This article was downloaded by: [Michigan State University]

On: 20 January 2015, At: 19:11

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41
Mortimer Street, London W1T 3JH, UK

The American Statistician

Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/utas20

Geostatistics
a
Noel Cressie
a
Department of Statistics , Iowa State University , Ames , IA , 50011 , USA
Published online: 27 Feb 2012.

To cite this article: Noel Cressie (1989) Geostatistics, The American Statistician, 43:4, 197-202

To link to this article: http://dx.doi.org/10.1080/00031305.1989.10475658

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained in the
publications on our platform. However, Taylor & Francis, our agents, and our licensors make no representations or
warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of the authors, and are not the views of or endorsed
by Taylor & Francis. The accuracy of the Content should not be relied upon and should be independently verified with
primary sources of information. Taylor and Francis shall not be liable for any losses, actions, claims, proceedings,
demands, costs, expenses, damages, and other liabilities whatsoever or howsoever caused arising directly or indirectly
in connection with, in relation to or arising out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to anyone is
expressly forbidden. Terms & Conditions of access and use can be found at http://www.tandfonline.com/page/terms-
and-conditions
 Geostatistics
NOEL CRESSIE”

Most data have a space and time label associated with them; is dependence between measurements at different locations.
data that are close together are usually more correlated than The models need to be more flexible than their temporal
those that are far apart. Prediction (or forecasting) of a counterparts; for example, it is not reasonable to assume
process at a particular label where there is no datum, from that spatial locations of data occur regularly, and “past,”
observed nearby data, is the subject of this article. One “present,” and “future” make way for (possibly aniso-
approach, known as geostatistics, is featured, from which tropic) dependence in a multitude of directions. Cliff and
linear methods of spatial prediction (kriging) will be con- Ord (1981), Ripley (1981), and Upton and Fingleton (1985)
sidered. Brief reference is made to other lineadnonlinear, have written books concerned with the statistical analysis
stochasticldeterministic predictors. The (linear) geostatis- of spatial data.
tical method is applied to piezometric-head data around a A scientific problem that is both temporal and spatial in
potential nuclear-waste repository site. nature is the study of the effects of atmospheric pollution,
and in particular acid rain (e.g., see Bilonick 1985; Cressie,
Downloaded by [Michigan State University] at 19:12 20 January 2015

KEY WORDS: Anisotropy; Covariance function; Isotropy;

Gotway, and Grondona, in press). Daily data collection over
Kriging; Spatial statistics; Stationarity; Trend; Variogram.
a number of years at various locations throughout the north-
east United States yields a massive data set. But most of it
is temporal so that, spatially speaking, the data are still
1. INTRODUCTION rather sparse. Nevertheless, spatial prediction is just as im-
Statistics, the science of uncertainty, attempts to find portant as temporal prediction, since people living in those
structure in chaos. The simplest structure imposed assumes cities and rural districts without monitoring stations have
that observations on a phenomenon are taken under identical the same right to know whether their water or their air is
conditions, and independently from one observation to an- polluted. Purely spatial problems are considered here, and
other; that is, the data are independent and identically dis- to provide a clear comparison between spatial methods and
tributed (iid). Traditionally, this is generalized by allowing temporal methods a data set that does not have a temporal
nonidentical means but retaining iid errors. Standard statis- component will be chosen in Section 4.
tical techniques (e.g., see Hogg and Craig 1978) can then Geostatistics is usually concerned with spatial prediction,
be used to build a statistical model and to estimate its pa- but there are other important areas (such as stationary-dis-
rameters. But the iid assumption should not be taken for tribution estimation, effect of aggregation, and spatial de-
granted, particularly when there are good physical reasons sign) that offer fruitful open problems. These could all be
to abandon it. It is the relaxation of the independence part lengthy articles in themselves; however, in this article I
of the assumption that I shall address in this article. present the basics of a spatial-prediction method known as
We have not yet been able to escape the three-dimensional (ordinary) kriging. For a comparative survey of various
world in which we live, nor the unidirectional flow of time forms of kriging with other stochastic and nonstochastic
through which we live. The notion that data close together, methods of spatial prediction, see Cressie (1989a), where
in time or space, are likely to be correlated (i.e., cannot be 15 methods are considered. These include disjunctive krig-
modeled as independent) is natural and has been used suc- ing, Markov-random-field prediction, inverse-distance-
cessfully by statisticians to model the processes generating squared weighted averaging, and spline smoothing.
the data. Pure temporal models, or time series models as
they have come to be known, are usually based on identi- 2. THE VARIOGRAM
cally distributed errors that are dependent and occur at equally First, a measure of the (second-order)spatial dependence
spaced time points [e.g., the autoregressivemoving average exhibited by the data is needed. I define a model-based
models of Box and Jenkins (1970)l. The equal spacing and parameter (which is a function) known as the variogram;
unidirectional flow of time underlie the construction of these then its estimate will provide such a measure. Statisticians
models. are used to dealing with autocovariance functions. It will
Spatial models are a more recent addition to the statistics be demonstrated that the class of processes with variogram
literature. Geology, soil science, crop science, forestry, as- functions includes the class of processes with autocovari-
tronomy, or simply any discipline that works with data ance functions; hence kriging can be carried out on a wider
collected from different spatial locations needs to develop class of processes than the one traditionally used.
(not necessarily statistical) models that indicate when there Let {Z(s): s E D C Rd} be a real-valued stochasticprocess
defined on a domain D of the d-dimensional space Rd, and
*Noel Cressie is Professor, Department of Statistics, Iowa State Uni- suppose that differences of variables lagged h apart vary in
versity, Ames, IA 5001 1. Helpful discussions with I. Clark are gratefully a way that depends only on h. That is,
acknowledged, as are the comments of the referees and the research as-
sistance of C. Gotway. This work was partially supported by National
Science Foundation Grant DMS-8703083.
var(Z(s + h) - Z(s)) = 2y(h) for all s, s + h E D;
(1)

