2003 Joint Statistical Meetings - Section on Survey Research Methods
A NEW PERSPECTIVE ON CALIBRATION ESTIMATORS
Victor M. Estevao and Carl-Erik Srndal
Statistics Canada, Ottawa, Ontario, Canada, K1T 3L2
KEY WORDS: Auxiliary Information, Calibration
Weights, Automated Linearization, Population
Residuals.
1 Introduction
Complex sampling design and complex parameter are
familiar concepts in survey sampling. Another survey
feature that often extends beyond the simple ordinary
formulation is the use of auxiliary information at the
estimation stage. As the recent literature has shown,
auxiliary information can be more or less complex,
depending on the survey design. Estimation for
complex cases is not well-covered by standard textbook
techniques. A broader framework for estimation is
needed with such auxiliary information. We use the
term complex auxiliary information for the several
non-standard cases examined in this paper.
The most basic use of auxiliary information occurs
for a single stage, single phase sampling design, where
a known set of auxiliary variables and their
corresponding totals are used to compute calibrated
weights for the estimate of a population total. This
procedure is reviewed in section 2. More complex cases
considered in later sections arise when auxiliary
information is available at different stages and phases
of sampling. Then it is not always evident how to make
efficient use of the available auxiliary information. In
this paper, we look at different ways of using complex
auxiliary information to produce efficient calibration
estimators in two-stage and two-phase sampling. The
derivation of the variance of these estimators requires a
simple procedure to linearize the expression for a
nonlinear calibration estimator. This simple procedure
is introduced as the method of automated linearization.
The paper is arranged as follows. In section 2 we
explain automated linearization for calibration
estimators in a one-stage, one-phase sampling design.
Section 3 examines estimation for two-phase sampling
designs, and section 4 looks at estimation in two-stage
sampling with and without integrated weighting. A
brief summary and comments are given in the
concluding section 5.
2 Automated linearization
We first look at the simple case of auxiliary information
for a one-stage, one-phase unit sampling design.
Consider a finite population U = { 1, 2, ..., k , ..., N }
from which a probability sample s is drawn. We denote
1346
by k the inclusion probability of unit k and by
a k = 1 / k the sampling weight of k. Let y be the
variable of interest. Its value for unit k, y k , is observed
for k s . The unknown total to be estimated is
Y = U y k .
We denote by x an auxiliary vector of dimension
J 1 , and by x k its value for unit k. We assume that
we have the following auxiliary information:
(i) The population vector total X = U x k is known.
(ii) The vector value x k is known for every k s .
Here, X is assumed known from an outside source
such as a census. If we know the value x k for every
k U , as when x k is on the population frame U for
every k, then both (i) and (ii) are met. We can compute
the simple unbiased estimator of the known X as
= a x . Under general conditions N 1 ( X
X X) is
s k k
O p (n 1/ 2 ) .
Our objective is to estimate Y = U y k . One
possibility is the simple unbiased Horvitz-Thompson
(HT) estimator Y = s ak y k . However, a more
efficient weighting of the observed yk is one that takes
the auxiliary information into account. Let us consider
instead YCAL = s wk yk , where the weights
{ wk ; k s } satisfy the calibration equation
s wk x k = X . We say that the weights { wk ; k s } are
calibrated to X = U x k .
Alternative sets of calibrated weights can be
derived by the distance measure approach, as for
example in Huang and Fuller (1978) and Deville and
Srndal (1992). The minimization of each distance
measure produces a different set of calibrated weights.
However, the proposed distance measures are fairly
similar so they tend to produce estimators with similar
properties. Instead we use the instrument vector
approach, also called generalized calibration, as in
Deville (2002) and Le Guennec and Sautory (2002).
This method allows a more general parameterization of
 2003 Joint Statistical Meetings - Section on Survey Research Methods
the calibration weights. We specify a vector z k of the
same dimension as x k and compute the weights
wk = a k (1 + Ts z k ) , k s
) T ( a z xT ) 1 . The mapping from
where Ts = (X X s k k k
z k to wk is not one-to-one. Different choices for z k
produce the same weights wk . We are free to choice
the form of z k as long as the J J matrix
(s a k z k xTk ) has an inverse for every possible sample
s. We refer to such a z k as a valid instrument vector.
The standard choice of z k = x k produces the
generalized regression estimator, although as explained
later, this choice is not necessarily optimal for any
given design. For any valid instrument vector z k , the
weights satisfy s wk x k = X , and the estimator can be
written as YCAL = s wk y k = Y + ( X X
)T B , where
B = ( a z xT ) 1 a z y .
s k k k s k k k
Here B is a nonlinear design-weighted statistic,
thus, it is not what we call a HT statistic. Although Y is
a linear statistic, the term ( X X )T B
makes Y
CAL a
nonlinear estimator. This causes no problem for point
estimation since YCAL can be readily computed. But the
nonlinear form of the estimator creates a problem for
obtaining a simple exact expression for the variance of
YCAL and for finding a corresponding sample-based
estimate of this variance. Linearization is the usual
technique for circumventing this difficulty with
nonlinear statistics. Woodruff (1971) is a basic
reference. Since then, many papers have appeared on
the linearization of complex statistics of interest in
survey sampling, for example, Binder (1996), Binder
and Kovaevi (1995), and Deville (1999). The
emphasis in these references is on linearization of
statistics for estimating complex parameters, a purpose
somewhat different from ours, which is the study of
calibration estimators of a total. Thberge (1999)
presents a linearization approach similar to the one
given here. His development is based on the use of
distance functions rather than an instrument vector.
Linearization of the nonlinear YCAL involves a
power series expansion, including an evaluation of
partial derivatives. The rather lengthy derivation is
given for example in Srndal, Swensson and Wretman
(1992). This method isolates a main term, YCAL ,lin ,
which is a linear statistic. The remainder term is of
lower order in probability and assumed negligible
compared to the main term. The expression for the
remainder term is usually not made explicit in
(2.1) Woodruff linearization. This is not a serious drawback,
because standard practice is to discard this term and
simply take YCAL ,lin to be a sufficiently good linear
approximation to YCAL . Under general conditions,
N 1 (YCAL YCAL ,lin ) is O p (n 1 ) , not just O p (n 1/ 2 ) ,
permitting the easily derived variance of Y to be
CAL ,lin
used as an accurate approximation of the variance of
YCAL , even for modest sample sizes.
Instead of the standard linearization approach, we
introduce the method of automated linearization. This
simple two-step procedure automatically makes
explicit both the linearized statistic and the lower order
term. In contrast to Woodruff linearization, automated
linearization requires no evaluation of partial
derivatives. For the case of simple auxiliary information
in this section, we confirm the well-known expression
for the variance of YCAL ,lin . Automated linearization has
two steps:
Step 1. In the expression YCAL = Y + ( X X
)T B
, create
a term of lower order in probability by centering B on
the population vector B = (U z k xTk ) 1 U z k yk to
which B converges in probability. Then B B is
O p (n 1/ 2 ) , and we have
YCAL = Y ( X
X) T B ( X
X)T (B B) (2.2)
X)T (B B) is O (n 1 ) , a lower order
where N 1 ( X p
compared to N (X X) B which is O p (n 1/ 2 ) .
1 T
Step 2. Rewrite (2.2) as
YCAL = (Y X
T B) + X T B ( X
X)T (B B) . (2.3)
The calibration estimator is the sum of three terms: the
constant term XT B , the design-based linear term
Y X
T B , and the design-based nonlinear term
(X X)T (B B) of lower order. The first two terms
on the right hand side of (2.3) define the linearized
statistic
T B) + X T B = a e + X T B
YCAL ,lin = (Y X s k k (2.4)
where ek = yk xTk B .
1347
 2003 Joint Statistical Meetings - Section on Survey Research Methods

