See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/323975467

Pairwise sample comparisons and multidimensional scaling of detrital zircon

ages with examples from the North American platform, basin, and passive
margin settings

Article in Lithosphere · March 2018

DOI: 10.1130/L700.1

CITATIONS READS
12 172

3 authors, including:

Gregory Wissink Gregory D. Hoke

Syracuse University Syracuse University
5 PUBLICATIONS 131 CITATIONS 86 PUBLICATIONS 2,142 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

High-Precision U-Pb Zircon Geochronology of the Late Triassic Chinle Fluvial System of the Colorado Plateau View project

Geology of the Andes, central Chile and Argentina View project

All content following this page was uploaded by Gregory D. Hoke on 26 March 2018.

The user has requested enhancement of the downloaded file.

 RESEARCH

Pairwise sample comparisons and multidimensional

scaling of detrital zircon ages with examples from the
North American platform, basin, and passive margin
settings
Gregory K. Wissink*, Bruce H. Wilkinson, and Gregory D. Hoke
DEPARTMENT OF EARTH SCIENCES, SYRACUSE UNIVERSITY, SYRACUSE, NEW YORK 13244, USA

ABSTRACT

Over the past several decades, the development of highly efficient and low-cost techniques for the acquisition of U-Pb ages has led to
the rapid expansion of detrital zircon analysis and interpretation. Multidimensional scaling (MDS) is a potentially powerful approach for
visualizing and interpreting such age data. This study suggests that for any study applying MDS to detrital zircons: (1) the choice of rep-
resenting age distributions as either probability density plots or kernel density estimations offers no particular advantage when subjected
to many inter-sample comparison measures of grain ages; (2) among the various approaches for such sample pairwise comparisons, like-
ness, similarity, the Kuiper test statistic, and the Sircombe-Hazelton metric are more effective than the Kolmogorov-Smirnov test statistic
or the cross-correlation coefficient; (3) variable proportions of both shared and unique age components have significant effects on MDS
mapping; and (4) three-dimensional MDS projections are often superior to the two-dimensional projections commonly used in detrital
provenance studies. We demonstrate these findings by representing age distributions of samples from platform, basin, and passive mar-
gin settings of North America as colored surfaces defined on the basis of sample depositional age, geographic position, and importance
of different detrital zircon components, and then compare these with three-dimensional MDS maps. These approaches suggest that the
application of MDS techniques to detrital zircon data affords easily attainable and significant advantages in the geologic interpretation
of detrital grain ages.

LITHOSPHERE GSA Data Repository Item 2018142 https://doi.org/10.1130/L700.1

INTRODUCTION of pairwise differences. Methodologies successfully employed thus far

Zircon is a common accessory mineral in clastic sedimentary deposits
and is widely used for provenance analysis. Its durability, resistance to
chemical and physical weathering, high initial concentrations of U and
Th, and low initial concentrations of Pb make it a robust geochronometer
and a useful mineral for sedimentary provenance studies. The widespread
use of single-grain detrital zircon age data is largely due to the develop-
ment of rapid and inexpensive age acquisition techniques (e.g., Gehrels,
2000; Gehrels et al., 2008) to determine ages for hundreds of zircon grains
per day (e.g., Pullen et al., 2014). This volume of data, however, presents
challenges for its management, visualization, and interpretation. Efficient,
effective, and meaningful procedures to compare and evaluate zircon
U-Pb age distributions have, therefore, garnered considerable attention
(e.g., Sircombe and Hazelton, 2004; Vermeesch, 2013, 2017; Saylor and
Sundell, 2016; Sundell and Saylor, 2017).
Differences in sample grain age frequency distributions are often quali-
tatively evaluated by visual comparison, but the increasing size of such
data sets has resulted in a growing desire for more robust characterization
Gregory Wissink http://orcid.org/0000-0003-1723-6231
*Corresponding author: gkwissin@syr.edu
include quantification of inter-sample variations (Gehrels, 2000; Amidon
et al., 2005a, 2005b; Saylor et al., 2012; Satkoski et al., 2013); mix-
ture modeling (e.g., Sambridge and Compston, 1994; Licht et al., 2016;
Sundell and Saylor, 2017; Sharman and Johnstone, 2017), hierarchical
cluster analysis (e.g., Sircombe and Hazelton, 2004), principal compo-
nent analysis (e.g., Sircombe, 2000; Fedo et al., 2003), and age spectrum
deconvolution (Sambridge and Compston, 1994). Here, we focus on the
multidimensional scaling (MDS) approach (e.g., Vermeesch, 2013).
The multivariable ordination method, or multidimensional scaling,
is a quantitative comparison technique that enhances the capability to
visualize variation among samples based on quantified pairwise com-
parisons of their zircon ages (Vermeesch, 2013). Here, using zircon ages
from both synthetic-model and real-world data sets, we: (1) consider
different representations of sample ages and metrics of intra-sample dis-
parities; (2) examine strengths and limitations of these approaches by
comparison of different metrics and by exploring their impact on MDS
ordination; (3) assess the impact of variable proportions of shared and
different age components on MDS; and (4) evaluate the relative merits
of two-dimensional (2-D) and three-dimensional (3-D) MDS projections
in detrital provenance studies. We then apply this methodology to detri-
tal zircon age distributions of samples from platform, basin, and passive