0 1989 American Statistical Association The American Statistician, November 1989, Vol. 43, No. 4 197
 typically the spatial index s is two- or three-dimensional 2y(h) = Ihl ( - m < h < m), but cov(W(t), W ( U ) )=
(i.e., d = 2 or 3). The quantity 2y(.), which is a function min(t, u), which is not a function of It - uI. Thus the class
only of the dzfference between the spatial locations s and of intrinsically stationary processes strictly contains the class
s + h, has been called the variograrn by Matheron (1963), of second-order stationary processes.
although earlier appearances in the scientific literature can Now I turn to finding an estimator of the variogram,
be found. It has been called a structurefunction by Yaglom which also serves as a measure of spatial dependence among
(1957) in probability and Gandin (1963) in meteorology, the data {Z(s,): i = 1 , . . ., n}. Suppose these are obser-
and a mean-squared difference by Jowett (1952) in time vations on an intrinsically stationary process [i.e., a process
series. Kolmogorov (1941a) introduced it in physics to study that satisfies (1) and (3)], taken at the n spatial locations
the local structure of turbulence in a fluid. Nevertheless, it +
{s;: i = 1 , . . ., n}. Because of (3), var(Z(s h) - Z ( s ) )
has been Matheron’s mining terminology that has persisted. = E ( Z ( s + h) - Z(s))*; hence the method-of-moments
The variogram must satisfy the conditional negative semi- estimator of the variogram 2y(h) is
definiteness condition, Xf= Xjk= a,aj2y(si - sj) 5 0, for
any finite number of spatial locations {s;: i = 1, . . ., k}, 2&h) =
Wh)
C ( Z ( S i ) -z(sj))*/IN(h)I, h E Rd, ( 5 )
and real numbers {a;: i = 1 , . . ., k} satisfying Xf=, ai =
0. When 2y(h) = 2f(llhll), the variogram is said to be where the average in (5) is taken over N(h) = {(s;, sj): s;
isotropic. Various parametric models were presented by - sj = h} and (N(h)l is the number of distinct elements
Journel and Huijbregts (1978, sec. II1.B). in N(h). For irregularly spaced data, N(h) is usually mod-
Downloaded by [Michigan State University] at 19:12 20 January 2015