Our point estimator of Y is YCAL . It has a small
bias, E(YCAL ) Y = E{ ( X
X ) T (B
B) } , since
N 1 E{ ( X
X)T (B B) } is of order O(n -1 ) .
Therefore, the variance of YCAL is approximately the
variance of the linearized statistic Y . Since XT B
CAL ,lin
is a constant, the use of auxiliary information reduces
the variance of the estimator from Var (s ak yk ) to
approximately Var (s ak ek ) . It is important to note
that the ek are fixed but unknown values and that
s ak ek is a HT statistic in the ek . Although the ek
resemble regression residuals, they arise automatically
from steps 1 and 2, without any explicit regression
model or fit. Because s ak ek is a HT statistic, we
obtain immediately,
Var(YCAL ) Var(YCAL ,lin ) = U Fkl ek el
aa
where Fkl = k l 1 for l k , Fkl = Fkk = a k 1 for
a kl
l = k , with a kl = 1 / kl where kl is the joint inclusion
probability of k and l . We use U as shorthand
for the double sum . To estimate the variance of
kU lU
(1987) determination of B so as to minimize the
variance of the unbiased difference estimator
Y + ( X X
)T B .
To see the features of the weights, let us write them
as wk = a k { 1 + (U x k s a k x k )T (s ak z k x Tk ) 1 z k } .
We note the following:
(i) The computation of the weights wk for k s
requires the design weights a k , the auxiliary
vector values x k , the instrument vector values z k ,
and the known auxiliary vector of totals U x k .
(ii) The a k are fixed by the design.
(iii) We are free to choose the z k as long as the matrix
s a k z k xTk is invertible.
(iv) The weights wk calibrate to the known totals
(2.5) U x k for any valid instrument vector z k .
(v) The weights wk are not dependent on y or on any
presumed relationship between y and x, as in a
model dependent approach.
Some choices of z k are better than others. The
optimal choice, as noted above, is
z k = z k = lU Fkl x l . It makes sense that the optimal
0