LITHOSPHERE
© | Volume
2018 The Authors. Gold 10 | Number
Open | www.gsapubs.org
Access:3This paper is published under the terms of the CC-BY-NC license. 1
WISSINK ET AL. | Pairwise sample comparisons and multidimensional scaling of detrital zircon ages
margin settings of North America (zircon U-Pb ages and sample locations
can be found in Table DR11). Throughout, we employ the terms “age
component” to represent a suite of grain ages comprising a single age
mode, “age population” as a suite of grains encompassing one or more
age components, and “sample” as ages determined for a limited number
of grains drawn at random from some population of assumed infinite
size. In this construct, an age component is analogous to a single Gauss-
ian age mode, whereas age populations comprise one or more Gaussian
components typically found a sand or sandstone sample.
MULTIDIMENSIONAL SCALING
Multidimensional scaling (MDS) is a typically iterative ordination
technique that graphically represents the relative dissimilarities among
samples arrayed in N-dimensional space. It is not new to geologic appli-
cations (e.g., Doveton, 1976; Hayward and Smale, 1992; Hounslow and
Morton, 2004; Honarkhah and Caers, 2010); however, more widespread
application to detrital zircon data sets is relatively recent (e.g., Vermeesch,
2013; Spencer and Kirkland, 2015; Vermeesch and Garzanti, 2015; Arboit
et al., 2016; Saylor et al., 2017; Xu et al., 2017). Although a detailed
description of the approach is beyond the scope of this paper, we offer
a brief overview.
General Considerations
In detrital geochronology, MDS transforms a matrix of pairwise simi-
larities among zircon age distributions into Cartesian coordinates in 2-D
or 3-D space, with the distance between sample points representing the
degree of dissimilarity to one another. Greater differences among inter-
sample ages lead to a greater spread of points on an MDS “map.” MDS
serves to reduce complex age spectra into relatively simple visualizations
that describe the variation between sample age frequency distributions.
Because n variables are typically collapsed into a 2-D or 3-D space, some
distortion must occur; the degree of this distortion is typically measured by
a goodness-of-fit statistic. Distances in MDS maps are unitless and orien-
tations are arbitrary; only the relative proximities and distances between
points have meaning.
In MDS, a matrix of inter-sample dissimilarities (δ) is transformed
by a function (f) into a disparity matrix of fitted Euclidean (straight-line)
distances (d) which, for two samples i and j, can be represented as:
dij ≈ f (δ ij ), (1)
where f(δij) is a monotonically increasing function that transforms the
dissimilarities to “disparities” or fitted distances. MDS uses these dispari-
ties to produce a configuration of points in two (e.g., Vermeesch, 2013)
or three (Saylor and Sundell, 2016) dimensions. Two variants on the
approach are nonmetric and metric MDS. In nonmetric MDS, the ranks of
the dissimilarities (ordinal data, not the absolute dissimilarities) are used
to determine the ordination, which is iteratively calculated (e.g., isotonic
regression; Kruskal, 1964a, 1964b). In metric MDS, a superset of classical
MDS (Torgerson, 1952), absolute values of the dissimilarities are used to
determine MDS configurations. Metric MDS simultaneously calculates
configurations of the disparities matrix and the fits of the transformation
via Eigenanalysis (Torgerson, 1952; Borg and Groenen, 2005). If f(δij) =
1
GSA Data Repository Item 2018142, Table DR1: Zircon U-Pb ages and spatial and
temporal information from the Colorado Plateau, Bighorn Basin, and Gulf Coast data
sets, is available at http://www.geosociety.org/datarepository/2018, or on request
from editing@geosociety.org.
LITHOSPHERE | Volume 10 | Number 3 | www.gsapubs.org
RESEARCH
δij and the dissimilarities are metric, then the configuration of points can
be solved via linear algebra (Carroll and Arabie, 1980).
How well the MDS configuration preserves the original inter-sample
dissimilarities is evaluated by a loss function. In nonmetric MDS, the loss
function is minimized numerically to optimize the configuration solution.
For metric MDS, the loss function is calculated simultaneously with the
calculated configuration of f(δ); a commonly used loss function is stress
(S; Kruskal, 1964a). In addition to loss functions, Shepard plots display
inter-point distances (d) against the non-transformed dissimilarities (δ)
and are often used to evaluate the goodness of an MDS fit. In a Shepard
plot, the closer points fall along a 1:1 line for metric scaling or along the
nonparametric function for nonmetric scaling, the better the MDS solu-
tion represents the differences among samples. A strength of MDS is that
neither the choice of metric or nonmetric MDS nor the choice of the loss
function drastically changes the resulting transformation for detrital sample
dissimilarities. However, metric MDS provides a global solution that will
not potentially collapse into a local optimum (stress minimum) that can
complicate finding the optimal solution. For simplicity, we therefore focus
on the application of metric MDS and utilize the stress (S) loss function.
Variation in Age Frequency Distributions
With this as a general background, we first explore two aspects of
detrital zircon data that potentially impact characterizing spatial and tem-
poral differences in zircon populations using MDS. These are: (1) the
implementation of one of several methodologies for representing the
relative abundances of different age components in an individual sample,
and (2) the use of one of several quantitative measures of similarity-
dissimilarity between pairs of sample ages. Once an approach to the
representation of relative age frequencies and an approach to quantify
inter-sample similarities are adopted, MDS can be used to visualize the
inter-sample variability.
Representation of Sample Ages
Intra-sample abundances of grain ages can be represented as age fre-
quency distributions (Fig. 1A; e.g., Machado et al., 1996), probability den-
sity plots (PDPs; e.g., Dodson et al., 1988, Ludwig, 2003; Fig. 1B), or as
kernel density estimations (KDEs; e.g., Silverman, 1986; Vermeesch, 2012;
Fig. 1C). Both PDPs and KDEs strive to identify the proportional contribu-
tion of one or more age components (e.g., Fig. 1D) within a given sample,
each presumably representing some time interval during which zircon crys-
tallized. Each necessitates the imposition of certain assumptions and biases.
A PDP function specifies the probability that a random variable will
occur within a particular range of values; it is nonnegative and is normal-
ized to unity. Determination of PDPs involves the summed probabilities of
measured grain ages and their associated Gaussian analytical uncertainties
(errors). Implicit in the method is that calculated age abundances (e.g.,
probability density) are related to analytical error (Fig. 1B). A KDE is a
nonparametric estimate of the probability density function of a random
variable but discards uncertainties associated with data acquisition. The
calculation of KDEs also involves the summed probabilities of measured
grain ages but utilizes some specified “bandwidth” (e.g., Fortmann-Roe
et al., 2012) rather than an analytical error (Fig. 1C). However, the selec-
tion of bandwidth is critical, as different choices may lead to strikingly
dissimilar density estimates. Bandwidth can be constant; Botev (2007)
presents a diffusion-based approach for its approximation that minimizes
the asymptotic mean squared error of the estimate (“kernel density estima-
tion”). Bandwidth can also vary (inversely) with data density over some
local sample space; these are referred to as “locally adaptive variable
bandwidth estimations” or “kernel density functions” if the bandwidth
2
WISSINK ET AL. | Pairwise sample comparisons and multidimensional scaling of detrital zircon ages RESEARCH

A C
Relative Frequency

0.04

Kernel Density
0.04
Frequency KDE
0.02
0.02

B D

Gaussian Density
Probability Density

0.04
0.04

PDP Normal
0.02
0.02

500 1,000 1,500 2,000 2,500 500 1,000 1,500 2,000 2,500

Age (Ma) Age (Ma)

Figure 1. Four representations of intra-sample variations in detrital zircon ages, in this case within a sample from the top of the Devo-
nian Temple Butte Formation (n = 100) exposed in the Grand Canyon (Gehrels et al., 2011). (A) Age frequency distribution binned at 5
Ma. (B) Probability density plot (PDP), where peaks (e.g., ca. 1430 and 1725 Ma) reflect abundant (age component) grains presumably
derived from different geologic provinces (e.g., Amarillo-Wichita and Yavapai-Mazatzal). (C) Kernel density estimation (KDE) derived using
a fixed bandwidth of 40 Ma (Botev, 2007). (D) Best-fit Gaussian distributions to four of the apparent age components (1722, 1426, 1913,
and 509 Ma), which account for 99% (88.3%, 3.4%, 4.6%, and 0.3%, respectively) of the observed variance in zircon ages for this sample.

varies inversely to the data point analytical uncertainties (KDFs; e.g.,
Sircombe and Hazelton, 2004; Vermeesch, 2017).
Calculation of Differences among Sample Ages
Vermeesch (2013) outlined the requirements for good measures of age
population dissimilarity, positing that the chosen metric should be: (1)
non-negative; (2) symmetric (δij = δji); and (3) exhibit triangular inequality
(δik ≤ δij + δik). A variety of approaches that meet these criteria to varying
degrees have been employed to quantitatively represent the differences in
grain age distributions among samples pairs. Several metrics have been
employed to assess inter-sample similarities. This includes the cross-
correlation coefficient, which is the coefficient of determination (R2) of a
cross-plot of age frequencies, PDPs, or KDEs of two samples (e.g., Saylor
et al., 2012, 2013; Saylor and Sundell, 2016). It is similar to quantile-
quantile or probability-probability plots (e.g., Wilk and Gnanadesikan,
1968), except PDPs or KDEs, rather than cumulative distribution func-
tions, are typically used in the cross-plot. Identical age spectra yield an R2
of 1; those that share no age peaks yield an R2 of 0. An alternative metric
is the nonparametric two-sample Kolmogorov-Smirnov (K-S) test. Used
to evaluate the null hypothesis that variation between two age populations
is within the expected variation assuming random sampling of a parent
population (Guynn and Gehrels, 2010; Razali and Wah, 2011), the K-S test
is based on the K-S statistic, which is the maximum difference between
the empirical cumulative distribution functions of two samples. The Kui-
per’s test statistic (Kuiper, 1960; Press et al., 2007; Saylor and Sundell,
2016) is closely related to the K-S statistic but represents the sum of the
absolute sizes of the most positive and most negative differences between
the two cumulative distribution functions being compared. Similarity is a
measure of resemblance between two PDPs or KDEs (Gehrels, 2000) and
is calculated by summing, over the time interval of interest and at some
temporal scale of resolution, the square root of the product of each pair
of common probabilities. It yields a value of 100% for identical curves
and a value of 0% for those that share no age components. Likeness is a
complement of the area mismatch metric of Amidon et al. (2005a, 2005b).
It represents the percent of “sameness” between two unitized PDPs or
KDEs (e.g., Satkoski et al., 2013) and is calculated as one minus the sum-
mation of absolute differences between all PDP or KDE values divided by
two. Values range from 100 to 0%. The Sircombe-Hazelton dissimilarity
measure assesses the “distance” between age distributions based on their
PDPs or KDEs. The difference between two age distributions is measured
as the integral of the squared difference of the two functions when plotted
on the same scale (Sircombe and Hazelton, 2004; Vermeesch et al., 2016).
Given the nontrivial number of approaches that have been taken to
quantify intra- and inter-sample age differences, one might anticipate
arriving at some sort of consensus. That this is not the case is illustrated
by two recent papers dealing specifically with quantifying comparison
of large data sets (Saylor and Sundell, 2016) and dissimilarity measures
(Vermeesch, 2017) in detrital geochronology. In a study examining the
effects of increasing the number of ages selected from a single popula-
tion, Saylor and Sundell (2016) suggest that metrics for differentiation
of sample zircon ages should: (1) increase with increasing sample size;
(2) maximize sensitivity by using the full range of possible coefficients;
and (3) minimize artifacts resulting from sample-specific complexity.
Based on these considerations, they conclude that cross-correlation coef-
ficients among PDPs pass all three criteria, that likeness and similarity
coefficients of PDPs (as well as K-S and Kuiper test D and V values)
pass two of the criteria, and that all coefficients calculated from KDFs

