Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Journal of Microscopy, Vol. 193, Pt 3, March 1999, pp.

199–211
Received 9 February 1998; accepted 12 October 1998

The efficiency of systematic sampling in


stereology – reconsidered

H. J. G. GUNDERSEN,* E. B. V. JENSEN,† K. KIÊU‡ & J. NIELSEN§


*Stereological Research Laboratory, University of Aarhus, DK-8000 Aarhus C, Denmark
†Department of Theoretical Statistics and Laboratory of Computational Stochastics,
University of Aarhus, DK-8000 Aarhus C, Denmark
‡Unité de Biométrie, Institut National de la Recherche Agronomique, Route de Saint-Cyr,
F-78000 Versailles, France
§Biometry Research Unit, Danish Institute of Agricultural Sciences, Research Centre Foulum,
DK-8830 Tjele, Denmark

Key words. Cavalieri estimator, covariogram, efficiency, Euler–MacLaurin formula,


point-counting, resampling, stereology, systematic sampling, transitive methods,
variance approximation.

Summary methods in stereology has been reconsidered by a French


group of statisticians, cf. Souchet (1995), Kiêu (1997) and
In the present paper, we summarize and further develop
Kiêu et al. (1999). Their main contribution has been to
recent research in the estimation of the variance of
formulate a set of sufficient conditions under which the
stereological estimators based on systematic sampling. In
relevant variance approximations hold.
particular, it is emphasized that the relevant estimation
In the present paper, we will present and further develop
procedure depends on the sampling density. The validity of
the work by the French group and discuss the practical
the variance estimation is examined in a collection of data
implications. In Section 2, we describe some of the most
sets, obtained by systematic sampling. Practical recommen-
common examples of systematic sampling in stereology. The
dations are also provided in a separate section.
transitive methods are presented in Section 3 (some
technicalities are deferred to the Appendix), and the
1. Introduction regularity conditions in Section 4. As a new development,
we show in Section 5 that the relevant variance approx-
Systematic sampling is widely used in stereology, partly
imation depends on the sampling density. In Section 6, we
because it is a convenient way of introducing replications in
discuss how to include measurement error in the estimation
the sampling and partly because the efficiency of the
procedure. As an alternative to the transitive methods,
estimators based on systematic sampling is often much
resampling from a large sample is discussed in Section 7.
higher than that of estimators based on independent sampl-
The rest of the main paper is devoted to practical
ing. In Matheron (1965, 1971), the transitive methods for
investigations of the methods. The analysis of a variety of
estimating the variance of estimators based on systematic
data sets is discussed in Section 8 and some of the earlier
sampling have been developed. Within the last decade,
methods of analysis are reconsidered in Section 9. Practical
these methods have been introduced and applied in stereo-
recommendations are provided in Section 10. Open
logy, cf. Gundersen & Jensen (1987), Cruz-Orive (1989),
problems and ideas for future work are presented in
Kellerer (1989), Matérn (1989) and Mattfeldt (1989).
Section 11.
In Cruz-Orive (1993), some further developments were
presented. In particular, it was discussed how to treat the 2. Examples of systematic sampling in stereology
case where the measurement function is estimated with
error. Also, examples were reported where the variance In the present paper, we consider mainly the situation
estimate suggested in Gundersen & Jensen (1987) was too where the parameter of interest can be expressed as a one-
high and this finding was also supported by derivations for dimensional integral
…∞
‘quasi-ellipsoidal’ objects. Recently, the use of transitive
Q¼ f ðxÞdx;
¹∞
Correspondence to: E. B. V. Jensen. where f is a non-negative function defined on the real axis.

䉷 1999 The Royal Microscopical Society 199


200 H . J. G . G U N D ER SE N ET AL .

The 2D analogue to the Cavalieri estimator is used to


estimate area from intersection length on a systematic set of
parallel equidistant lines. In the planar case, however, we
have an alternative. Thus, if Q is the area of a bounded
planar region, this area may be expressed as a circular
integral where f(v) is equal to 1/2 times the so-called
squared ray distance in direction v, cf. Courant (1934, p.
275), see also Jensen (1998). The estimator of Q based on n
systematic rays is the so-called nucleator, cf. Gundersen
(1988). The 3D analogue may be derived from results in
Courant (1936, p. 267).

Fig. 1. Systematic sampling using a set of parallel planes with dis- 3. The transitive methods
tance T between neighbour planes (grey planes). The parallel white
plane goes through a fixed point O. The position of the grey (sam- We shall in this section consider mainly the case where Q
pling) planes is determined by U which is uniform random in the may be expressed as an integral along the real axis. The
interval [0, T). Only grey planes hitting the set are shown. modifications for the circular case will be indicated when
needed.
The general estimator of Q to be discussed is based on It may be shown that the variance of Q̂T can be expressed
observations at a systematic set of points with spacing T in terms of the covariogram g of f
…∞
between neighbour points. The estimator takes the form
gðyÞ ¼ f ðxÞf ðx þ yÞdx:
X
∞ ¹∞
Q̂T ¼ T f ðU þ kTÞ;
k¼¹∞
We thus have
X
∞ …∞
where U is uniform random in an interval of the real axis of Var ðQ̂T Þ ¼ T gð jTÞ ¹ gð yÞdy: ð1Þ
length T. j¼¹∞ ¹∞
A well-known geometric example of this set-up is the case
where Q is the volume of a bounded set X ⊆ R3 and f(x) the The variance of Q̂T may therefore be interpreted as the
area of the intersection between X and a plane with a fixed difference between the integral of g and a discrete
orientation and position x 僆 R1. Here, Q̂T becomes the approximation of this integral. A similar formula holds in
famous Cavalieri estimator, which is named after the Italian the circular case, cf. the Appendix.
mathematician Bonaventura Cavalieri, cf. Fig. 1. Another As suggested originally by Matheron (1965, 1971),
very important example is obtained by letting Q be the the difference (1) may be evaluated, using the so-called
number of points in a finite set X ⊆ R3 and f(x) ¼ n(x)/h, Euler–MacLaurin formula. The resulting approximation
where n(x) is the number of points counted in a disector is valid for T → 0 and is based on the behaviour of g near
with position x 僆 R1 and height h. Then, Q̂T is the disector the origin. For instance, if g has an expansion near the
estimator of number based on systematic sampling. The origin as
counting principle was introduced as late as 1984 (Sterio, gðhÞ ⬇ b0 þ b1 jhj þ b2 h2 þ b3 jhj3 ; ð2Þ
1984). Another very useful design closely related to the
then for small T
disector design is the fractionator design, cf. Gundersen
(1986). b1 2 b3 4
VarðQ̂T Þ ⬇ ¹ T þ T ; ð3Þ
In some cases, the parameter of interest Q may instead be 6 60
expressed as a circular integral cf. e.g. Cruz-Orive (1993, (2.4) and (2.5)). The actual
… 2p calculation of the constants in (3) will be explained later in
Q¼ f ðvÞdv; this section.
0
It was by the work of Souchet (1995), Kiêu (1997) and
where f is a non-negative function defined on [0, 2p). The Kiêu et al. (1999) that it was first made explicit how the
estimator to be considered is then smoothness of g near the origin depends on the smoothness
n¹1  
2p X 2p of the measurement function f. Under the regularity
Q̂n ¼ f ⌰þi ; conditions stated in the next section, it may be shown
n i¼0 n
that if m is the smallest integer such that f (m) may have
where ⌰ is uniform random in [0, 2p/n) and n is a positive jumps, then g is 2m times continuously differentiable, cf.
integer. Kiêu (1997).