Replacing (1) with the stronger assumption ified to {&, sj):si - sj € T(h)}, where T(h) is a tolerance
region of Rd surrounding h (see Sec. 4). Other estimators,
cov(Z(s + h), Z ( s ) ) = C(h) for all s, s + h € D, (2) more robust than ( 5 ) , are given in Cressie and Hawkins
and specifying the mean function to be constant, (1980). Parametric models can be fit to the estimator (5)
by various means; as a compromise between efficiency and
E ( Z ( s ) ) = p for all s E D, (3) simplicity, Cressie (1985) proposes minimizing a weighted
defines the class of second-order (or wide-sense) stationary sum of squares,
processes in D, with covariance function C(.). Time series
analysts usually assume (2) and work with the quantity p( * )
= C(.)/C(O).Conditions ( 1 ) and (3) define the class of
intrinsically stationary processes, which is now shown to with respect to variogram model parameters 8. The sequence
contain the class of second-order stationary processes. h(l), . . ., h(K) denote the “lags” at which an estimator
Assuming only (2), ( 5 ) was obtained, and that satisfy conditions such as those
given by Journel and Huijbregts (1978, p. 194, sec. 111.42).
y(h) = C(0) - C(h); (4) Zimmerman and Zimmerman (1989) summarize and com-
that is, the sernivariogram (half of the variogram) is related pare several methods of variogram-parameter estimation.
very simply to the covariance function. An example of a They find that no method dominates, but their simulation
process for which 2y( .) exists but C( .) does not is a one- results indicate that the weighted least squares approach
dimensional standard Wiener process {W(t):t 2 0). Here usually performs well and never does poorly.

I I I‘

Figure 7. Location and Levels of Piezometric-Head Data in the Wolfcamp Aquifer: 9, Location of Amarillo, Texas (Potter County).

198 The American Statistician, November 1989, Vol. 43, No. 4

 Table 7. Wolfcamp-Aquifer Data

42.78275 127.62282 1,464 103.26625 20.34239 1,591

- 27.39691 90.78732 2,553 - 14.31073 31.26545 2,540
- 1.1 6289 84.89600 2,158 - 18.13447 30.181 1 8 2,352
- 18.61823 76.45199 2,455 -18.12151 29.53241 2,528
96.46549 64.58058 1,756 - 9.88796 38.14483 2,575
108.56243 82.92325 1,702 -12.16336 39.11081 2,468
88.36356 56.45348 1,805 1 1.65754 18.73347 2,646
90.04213 39.25820 1,797 61.69122 32.49406 1,739
93.17269 33.05852 1,714 69.57896 33.80841 1,674
97.61099 56.27887 1,466 66.72205 33.93264 1,868
90.62946 35.08169 1,729 - 36.65446 150.91456 1,865
92.55262 41.75238 1,638 - 19.55102 137.78404 1,777
99.48996 59.15785 1,736 - 21.29791 131.82542 1,579
- 24.06744 184.76636 1,476 - 22.36166 137.13680 1,771
- 26.06285 1 14.07479 2,200 21.14719 139.26199 1,408
56.27842 26.84826 1,999 7.68461 126.83751 1,527
73.03881 18.88140 1,680 - 8.33227 107.77691 2,003
80.26679 12.61593 1,806 56.70724 171.26443 1,386
80.23009 14.61795 1,682 59.00052 164.54863 1,089
68.83845 107.77423 1,306 68.96893 177.24820 1,384
Downloaded by [Michigan State University] at 19:12 20 January 2015

76.39921 95.99380 1,722 70.90225 161.38136 1,030

64.46148 1 10.39641 1,437 73.00243 162.98960 1,092
43.39657 53.61499 1,828 59.66237 1 70.10544 1,161
39.07769 61.99805 2,118 61.87429 174.30178 1,415
1 12.80450 45.54766 1,725 63.70810 173.91453 1,231
54.25899 147.81987 1,606 5.62706 79.08730 2,300
6.13202 48.32772 2,648 18.24739 77.39191 2,238
- 3.80469 40.40450 2,560 85.68824 139.81701 1,038
- 2.23054 29.91113 2,544 105.07646 132.03181 1,332
- 2.36177 33.82002 2,386 - 101.64278 10.65106 3.510
-2.18890 33.68207 2,400 - 145.23654 28.02333 3,490
63.22428 79.49924 1,757 - 73.99313 87.97270 2,594
- 10.77860 175.11346 1,402 - 94.48182 86.62606 2,650
- 18.98889 171.91694 1,364 - 88.84983 76.70991 2,533
- 38.57884 158.52742 1,735 - 120.25898 80.76485 3,571
83.14496 1 59.11 558 1,376 - 86.02454 54.36334 2,811
- 21.80248 15.02551 2,729 - 72.79097 43.09215 2,728
- 23.56457 9.41441 2,766 - 1 00.17372 42.89881 3,136
- 20.11299 22.09269 2,736 - 78.83539 40.82141 2,553
- 16.62654 17.25621 2,432 - 83.69063 46.50482 2,798
29.90748 175.12875 1,024 - 95.61661 35.82183 2,691
100.91568 22.97808 1,611 - 87.55480 29.39267 2,946
101.29544 22.96385 1,548
NOTE The first two columns are data b a t o n s 4 = (x,, n), in miles from an arbitrary ongin. the third column is plezomelnc head Z(8,). in feat above sea level Here I = 1, ., 85.