YCAL , we use the sample-based analogue of (2.5), choice depends on the sampling design. The
sample-based choice corresponding to z 0k is
V (YCAL ) = s (ak al akl )ek el (2.6)
z k = z *k = a k1 ls a kl Fkl x l . The weights wk do not
where ek = yk xTk B and s stands for . depend on the values yk of the variable of interest y
ks ls
and thus the optimal weights do not depend on yk .
The weights wk in YCAL = s wk y k depend on the Once the z k are specified, the same weights can be
used for all y-variables in the survey. The estimator
instrument vector z k . For every choice of z k for
YCAL is free of any unverifiable assumptions about a
k U , there corresponds a vector B satisfying the
possible regression of y on x. In the application of this
equation (U z k xTk ) B = U z k yk . We can find an approach it does not matter whether there exists a linear
optimal B, and a corresponding z k , by minimizing relationship between y and x. Furthermore, no
assumptions are required on the properties of the
Var(Y ) given by (2.5). This z is asymptotically
CAL ,lin k
residuals ek . These are treated as fixed but unknown
optimal for YCAL in that it minimizes values over the population U rather than random
Var(Y ) Var(YCAL ) . The optimal B is B 0 , defined variables from a hypothetical superpopulation model.
CAL ,lin

as the solution of the normal equation As a simple illustration, consider Simple Random
Sampling (SRS) from U with the sampling fraction
(U Fkl x l x ) B = U Fkl x l y k .
T
k
0
(2.7)
f = n / N and consider x k = (1, xk )T , where xk is a
A comparison with the general form scalar variable value. The required population
(U z k x Tk ) B = U z k y k defining B, suggests that an information is U x k = ( N , U xk )T . Then the optimal
optimal instrument vector is z k = z 0k , where instrument is found to be z k = z 0k = lU Fkl x l =
z 0k = lU Fkl x l . The result agrees with Montanaris