䉷 1999 The Royal Microscopical Society, Journal of Microscopy, 193, 199–211


S YS T E M ATI C S A M P L I N G I N S T E R E O L O G Y 201

Since the covariogram is a symmetric function,


gðyÞ ¼ gð¹yÞ; y 僆 R1 ;
it follows that
gðjÞ ð¹yÞ ¼ ð¹1Þj gðjÞ ðyÞ; y 僆 R1 ;
j ¼ 0, 1, . . ., 2m, which in turn implies that all uneven
derivatives of g of order up to 2m ¹ 1 are zero at the origin,
i.e.
gð2kþ1Þ ð0Þ ¼ 0; k ¼ 0; 1; . . . ; m ¹ 1: ð4Þ
Note that, in particular, this implies that if f is continuous
(m ⱖ 1), then the derivative of g at the origin is 0 and
therefore b1 ¼ 0 in (2) and (3). It had earlier been believed
that a zero slope of g at the origin could be expected only in
special cases, e.g. for measurement functions associated
with volume estimation of quasi-ellipsoidal objects, cf. Cruz-
Orive (1993).
The parameter m will from now on be used to denote the
smoothness class of f. In the case where f indicates the area
of intersection between X ⊆ R3 and a plane, the smoothness Fig. 2. The exact squared coefficient of error (dots) as a function of
of f should not be confused with the smoothness of X. the number of observations in a systematic sample. The measure-
Furthermore, if f (m) jumps at a1, . . . , ap, then, cf. Kiêu ment functions considered are a right triangle (smoothness class
(1997), m ¼ 0), a symmetric triangle (m ¼ 1), and powers of sinus, viz.
sin2v (m ¼ 2), sin3v (m ¼ 3) and sin4v (m ¼ 4), respectively. The
1X
p
gð2mþ1Þ ð0þ Þ ¼ ð¹1Þmþ1 i Þÿ ;
ðmÞ 2
½f ðaþ
i Þ¹f
ðmÞ
ða¹ ð5Þ approximate squared coefficient of errors based on (6) are also
2 i¼1 shown as straight lines in the double logarithmic plot, with slopes
where þ and ¹ indicate limits from the right and left, ¹2, ¹4, ¹6, ¹8 and ¹10, respectively.
respectively.
Let us illustrate the relationship between the smoothness where B2mþ2 is the Bernoulli number of order 2m þ 2, cf.
of f and g by a simple example. Let Abramovitz & Stegun (1964), Kiêu (1997) and the
( Appendix. Note that g(2mþ1)(0þ) is determined by the
pð1 ¹ x2 Þ jxj ⱕ 1 ‘transitions’ (jumps) of f (m), cf. (5). Hence the name
f ðxÞ ¼
0 jxj > 1: ‘transitive methods’, cf. Matheron (1971, p. 9).
The order of magnitude of the variance is T2mþ2. If the
The function f is the area of the intersection between the
length of the support of the function f is l , then the mean
unit ball in R3 and a plane at distance |x| from O. Note that
number of measurements inside the support is n ¼ l /T such
the smoothness class of f is 1 and that f is not differentiable
that the variance is of the order n¹(2mþ2). The approxima-
at p ¼ 2 points, viz. a1 ¼ ¹1, a2 ¼ 1. By elementary, but
tion (6) also holds in the circular case where l ¼ 2p, cf. the
somewhat tedious, calculations one may directly derive the
Appendix.
covariogram in this case:
( In Fig. 2, a collection of measurement functions is shown,
15 j yj ⱕ 2
j yj5 þ 23 j yj3 ¹ 43 j yj2 þ 16
1 together with a plot of the true variances and the
¹ 30
gð yÞ ¼ p ×2
0 j yj > 2: approximate variances (6) as a function of n. The true
variances have been determined by computer simulations.
Using (5) or direct differentiation, we find that The results shown in Fig. 2 confirm the asymptotic theory.
gð3Þ ð0þ Þ ¼ 1
2 ð½ f
ð1Þ
ð¹1þ Þ ¹f ð1Þ
ð¹1¹ Þÿ2 Let us look in more detail at the cases m ¼ 0, 1. For m ¼ 0,
it is well known that for small T we may, instead of (6), use
þ ½ f ð1Þ ð1þ Þ ¹ f ð1Þ
ð1¹ Þÿ2 Þ the following approximation
¼ 4p2 : T
m ¼ 0 : VarðQ̂T Þ ⬇ f3gð0Þ ¹ 4gðTÞ þ gð2TÞg; ð7Þ
12
For a measurement function of smoothness class m, the
variance of Q̂T can for small T be approximated as follows cf. Gundersen & Jensen (1987) and Cruz-Orive (1993). For
m ¼ 1, the expansion (2) of g at the origin will appear as
B2mþ2 ð2mþ1Þ þ
VarðQ̂T Þ ⬇ ¹2T 2mþ2 g ð0 Þ; ð6Þ
ð2m þ 2Þ! gðhÞ ⬇ b0 þ b2 h2 þ b3 h3 ;