To derive the kriging equations in Section 3, I assume where A = ( A l , A,, . . ., A,,, m ) ‘ , y = (y(s, - so), . . .,
that the variogram is known. In practice the estimation and y(s, - so), I)’, and
fitting stage of the spatial analysis has to be carried out first,
and this I do for the example in Section 4. Discussion of I’= y ( s , - s j ) , i = 1, . . . , n , j = 1, . . . , n
the effect of this initial variogram fitting is given in Section = 1, i =n + 1, j = 1 , . . . , n
5.
= 0, i = n + l , j = n + l ;
3. KRIGING
In Matheron’s geostatistics and Gandin’s objective anal-
+ +
r is a symmetric (n 1) X (n 1) matrix. The minimized
root mean squared prediction error (sometimes called kng-
ysis, the variogram is used to define coefficients in an op- ing standard error) is given by
timal linear predictor. Suppose it is desired to predict Z(%)
at some unsampled spatial location s,using a linear function
of the data
n
Matheron (1963) called this spatial prediction method
kriging, after D. G. Krige (a South African mining engineer
who, in the 1950s, developed empirical methods for deter-
which is unbiased and minimizes the mean squared predic-
mining true ore grade distributions from distributions based
tion error E(Z(s,,) - 2 ( ~ ~Then) ) ~a .straightforward min-
on sampled ore grades), whereas Gandin (1963) called it
imization using Lagrange multipliers to ensure unbiasedness
optimal interpolation. In fact, derivation of an optimal linear
+
yields optimal A’s that satisfy an ( n 1)-dimensional equa-
predictor (with known mean p ) can be found earlier in the
tion depending on 2y(si - s j ) , i, j = 0 , l , . . ., n:
literature in works of Wold (1938), Kolmogorov (1941b),
rh = y, (7) and Wiener (1949); see Cressie (1989b) for further historical

The American Statistician, November 1989, Vol. 43, No. 4 199

 made in December 1987 by the U.S. Congress to locate the
site in Nevada, probably at Yucca Mountain.) From pie-

4-1 I*? f zometric-head data for the Wolfcamp Aquifer, a variogram

will be fit and kriging will be used to predict the head surface
throughout a region in West Texas/New Mexico.
Figure 1 shows the 85 data and their locations (in prin-
ciple, obtained by drilling a narrow pipe into the aquifer
and letting the water find its own level in the pipe); the
measurements are in units of feet above sea level. This figure
is essentially that found in Harper and FUK(1986), which
was my source for the piezometric-head data along with
their spatial coordinates. The data are repeated here in Table
1, for completeness.
The proposed nuclear-waste site was located in Deaf Smith
County, seen in Figure 1 to border with New Mexico, in
the Texas panhandle. For country-music listeners who would
like to get their bearings, Amarillo, Texas, is also shown
Figure 2. Scatterplot of Piezometric-Head Data. Verticle lines on the map.
Downloaded by [Michigan State University] at 19:12 20 January 2015