1348
 2003 Joint Statistical Meetings - Section on Survey Research Methods
N 1
( 1)(x k xU ) where xU = U x k N . The
N 1 f
corresponding sample based choice is z k = z *k =
n 1
( 1)(x k x s ) where x s = s x k n . However
n 1 f
both z 0k and z *k are invalid because the first
component of these vectors is always zero, leading to a
singular matrix s a k z k xTk . We drop the first auxiliary
variable with the known total N and work instead with
the vector x k = xk . This gives
n 1
z k = z k* = ( 1)( xk xs ) . The result is the
n 1 f
familiar YCAL = N { y s + ( xU xs ) b }
b = { s ( xk xs )( y k y s ) s ( xk x s ) } . As is easily
2
verified, this estimator gives YCAL = N when y k = 1 for
all k, and YCAL = N xU = U xk when y k = xk . That is,
even though we must reduce the auxiliary vector from
x k = (1, xk )T to x k = xk , the resulting set of weights
still reproduce the two known quantities N and U xk .
No loss of information is incurred from the
non-invertibility of s a k z k xTk with x k = (1, xk )T .
We end this section by listing the steps of the
preceding argument. These important steps are applied
in each of the subsequent sections, where the auxiliary
information is more complex.
Step 1 The auxiliary information: Specify an
x k -vector with known totals.
Step 2 Point estimation: Specify a valid z k , compute
the calibrated weights and the resulting point
estimate.
Step 3 Variance and variance estimation: Use
automated linearization to (a) identify the
linearized statistic, (b) obtain the residuals that
determine the variance, and (c) transform that
variance into an estimated variance.
3 Calibration estimation in two-phase sampling
We now consider the setup for sampling in two phases.
From the population U = { 1, 2, ..., k , ..., N } , a large
probability sample, s1 , is drawn with known first-phase
inclusion probabilities 1k . The first-phase sampling
weights are a1k = 1 / 1k for k s1 . One or more
variables are observed for k s1 . Then, from s1 , a
sub-sample, s, is drawn with known conditional
probabilities 2 k . The second-phase sampling weights
are a 2 k = 1 / 2k , conditionally on the realized s1 . We
denote by ak = a1k a2 k the overall sampling weight for
unit k. The value yk of the variable of interest is
observed for all k s . The objective is to find a more
efficient alternative for estimating Y = U yk than the
two-phase double expansion estimator Y = s ak y k .
We need to consider two auxiliary vectors for each
unit k. We denote these by x1 and x 2 , with x1k and
x 2 k representing their respective values for unit k.
Their dimensions are J1 1 and J 2 1 respectively.
The auxiliary information for x1 and x 2 is as follows:
with
(i) The population vector total X1 = U x1k is known
while the population vector total X 2 = U x 2k is
not known.
(ii) For every k s1 , the vector values x1k and x 2 k
are known.
(iii) For every k s , the vector values x1k and x 2 k
are known.
The information given by (i), (ii) and (iii) is used to
compute the weights for the calibration estimator
YCAL = s wk y k in an effort to improve on
Y = s ak y k . There are different ways to produce these
weights wk , depending on how we use (i) to (iii). For
example, we can carry out a single calibration step, or
arrive at the wk in two steps by first producing a set of
first-phase weights w1k . Each step requires starting
weights, an auxiliary vector and a valid instrument
vector. We consider the following alternatives:
(a) One step calibration. Starting from ak = a1k a2 k ,
compute directly final weights wk for k s ,
U x1k
calibrated to satisfy s wk x k = , with
s1 a1k x 2k
x
x k = 1k of dimension J1 + J 2 . This is case B1
x 2k
in Estevao and Srndal (2002).
(b) Two step calibration. In step one, starting from
a1k , compute first-phase weights w1k for k s1 ,
such that s1 w1k x1k = U x1k . In step two, starting
from ak = a1k a2 k , and using the w1k from step (i),
compute final weights wk for k s , such that
1349
 2003 Joint Statistical Meetings - Section on Survey Research Methods

s wk x k = s1 w1k x k . The final weights satisfy YCAL = s a k ( y k x T B)

(3.1)
U x1k + s1 a1k (xTk B x1Tk B1* ) + X1T B1* + R
s wk x k = w x . This is case A1 in
s1 1k 2 k where R is the lower order term given by
Estevao and Srndal (2002). T
B 1 B1
X X
R = 1 1 (X
X )T (B
* B* )
The two procedures make slightly different use of B B 1 1 1 1

the auxiliary information and in general, they produce X

2 X 2 2 2

different weights wk for k s . The use of information

is somewhat more extensive in (b) than (a), in that it
requires the information about the individual values x1k
for k s1 . This may or may not lead to an increase in
efficiency, depending on the relation between x1k , x 2 k
and yk . These questions are discussed in Estevao and
Srndal (2002). We can use automated linearization to
obtain the form of the residuals and the variance of each
estimator in (a) and (b). We examine case (b) below.
The first-phase calibrated weights for case (b) are
computed as
w1k = a1k (1 + Ts1 z1k ) with
Ts1 = (U x1k s1 a1k x1k )T (s1 a1k z1k x1Tk ) 1
for some valid instrument vector z1k . The w1k are then
used as input to compute the final calibrated weights as
= a x and X
with X j s1 1k jk j = s a k x jk , j = 1 , 2 . The
term of main interest is the linear statistic
YCAL ,lin = s ak ek + s1 a1k ek* + X1T B1* (3.2)
where X1T B1* = ( U x1k )T B1* is a constant, and the
residuals in the two random terms are
e k = y k x Tk B = y k x 1Tk B 1 x T2k B 2 and
ek* = xTk B x1Tk B1* .
By ignoring the lower-order term in (3.1), we can
use the linear design-weighted statistic (3.2) to obtain
the approximate variance of YCAL . Then, conditioning
on s1 , we obtain
Var(YCAL ) Var(YCAL ,lin ) = V1 (E c ) + E1 (Vc )