䉷 1999 The Royal Microscopical Society, Journal of Microscopy, 193, 199–211


202 H . J. G . G U N D ER SE N ET AL .

and for small T we have the approximation


b3 4
VarðQ̂ T Þ ⬇
T :
60
(Compare with (6), B4 ¼ ¹1/30.) If T is small enough,
ð4T 3 Þ¹1 f3gð0Þ ¹ 4gðTÞ þ gð2TÞg
may be used as an approximation of b3 and we get
T
m ¼ 1 : VarðQ̂ T Þ ⬇ f3gð0Þ ¹ 4gðTÞ þ gð2TÞg; ð8Þ
240
Fig. 3. 2D illustration of the modified ball used for simulation in
cf. Cruz-Orive (1993). Fig. 4. For details see the text.
An unbiased estimator of the right-hand sides of (7) and
(8) can be obtained by replacing g( jT ) by its estimate The simulations shown in Fig. 2 for very simple
X measurement functions f indicate that already for n ¼ 10,
ĝð jTÞ ¼ T f ðU þ kTÞ f ðU þ ðk þ j ÞTÞ; j ¼ 0; 1; 2:
k the approximation works well. However, some caution is
In the circular case, the covariogram may also be estimated necessary here. Let us consider an example. Let X ⊆ R3 be a
in this way, if a periodic continuation of f is used. Recall modified ball of radius R. The modification is as follows: a
that in this case T ¼ 2p/n. cone with circular base is added at two opposite points on the
surface of the ball, cf. the 2D illustration in Fig. 3. The total
height of the cone is e, of which only p1R is outside the ball,
4. Regularity conditions
cf. Fig. 3. From each cone, a top of height p2e is removed.
The results reported in the previous section hold if f is a so- In Fig. 4, the true squared coefficient of error of the
called (m, 1)-piecewise smooth function, cf. Kiêu (1997). Cavalieri estimator (black dots in the figure) is shown as a
this means that the support of f is bounded and f is m þ 1
times piecewise differentiable such that
X all the derivatives of f of smaller order than m are

continous on R1,
X the derivatives of order m and m þ 1 need not be

continuous on a finite set, but their jumps are required


to be finite.
These conditions are expected to hold for the 3D examples
mentioned in Section 2. In particular, in the case of volume
estimation from section areas, the jumps of the derivative of
f will be finite at the points where the section plane touches
the set X, if the surface of X at these points is ellipsoidal.
For the 2D analogue where areas are estimated from a
systematic set of parallel lines, the regularity conditions do
not hold when the boundary of the planar set may be
approximated locally by circular arcs. The jumps of the first
derivative of the measurement function are thus not finite at
the points where a line touches the planar set. It is still an
open problem to find variance approximations when the
regularity conditions are not fulfilled, see Section 11.

5. The relevant variance approximation depends on the


sampling density
The variance approximation (6) is valid for small T, or Fig. 4. The squared coefficient of error (dots) as a function of the
equivalently, for a large mean number of observations n ¼ l /T. number of observations in a systematic sample. The measurement
Expressed in terms of n, the variance approximation is function is the area of intersection between the modified ball indi-
cated in Fig. 3 and a plane. The straight lines correspond to the
B2mþ2 ð2mþ1Þ þ ¹ð2mþ2Þ
VarðQ̂ T Þ ⬇ ¹2l 2mþ2 g ð0 Þ·n : ð9Þ squared coefficient of error for an approximating measurement
ð2m þ 2Þ! function at the given sampling density. The cascade of insets
In a real situation, the question is of course how large n shows how more and more of the details of the object are available
needs to be before the approximation can be used. at higher and higher magnifications.

䉷 1999 The Royal Microscopical Society, Journal of Microscopy, 193, 199–211


S YS T E M ATI C S A M P L I N G I N S T E R E O L O G Y 203

the missing tops of the cones. More interestingly, the


variance for sample sizes up to 100, say, may very well be
approximated by using the measurement function for a ball.
Note that a sample size of 100 corresponds to about one
observation in one of the two cones. The intermediate
behaviour corresponds to the expected behaviour for a ball
modified with two circular cones. So the variance may, as n
increases, be described by a series of asymptotic behaviours,
corresponding to approximating measurement functions
which take into account more and more fine details of the
actual measurement function.
In Fig. 5, another example of a similar type is presented. It
concerns a whole series of measurement functions, indexed
by a positive integer k. All measurement functions fk have the
unit interval as support and are obtained from f1 by a simple
geometric procedure. The graph of f1 is a right triangle,
f1 ðxÞ ¼ x; x 僆 ½0; 1ÿ:
Now, the graph of fk is obtained by cutting the graph of f1
Fig. 5. The squared coefficient of error as a function of the number into 2k ¹ 1 pieces of constant breadth and rearranging the
of observations in a systematic sample. The oscillating behaviour is pieces such that the piece with the largest function value is
not due to noise in the computer simulations, but shows the actual placed at 12 and then the other pieces in descending order,
behaviour (Zitterbewegung) of the squared coefficient of error. The symmetrically around 12, cf. Fig. 5. Evidently, fk will converge,
measurement functions are fk, k ¼ 1, 10, 501, 50001, ∞. For as k → ∞, to f∞ which has a symmetric triangle as graph,
details, see text. In the figure, the squared coefficients of error are (
in the same order indicated by ◊, W, K, þ and ×. The approximate 2x 0 ⱕ x ⱕ 12
f∞ ðxÞ ¼
squared coefficients of error based on (6) are also shown for f1 and 2 ¹ 2x 2 ⱕ x ⱕ 1:
1