are sited at the spatial locations, and their heights are equal to the The goal in analyzing this (purely) spatial data set is to
corresponding piezometric heads (in feet above sea level). draw a map of a predicted surface, based on the (irregularly
located) 85 data available. An advantage of using a sto-
observations. The kriging predictor given by (6) and (7)
chastic method of prediction (kriging) is that a map of root
assumes that the process Z contains no measurement error;
mean squared prediction error can also be drawn, quanti-
Cressie (1988) has details on how to predict noiseless ver-
fying the uncertainty in the predicted piezometric-head sur-
sions of Z.
face.
In the notation of Section 3, it is desired to use the pre-
4. FLOW OF GROUNDWATER FROM A
dictor 2(s,) defined by (6), along with a&,) defined by
NUCLEAR-WASTE SITE
(8), as s, varies over the extent of the Wolfcamp Aquifer.
Several years ago there were three potential high-level The kriging equations (7) are straightforward to solve once
nuclear-waste sites proposed in Nevada, Texas, and the state the variogram (1) is known.
of Washington. This section contains a geostatistical anal- Assume that the data { Z ( s , ) , . . ., Z(sss)} given in Table
ysis of groundwater flow in the surrounding regions of one 1 are observed without measurement error, and that they
of those sites. are a sampling from an intrinsically stationary stochastic
The chosen site will eventually contain more than 68,000 process. [In fact, Harper and Furr (1986) did not make
high-level waste canisters placed about 30 feet apart in holes Assumption (3), assuming instead linear deterministic trend
or trenches surrounded by salt, covering an area of about (see Sec. 5 for further discussion).] Estimation of the var-
two square miles, and buried deep in the ground. Leaks iogram defined by (1) allows (7) to be solved and hence
could occur, however, or the radioactive heat could cause maps based on (6) and (8) to be drawn.
the tiny quantities of water in the salt to migrate toward the It is clear from the data shown in Figure 2 that the behavior
heat until eventually each canister is surrounded by water of the process in the NE-SW direction is different from that
(about six gallons). The chemical reaction of salt and water in the NW-SE direction. For this reason variogram esti-
would create hydrochloric acid that could slowly corrode mators based on ( 5 ) were calculated in these two directions.
the canisters. Figure 3a shows 2T(h) for h E {h(l)e, . . ., h(24)e);
Using geostatistics, I address the question Where would e is a vector of length llell = 5 miles, and direction d 4 ;
radionuclide contamination flow for the site in Deaf Smith h(k) = k - Yz for k = 1, . . ., 24. Figure 3b shows a
County, Texas? (As a matter of interest, a decision was similar plot, but e has direction 37~14.Since the data are

0 I 10 I5 20 25 0 I 10 I5 20 21
h h

Figure 3. Experimental Variogram in the (a) NE-SW and (b) NW-SE Directions. The dashed line is the weighted-least-squares fit of a
power model: 2y(h; co, b, p) = 2{c0 + b ( h ( p }(h > 0). On the horizontal axis, 7 unit = 5 miles; on the vertical axis, 7 unit = 705ft2.

200 The American Statistician, November 1989, Vol. 43, No. 4

 Parch
I I 1 r -n

160.00

12O.OcI 120.00

80.00 w0.t 83.M

40.00

I I l ! l I I I I I I I I 0.00 I I I I I I I I l - i o , I o . w
-135.W -95.00 -55.00 -15.00 25.00 65.00 105.00 --135.00 -95.00 -55.W -15.00 25.00 65.03 105.00

Figure 4. (a) Predicted Surface (6) and (b) Standard Error Surface (8) (in ft.) Obtained by Ioiging: x, Center of Deaf Smith County;
8, Location of Amarillo, Texas.

not on a regular grid, tolerance regions for N(h) in (5) have 5. CONCLUDING REMARKS
Downloaded by [Michigan State University] at 19:12 20 January 2015