LITHOSPHERE | Volume 10 | Number 3 | www.gsapubs.org 3

 WISSINK ET AL. | Pairwise sample comparisons and multidimensional scaling of detrital zircon ages RESEARCH

fail at least two of the criteria. In contrast, Vermeesch (2017) suggests
that when likeness and cross-correlation coefficients are based on PDPs,
they employ a too-narrow smoothing bandwidth to precise data and a
too-wide bandwidth to imprecise data, such that results are “insensible.”
He suggests that the Sircombe-Hazelton dissimilarity measure using a
KDF (bandwidth is allowed to vary) produces more “functional” results.
Choices of Representation of Sample Ages and Metrics of Their
Differences
Given the number of approaches to characterizing sample zircon ages
and their differences, it is first necessary to attempt to arrive at a suitable
measure before proceeding to their application in MDS. Disagreement
about the most suitable similarity metric might derive from the fact that
both Saylor and Sundell (2016) and Vermeesch (2017) rely heavily on
model age distributions to make various points. A first step in this direction
is to examine different descriptions of sample ages and measures of their
differences based on actual real-world measured zircon ages. To this end,
we consider detrital zircon ages from the Paleozoic (Gehrels et al., 2011)
and Mesozoic (Dickinson and Gehrels, 2009, 2008) succession across the
Colorado Plateau, the subject of MDS mapping later in the paper (Table
DR1). These data are represented by 6173 zircon ages determined by laser
ablation–multicollector–inductively coupled plasma–mass spectrometry
among 65 samples (average grains per sample n = 95) extending from
the Cambrian Tapeats Sandstone at the base section through the Jurassic
Morrison Formation at the top. For each sample, we first calculated their
PDP and fixed bandwidth KDE (e.g., Botev, 2007). A total of 65 samples
comprise 2080 possible sample pairs; for each of these, we calculate the 6
similarity/dissimilarity metrics noted above (cross-correlation coefficient,
K-S statistic, Kuiper’s statistic, similarity, likeness, and the Sircombe-
Hazelton dissimilarity measure) at a temporal resolution of 1 Ma.
Several conclusions about the choice of age representation and metric
of age difference can be drawn from this exercise. First, plots of similar-
ity, likeness, and the Sircombe-Hazelton dissimilarity measure based on
PDPs are virtually identical to those based on KDEs (Figs. 2A, 2B, and
2C). The standard deviation of each metric based on KDE pairs compared
to that of each metric based on PDP pairs (standard deviation of residu-
als of the PDP-KDE regressions) is only 2.3%, 2.8%, and 1.6% for the
similarity, likeness, and the Sircombe-Hazelton metrics, respectively.
Second, greatest agreement among these measures of pairwise sample
differences exists between values of likeness and similarity (Fig. 2D, R2
= 0.94) and likeness and the Kuiper’s test statistic (Fig. 3A; R2 = 0.85).
Values of likeness and the K-S statistic (Fig. 3B; R2 = 0.77) and likeness
and the Sircombe-Hazelton dissimilarity measure (Fig. 3C; R2 = 0.77)
are somewhat less similar, and agreement between values likeness and

0.9 A 0.9 B
0.8 0.8
0.7 0.7
KDE Similarity

KDE Likeness

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3
y = -0.0757 + 1.12x y = -0.0281 + 1.08x
0.2 0.2
R2 = 0.966 R2 = 0.981
0.1 0.1

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
PDP Similarity PDP Likeness
KDE Sircombe-Hazelton Dissimilarity

0.01 C 0.9 D
0.02 0.8
0.03 0.7
PDP Similarity

0.04 0.6
0.05 0.5
0.06 0.4
0.07 0.3 y = 0.288 + 0.871x
y = -0.0043 + 1.02x 0.2
R2 = 0.942
0.08
R2 = 0.971
0.09 0.1

0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
PDP Sircombe-Hazelton Dissimilarity PDP Likeness
Figure 2. (A, B, and C) Comparison of similarity, likeness, and the Sircombe-Hazelton dissimilarity metric among detrital zircon age data determined
for sample probability density plots (PDPs; x-axes) versus sample kernel density estimations (KDEs; y-axes). Points are measures of difference between
each of 2080 possible pairs of 65 samples from the Paleozoic (Gehrels et al., 2011) and Mesozoic (Dickinson and Gehrels, 2009, 2008) succession
extending across the Colorado Plateau. Note that for each of these three metrics, similarity/dissimilarity values are virtually identical for PDP and KDE
representations of sample ages. (D) Comparisons of values of PDP likeness and of PDP similarity among the 65 Colorado Plateau samples (R2 = 0.94).

LITHOSPHERE | Volume 10 | Number 3 | www.gsapubs.org 4

WISSINK ET AL. | Pairwise sample comparisons and multidimensional scaling of detrital zircon ages RESEARCH

0.1 A 0.1 B
Kuiper's test statistic (V)

0.2 0.2

0.3 0.3

K-S statistic (D)

0.4 0.4

0.5 0.5

0.6 0.6
0.7 0.7
y = 0.845 - 1.17x y = 0.773 + 1.07x
0.8 0.8 R2 = 0.773
R2 = 0.854
0.9 0.9
PDP Sircombe-Hazelton Dissimilarity

C D

PDP Cross-correlation Coefficient

0.9
0. 02
0.8
y = -0.286 + 1.06x
0.7 R2 = 0.674
0. 04
0.6
0. 06 0.5
0.4
0. 08 y = 00823 + 10.0908x
0.3
R2 = 0.768
0.2
0. 10
0.1

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
PDP Likeness PDP Likeness
Figure 3. Comparisons of probability density plot (PDP) likenesses and (A) Kuiper’s test statistics, (B) Kolmogorov-Smirnov (K-S) statistic,
(C) Sircombe-Hazelton dissimilarity metrics, and (D) PDP cross-correlation coefficients among each of 2080 possible pairs of 65 samples
from Colorado Plateau succession. Note progressively poorer correlation between likeness and similarity (R2 = 0.94, Fig. 2D), Kuiper’s test
statistics (R2 = 0.85, Fig. 3A), Kolmogorov-Smirnov statistic (R2 = 0.77, Fig. 3B), Sircombe-Hazelton dissimilarity metrics (R2 = 0.77, Fig. 3C),
and PDP cross-correlation coefficients (R2 = 0.67, Fig. 3D).

the cross-correlation coefficient (Fig. 3D; R2 = 0.67) are the least-well
correlated. In other words, when applied to these real-world data, and
when compared to likenesses as a measure of pairwise sample differ-
ences, similarity yields the greatest agreement, followed by the Kuiper’s
test statistic and the Sircombe-Hazelton dissimilarity measure, and then
the cross-correlation coefficient. In retrospect, such an outcome is perhaps
not surprising in that likeness, similarity, and the Sircombe-Hazelton dis-
similarity metrics are based on summations of many PDP or KDE differ-
ences at some designated scale of resolution (here, 1 Ma).
Evaluating Dissimilarity Measures in MDS
Based on consideration of these real-world data, it seems apparent that
representation of sample ages as either PDPs or KDEs has little impact on
the veracity of various measures of sample difference. Moreover, based
on similarity of values, it seems that likeness and similarity may be the
most efficient measures of pairwise sample differences followed by the
Kuiper’s, the Sircombe-Hazelton, and the K-S statistics; cross-correlation
coefficients are the least similar to other measures of sample differences.
However, it is not immediately apparent how these measures directly
translate into the MDS ordinations that best represent dissimilarity within
a data set. Although a more thorough analysis of these various options is
beyond the scope of this paper, it is informative to examine the impact
of three of these metrics of dissimilarity on MDS mapping. To this end,
we consider likeness, the K-S test statistic, and cross-correlation coef-
ficients to evaluate which appears, on the basis of agreement with the
other metrics, to represent the best, better, and the poorest measures of
sample differences. To do this, we construct a simple, three-component
model defined by Gaussian distributions of 100 ± 10, 200 ± 10, and 300
± 10 Ma components, and mix these in increments of thirds, resulting in
10 populations (Fig. 4). The 2-D MDS map generated from such a model
resembles a ternary diagram with each single-component sample, repre-
senting 100% of one component, positioned at the vertices and with all
remaining sample points representing linearly correlated dissimilarities
according to the percentage of each component (Fig. 4).
Using metric MDS and the stress (S) loss function, we produced MDS
configurations based on three dissimilarities measures of likeness (Fig.
4A), PDP cross-correlation (Fig. 4B), and the K-S D statistics (Fig. 4C). Of
these measures, likeness clearly results in the best ternary approximation
of the modeled data set, with nearly uniform spacing between all points.
The configurations for PDP cross-correlation and the K-S D statistics both
define the components as vertices and have stresses comparable to likeness
(S = 0.08); however, both are less accurate representations of the modeled
data. For multi-modal populations, cross-correlation results in samples
clustering toward the vertices containing the component with the highest
proportions and results in unequal spacing among points. For the K-S D
statistic, the end-member vertices are established, but the mixed data are
clearly displaced toward the 200 Ma component (Fig. 4C). That likeness
yields a better ternary approximation of the modeled data set than either
cross-correlation or the K-S D statistic is in general agreement with the
observation that values of likeness exhibit considerably less PDP-KDE