f∞. The graphs of f1, f10 and f∞ are shown to the right. At the reso-
lution used in the figure, the graphs of f501 and f50001 cannot be Formally, we can define fk as follows. For i ¼ 1, 2, . . . , 2k ¹ 1,
8 i¹1
distinguished from that of f∞.   >
<x þ iⱕk
i¹1 i 2k ¹1
function of the mean number of sections n. In the x僆 ; ⇒ fk ðxÞ ¼
2k ¹ 1 2k ¹ 1 >
: x þ 4k ¹ 3i i > k:
simulations, p1 ¼ 0·006 and p2 ¼ 0·003. The straight lines
2k ¹ 1
have slopes (from left to right) of ¹4, ¹6 and ¹2 and their
All fks are of smoothness class 0 because they are not
levels have been calculated using an approximating
continuous while f∞ is of smoothness class 1. In Fig. 5, the
measurement function at the given sampling density, i.e.
squared coefficient of error of Q̂T is shown as a function of n,
(from left to right) the measurement function associated
for measurement functions fk, k ¼ 1, 10, 501, 50001, ∞. As
with a ball (smoothness class m ¼ 1), a ball with two small
can be seen from Fig. 5, the expected asymptotic behaviour
circular cones (m ¼ 2), and a ball with two small circular
of the order of n¹2 is first reached at a sampling density
cones with removed tops (m ¼ 0). The formulae for the lines
which depends on k. The larger k is, the higher sampling
have been derived from (9) and are as follows:
density or ‘resolution’ is required to discover that fk is
1 ¹4 1 1 þ ð1 þ p1 Þ4 ¹6 3 p21 p42 ð2 þ p1 Þ2 ¹2 actually different from f∞. For smaller sampling densities,
·n ; · ·n ; · ·n
10 105 p21 ð2 þ p1 Þ2 8 ð1 þ p1 Þ4 the squared coefficient of error obtained for fk can be
ð10Þ approximated by that obtained for f∞.
In (10), we have for simplicity used a support of length 2R
6. Measurement error
and a volume of 4/3pR3, also for the two modified balls,
which of course does not affect the results for the choices of In this section, we discuss the quite realistic situation where
p1 and p2, considered in the example. the measurement function f is estimated with error. Let us
In this remarkable example, the expected asymptotic suppose that we observe Fk at position U þ kT, where
behaviour, which is of order n¹2, is first reached for sample E½Fk jUÿ ¼ f ðU þ kTÞ:
sizes of about 104 or more. The intuitive reason for this is
Furthermore, given U, we assume that Fk1 and Fk2 are
that the jumps of the measurement function are of
stochastically uncorrelated for k1 ⫽ k2 and
negligible size and can only be ‘seen’ at the ‘resolution’
provided by a very dense sampling. In fact, the order of n¹2 Var½Fk jUÿ ¼ j2 ðU þ kTÞ;
appears first when there is at least one observation in one of say.

䉷 1999 The Royal Microscopical Society, Journal of Microscopy, 193, 199–211


204 H . J. G . G U N D ER SE N ET AL .

The estimator of Q becomes with position x 僆 R1 and height h. Let us suppose that h p T
X
∞ such that (12) and (13) are expected to provide good
Q̂ˆ T ¼ T Fk : approximations to the variance. Furthermore, suppose that
k¼¹∞ not all points are counted in a disector but only a subset
The variance of oil Q̂ˆT can be decomposed into a component, corresponding to a sampling fraction of J. If J p 1, the
due to variation of U, and a component, due to measure- count in the kth disector is expected to be distributed as
ment error, Ck jU ⬃ PoissonðhJf ðU þ kTÞÞ
Var½Q̂ˆ T ÿ ¼ Var½E½Q̂ˆ T jUÿÿ þ E½Var½Q̂ˆ T jUÿÿ and Fk ¼ Ck/(hJ). Then,
…∞ 1
¼ Var½Q̂T ÿ þ T j2 ðxÞdx ð11Þ Var½Fk jUÿ ¼ f ðU þ kTÞ ¼ j2 ðU þ kTÞ
¹∞ hJ
¼ Var½Q̂T ÿ þ Tj ; 2 and
…∞
1
where j2 ¼ j2 ðxÞdx ¼ Q:
…∞ ¹∞ hJ
j2 ¼ j2 ðxÞdx; 2
It follows that j can be estimated by
¹∞
T X
cf. Kiêu (1997). The parameter j2 will be called the S2 ¼ F:
cumulative measurement error. Note that the component hJ k k
originating from the measurement error is of the order of T, An estimation procedure of this type has been used in e.g.
while the other component is of the order of at least T2, cf. Bagger et al. (1993), West et al. (1996), Mulders et al.
(6). (1997) and West & Slomianka (1998).
The covariogram also plays a central role in this more Secondly, let us consider error due to point-counting. In
general case where the measurement error is taken into this case, f(x) is the area of the intersection between a
account. Let bounded set X⊆R3 and a plane with fixed orientation and
X
Gj ¼ T Fk Fkþj ; j ¼ 0; 1; . . . ; position x 僆 R1. We suppose that the areas are estimated,
k using a quadratic point grid with shortest distance u
between neighbour grid points. Then, according to Matheron
be the empirical covariogram. Then, it is easy to show that
(1965, p. 88),
(
gð0Þ þ j2 j ¼ 0 j2 ðxÞ ⬇ c × u3 × BðxÞ;
EðGj Þ ¼
gð jTÞ j ⫽ 0:
where c ¼ (2p3)¹1z(3) þ (6p)¹1 ¼ 0·0724, z is Riemann’s
Thus, except at the origin, the empirical covariogram is still zeta function and B(x) is the boundary length of the
an unbiased estimator of the theoretical covariogram of the intersection of X with a plane with position x. Accordingly,
…∞ …∞
measurement function.
j2 ¼ j2 ðxÞdx ⬇ c × u3 × BðxÞdx
As originally suggested by A. J. Baddeley (personal ¹∞ ¹∞
communication), the empirical covariogram may be bias- 2
and j may be estimated by
corrected at the origin if an estimate S2 of the error j2 is
available. The method of estimating the variance, applicable X

S2 ¼ c × u3 × T BðU þ kTÞ:
in the case of no measurement error, may then be used on k¼¹∞
the bias-corrected covariogram. For instance, for m ¼ 0, 1,
the resulting variance estimate becomes The boundary lengths B(U þ kT) may in turn be estimated,
using intersection counts with planar line grids.
T Suppose that the section through X with position U þ kT
m¼0: f3ðG0 ¹ S2 Þ ¹ 4G1 þ G2 g þ TS2 ð12Þ
12 consists of Nk profiles of identical shape. If Bki and Aki are the
boundary length and area of the i th profile inpthe  k th
T
m¼1: f3ðG0 ¹ S2 Þ ¹ 4G1 þ G2 g þ TS2 : ð13Þ section, i ¼ 1, . . . , Nk, we thus assume that Bki/ Aki ¼ a.
240 Then,
The estimates (12) and (13) are obtained by combining (7) ! !
X XX
and (8) with (11). E BðU þ kTÞ ¼ E Bki
Let us consider two situations where an independent k k i
estimate of the measurement error is possible. First, let Q be !
X X p
the number of points in a finite set X⊆R3 and f(x) ¼ n(x)/h ¼ aE Aki :
where n(x) is the number of points counted in a disector k i

