Thus, the large sample size in this
study produced PCA and PLS axes that may be interpreted very
cleanly as distinct patterns of dental characteristics and malalignment patterns. At smaller sample sizes, one or more outlier specimens may heavily influence one or more of the PC or PLS
axes. The patterns in both PLS and PCA are straightforward and
simple to interpret because of the limited influence of outliers in
such a large data set. The first PCA and PLS axes both show
the contrast between relatively wide and narrow arches and indicate that this is the dominant shared pattern between maxillary
and mandibular incisal edges. Interestingly, this pattern is large

68

JOURNAL OF FORENSIC SCIENCES

percentage of covariation (65.4%) but a lesser percentage of variation in the mandible (54%) and particularly in the mandible
(39.8%). This seems to indicate that the two arches strongly
share an overall pattern of wide or narrow structure, but other
characteristics are not as strongly shared between the two, as the
percentage of covariance explained by subsequent PLS axes
drops more rapidly than the percentages of variance along the
PCA axes do.
In conclusion, precise dental shape measurement of a large
population permits conclusions concerning shape variation and
match rates to be drawn. An earlier study showed the principal
variation in multiple bitemarks inflicted with the same dentition
on cadaver skin was in arch width and that size of bitemarks
varied by a range of roughly 24% (13). The dramatic increase in
match rate encountered when measurement error was increased
by a similar amount in this study must be considered a cautionary finding for investigators seeking to make statements of individualization to a reasonable degree of dental certainty.
References
1. Committee on Identifying the Needs of the Forensic Sciences Community,
National Research Council. Strengthening forensic science in the United
States: a path forward. Washington, DC: National Academies Press, 2009.
2. Pretty IA, Sweet D. The scientific basis for human bitemark analyses: a
critical review. Sci Justice 2001;41(2):8592.
3. Pretty IA. The barriers to achieving an evidence base for bitemark analysis. Forensic Sci Int 2006;159S(Suppl 1):S11020.
4. Bowers CM. Problem-based analysis of bitemark misidentifications: the
role of DNA. Forensic Sci Int 2006;159S:S1049.
5. Rothwell BR. Bitemarks in forensic dentistry: a review of legal, scientific
issues. J Am Dent Assoc 1995;126(2):22332.
6. Cole S. Forensics without uniqueness, conclusions without individuation:
the new epistemology of forensic identification. Law, Probability and
Risk 2009;8:23355.
7. Page M, Taylor J, Blenkin M. Uniqueness in the forensics identification
sciencesfact or fiction? Forensic Sci Int 2011;206(13):128.
8. Saks MJ, Koehler JJ. The individuation fallacy in forensic science evidence. Vanderbilt Law Rev 2008;61:199219.
9. Herschaft EE, Alder ME, Ord DK, Rawson RD, Smith ES. Manual of
forensic odontology, 4th edn. Colorado Springs, CO: American Society
of Forensic Odontology, 2007.
10. Rawson RD, Ommen RK, Kinard G, Johnson J, Yfantis A. 1984 Statistical evidence for the individuality of the human dentition. J Forensic Sci
1984;29(1):24553.
11. Kieser JA, Bernal V, Waddell JN, Raju S. The uniqueness of the human
anterior dentition: a geometric morphometric analysis. J Forensic Sci
2007;52(3):6717.
12. Bush MA, Bush PJ, Sheets HD. Statistical evidence for the similarity of
the human dentition. J Forensic Sci 2011;56(1):11823.
13. Bush MA, Bush PJ, Sheets HD. A study of multiple bitemarks inflicted
in human skin by a single dentition using geometric morphometric analysis. Forensic Sci Int 2011;211:18.
14. Bush MA, Bush PJ, Sheets HD. Similarity and match rates of the human
dentition in three dimensions: relevance to bitemark analysis. Int J Legal
Med 2011;125(6):77984.

15. Sheets HD, Bush PJ, Brzozowski C, Nawrocki LA, Ho P, Bush MA.
Dental shape match rates in selected and orthodontically treated populations in New York State: a 2 dimensional study. J Forensic Sci 2011;56
(3):6216.
16. Sheets HD, Bush MA. Mathematical matching of a dentition to bitemarks: use and evaluation of affine methods. Forensic Sci Int
2011;207:1118.
17. American Board of Forensic Odontology. ABFO manual. March 2011
ed., http://www.abfo.org (accessed July 28, 2011).
18. Bookstein FL. Morphometric tools for landmark data: geometry and biology. Cambridge, UK: Cambridge University Press, 1991.
19. Rohlf FJ, Bookstein FL. Proceedings of the Michigan morphometrics
workshop. Ann Arbor, MI: University of Michigan Museum of Zoology,
1990;Special Publication Number 2.
20. Landmark software system. Institute for data analysis and visualization.
Sacramento, CA: UC Davis, http://graphics.idav.ucdavis.edu/research/projects/EvoMorph (accessed April 9, 2011).
21. Dryden IL, Mardia KV. Statistical shape analysis. New York, NY:
Wiley, 1998.
22. Slice DE. Geometric morphometrics. Annu Rev Anthropol 2007;36:261
81.
23. Zelditch ML, Swiderski DL, Sheets HD, Fink WL. Geometric morphometrics for biologists: a primer. San Diego, CA: Elsevier, 2004.
24. Pearson K. On lines and planes of closest fit to systems of points in
space. Phil Mag 1901;2(6):55972.
25. Rohlf FJ, Corti M. Use of two-block partial least-squares to study covariation in shape. Syst Biol 2000;49:74053.
26. Lay DC. Linear algebra and its applications, 3rd edn. New York, NY:
Addison-Wesley, 2003.
27. Eckart C, Young G. The approximation of one matrix by another of
lower rank. Psychometrika 1936;1:2118.
28. Jackson JE. A users guide to principal components. New York, NY:
Wiley and Sons, 1991.
29. Streissguth AP, Bookstein FL, Sampson PD, Barr HM. The enduring
effects of prenatal alcohol exposure on child development: birth through
seven years, a partial least squares solution. Ann Arbor, MI: University
of Michigan Press, 1993.
30. Bookstein FL, Streissguth A, Sampson P, Barr H. Exploiting redundant
measurements of dose and behavioral outcome: new methods from the
teratology of alcohol. Dev Psychol 1996;32:40415.
31. Bernitz H, van Heerden WF, Solheim T, Owen JH. A technique to capture, analyze, and quantify anterior teeth rotations for application in court
cases involving tooth marks. J Forensic Sci 2006;51(3):6249.
32. Barsley RE, Lancaster DM. Measurement of arch widths in a human
population: relation of anticipated bite marks. J Forensic Sci 1987;32
(4):97582.
Additional information and reprint requests:
Mary A. Bush, D.D.S.
Department of Restorative Dentistry
Laboratory for Forensic Odontology Research
School of Dental Medicine
SUNY at Buffalo
B1 Squire Hall, South Campus
Buffalo, NY 14214
E-mail: bushma@buffalo.edu