LITHOSPHERE | Volume 10 | Number 3 | www.gsapubs.org 5

WISSINK ET AL. | Pairwise sample comparisons and multidimensional scaling of detrital zircon ages
C δ = K-S D
S = 0.08
-0.2 0 0.2
100 200
-0.4
100 200 300
-0.2
100 200 300
100 200 300
0.4
A 100 200 300 100
-0.8 -0.6 -0.4 -0.2 0 0.2
scatter than either cross-correlation or the K-S statistic. They are also in
accord with the observation that likeness values are closest to calculated
values of similarity (Fig. 2D) and the Kuiper’s test statistic (Fig. 3A) and
are least similar to coefficients of cross-correlation (Fig. 3D).
Effect of Variable Proportions of Shared and Unique
Components
Employing likeness as the measure of dissimilarity and stress (S) as
the loss function in metric MDS mapping, we can now explore the influ-
ences of several variables common to detrital geochronology on MDS.
These include the effects of variable proportions of shared and different
age components among pairs of age frequency distributions. The effect of
shared but unequal proportions of age components is one of the challenges
of detrital geochronological data. To demonstrate how this impacts MDS,
we generate five sets of population pairings from two age components (100
± 10 Ma and 200 ± 10 Ma; Fig. 5). For simplicity, we fix one population
at equal proportions of the two age components, while components in
the second population vary in increments of 5% per MDS analysis. Three
100-age randomly drawn samples are then created for each population
and subjected to MDS analysis (Fig. 5A–5E). Based on these five model
sets, visual separation of samples in MDS space becomes significantly
more apparent when likeness values exceed 0.20.
The sensitivity of MDS to the presence of unique components is also
highly relevant to its use as a provenance tool. We consider three age
components (100 ± 10 Ma, 200 ± 10 Ma, and 300 ± 10 Ma) and allow
the 200 ± 10 Ma component to be shared in equal proportion between the
two populations (Fig. 6). The average degree of differentiation among 10
iterations of each model set demonstrates that MDS differentiation only
LITHOSPHERE | Volume 10 | Number 3 | www.gsapubs.org
RESEARCH
B δ = CC
S = 0.08
Figure 4. Two-dimensional metric multi­
dimensional scaling (MDS) images employing
different measures of dissimilarity (δ) based
300 on variable mixing of three components (100
± 10, 200 ± 10, and 300 ± 10 Ma) at 33.3% inter-
vals. Rectangular insets are probability curves
representing population ages for the near-
est point in the MDS plot. Nearest neighbor
-0.4 lines (with respect to calculated dissimilarity)
are solid back; second nearest neighbors are
100 200 300
connected with dashed gray lines. S = stress
value. (A) The central ternary shows the MDS
-0.2
map using likeness (LK). (B) The same model
data set and parameters but using the cross-
correlation coefficient (CC) as the measure
100 200 300
of dissimilarity. (C) As A and B, but using the
Kolmogorov-Smirnov maximum difference
between cumulative distribution functions
(K-S D statistic) as the measure of dissimi-
larity. Note that among these three metrics,
only likeness (A) produces the theoretically
100 200 300 correct (and desired) configuration.
0.4
200 300
δ = LK
S = 0.07
0.4 0.6 0.8
becomes consistently apparent as likeness values exceed ~0.10 (mixtures
B and C) and highlights its sensitivity to the inclusion of unique age
components (Fig. 6).
n-Dimensionality and Visualizations of MDS
MDS is often characterized as a useful tool by which to evaluate the
generalities of large data sets containing multiple components. In prac-
tice, n-1 dimensions are required to project the data in MDS space with-
out distortion. However, typical detrital zircon data often contain greater
than four components and are never as simple as our synthetic examples.
Here, we briefly show how a four-component model projected onto a
2-D MDS surface can result in poorly transformed configurations, and
further show how this problem is significantly resolved when scaling to
three dimensions. To do so, we use a mixture model based on fourths of
the four age components (100 ± 10 Ma, 200 ± 10 Ma, 300 ± 10 Ma and
400 ± 10 Ma), and translate the dissimilarity matrix of this data set into
2-D (Fig. 7A) and 3-D (Fig. 7C) MDS configurations. Theoretically, a
properly translated four-component data set plot should resemble a tetra-
hedron with equally spaced vertices. Distances between points should be
some multiple of one-fourth the distance between any of the four vertices.
Two-dimensional MDS projects the tetrahedron into the two available
dimensions. In this instance, nearest neighbor lines seem directed toward
one of the tetrahedron’s corners, the projected peak of the pyramid (Fig.
7A). The poor representation of the four components in 2-D is apparent in
the Shepard plot (Fig. 7B) and the elevated (poor) stress value (S = 0.23).
While the vertices do indeed represent each of the four age components,
the placement of samples relative to those endmembers is highly suspect,
as the distances between points do not correlate well with dissimilarities
6
WISSINK ET AL. | Pairwise sample comparisons and multidimensional scaling of detrital zircon ages RESEARCH

50-50% 50-50% 50-50% 50-50% 50-50%

45-55% 40-60% 35-65% 30-70% 25-75%
Frequency
Stacked

100 150 200 100 150 200 100 150 200 100 150 200 100 150 200
LK = 0.05 LK = 0.10 LK = 0.15 LK = 0.20 LK = 0.25
0.2
0.1
A 0.05
B 0.1
C D 0.2 E
0.0 0.0 0.0
0.0 0.0

-0.05 -0.2
-0.1
S = 0.14 S = 0.11 -0.1 S = 0.11 -0.2
S = 0.05 S = 0.03
-0.1 0 0.1 -0.1 0 0.1 -0.1 0 0.1 -0.2 0 0.2 -0.2 0 0.2
Figure 5. Models incorporating variable proportions of two age components. Probability curves afford five possible mixtures (columns) of two (upper
row) age populations with variable contributions of 100 ± 10 Ma and 200 ± 10 Ma components. LK = likeness values between population age fre-
quencies. Averages of 10 metric multidimensional scaling (MDS) results (bottom row), each comprising three random, 100-age draws from the two
distributions; S = stress. Note that actual sample differences manifest as MDS separation becomes increasingly apparent at likeness values in excess
of ~0.2 (mixtures D and E).

5% 100 Ma 10% 100 Ma 15% 100 Ma

Probabilities

95% 200 Ma 90% 200 Ma 85% 200 Ma

LK = 0.05 LK = 0.10 LK = 0.15

Probabilities

95% 200 Ma 90% 200 Ma 85% 200 Ma

5% 300 Ma 10% 300 Ma 15% 300 Ma

100 200 300 100 200 300 100 200 300

0.2
A 0.1
B C
0.04
0.1

0.00
0.0
0.0

-0.04
-0.1
S = 0.13 S = 0.03 -0.1 S = 0.07

-0.1 -0.05 0 0.05 0.1 -0.1 0 0.1 0.2 -0.1 0 0.1 0.2

Figure 6. Models incorporating variable proportions of three age components. Probability curves affording three possible
mixtures (columns) of two (upper and middle rows) bimodal age distributions with variable contributions of 100 ± 10, 200 ± 10
Ma, and 300 ± 10 Ma. LK = likeness values between top and middle row frequencies. Averages of 10 metric multidimensional
scaling (MDS) results (bottom row), each comprising three random, 100-age draws from each of the two distributions in
each of the five sets of mixtures; S = Stress. Note that actual sample differences manifest as MDS separation only becomes
apparent as likeness approaches a value of ~0.1 (mixtures B and C).

LITHOSPHERE | Volume 10 | Number 3 | www.gsapubs.org 7

WISSINK ET AL. | Pairwise sample comparisons and multidimensional scaling of detrital zircon ages RESEARCH

Likeness Likeness
0.8 0.6 0.4 0.2 0.8 0.6 0.4 0.2

B D

Distance/Disparity

Distance/Disparity
0.8 0.8

0.6 0.6

0.4 0.4

A 0.2 0.2
0.6
C
0.4 0.8

0.6
0.2
0.4
0.0 0.2

-0.2 0.0

-0.2
-0.4
-0.4
-0.6 -0.6 -0.5
Component
S = 0.23 Mixture S = 0.09 0
0.5
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.6 -0.4 -0.2
0 0.2 0.4 0.6
Figure 7. Two- versus three-dimensional multidimensional scaling (MDS) for four component mixtures. (A) Two-dimensional MDS for a four-component
(100 ± 10, 200 ± 10, 300 ± 10, and 400 ± 10 Ma) mixture model (solid and dashed gray lines = nearest and next-nearest neighbors). (B) Shepard plot for
two-dimensional MDS. The poor linearity between dissimilarities and distances indicates poor translation of the dissimilarity matrix. (C) Three-dimen-
sional MDS for the four-component mixture model (circle sizes are scaled to reflect distance of the circle from point of view of viewer; smaller is more
distant). (D) Shepard plot for three-dimensional MDS. The improved fit demonstrates the value of the added dimension in resolving sample differences.