䉷 1999 The Royal Microscopical Society, Journal of Microscopy, 193, 199–211


S YS T E M ATI C S A M P L I N G I N S T E R E O L O G Y 205

The resulting estimate of j2 becomes The resampling estimate of Var [Q̂ˆ T ] is simply
X p
1 X
k0 ¹1
S2 ¼ ca × u3 T × Fprofile ; ðQ̂ˆ ¹ Q̂ˆ T;· Þ2 ;
profiles k0 i¼0 T;i

where the sum is over all profiles in all sections and Fprofile is where
1 X
k0 ¹1
the point-count estimate of the profile area; for a real
Q̂ˆ T;· ¼ Q̂ˆ ¼ Q̂ˆ T0 :
example, see Geinisman et al. (1996). It should be noted k0 i¼0 T;i
that a refers to the shape of a single profile. The shape factor
Note that
a may be judged from the nomogram in Gundersen &
1 X
k0 ¹1
Jensen (1987,pFig. 18). In the latter figure, the shape factor E ðQ̂ˆ ¹ Q̂ˆ T;· Þ2 ¼ Var½Q̂ˆ T ÿ ¹ Var½Q̂ˆ T0 ÿ: ð16Þ
is called B/ A. k0 i¼0 T;i
If an estimate of j2 cannot be constructed using special Therefore, the resampling estimate is approximately unbiased
knowledge of the error source, the relevant information if Var½Q̂ˆ T ÿ q Var½Q̂ˆ T0 ÿ. This will be the case if the subsample
may be extracted using an extra term of the empirical size is not too large compared to the size of the large sample.
covariogram, cf. Kiêu (1997). Using this technique, the We have studied the bias by simulation too. As
resulting variance estimate becomes measurement function, we have used the function shown
T in Fig. 6. (It is a smoothed version of the data set V5,
m¼0: f12G0 ¹ 31G1 þ 28G2 ¹ 9G3 g ð14Þ
12 discussed in the next section.) On top of this measurement
function we have independent multiplicative log-normal
T
m¼1: f1320G0 ¹ 2155G1 þ 1072G2 ¹ 237G3 g: distributed noise. Given U, Fk is thus log-normal distributed
1320
with
ð15Þ
Var½F jUÿ
The formula for m ¼0 was also derived by Cruz-Orive E½Fk jUÿ ¼ f ðU þ kT0 Þ and 2 k ¼ t2 :
E ½Fk jUÿ
(1993).
More specifically, we have
In some cases, the set X is the disjoint union of a large
number of small subsets with an individual size much ln Fk jU ⬃ Nðyk ; j2 Þ;
smaller than T, like sparsely distributed cells, pancreatic where yk and j2 are chosen such that
islands or kidney glomeruli. In this situation, the empirical
covariogram may appear to have an extra jump at the expðyk þ 12 j2 Þ ¼ f ðU þ kT0 Þ and expðj2 Þ ¹ 1 ¼ t2 :
origin. This effect is sometimes called the small scale effect. In Fig. 7, the results of the simulations are shown for
In such cases formulae (14) and (15) may still be used. noises of various sizes. The noise is represented by t. The fine
spacing T0 has been chosen such that on the average 50
7. Resampling from a large sample observations is in the large sample. Note that the bias is most
important in the case where the noise is most substantial.
In order to investigate the applicability to real data sets of
the transitive methods, it is of course important to be able to
judge the validity of the approximations. This can be done
by comparing with a more laborious method based on
resampling from a large sample.
In this section, we will briefly explain the resampling
method and comment on its statistical properties. Let the
large data set be collected with spacing T0 and consist of the
measurements
Fk ; k ¼ 0; ⫾1; ⫾2; . . . ;
where
E½Fk jUÿ ¼ f ðU þ kT0 Þ;
and U is uniform random in [0, T0). Let us suppose that
T ¼k0T0, where k0 is a positive integer. On the basis of the
data, we may construct k0 estimates of Q with spacing T, viz. Fig. 6. The smooth curve is the measurement function. The obser-
X
∞ vations shown have in addition a multiplicative, log-normal distrib-
Q̂ˆ T;i ¼ T Fjk0 þi ; i ¼ 0; 1; . . . ; k0 ¹ 1: uted noise with coefficient of variation of t ¼ 0·1. For further
j¼¹∞ details, see text.

䉷 1999 The Royal Microscopical Society, Journal of Microscopy, 193, 199–211


206 H . J. G . G U N D ER SE N ET AL .

Fig. 7. The squared coefficient of error as a function of the number of observations in systematic samples with coefficient of variation of t ¼ 0,
t ¼ 0·001, t ¼ 0·01 and t ¼ 0·1, respectively. The oscillating curve is the true squared coefficient of error while the other curve is the approx-
imation based on (6) with m ¼ 1 and (11). The dots are average resampling estimates when using a large sample of size 50.