wk = ak (1 + Ts z k ) where where E c = s1 a1k ( yk x1Tk B1* ) and Vc is the

T 1
= (sI w1k x k s ak x k ) (s ak z k x )
T
s
T
k conditional variance of s ak ek , given s1 . The
expressions for V1 (E c ) and E1 (Vc ) , and for their
x
where x k = 1k and z k is another valid instrument respective estimates are not detailed here. They follow
x 2k well-known patterns for two-phase sampling as shown
vector. We derive the variance by using automated for example in Estevao and Srndal (2002). For the
linearization. First, we insert into YCAL = s wk yk the estimated variance, we use B and B1* instead of B
B and B1* . Note that E1 (Vc ) = 0 if sample selection stops
expression for wk . Then, we define B = 1 =
after the first phase.
B2
( a z x )
s k k
T 1
k (s a k z k yk ) and center it on its The first term, V1 (E c ) , is reduced by the presence
population counterpart, the non-random vector in ek* of the regressor x1k only, whereas the second
B
( 1
)
B = 1 = U z k xTk (U z k y k ) . After some algebra,
term, E1 (Vc ) , gets reduced by both regressors, x1k and
B2 x 2 k . These features seem logical under the survey
we then define the statistic B1* = conditions. An interesting question, which we leave
unresolved here, is the jointly optimal choice for the
( s1 a1k z1k x1Tk ) 1 ( s1 a1k z1k (xTk B
)) , which we center
two instruments, z1k and z k . The simple standard
on its population counterpart B1* = choices are z1k = x1k and z k = x k .
( U z1k x1Tk ) 1 ( U z1k (xTk B)) . The result is
We comment briefly on the automated linearization
in case (a). The outcome is also a variance of the form
V1 (E c ) + E1 (Vc ) , with one residual for the first