(Fig. 7A). Conversely, a 3-D MDS translation of the same data reproduces
the predicted pyramidal shape and uniform spacing of sample points. The
stress value is reduced, as will always occur with increasing dimension-
ality (Kruskal and Wish, 1978), to S = 0.9. The Shepard plot (Fig. 7D)
also shows the far better correlation between distances and dissimilari-
ties, thus highlighting the more appropriate dimensionality for the data.
What is clearly demonstrated here is that despite the simplicity of the data
set, the 3-D projection quickly outpaces the 2-D projection most typi-
cally used in detrital provenance studies. If the desired result of a detrital
zircon provenance study is to characterize a single shift in provenance
(e.g., Vermeesch, 2013), increased dimensionality and low S values might
not be necessary. In more complex cases, the consequences of distortion
related to 2-D or 3-D projection of n-dimensional data require careful
consideration.
PHANEROZOIC EXAMPLES FROM NORTH AMERICA
While modeled data provide a valuable baseline against which to visu-
alize and evaluate MDS mapping, we conclude by considering MDS
analyses of three real-world data sets. Our intent here is to demonstrate
the utility of the approach in highlighting sample age variation in both
time and space. However, we do not intend to expand upon the previously
published data sets nor published conclusions, merely present how visual
representations can provide clarity within detrital data. The particulars
of such differences would, of course, then serve as the basis for further
investigation. In each instance, we represent zircon age data as a color-
shaded age frequency “heat map” developed on an interpolated surface
comprising PDPs ordered by stratigraphic position (e.g., the Colorado
Plateau), relative age (e.g., the Bighorn Basin), or geographic position
(e.g., the Gulf Coast). Such representations allow for easy visualization
of changes in different age components in time and space. In addition,
we ordinate the MDS maps to best represent the relative sense (in time
or space) of inter-sample differences.
The Colorado Plateau
Remarkable exposures of Paleozoic and Mesozoic strata across the Col-
orado Plateau offer a nearly unparalleled opportunity to examine changes
in detrital zircon provenance of Cambrian through Jurassic sandstones
across of the southwestern United States. Recent LA-ICP-MS determina-
tions of 2529 zircon U-Pb grain ages from 26 Paleozoic samples (Gehrels
et al., 2011) and 3444-grain ages from 36 Triassic and Jurassic samples
(Dickinson and Gehrels, 2008, 2009) afford an excellent opportunity to
compare more traditional representations of grain ages (Fig. 8A) with
MDS plots (Fig. 8B). Detrital zircons from Cambrian through Devonian
strata yield mainly 1.7 and 1.4 Ga ages derived from basement rocks of

LITHOSPHERE | Volume 10 | Number 3 | www.gsapubs.org 8

 WISSINK ET AL. | Pairwise sample comparisons and multidimensional scaling of detrital zircon ages RESEARCH

Total PDP

3000
A 2 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
2800

2600 B
0

2400
0.4
Stratigraphic Height (meters)

2200 0.6
1
2000
1 0.4
1800

1600 0.2

0.03
1400
0
1200

1000
0.02 2 -0.2
2
800 C 0.8

Distance/Disparity
0.01
600
0.6
-0.6
400
0.4
200 Density
0.2
500 1000 1500 2000 S = 0.17
Grain Age (Ma)
0.8 0.6 0.4 0.2
Likeness
Figure 8. Detrital zircon ages from the Paleozoic (Gehrels et al., 2011) and Mesozoic (Dickinson and Gehrels, 2009, 2008) succession across the Colorado
Plateau. (A) Age probability densities (PDP) of 65 samples (gray horizontal lines) ranging from the Cambrian Tapeats Sandstone at the base through the
Jurassic Morrison Formation at the top. Predominant modal age components as vertical tan lines. (B) Multidimensional scaling (MDS) plot of sample
age likenesses colored by stratigraphic height. Sample points are scaled for distance into the page. Solid and dashed gray lines indicate nearest and
next-nearest neighbors. Three sample clusters are: (stratigaphic base) dominated by ca. 1.7 Ga Yavapai-Mazatzal and ca. 1.4 Ga Amarillo-Wichita prov-
ince grains (below ~420 m); (stratigraphic center) dominated by ca. 1.7 Ga Yavapai-Mazatzal and ca. 1.1 Ga Grenvillian grains (between ~420 and 1100
m); and (stratigraphic top) dominated by ca. 1.1 Ga Grenvillian and younger province grains (above ~1100 m). Note several obvious anomalies in both
probability density plot (A) and MDS (B) visualizations. These are an early Jurassic Moenave sample that contains numerous late Paleozoic grains
(“1” and blue arrow in A and B), and two Late Cretaceous samples that are dominated by ca. 1.7 Ga Yavapai-Mazatzal ages (“2” and red arrows in A
and B). (C) Shepard plot for the three-dimensional MDS.