8. Analysis of 20 data sets where k0 is the integer closest to N/10. There are k0 of these
samples, of average size N/k0.
We have investigated the different methods of variance In each of the plots shown in Fig. 9, R indicates the
estimation in 20 data sets from a range of different 3D resampling estimate as described in Section 7. The
objects. There are 13 data sets, V1, . . . , V13, dealing with remaining four estimates are based on the transitive
volume estimation, based on section data obtained by either methods described in Section 6. Here, (m, þ) and (m, ¹)
magnetic resonance imaging or physical sectioning. In some refer to estimates of the squared coefficient of error under
of these data sets, the areas have been estimated by point- smoothness class m, using either an estimate S2 of j2, as
counting as described in Section 6. The remaining seven described in Section 6, together with the empirical
data sets, N1, . . ., N7, deal with 3D-number estimation from covariogram (þmethod) or the covariogram alone
disector counts. In Fig. 8, some of the data sets are shown, (¹method). Thus, (1, þ), (1, ¹), (0, þ) and (0, ¹) refer
together with their empirical covariograms. For ease of to the formulae (13), (15), (12) and (14), respectively. Note
presentation, the data points as well as the points at which that there is only one resampling estimate while k0
the covariogram is known have been joined by line estimates for the remaining four methods. Note that for
segments. Note that this does not mean that the functions some data sets an estimate S2 of j2 is not available and
are actually continuous. therefore the þmethods cannot be used here.
The different methods of estimating the variance are For the data dealing with volume estimation, the
illustrated in Fig. 9, for all 20 data sets. The estimates of the smoothness class is expected to be m ¼ 1, cf. Section 4,
squared coefficient of error are shown, using five different and this also seems to be the appropriate choice judged from
methods of estimation. The data sets contain a variable Fig. 9, using the resampling estimates as yardsticks. For
number of observations (from 30 to 202). In order to make number estimation from counts, the suitable choice of
the results comparable between data sets, the variance has smoothness class is less clear, in general. For the data
been estimated for a spacing corresponding to roughly the shown here, the jumps of the measurement function are of
same number of non-zero observations, viz. 10 observa- negligible size compared to the sampling density, and we
tions. Thus, if the data set has N observations different from can see from Fig. 9 that m ¼ 1 is an appropriate choice, in
0, with distance T0 between neighbour observations, then terms of both level and variation. Note also that the
the variance is estimated for a sample with spacing k0T0 þmethod provides a lower variability than the ¹method

䉷 1999 The Royal Microscopical Society, Journal of Microscopy, 193, 199–211


S YS T E M ATI C S A M P L I N G I N S T E R E O L O G Y 207

and that the ¹method does not work very well for m ¼ 0;
see the percentage of negative estimates. As expected,
estimates based on knowledge of the mechanism of noise, cf.
(12) and (13), are more stable than those using the
observations alone, cf. (14) and (15). We also tried to
estimate m from the data sets themselves, using the methods
suggested in Kiêu et al. (1999), but we got satisfactory
results only for very long observation series.

9. Discussion of earlier methods of analysis


In Gundersen & Jensen (1987), it was discussed how to
implement the transitive methods suggested by Matheron. It
was in that paper tacitly assumed that m ¼ 0 and,
furthermore, the measurement error was not taken into
account. The resulting variance approximation took the form
T
VarðQ̂T Þ ⬇ f3gð0Þ ¹ 4gðTÞ þ gð2TÞg: ð17Þ
12
If instead m had been chosen to be 1, then the resulting
approximation would have been a factor 20 lower, cf. (8),
T
VarðQ̂T Þ ⬇ f3gð0Þ ¹ 4gðTÞ þ gð2TÞg: ð18Þ
240
In Fig. 10, the resampling estimates are shown, together
with the estimates (17) and (18) of the squared coefficient of
errors, obtained by erroneously neglecting the error. To the
right, estimates of the squared coefficient of error for
independent sampling are also shown. The figure thereby
has the same layout as Gundersen & Jensen (1987, Fig. 7).
The conclusion is that the estimate based on the assumption
m ¼ 0, cf. (17), matches the resampling estimate best, as
was also proposed by Gundersen & Jensen (1987).
The reason is probably that two errors cancel. To be more
specific, recall that under the model described in Section 6,
the variance has two components
Var½Q̂ˆ ÿ ¼ Var½Q̂ ÿ þ Tj2 :
T T ð19Þ
As indicated by earlier investigations in the present paper,
m ¼ 1 is often a suitable choice of smoothness class. By
using (17) instead of (18), Var[Q̂T] is thereby overestimated
and at the same time the error term Tj2 is neglected.
In Gundersen & Jensen (1987, Fig. 8), it was also shown
in a collection of examples that the variance was of order
n¹2. In fact, this order of magnitude can be explained as a
mixture of magnitudes. Thus, if we assume that m ¼ 1, then
combining (9) and (19), we get
l 4 ð3Þ þ ¹4
Var½Q̂ˆ T ÿ ⬇ g ð0 Þ·n þ l j2 ·n¹1 :
360
The variance thereby becomes a mixture of a term of order
n¹4 and a term of order n¹1.
Fig. 8. Some of the 20 data sets (left column) and their empirical
covariograms (right column). The first four data sets (V1, V2, V5, 10. An example and some practical recommendations
V7) deal with volume estimation while the last two (N1, N3) deal
with 3D-number estimation from disector counts. In V1 and V2, The method of estimating the coefficient of error that
point-counting has been used, as described in Section 6. generally seems to work best for systematic sampling along

䉷 1999 The Royal Microscopical Society, Journal of Microscopy, 193, 199–211


208 H . J. G . G U N D ER SE N ET AL .

Fig. 9. The squared coefficient of error for sample size 10 estimated in each of the 20 data sets by five different methods. The symbol R indi-
cates the resampling estimate, cf. Section 7, while (1, þ), (1, ¹), (0, þ) and (0, ¹) correspond to formulae (13), (15), (12) and (14), respec-
tively, in Section 6. The numbers at the top of some of the plots are the percentage of negative estimates. A percentage of 0 (all estimates are
positive) is not indicated.

an axis can be summarized as follows. Let the series of distance between neighbour sampling points is T ¼ 1. If an
measurements be denoted by f1, f2, . . . , fn and let estimate S2 of the cumulative measurement error is
available, then the coefficient of error due to systematic
X
n X
n¹1 X
n¹2 X
n¹3
sampling and measurement error can be estimated by
A¼ f 2i ; B ¼ fi fiþ1 ; C ¼ fi fiþ2 ; D ¼ fi fiþ3 :
i¼1 i¼1 i¼1 i¼1 p
½3ðA ¹ S2 Þ ¹ 4B þ Cÿ=240
CEsys ¼ P ð20Þ
Let us suppose that the scale has been chosen such that the fi