to be declared. The pair of locations (s, u) E N(h) if 11s - uII

In attempting to paint a picture of the geostatistical method
- ((h((is bounded between 2 2 . 5 miles, and the direction in broad brushstrokes, many subtleties have been glossed
of (s - u) minus the direction of h is bounded between 2 d over. David (1977) and Journel and Huijbregts (1978) are
4 radians.
good sources for these, or patient readers might wait to
Superimposed on Figure 3 are weighted-least-squares fits
consult chapters 2-5 of Cressie (in press).
of the variogram model
One of the least subtle observations one might make from
2 y( h ; 0) = 0 , h=O Section 4 is that the constant-mean assumption (3) may not
hold. Write
= 2{c0 + blhl”), h > 0 , c, > 0 ,
b > 0 , o d p < 2. (9)
where E(Z(s)) = ~ ( s and) S( .) is a zero mean, intrinsically
In Figure 3a (NE-SW direction), stationary stochastic process with var(S(s h) - S(s)) = +
+
var(Z(s h) - Z(s)) = 2y(h). The “large-scale variation”
6 = (20, &,b) = (14 x IO3ft.’, 38, 1.99); p ( . ) and the “small-scale variation” a(-) are modeled as
deterministic and stochastic processes, respectively, but
in Figure 3b (NW-SE direction), without any way of making each of them identifiable. What
6 = (co, &,;) = (14 x 103ft.’, 15, 1.99). is one person’s mean structure could be another person’s
correlation structure.
Thus the process is exhibiting anisotropy; the fitted two- Often this problem is resolved in a substantive application
dimensional anisotropic semivariogram model used for krig-
ing is

y(h; 6) = to+ (6;’; r2 cos2(: - 4)

where h = ( h , , h2) = (r cos 4, r sin 4). Geometrical

anisotropy (Joumel and Huijbregts 1978, p. 179) was as-
sumed in deriving (10).
From the (fitted) variogram (lo), contour maps of the
kriging predictor (6) and the kriging standard error (8) can
be drawn. The Toolkit package by Geostokos Ltd., London,
was used; it is interactive and well-documented, has the
anisotropy option built in, and outputs prediction and stan-
dard error values for easy contouring. The results are dis-
played in Figure 4. Figure 5 shows a three-dimensional plot
of the ordinary-kriging surface of Figure 4a, viewed from
the northeast comer of the region under study. It can be
concluded from these maps that contaminated groundwater
from Deaf Smith County, Texas, would flow directly Figure 5. Three-dimensional View of Ordinary-Kriging Surface
“downhill” toward Amarillo, Texas. {Z(so): so E D}, From the Nonheast Corner of 0.

The American Statistician, November 1989, Vol. 43, No. 4 201

 by relying on scientific or habitual reasons for determining Journal of the lnternational Associationfor Mathematical Geology, 17,
the mean structure. Cressie (1986) discussed various ways 563-586.
(1986). “Kriging Nonstationary Data,” Journal of the American
of kriging in the presence of nonconstant mean. The contour Statistical Association, 8 1, 625-634.
map of predicted values {&so): so E D } typically does not (1988). “Spatial Prediction and Ordinary Kriging,” Mathematical
change much, but the map of kriging standard errors does; Geology, 20, 405-421.
Cressie (in press, sec. 4.1) illustrates this point on the Wolf- (1989a), “The Many Faces of Spatial Prediction,” in Geostatistics
camp aquifer data. (Vol. 1). ed. M. Armstrong, Dordrecht, Holland: Kluwer, pp. 163-
176.
Another problem that geostatistics has yet to address prop- -(1989b), “The Origins of Kriging,” unpublished manuscript sub-
erly is the effect of estimating variogram parameters on mitted to Mathematical Geology.
mean squared prediction errors. Intuitively, this estimation -(in press), Statisticsfor Spatial Data, New York: John Wiley.
should lead to larger actual prediction standard errors than Cressie, N., Gotway, C. A,, and Grondona, M. 0. (in press), “Spatial
u&) given by (8). Zimmerman and Cressie (1988) have Prediction From Networks,” Chemometrics and Intelligent Laboratory
Systems, 6.
been able to quantify this intuition in several special cases. Cressie, N., and Hawkins, D. M. (1980), “Robust Estimation of the Var-
In this article I have summarized the geostatistical method, iogram: I,” Journal of the Internutional Association for Mathematical
featuring a particular method of spatial prediction known Geology, 12, 115-125.
as kriging. Many other prediction methods, both stochastic David, M. (1977), Geostatistical Ore Reserve Estimation, Amsterdam:
and nonstochastic, are available, and are summarized in Elsevier.
Gandin, L . S. (1%3), Objective Analysis of Meteorological Fields, Len-
Cressie (1989a). Confronted with a set of data, which is the
Downloaded by [Michigan State University] at 19:12 20 January 2015