LITHOSPHERE | Volume 10 | Number 3 | www.gsapubs.org 9

WISSINK ET AL. | Pairwise sample comparisons and multidimensional scaling of detrital zircon ages
the Yavapai-Mazatzal Province and the Amarillo-Wichita uplift, respec-
tively (Gehrels et al. 2011). Mississippian through Permian rocks contain
predominantly 1.1 Ga and younger Paleozoic grains derived from the
Grenville and Appalachian orogen, respectively. Mesozoic sands record
a greater number of age components but exhibit little directional upsec-
tion stratigraphic/secular variation (Dickinson and Gehrels, 2008, 2009).
With multiple age components throughout the data set that vary in their
presence and proportions upsection, as well as the relatively large number
of samples, a 3-D MDS plot affords an opportunity to view clear changes
in provenance within the stratigraphic sequence in a single plot. MDS
mapping of the 65 samples reveals that differences among samples, as
revealed on axis one, largely correlate with stratigraphic position (Fig. 8B).
Sample ages also define three general clusters, which correspond closely
with age groups that are discerned qualitatively from sample age abun-
dances. In particular, the three clusters in the MDS visualization (Fig. 8B)
are almost exactly the same as the groupings of Cambrian through Devo-
nian, Mississippian through Permian, and Mesozoic that are apparent in
the PDP heat map (Fig. 8A; Gehrels et al. 2011; Dickinson and Gehrels,
2008, 2009). Moreover, the samples that exhibit somewhat anomalous dis-
tances in MDS mapping are also quickly identifiable in the sample PDPs.
Among these are an early Jurassic (Moenave Formation, Utah) sample
which contains an anomalously high concentration (40 of 93 grains, 43%)
spanning only 40 Ma of the late Paleozoic (202–242 Ma; labeled 1 in
Fig. 8B), and two Late Cretaceous samples (northeastern New Mexico and
southwestern Utah) that are dominated by ca. 1.7 Ga Yavapai-Mazatzal
ages and are the only two Mesozoic samples in this compilation contain-
ing significant pre-Grenville age grains (labeled 2 in Fig. 8B). While the
specifics of the geologic history of these localities have surely served to
give rise to these deviations, it is their clear deviance in MDS that serves
to bring their differences into sharper view.
The Bighorn Basin
Dissimilarities in abundances of different age components in any popu-
lation of zircon grains potentially reflect changes in the tectonic history
of basins or of variations in climate and sea level, all of which possibly
impact paleogeography and sediment dispersal patterns. May et al. (2013)
compiled and qualitatively interpreted zircon age data from Phanerozoic
units exposed in the Bighorn Basin of Wyoming to better understand the
tectonostratigraphic evolution of that region. Based on 4104 U-Pb detrital
zircon ages from 44 samples, May et al. subdivided Upper Mississippian
through Paleocene units into four “tectonostratigraphic assemblages”
based on perceived patterns in sample zircon age frequency distribu-
tions: (1) a Paleozoic–Triassic proximal continental margin assemblage;
(2) a Jurassic–early Cretaceous (early Albian) assemblage associated with
organization of Cordilleran orogeny; (3) a late Cretaceous (late Albian
through Maastrichtian) interior seaway foreland basin assemblage; and
(4) a Paleogene assemblage accumulated during structural segmentation
during the Laramide orogeny (Fig. 9A). Because their interpretations are
based on qualitative inspection of probability density functions, it serves as
an excellent case study to which to apply MDS. A 3-D MDS plot should
prove adequate in resolving such a four-component system.
Quantitative evaluation by MDS largely confirms May et al.’s (2013)
visual examination of sample PDPs with three of the four proposed
assemblages being well-resolved (Fig. 9B). However, the posited fourth
assemblage, comprising three Paleogene samples, does not separate as a
distinct group or cluster. Each of the resolved assemblages contains either
a unique component or component with anomalously high proportions.
The samples of the fourth assemblage, however, share components with
each of the stratigraphically older assemblages, likely contributing to its
LITHOSPHERE | Volume 10 | Number 3 | www.gsapubs.org
RESEARCH
poor differentiation. In addition, MDS mapping reveals that the population
differences differentiating the first (Paleozoic) and second (lower Meso-
zoic) assemblages are smaller than those distinguishing the second from
the third assemblage (upper Cretaceous). Qualitative differences among
sample PDPs support this conclusion; abundant Grenvillian (ca. 1.1 Ga)
grains persist upsection through Paleozoic and lower Mesozoic units but
are largely absent above the early Albian Greybull Sandstone at the top
of association (sample 22, Fig. 9A).
In addition to largely verifying the tectonostratigraphic assemblages
identified by May et al. (2013), MDS provides additional insights into
the interpretation of detrital zircon data. It serves to identify those sam-
ples within any assemblage that are least dissimilar to those in any other,
and it allows for a visualization of the relative degrees of difference
between groups of samples; both provide motivation for further examina-
tion of spatial and temporal attributes that might not be apparent in the
absence of MDS.
The Gulf Coast
In addition to better understanding secular differences among various
zircon samples, such as those upsection in the Colorado Plateau and Big-
horn Basin, MDS should be equally useful in the discrimination of lateral
heterogeneities within stratal units of similar age. This approach has been
explicitly taken by Xu et al. (2017) with respect to early Miocene sediment
supply to the Gulf of Mexico Basin. Here, we conclude by examining age
data from samples of the Paleocene Wilcox Formation collected along
~2000 km of outcrop belt, extending northwest from western Alabama
to the head of the Mississippi embayment south of St. Louis, Missouri,
and then southwest toward San Antonio, Texas (Blum and Pecha, 2014).
Cenozoic-scale persistence of several major fluvial axes supplying
sediment to the northwestern Gulf of Mexico coastal plain (e.g., Gallo-
way et al., 2011) served to impart prominent along-strike variation in the
dominance of the same major age components observed comprising grains
in the Colorado Plateau and Bighorn Basin (Fig. 10A). Blum and Pecha
(2014) report 2564 ages from 27 along-strike samples of the Wilcox For-
mation. Dominant age components consist of ca. 1.7 Ga Yavapai-Mazatzal,
ca. 1.4 Ga Amarillo-Wichita, and younger Western Cordillera (ca. 170
Ma) and Laramide (ca. 60–75 Ma) province grains along the western
third of the belt, and ca. 1.1 Ga Grenvillian grains along the eastern two-
thirds of the section (Fig. 10A). Once again, the multiple components and
numerous samples provide an excellent application of 3-D MDS configu-
rations; 2-D will not adequately represent the component variance of the
detrital data set. MDS mapping of these age data results in an excellent
agreement between qualitative assessments of lateral variations based on
PDP inspection and variations based on quantified differences (Fig. 10B).
Variation in the MDS distribution of these 27 samples corresponds closely
with position along the outcrop belt. Eastern samples are dominated by
Grenville grains and are more closely clustered than other samples; more
western samples are noticeably more dispersed, an aspect reflecting the
incorporation of a larger number of components (Fig. 10A).
CONCLUSIONS
MDS has the potential to be a valuable resource for detrital geochro-
nology. Like any tool, proper recognition of the strengths and limitations
of the technique is necessary if valid conclusions are to be drawn from
MDS data configurations. Our analyses suggest the following. (1) In a
context of input representations of age frequencies for the derivation of
“sameness” metrics, both PDPs and KDEs serve equally well. Although
the latter may afford some better representation on means and durations
10
 WISSINK ET AL. | Pairwise sample comparisons and multidimensional scaling of detrital zircon ages RESEARCH

Likeness
0.8 0.6 0.4 0.2 Total PDP

C
0.8
A
Distance/Disparity

48
Upper Mississippian - Triassic 46
0.6
Jurassic- Lower Albian 44

Upper Albian - Cretaceous 42

0.4
40
Paleocene - Eocene
38
0.2
36
S = 0.15 B

Sample Number
34

32

30

0 28
0.05 26

24
0.2
0.04 22

20

0.4 18
0.03
16

14
0.02
12

10
0.4
0.01 8
0 6

4
-0.4 Density

-0.4 -0.2 0 0.2 0.4 500 1000 1500 2000 2500 3000 3500

Grain Age (Ma)

Figure 9. Upper Mississippian to Paleocene detrital zircon ages from the Bighorn Basin, Wyoming (May et al., 2013). (A) Probability densities (PDP) of
44 samples (black horizontal lines). Dominant age components as vertical tan lines. (B) Multidimensional scaling (MDS) plot of sample age likeness
colored by tectonostratigraphic assemblage membership as inferred by May et al. (2013). As in previous figure, sample points are scaled to distance
into the page and nearest neighbor lines are represented as solid and dashed gray lines. Three of the four assemblages are well-resolved by MDS:
(1) Upper Mississippian–Triassic samples (blue bar in A and points in B), (2) Jurassic–lower Albian samples (green bar in A and points in B), and (3) upper
Albian–Cretaceous samples (yellow bar in A and points in B). A posited fourth assemblage comprising three Paleogene samples (red bar in A and
points in B) does not well segregate as a distinct population. The persistence of ca. 1.1 Ga Grenvillian grains visible in A creates a closer association
between the upper Mississippian to Triassic (blue) and Jurassic to lower Albian (green) in MDS than the upper Albian to end Cretaceous. (C) Shepard
plot for the three-dimensional MDS.

LITHOSPHERE | Volume 10 | Number 3 | www.gsapubs.org 11

 WISSINK ET AL. | Pairwise sample comparisons and multidimensional scaling of detrital zircon ages RESEARCH

San Antonio,TX Texarkana, AR St. Louis, MO Meridian, MS

500
A
Grain Age (Ma)

1000

1500

Total PDP
2000 0.02

2500
0.01
3000

3500
Density

2000 1800 1600 1400 1200 1000 800 600 400 200 0
Distance Along Strike (km)
Likeness 0.4
0.8 0.6 0.4 0.2
Distance/Disparity

0.3
0.8
C B 0.2

0.6 0.1

0.4 0

-0.1
0.2 S = 0.13
-0.2

-0.3

-0.4
0.4
0
-0.4
0.6 0.4 0.2 0 -0.2 -0.4
Figure 10. Lateral variation in zircon ages along ~18,000 km of outcrop of the Paleocene Wilcox Formation from near Meridian, Mississippi (MS), to St.
Louis, Missouri (MO), to Texarkana, Arkansas (AR), to San Antonio, Texas (TX) (Blum and Pecha, 2014). (A) Age probability density (PDP) of 27 samples
(vertical colored lines); dominant modal age components as horizontal tan lines. Qualitatively, note abundant ca. 1.7 Ga Yavapai-Mazatzal, ca. 1.4 Ga
Amarillo-Wichita, and Western Cordillera and Laramide grains along the western third of the transect (left), and ca. 1.1 Ga Grenvillian grains along the
eastern two-thirds of the section. (B) Multidimensional scaling (MDS) plot of sample age likenesses colored by lateral geographic location (solid and
dashed gray lines indicate nearest and next-nearest neighbors; circles size scaled to distance into page from the viewer). Note strong correspondence
of MDS and geographic separation among most samples. (C) Shepard plot for the three-dimensional MDS.