䉷 1999 The Royal Microscopical Society, Journal of Microscopy, 193, 199–211


S YS T E M ATI C S A M P L I N G I N S T E R E O L O G Y 209

Fig. 10. The squared coefficient of error for sample size 10 estimated in data sets dealing with volume and number in 3D objects, respectively,
by four different methods. The symbol R indicates the resampling estimate, while m ¼ 0 and m ¼ 1 indicate the estimates obtained by assum-
ing (erroneously) no measurement error and the indicated smoothness class, cf. (17) and (18). The symbol I indicates the estimate obtained if
the sections are assumed to be independent. The dashed curves indicate four data sets (V6 ¹ V9) where the small scale effect is pronounced.
Two of these curves concern a Poisson-like structure, and here R and I give approximately the same result. The dashed horizontal lines indi-
cate CE ¼ 0·05.

and (iii) Consider carefully the noise mechanism and reduce the
p noise variance to the desired level.
S2 The recommendation (i) is based on empirical evidence
CEnoise ¼ P ; ð21Þ
fi collected from data sets like those analysed in Sections 7
and 8, but evidently need not be valid for artificial or man-
respectively. The estimate of the total coefficient of error
made real objects.
becomes
It follows from the considerations and examples in
p Sections 5 and 8 that (20) is expected to be a realistic
½3ðA ¹ S2 Þ ¹ 4B þ Cÿ=240 þ S2
CEtotal ¼ P ð22Þ estimator of CEsys in a broad class of cases, as recommended
fi
in (ii). Real 3D objects may have artefacts with corners,
Let us consider the example given in Table 1. Suppose edges or flat faces. Such features are not expected to affect
that an estimate S2 is available, S2 ¼ 7·8, say. Applying (20),
Table 1. An example.
(21) and (22), we obtain
CEsys ¼ 0·0091; CEnoise ¼ 0·0548; CEtotal ¼ 0·0555:
Section
The example illustrates nicely what is expected for no, i fi fi fi fi fiþ1 fi fiþ2 fi fiþ3
stereological data from systematic designs of the type
discussed in the present paper, viz. the random noise in
the data will be responsible for the dominant part of the 1 2 4 6 20 16
2 3 9 30 24 24
coefficient of error. Theoretical support for this statement
3 10 100 80 80 90
can be found in Section 6. See also the example
4 8 64 64 72 48
illustrated in Fig. 6. 5 8 64 72 48 32
For systematic designs as described in this paper, we may 6 9 81 54 36 9
therefore make some broad recommendations (with one 7 6 36 24 6
practical exception, discussed at the end of this section): 8 4 16 4
(i) Use a distance between sampling points (sections, 9 1 1
typically) in the systematic design which yields n ⬃ 10. S 51 375 334 286 219
(ii) Use (20) and (22) to check that the contribution to Sfi A B C D
CEtotal due to systematic sampling is negligible.

䉷 1999 The Royal Microscopical Society, Journal of Microscopy, 193, 199–211


210 H . J. G . G U N D ER SE N ET AL .

more demanding. The available estimators of CEtotal are


those based on the variance estimators (14) and (15). Our
experience based on the data analysis with these estimators
is that they are less stable than those based on knowledge of
the error mechanism. The best strategy may be to increase
the sample size n well above 10.

11. Open problems and further research


In the present paper, we have concentrated on systematic
sampling along an axis. There are a number of other cases
which are of interest in practice, most importantly
systematic sampling in the plane and on the unit sphere.
It still remains to study these cases to the same depth as the
one-dimensional case.
Also, further research is needed for one-dimensional
systematic sampling of measurement functions which do
not satisfy the regularity conditions, discussed in Section 4.
Simulations show that in such cases the variance due to
systematic sampling may be of the order of n¹k, where k is
an odd integer. This is also supported by ongoing theoretical
work. As mentioned in Section 4, it is an open problem to
Fig. 11. The squared coefficient of error as a function of the number find variance approximations in this situation.
of observations in a systematic sample. The measurement function Finally, observe that variance approximations taking into
is the length of the intersection between a vertical line and a tilted account other features than the spacing between probes
rectangle (dimension 2 × 1/2). The tilted rectangle makes an angle
have been provided in recent papers. In Cruz-Orive & Gual
a with the vertical direction. In the figure, the squared coefficients
Arnau (1998), volume estimation based on systematic
of error are shown for a ¼ 0⬚ (W), 0·011⬚ (K), 0·040⬚ (þ), 0·29⬚
(×), 1·8⬚ (◊) and 14⬚ (L), respectively. To the right, a-tilted rectan-
slices is considered and variance approximations involving
gles are shown for a ¼0⬚, 1·8⬚ and 14⬚, respectively. The approxi- the height of the slices are suggested. Similar methods can
mate squared coefficient of errors based on (6) are also shown for a be found in Kiêu et al. (1998) for counts based on
tilt of a ¼0⬚ and 14⬚, respectively. systematic disectors.
The data analysed in this paper are available via
CEsys if the sampling density is not high enough to reveal anonymous ftp from ftp://ftp.imf.au.dk/pub/dist/stoclab.
these artefacts, cf. Figs 4 and 5. If sections are taken
perfectly parallel to a flat face, CEsys will decrease only as
Acknowledgements
n¹2, also for small n, cf. Fig. 11. If the face is tilted a bit,
compared to the section direction, CEsys will, however, We are grateful to Jakob Goldbach and Michael Kjægaard
decrease as n¹4. For instance, for the rectangle shown in Sørensen for their skilful technical assistance. We also
Fig. 11, a tilt of about 2⬚ ensures n¹4 behavior already for thank Luis M. Cruz-Orive, Marta G. Fiñana and Ivan Saxl for
n ¼ 10. constructive criticism of an earlier version of the manuscript.
An important practical example of this is a large natural
object that has to be cut into slabs before embedding. To
References
reduce CEsys, it is a good idea to tilt the slab a bit before
sectioning so that one typically obtains a few incomplete Abramovitz, M. & Stegun, I.A. (1964) Handbook of Mathematical
(and small) sections at the beginning and end of each slab. Functions with Formulas, Graphs and Mathematical Tables, Dover
Noise is a very different story. If the data are disector Publications, New York.
counts which, for very small sampling fractions, may be Bagger, P.V., Bang, L., Christiansen, M.D., Gundersen, H.J.G. &
Mortensen, L. (1993) Total number of particles in a bounded
regarded as Poisson distributed then a CEnoise ¼ 0·01 will
region estimated directly with the nucleator: granulosa cell
require a count of about 10 000 particles. More realistic
number in ovarian follicles. Am. J. Obstet. Gynecol. 168, 724–
figures are counts of 100–200, providing a corresponding 731.
CEnoise of 0·1 to 0·07. For point-counting of areas, CEnoise Courant, R. (1934) Differential and Integral Calculus, Vol. I. Blackie,
for a total count of 200 points will rarely exceed 0·02– London.
0·04, cf. Gundersen & Jensen (1987, Fig. 18). Courant, R. (1936) Differential and Integral Calculus, Vol. II. Blackie,
If the noise mechanism is not well defined, the situation is London.