1350
 2003 Joint Statistical Meetings - Section on Survey Research Methods
component, V1 (E c ) , and another residual for the
second, E1 (Vc ) . But these residuals are somewhat
different in (a) and (b). For (b), we stated earlier in this
section the first component residuals as
ek = yk x1k B1 x 2k B 2 , showing a removal of the
T T
influence of both x1k and x 2 k , and those of the second
component as e = x B x B , showing a removal of
*
k
T
k
T
1k
*
1
the influence on xTk B (rather than on yk ) of x1k alone.
The same pattern holds for (a), in that both x1k and x 2 k
are removed in the first residual, and x1k alone in the
second. Cases (a) and (b) differ in the B-coefficients of
the two kinds of residuals. The automated linearization
of (a) readily reveals the form of these B -coefficients.
We do not show them here. The important point is that
the influence of the x-vectors is removed according to a
common pattern, although the values of ek and ek* are
not the same. Thus we can expect that (a) and (b) will
usually generate rather small differences in the variance
of the corresponding YCAL estimators. This is confirmed
by the simulations in Estevao and Srndal (2002). For
unusual relationships between yk , x1k and x 2 k , the
differences can be more significant. Further studies are
needed to examine this.
4 Calibration estimation in two-stage sampling
We start from the usual formulation of sampling in two
stages. A sample of units is realized by two-stage
selection from a population U = { 1, 2, ..., k , ..., N }
grouped into clusters. This design involves sampling
from two distinct populations of interest: (i) the
population of first stage clusters,
U I = { 1, 2, ..., i, ..., N I } , and (ii) the population of
second stage units U = { 1, 2, ..., k , ..., N } . For
simplicity, we refer to them as the population of
clusters and the population of units respectively. The
population U = { 1, 2, ..., k , ..., N } is the union of all
the units U i , i U I in the N I clusters.
First, we draw a sample of clusters sI from U I ,
with known first-stage inclusion probabilities 1i . The
first-stage sampling weights are aI i = 1/ I i for i sI .
At the second stage, we sample units within each of the
selected clusters. From U i , we draw a sample si of
units, with known second-stage probabilities k i
conditional on si . The conditional sampling weights
are ak i = 1/ k i for k si . Thus, a k = a I i a k i is the
1351
overall sampling weight for unit k, and s = U si is the
isI
sample of units. The value yk of the variable of interest
is observed for all units k s . We want to estimate the
total Y = U yk , but more efficiently than with the
simple unbiased Y = s ak y k = sI a I i (si a k i y k ) .
In general, auxiliary information exists for both the
units and the clusters. We denote by x I i an auxiliary
vector value associated with cluster i, and by x k an
auxiliary vector value associated with unit k. We
consider the following information to be available:
(i) The cluster population vector total U I x I i is
known.
(ii) For every i sI , the cluster vector x I i is known.
(iii) The unit population vector total U x k is known.
(iv) For every k s , the unit vector x k is known.
If x I i is known for every i U I , then (i) and (ii)
are met. This occurs, for example, in area sampling
where each cluster is a geographical entity for which
we have a useful auxiliary measurement vector, for
example, the surface area and/or the number of
inhabitants. On the other hand, it is unlikely that we
would have information x k about every unit k U in
a survey where the absence of a list frame of units
precludes single stage sampling and forces us to use
two-stage sampling. But conditions (iii) and (iv) are
met if x k is recorded for all sampled units and the total
U x k can be imported from an accurate outside
source, a census or a census projection, as it is, for
example, in the Canadian Labour Force survey. This
section examines calibration estimators derived from
some or all of the information (i) to (iv).
The information is somewhat different when there
is a known value x k for every unit k U i , where U i is
a selected cluster, i sI . This case is covered by (i) to
(iv) and we need not consider it, because the known
cluster total t x i = U i x k for i sI can then be entered
into x I i in (ii), assuming U I t x i = U x k is also
known.
Surveys involving sampling of clusters often have
the double objective of computing estimates of totals
for both the population of units U (referred here as unit
statistics) and the population of clusters U I (cluster
statistics). Then, we observe both the value of a cluster
 2003 Joint Statistical Meetings - Section on Survey Research Methods
variable of interest, yI i for i sI , and the value of a
unit variable of interest, yk for k s . For example, if
households are clusters, yI i may be the value for
household i of the variable yI = household income; and
if units are persons in the households, yk may be the
value of the variable of interest y = employment status
(0 if employed, 1 if unemployed).
The totals to be estimated are then YI = U I y Ii for
statistics on household income, and Y = U y k for
statistics on individuals employment. We thus examine
the calibration estimators YI,CAL = sI wI i y Ii and YCAL =
s wk y k with cluster weights wI i satisfying
sI wI i x I i = U I x I i
and unit weights wk satisfying
s wk x k = U x k .
We also allow for the fact that many two-stage
designs call for some form of integrated weighting. Its
objective is to impose a simple relation between a
cluster weight wI i and the weights wk for the selected
units k of that cluster. The interest in integrated
weighting is promoted by Eurostat in its efforts to
harmonize the estimation methods used by the member
states of the European Union. Also, integrated
weighting schemes are of interest for the further
development of generalized estimation systems such as
Bascula, CLAN and GES. We examine two options for
integrated weighting:
(1) si wk = N i wI i for every i sI , where N i is the
known size of cluster i.
(2) wk = ak i wI i for the selected units k in cluster
i sI .
Each option imposes a simple relationship between
the wk and the wI i . Depending on the option selected,
we can write (4.1) and (4.2) in terms of either wk or
wI i . We assume that the resulting set of equations is
consistent. Depending on the choice, there is some
effect on the precision of the resulting calibration
estimates, as discussed in this section. Option (1) is
based on the requirement that the estimated number of
units within any group of clusters must be the same
whether the cluster weights or the unit weights are used
to create that estimate. Option (2) preserves the
conditional design weights. One can argue that option
(2) is slightly simpler than (1) but it actually imposes
more severe restrictions on the unit weights. As we see
later, this has implications on the variance of the
estimators.
A special case of (2) that has drawn considerable
attention occurs for single stage cluster sampling, see
for example Lematre and Dufour (1987), Andersson
(1997), and Nieuwenbroek (1993). Since all k in cluster
i are observed, a k i = 1 and (2) implies wk = wI i . It is
practical to assign the same weight to all units in a
cluster, for the calculation of unit statistics, and this
common weight is the cluster weight for cluster
statistics. The approach of Lematre and Dufour (1987)
differs from ours. They find wk to satisfy (4.2) but in
(4.1) such a way that the known auxiliary vector value x k is
replaced by one and the same constructed value,
Ui x k / N i , for every k in cluster i. By contrast, we
keep the individual x k and use one of the integrated
(4.2)
weighting options to set up the calibration problem,
leading to integrated wk and wI i . The calculation of
the weights of the calibration estimators YCAL = s wk yk
and Y = w y is described below.
I ,CAL sI Ii Ii
(a) Non-integrated calibration: Starting from aIi ,
compute cluster weights wI i for i sI , calibrated
to the cluster information in the manner of (4.1); in
an independent second calibration, starting from
a k = a I i a k i , compute unit weights wk for k s
calibrated to the unit information as stated in (4.2).
(b) Calibration with integration option (1): In (4.1),
replace wI i by si wk / N i , making that equation a
function of the wk . Assign the equal shares
value x i k = x I i / N i to every selected unit k in
cluster i. Then starting from a k = a I i a k i , compute
unit weights wk for k s , calibrated to satisfy
x i k U I x I i
s wk = . Then compute the cluster
x k U x k
weights as wI i = si wk / N i .
(c) Calibration with integration option (2): In (4.2),
replace wk by a k i wI i , making that equation a
function of the wI i . Starting from a I i , compute
cluster weights wI i for i sI , calibrated to satisfy
1352
 2003 Joint Statistical Meetings - Section on Survey Research Methods