LITHOSPHERE | Volume 10 | Number 3 | www.gsapubs.org 12

WISSINK ET AL. | Pairwise sample comparisons and multidimensional scaling of detrital zircon ages
of zircon generation (crystallization) episodes, it affords no advantage
over the former insofar as quantifying pairwise sample age differences.
(2) Among the more commonly used metrics of dissimilarity, the likeness
and similarity yield the most consistent values when calculated from prob-
ability and kernel density estimates; values of the Kuiper’s test statistic
and the Sircombe-Hazelton dissimilarity measure are more variable; the
Kolmogorov-Smirnov statistics and cross-correlation coefficients exhibit
the least agreement with other sameness metrics. (3) Among likeness, the
K-S statistic, and cross-correlation coefficients, use of the latter two as
dissimilarity measures can significantly distort MDS configurations while
the use of likeness results in a strong correlation between percentages of
age component incorporation into various populations and their sample-to-
sample dissimilarity. (4) MDS is more sensitive to the presence of unique
age components than to variable proportions of shared components with
differences in as little as 5% leading to appreciable MDS differentiation in
simple low component samples. MDS is also relatively sensitive to degrees
of overlap of different age modes. (5) More than two dimensions may be
required to properly transform the dissimilarity matrix. More sources and
more components require increasing dimensions to effectively resolve
variation in MDS, and insufficient dimensions can lead to oversimplified
solutions where inter-sample distances do not accurately represent inter-
sample dissimilarities. (6) The application of MDS, in conjunction with
other analytical tools, serves to improve the accuracy of interpretations
and conclusions of detrital zircon data.
Application of MDS mapping to the three data sets representing sedi-
ment accumulation in the North American platform, intermontane basin,
and passive margin settings reveals that this approach provides informa-
tive visualizations of relative differences among sample zircon ages and
the degree to which these differences can be related to their disposition in
time and space. In each application, MDS configurations in three dimen-
sions improved the resolving power due to the complexities of the detrital
samples. MDS affords an important tool for understanding the nature of
the distributions of zircon ages and ultimately, to the geologic processes
and histories reflected in these data sets.
ACKNOWLEDGMENTS
Early versions of this work benefited greatly from critical reviews by Elena Belousova, James
Brower, Aaron Cavosie, William Griffin, Linda Ivany, Christa Kelleher, Damian Nance, Joel
Saylor, and Pieter Vermeesch; to these people, we offer our sincere thanks. We gratefully
acknowledge our Syracuse University colleagues Mariana Bonich-Wissink, Scott Samson,
and Pedro Val for their insights during the formulation of many of the ideas presented herein.
We also commend Pieter Vermeesch, who first widely advocated the utility of MDS mapping
in detrital zircon geochronometric studies. This research was in part supported by National
Science Foundation Grant EAR-1019427 to G.D. Hoke.
REFERENCES CITED
Amidon, W.H., Burbank, D.W., and Gehrels, G.E., 2005a, Construction of detrital mineral popu-
lations: insights from mixing of U-Pb zircon ages in Himalayan rivers: Basin Research,
v. 17, p. 463–485, https://​doi​.org​/10​.1111​/j​.1365​-2117​.2005​.00279​.x.
Amidon, W.H., Burbank, D.W., and Gehrels, G.E., 2005b, U–Pb zircon ages as a sediment mix-
ing tracer in the Nepal Himalaya: Earth and Planetary Science Letters, v. 235, p. 244–260,
https://​doi​.org​/10​.1016​/j​.epsl​.2005​.03​.019.
Arboit, F., Collins, A.S., Morley, C.K., King, R., and Amrouch, K., 2016, Detrital zircon analysis
of the southwest Indochina terrane, central Thailand: Unravelling the Indosinian orogeny:
Geological Society of America Bulletin, p. 1024–1043, https://​doi​.org​/10​.1130​/B31411​.1.
Blum, M., and Pecha, M., 2014, Mid-Cretaceous to Paleocene North American drainage reorga-
nization from detrital zircons: Geology, v. 42, p. 607–610, https://​doi​.org​/10.1130​/G35513​.1.
Borg, I., and Groenen, P., 2005, Metric and Nonmetric MDS, in Modern Multidimensional
Scaling: Theory and Applications: New York, Springer-Verlag, p. 199–225, https://​doi​.org​
/10​.1007​/0​-387​-28981​-X.
Botev, Z.I., 2007, Nonparametric density estimation via diffusion mixing: Technical Report,
Postgraduate series, Department of Mathematics, University of Queensland, https://es-
pace​.library​.uq​.edu​.au​/view​.php​?pid​=UQ:120006.
Carroll, J.D., and Arabie, P., 1980, Multidimensional scaling: Annual Review of Psychology, v.
31, p. 607–649, https://​doi​.org​/10​.1146​/annurev​.ps​.31​.020180​.003135.
Dickinson, W.R., and Gehrels, G.E., 2008, Sediment delivery to the Cordilleran foreland basin:
Insights from U-Pb ages of detrital zircons in Upper Jurassic and Cretaceous strata of the
Colorado Plateau: American Journal of Science, v. 308, p. 1041–1082, 10.2475/10.2008.01.
LITHOSPHERE | Volume 10 | Number 3 | www.gsapubs.org
RESEARCH
Dickinson, W.R., and Gehrels, G.E., 2009, Use of U-Pb ages of detrital zircons to infer maximum
depositional ages of strata: A test against a Colorado Plateau Mesozoic database: Earth
and Planetary Science Letters, v. 288, p. 115–125, https://d ​ oi.​ org/​ 10.​ 1016/​ j.​ epsl.​ 2009.​ 09.​ 013.
Dodson, M.H., Compston, W., Williams, I.S., and Wilson, J.F., 1988, A search for ancient detrital
zircons in Zimbabwean sediments: Journal of the Geological Society, v. 145, p. 977–983,
https://​doi​.org​/10​.1144​/gsjgs​.145​.6​.0977.
Doveton, J.H., 1976, Multidimensional Scaling of Sedimentary Rock Descriptors, in Mer-
riam, D.F., ed., Quantitative Techniques for the Analysis of Sediments: An International
Symposium: Elsevier, p. 143–155, https://​doi​.org​/10​.1016​/B978​-0​-08​-020613​-4​.50017​-3.
Fedo, C.M., George, T., Sircombe, K.N., Rainbird, R.H., and Street, B., 2003, Detrital Zircon
Analysis of the Sedimentary Record: Reviews in Mineralogy and Geochemistry, v. 53,
p. 277–303, https://​doi​.org​/10​.2113​/0530277.
Fortmann-Roe, S., Starfield, R., and Getz, W.M., 2012, Contingent Kernel Density Estimation:
PLoS One, v. 7, no. 2, p. e30549, https://​doi​.org​/10​.1371​/journal​.pone​.0030549.
Galloway, W.E., Whiteaker, T.L., and Ganey-Curry, P., 2011, History of Cenozoic North American
drainage basin evolution, sediment yield, and accumulation in the Gulf of Mexico basin:
Geosphere, v. 7, p. 938–973, https://​doi​.org​/10​.1130​/GES00647​.1.
Gehrels, G.E., 2000, Introduction to detrital zircon studies of Paleozoic and Triassic strata
in western Nevada and northern California, in Soregham, M.J. and Gehrels, G.E. eds.,
Paleozoic and Triassic Paleogeography and Tectonics of Western Nevada and Northern
California: Geological Society of America Special Paper 347, p. 1–17, https://​doi​.org​/10​
.1130​/0​-8137​-2347​-7​.1.
Gehrels, G.E., Valencia, V.A., and Ruiz, J., 2008, Enhanced precision, accuracy, efficiency,
and spatial resolution of U-Pb ages by laser ablation-multicollector- inductively coupled
plasma-mass spectrometry: Geochemistry, Geophysics, Geosystems, v. 9, no. 3, https://​
doi​.org​/10​.1029​/2007GC001805.
Gehrels, G.E., Blakey, R., Karlstrom, K.E., Timmons, J.M., Dickinson, B., and Pecha, M., 2011,
Detrital zircon U-Pb geochronology of Paleozoic strata in the Grand Canyon, Arizona:
Lithosphere, v. 3, p. 183–200, https://​doi​.org​/10​.1130​/L121​.1.
Guynn, J., and Gehrels, G.E., 2010, Comparison of detrital zircon age distributions in the K-S
test: Tucson, University of Arizona, Arizona LaserChron Center, 16 p.
Hayward, B.W., and Smale, D., 1992, Heavy minerals and the provenance history of Waitemata