ingrad: Gidrometeorologicheskoe Izdatel’stvo (GIMIZ) (1%5 ed. avail-

best method to apply? Depending on which model is true, able from the Israel Program for Scientific Translations, Jerusalem).
the answer may change drastically. For this reason, several Harper, W. V., and Furr, J. M. (1986), “Geostatistical Analysis of Po-
authors have conducted investigations where the values to tentiometric Data in the Wolfcamp Aquifer of the Palo Duro Basin,
be predicted were actually known. They then compared the Texas,” Technical Report ONWI-587, Battelle Memorial Institute, Co-
lumbus, OH.
performance of various methods by determining the close- Hogg, R. V., and Craig, A. T. (1978). Introduction to Mathematical Sta-
ness of predicted to actual. For example, Laslett, Mc- tistics (4th ed.), New York: Macmillan.
Bratney, Pahl, and Hutchinson (1987) found interpolating Joumel, A. G., and Huijbregts, C. J. (1978), Mining Geostatistics, New
methods in general to be poor, and Laplacian smoothing York: Academic Press.
splines to be as good as kriging (for predicting soil pH); the Jowett, G. H. (1952), “The Accuracy of Systematic Sampling From Con-
veyor Belts,” Applied Statistics, I , 50-59.
relationship between kriging and splines was reviewed by Kolmogorov, A. N. (1941a). “The Local Structure of Turbulence in an
Watson (1984). In all comparisons, on real and simulated Incompressible Fluid at Very Large Reynolds Numbers,” Doklady Aka-
data, universal kriging generally did as well or better than demii Nauk SSSR, 30, 229-303.
the other methods. (1941b), “Interpolation and Extrapolation of Stationary Random
Intuitively, the prediction method must be flexible, adapt- Sequences,” lzvestiia Akademii Nauk SSSR, Seriia Matematicheskaia,
5 , 3-14.
ing to the underlying spatial variation in the data. Kriging Laslett, G. M., McBratney, A. B., Pahl, P. J., and Hutchinson, M. F.
has this flexibility, since the spatial dependence structure is (1987), “Comparison of Several Spatial Prediction Methods for Soil
first gauged from an initial data analysis, before the kriging pH,” Journal of Soil Science, 38, 325-341.
equations are solved. Unlike splines, kriging adapts to the Matheron, G. (1963), “Principles of Geostatistics,” Economic Geology,
quality and quantity of spatial dependence demonstrated by 58, 1246-1266.
Ripley, B. D. (1981), Spatial Statistics, New York: John Wiley.
the data, produces prediction standard errors, and easily Upton, G. J. G., and Fingleton, B. (1985), Spatial Data Analysis by Ex-
handles prediction of an unobservable whose support is dif- ample, Chichester, U.K.: John Wiley.
ferent from those of the data (Matheron 1963). Watson, G. S. (1984). “Smoothing and Interpolation by Kriging and
Splines,” Journal of the International Association for Mathematical
Geology, 16, 601-615.
[Received October 1987. Revised May 1989.1 Wiener, N. (1949), Extrapolation, Interpolation and Smoothing of Sta-
tionary Time Series, Cambridge, MA. MIT Press.
Wold, H. (1938), A Study in the Analysis of Stationary Time Series,
Uppsala, Sweden: Almqvist & Wiksells.
Yaglom, A. M. (1957), “Some Classes of Random Fields in n-Dimen-
REFERENCES sional Space, Related to Stationary Random Processes,” Theory of Prob-
Bilonick, R. A. (1985), “The Space-Time Distribution of Sulfate-Depo- ability and Its Applications, 2, 273-320.
sition in the Northeastern United States,” Atmospheric Environment, Zimmerman, D. L., and Cressie, N. (1988), “Improved Estimation of the
19, 1829-1845. Kriging Variance,” Reprint 88-37, Iowa State University, Statistical
Box, G. E. P., and Jenkins, G. M. (1970), Time Series Analysis, Fore- Laboratory.
casting and Control, San Francisco: Holden-Day. Zimmerman, D. L., and Zimmerman, M. B. (1989), “A Monte Carlo
Cliff, A,, and Ord, J. K. (1981), Spatial Processes: Models and Appli- Comparison of Spatial Variogram Esstimators, Kriging Predictors, and
cations, London: Pion. Kriging Variances, ” unpublished manuscript submitted to Technome-
Cressie, N. (1985), “Fitting Variogram Models by Weighted Least Squares,” trics.

202 The American Statistician, November 1989, Vol. 43, No. 4