Case Integrated Method Calibration Equation(s)

Weighting
Option
Using a I i as starting weights, compute wI i to
(a) None sI wI i x I i = U I x I i
satisfy (4.1). Independently, using ak as
s wk x k = U x k
starting weights, compute wk to satisfy (4.2).
In (4.1), replace wI i by si wk / N i . With ak as
(b) si wk = N i wI i x ik U I x I i
starting weights, compute wk to satisfy both s wk =

x k U x k
(4.1) and (4.2). Then compute the wI i .
In (4.2), replace wk by ak i wI i . With a I i as
(c) wk = ak i wI i x I i U x I i
starting weights, compute wI i to satisfy both sI wI i t = I x
x i U k
(4.1) and (4.2). Then compute the wk .

Table 1. Summary of Calibrated Weighting Methods for Two-Stage Estimation.

x I i U x I i component of variance. These residuals are summarized

sI wI i t = I x , where t x i = si ak i x k is in Table 2.
x i U k
the sample-weighted, unbiased estimator of Consider case (b) for unit statistics. The total to
t x i = U i x k . Then compute the unit weights as estimate is Y = U y k . The weights for YCAL = s wk y k
wk = ak i wI i . are computed for k s as wk = ak (1 + Ts z k ) with
1
x a x T T

The procedures for (a), (b) and (c) are summarized a z x i k ,
= U I I i s k i k
T
k k
x k
U x k s ak x k
s s
in Table 1. All three cases reduce to a weight
calculation of the form (2.1), just as in single-stage unit
sampling. That is, despite the two stages of sampling, where z k is any valid instrument, and x i k = x I i / N i for
the point estimation does not become any more every selected unit k in cluster. The estimator of the
complex than in single-stage unit sampling. A software total for units, is YCAL = s wk y k . Automated
programmed to compute formula (2.1), such as
CLAN97 or GES, can be used to compute the linearization gives YCAL = YCAL ,lin + R , where R is the
calibration estimators in (a), (b) and (c). However, the lower order term
two-stage design leads to a more complicated variance T
than in section 2. The variance has two components, ak x i k U I x I i
R = s (B B )

one for each stage of selection, as shown later in the s ak x k U x k
section. 1
B x
T

with B = 1 = s ak z k i k

(s ak z k yk ) , and
x k
We have three cases, (a) to (c), and for each, both
cluster statistics and unit statistics are examined. There B2
are thus 3 2 = 6 situations to examine. For each of the linearized statistic is
these, the approximate variance of the calibrated T
estimator (which equals the variance of the linearized x Ii B1
YCAL ,lin = s ak ek + U I . (4.3)
statistic) has the form V1 (E c ) + E1 (Vc ) , where V1 (E c ) U x k B2
is the first stage variance component and E1 (Vc ) the
second stage variance component. The latter is zero if The second term on the right hand side is a
there is no sampling at the second stage, that is, all units constant, and the preceding random term, s ak ek =
in selected clusters are observed (single stage cluster
sampling). It is straightforward to carry out the sI aI i (si ak i ek ) , has the residuals
automated linearization in the 6 situations. This leads to
the expression for the residuals, one for each

1353
 2003 Joint Statistical Meetings - Section on Survey Research Methods

T T
x B1 x I i B ( c)
ek = y k i k with eIi = y I i 1( c ) . The residuals for unit and
xk B2 t x i B2
1 cluster statistics, are summarized in Table 2.
B1 xi k
B = = U z k (U z k yk ) . (4.4)
B2 xk Another issue of interest in case (b) is the choice of
the instrument z k . The standard choice is to take
x
By conditioning on sI , and using Var(YCAL ,lin ) = z k = i k for k in cluster i. But one can derive a
V1 (E c ) + E1 (Vc ) , we find xk
B = B 0 that minimizes (4.5), with a corresponding
Var (YCAL ) Var(YCAL ,lin ) optimal z k = z 0k . Some algebra shows that z 0k =
(4.5)
= U I FI i j eI i eI j + U I a I i Vi x xIj
a I i lU i Fkl i i l + jU I FI i j for k in cluster i. It

x Ii
T xl t x j
B1
where eI i = U i ek = t y i is no surprise that z 0k depends on the sampling design
t xi B2
a k i al at both stages. Future work will examine when z k = z 0k
and Vi = U i Fkl i ek el with Fkl i =
i
1 and is a valid instrument and whether z k = z 0k gives any
a kl i
appreciable variance advantage over the simple
aI i aI j standard z k = x k . If this advantage is minimal, the
FI i j = 1 .
aI i j preferred choice in practice is the simple z k = x k .