䉷 1999 The Royal Microscopical Society, Journal of Microscopy, 193, 199–211


S YS T E M ATI C S A M P L I N G I N S T E R E O L O G Y 211

Cruz-Orive, L.M. (1989) On the precision of systematic sampling: a striatum: an in situ hybridization study using the optical
review of Matheron’s transitive methods. J. Microsc. 153, 315– fractionator method. J. Comp. Neurol. 370, 11–22.
333.
Cruz-Orive, L.M. (1993) Systematic sampling in stereology.
Proceedings of the 49th Session of the International Statistical Appendix
Institute (Book 2), pp. 451–468. In this appendix, we will indicate briefly the special features
Geinisman, Y., Gundersen, H.J.G., van der Zee, E. & West, M.J.
of the circular case. Here, an integral of the form
(1996) Unbiased stereological estimation of the total number of
… 2p
synapses in a brain region. J. Neurocytol. 25, 805–819.
Gual Arnau, X. & Cruz-Orive, L.M. (1998) Variance prediction Q¼ f ðvÞdv
0
under systematic sampling with geometric probes. Adv. Appl.
Prob. 28, 982–992. is to be estimated where f is a non-negative function defined
Gundersen, H.J.G. (1986) Stereology of arbitrary particles. A on [0, 2p). The estimator becomes
review of unbiased number and size estimators and presentation
n¹1  
of some new ones, in memory of William R. Thompson. J. 2p X 2p
Microsc. 143, 3–45.
Q̂n ¼ f ⌰þi ;
n i¼0 n
Gundersen, H.J.G. (1988) The nucleator. J. Microsc. 151, 3–21.
Gundersen, H.J.G. & Jensen, E.B. (1987) The efficiency of where ⌰ is uniform in [0, 2p/n).
systematic sampling in stereology and its prediction. J. Microsc. In the circular case, the suitable way to define the
147, 229–263. covariogram is as follows:
Jensen, E.B.V. (1998) Local Stereology, World Scientific, Singapore.
… 2p
Kellerer, A.M. (1989) Exact formulae for the precision of systematic
sampling. J. Microsc. 153, 285–300. gðaÞ ¼ f ðvÞ f ðv þ aÞ dv; a 僆 ½0; 2pÞ;
0
Kiêu, K. (1997) Three Lectures on Systematic Geometric Sampling.
Memoirs No. 13, Department of Theoretical Statistics, Institute where we use a periodic continuation of f. In analogy with
of Mathematics, University of Aarhus. (1), we can express the estimator variance as
Kiêu, K., Souchet, S. & Istas, J. (1999) Precision of systematic
n¹1   … 2p
sampling and transitive methods. J. Statist. Plan. Inf. in press. 2p X 2p
VarðQ̂n Þ ¼ g j ¹ gðvÞdv:
Kiêu, K., Xiong, W. & Trubuil, A. (1998) Precision of systematic n j¼0 n 0
counts. Rapport Technique 1998–1, Unité de Biométrie, INRA-
Versailles. If a periodic continuation of the function f satisfies the
Matérn, B. (1989) Precision of area estimation: a numerical study. regularity conditions stated in Section 4, except for the
J. Microsc. 153, 269–284. condition that f has compact support, we say in the circular
Matheron, G. (1965) Les Variables Régionalisées et leur Estimation.
case that f is of smoothness class m. The formula (6) still
Masson, Paris.
holds, i.e.
Matheron, G. (1971) The Theory of Regionalized Variables and its
Applications. Les Cahiers du Centre de Morphologie Mathéma-
 2mþ2
2p B2mþ2 ð2mþ1Þ þ
tique de Fontainebleau, No. 5. Ecole Supérieure des Mines de VarðQ̂n Þ ⬇ ¹2 g ð0 Þ:
n ð2m þ 2Þ!
Paris, Fontainebleau.
Mattfeldt, T. (1989) The accuracy of one-dimensional systematic In the above formula as well as in the main text, Bernoulli
sampling. J. Microsc. 153, 301–313. numbers are used. The Bernoulli numbers Bk, k ¼ 0, 1, . . . ,
Mulders, W.H.A.M., West, M.J. & Slomianka, L. (1997) Neuron are defined by
numbers in the presubiculum, parasubiculum, and entorhinal B0 ¼ 1
area of the rat. J. Comp. Neurol. 385, 83–94.
Souchet, S. (1995) Précision de l’estimateur de Cavalieri, Rapport k 
X 
kþ1
de Stage, D.E.A. de Statistiques et Modèles Aléatoires appliqués à la Bv ¼ 0; k ¼ 1; 2; . . .
Finance, Université Paris-VII. Laboratoire de Biométrie, INRA, v¼0
v
Versailles.
Note that
Sterio, D.C. (1984) The unbiased estimation of number and sizes of
arbitrary particles using the disector. J. Microsc. 134, 127–136. 1 1 1
B1 ¼ ¹ ; B2 ¼ ; B3 ¼ 0; B4 ¼ ¹ ; B5 ¼ 0
West, M.J. & Slomianka, L. (1998) Total number of neurons in the 2 6 30
layers of the human entorhinal cortex. Hippocampus, 8, 1–14.
West, M.J., Østergaard, K., Andreassen, O.A. & Finsen, B. (1996) 1 1 5
B6 ¼ ; B7 ¼ 0; B8 ¼ ¹ ; B9 ¼ 0; B10 ¼ :
Estimation of the number of somatostatin neurons in the 42 30 66

䉷 1999 The Royal Microscopical Society, Journal of Microscopy, 193, 199–211

You might also like