Basin sediments (early Miocene, Northland, New Zealand): New Zealand Journal of Ge-
ology and Geophysics, v. 35, p. 223–242, https://​doi​.org​/10​.1080​/00288306​.1992​.9514516.
Honarkhah, M., and Caers, J., 2010, Stochastic simulation of patterns using distance-based
pattern modeling: Mathematical Geosciences, v. 42, p. 487–517, https://​doi​.org​/10​.1007​
/s11004​-010​-9276​-7.
Hounslow, M.W., and Morton, A.C., 2004, Evaluation of sediment provenance using magnetic
mineral inclusions in clastic silicates: Comparison with heavy mineral analysis: Sedimen-
tary Geology, v. 171, p. 13–36, https://​doi​.org​/10​.1016​/j​.sedgeo​.2004​.05​.008.
Kuiper, N.H., 1960, Tests concerning random points on a circle: Koninklijke Nederlandse
Akademie van Wetenschappen, v. 63, p. 38–47.
Kruskal, J.B., 1964a, Multidimensional scaling by optimizing goodness of fit to a nonmetric
hypothesis: Psychometrika, v. 29, p. 1–27, https://​doi​.org​/10​.1007​/BF02289565.
Kruskal, J.B., 1964b, Nonmetric multidimensional scaling: A numerical method: Psychometrika,
v. 29, p. 115–129, https://​doi​.org​/10​.1007​/BF02289694.
Kruskal, J.B., and Wish, M., 1978, Multidimensional Scaling: Beverly Hills, California, USA,
and London, UK, Sage Publications, Quantitative Application in the Social Sciences, no.
11, https://​doi​.org​/10​.4135​/9781412985130.
Licht, A., Pullen, A., Kapp, P., Abell, J., and Giesler, N., 2016, Eolian cannibalism: Reworked
loess and fluvial sediment as the main sources of the Chinese Loess Plateau: Geological
Society of America Bulletin, v. 128, no. 5–6, p. 944–956, https://​doi​.org​/10​.1130​/B31375​.1.
Ludwig, K.R., 2003, User’s manual for Isoplot 3.00—A geochronological toolkit for Microsoft
Excel: Berkeley, California, Berkeley Geochronology Center, Special Publication 4, 70 p.
Machado, N., Schrank, A., Noce, C.M., and Gauthier, G., 1996, Ages of detrital zircon from Ar-
chean-Paleoproterozoic sequences: Implications for Greenstone Belt setting and evolution
of a Transamazonian foreland basin in Quadrilátero Ferrífero, southeast Brazil: Earth and
Planetary Science Letters, v. 141, p. 259–276, https://​doi​.org​/10​.1016​/0012​-821X​(96)00054​-4.
May, S.R., Gray, G.G., Summa, L.L., Stewart, N.R., Gehrels, G.E., and Pecha, M.E., 2013, De-
trital zircon geochronology from the Bighorn Basin, Wyoming, USA: Implications for
tectonostratigraphic evolution and paleogeography: Geological Society of America Bul-
letin, v. 125, p. 1403–1422, https://​doi​.org​/10​.1130​/B30824​.1.
Press, W.H., Teukolsky, S.A., Vetterling, W.T., and Flannery, B.P., 2007, Numerical Recipes: The
Art of Scientific Computing (third edition): New York, Cambridge University Press, 1256 p.
Pullen, A., Ibanez-Mejia, M., Gehrels, G.E., Ibanez-Mejia, J.C., and Pecha, M., 2014, What hap-
pens when n=1000? Creating large-n geochronological datasets with LA-ICP-MS for geo-
logic investigations: Journal of Analytical Atomic Spectrometry, v. 29, no. 6, p. 971–980,
https://​doi​.org​/10.1039​/C4JA00024B.
Razali, N.M., and Wah, Y.B., 2011, Power comparisons of Shapiro-Wilk, Kolmogorov-Smirnov,
Lilliefors and Anderson-Darling tests: Journal of Statistical Modeling and Analytics,
v. 2, p. 21–33.
Sambridge, M.S., and Compston, W., 1994, Mixture modeling of multi-component data sets
with application to ion-probe zircon ages: Earth and Planetary Science Letters, v. 128,
p. 373–390, https://​doi​.org​/10​.1016​/0012​-821X​(94)90157​-0.
Satkoski, A.M., Wilkinson, B.H., Hieptas, J., and Samson, S.D., 2013, Likeness among detri-
tal zircon populations—An approach to the comparison of age frequency data in time
and space: Geological Society of America Bulletin, v. 125, p. 1783–1799, https://​doi​.org​
/10​.1130​/B30888​.1.
Saylor, J.E., and Sundell, K.E., 2016, Quantifying comparison of large detrital geochronology
data sets: Geosphere, v. 12, p. 203–220, https://​doi​.org​/10​.1130​/GES01237​.1.
Saylor, J.E., Stockli, D.F., Horton, B.K., Nie, J., and Mora, A., 2012, Discriminating rapid exhu-
mation from syndepositional volcanism using detrital zircon double dating: Implications
13
WISSINK ET AL. | Pairwise sample comparisons and multidimensional scaling of detrital zircon ages
for the tectonic history of the Eastern Cordillera, Colombia: Geological Society of America
Bulletin, v. 124, no. 5–6, p. 762–779, https://​doi​.org​/10​.1130​/B30534​.1.
Saylor, J.E., Knowles, J.N., Horton, B.K., Nie, J., and Mora, A., 2013, Mixing of source popu-
lations recorded in detrital zircon U-Pb age spectra of modern river sands: The Journal
of Geology, v. 121, p. 17–33, https://​doi​.org​/10​.1086​/668683.
Saylor, J.E., Jordan, J.C., Sundell, K.E., Wang, X., Wang, S., and Deng, T., 2017, Topographic
growth of the Jishi Shan and its impact on basin and hydrology evolution, NE Tibetan
Plateau: Basin Research, p. n/a-n/a. https://​doi​.org​/10​.1111​/bre​.12264.
Silverman, B., 1986, Density Estimation for Statistics and Data Analysis: London, Chapman
and Hall, Monographs on Statistics and Applied Probability series, 22 p.
Sircombe, K.N., 2000, Quantitative comparison of large sets of geochronological data us-
ing multivariate analysis: A provenance study example from Australia: Geochimica et
Cosmochimica Acta, v. 64, p. 1593–1616, https://​doi​.org​/10​.1016​/S0016​-7037​(99)00388​-9.
Sircombe, K.N., and Hazelton, M.L., 2004, Comparison of detrital zircon age distributions
by kernel functional estimation: Sedimentary Geology, v. 171, p. 91–111, https://​doi​.org​
/10.1016​/j​.sedgeo​.2004​.05​.012.
Spencer, C.J., and Kirkland, C.L., 2015, Visualizing the sedimentary response through the
orogenic cycle: A multidimensional scaling approach: Lithosphere, v. 8, p. 29–37, https://​
doi​.org​/10​.1130​/L479​.1.
Sharman, G. R., and Johnstone, S. A., 2017, Sediment unmixing using detrital geochronology:
Earth and Planetary Science Letters, v. 477, p. 183–194, Supplement C.
Sundell, K.E., and Saylor, J.E., 2017, Unmixing detrital geochronology age distributions:
Geophysics, Geochemistry, Geosystems, v. 18, p. 2872–2886, https://​doi​.org​/10​.1002​
/2016GC006774.
LITHOSPHERE | Volume 10 | Number 3 | www.gsapubs.org
View publication stats
RESEARCH
Torgerson, W.S., 1952, Multidimensional scaling: I. Theory and method: Psychometrika, v. 17,
p. 401–419, https://​doi​.org​/10​.1007​/BF02288916.
Vermeesch, P., 2012, On the visualization of detrital age distributions: Chemical Geology,
v. 312–313, p. 190–194, https://​doi​.org​/10​.1016​/j​.chemgeo​.2012​.04​.021.
Vermeesch, P., 2013, Multi-sample comparison of detrital age distributions: Chemical Geol-
ogy, v. 341, p. 140–146, https://​doi​.org​/10​.1016​/j​.chemgeo​.2013​.01​.010 ​(corrigendum to
this paper:​ http://​dx​.doi​.org​/10​.1016​/j​.chemgeo​.2013​.12​.009; http://​dx​.doi​.org​/10​.1016​/j​
.chemgeo​.2016​.02​.005).
Vermeesch, P., 2017, Dissimilarity measures in detrital geochronology: Earth-Science Reviews,
https://​doi​.org​/10​.1016​/j​.earscirev​.2017​.11​.027.
Vermeesch, P., and Garzanti, E., 2015, Making geological sense of “Big Data” in sedimen-
tary provenance analysis: Chemical Geology, v. 409, p. 20–27, https://​doi​.org​/10​.1016​/j​
.chemgeo​.2015​.05​.004.
Vermeesch, P., Resentini, A., and Garzanti, E, 2016, An R package for statistical provenance
analysis: Sedimentary Geology, v. 336, p. 14–25.
Wilk, M.B., and Gnanadesikan, R., 1968, Probability plotting methods for the analysis of data:
Biometrika, v. 55, p. 1–17, 10.1093/biomet/55.1.1.
Xu, J., Snedden, J.W., Stockli, D.F., Fulthorpe, C.S., and Galloway, W.E., 2017, Early Miocene con-
tinental-scale sediment supply to the Gulf of Mexico Basin based on detrital zircon analysis:
Geological Society of America Bulletin, v. 129, p. 3–22, https://​doi​.org​/10​.1130​/B31465​.1.
MANUSCRIPT RECEIVED 18 AUGUST 2017
REVISED MANUSCRIPT RECEIVED 9 FEBRUARY 2018
MANUSCRIPT ACCEPTED 28 FEBRUARY 2018
14