The residual eIi in the first component V1 (E c ) = Consider now case (c). The calibrated estimator of
U I FI i j eI i eI j equals the cluster total of the residuals the cluster total YI = U I y Ii is YI ,CAL = sI wI i yI i , with
ek in the second component E1 (Vc ) = U I aI i Vi . Both the cluster weights wI i = a I i (1 + TsI z i ) for i sI , where
eIi and ek have their magnitude reduced by both the T 1
U x I i s a I i x I i xIi
cluster auxiliary and the unit auxiliary. The regressor is = I I sI a I i z i
T
U t x i
I sI I i x i
sI
a t t xi
(x TI i , t Tx i ) in eI i and (x Ti k , x Tk ) in ek . In particular, in
single stage cluster sampling, Vi = 0 for all i, and only and t x i = si a k i x k . The calibrated estimator of the
the first variance component remains. unit total Y = U y k is YCAL = s wk yk , where the
Consider now case (b) for cluster statistics. The integrated unit weights for k in cluster i are
wk = ak i wI i , using the computed wI i . We can use
total YI = U I y I i is estimated by YI,CAL = sI wI i yI i ,
where the wI i are computed from the already available automated linearization on YI ,CAL and YCAL to obtain
w as w = w / N . We can write Y as a sum the linearized statistic and the residuals that determine
k Ii si k i I ,CAL
the two components of the approximate variance. The
of unit values YI,CAL = s wk yi k , if we define details of the derivations are omitted. The residuals,
yi k = yI i / N i for all k in cluster i. To obtain its
given in Table 2, are expressed in terms of the vectors
1
B I1 x I i
T
variance, we simply change the variable of interest in
equations (4.3) to (4.5). We replace yk by yik , keeping
B I = = U z i
I t x i U I z i y I i( ) and
BI2
B ( c)
other quantities intact. Denote by B (c ) = 1( c ) the B (u )
B2 B (Iu ) = (I1u ) , where B (Iu ) is obtained by replacing yI i
BI2

result of replacing yk by yi k in B of (4.4). The
in B I by t y i = U i y k .
approximation to Var (YI,CAL ) is then given by (4.5)
T
x B1( c ) The residuals given in Table 2 for case (a) are
with ek = yi k i k and simple to explain. In case (a) for unit statistics, the
B ( c)
xk 2

1354
 2003 Joint Statistical Meetings - Section on Survey Research Methods
Case Integrated Estimation
Weighting of a total for
Option
(a) None Units
Clusters
(b) si wk = N i wI i Units
Clusters
(c) wk = ak i wI i Units
Clusters
Table 2. Summary of residuals in the Components of the Variance (4.5) for Two-Stage Sampling and
Estimation. The notation is explained in the text.
automated linearization of YCAL = s wk yk produces
B = (U z k x Tk ) 1 (U z k yk ) , and the residuals in Table
2 follow from case (b) for unit statistics by setting
x I i = 0 and x i k = 0 for all i and k, because case (a)
involves no cluster related information in estimating for
units. For case (a) for cluster statistics, the automated
linearization of YI,CAL = sI wI i yI i leads to B I =
(U z i xTI i ) 1 (U z i yI i ) , and the residuals follow from
case (c) for cluster statistics by setting x k = 0 and
t x i = 0 for all i and k, because case (a) uses no unit
related information in estimating for clusters.
An examination of the residuals in Table 2 leads to
some interesting conclusions. Let us first compare the
residuals for unit statistics. In (b) and (c), the residuals
eI i are adjusted for both x I i and t x i , but in (a) they are
only adjusted for t x i . Thus (b) and (c) are better than
(a) for the first component. The residual ek is adjusted
for both auxiliaries in (b), but not in (a) and (c), where
it is only adjusted for x k . Thus (b) has the best
potential for efficient estimation of unit statistics.
Compare now the residuals for cluster statistics. In (b)
and (c), the residual eIi is adjusted for both x Ii and t xi ,
but in (a) it is only adjusted for x Ii . By design, the
residual ek is always zero in (a). A particularly
Residual eI i Residual ek
t yi t Tx i B y k xTk B
y Ii x TI i B I 0
t yi x TI i B1 t Tx i B 2 y k xTi k B1 xTk B 2
y Ii x TI i B1( c ) t Tx i B (2c ) yi k xTi k B1( c ) xTk B (2c )
t y i xTI i B (I1u ) t Tx i B (I u2) yk xTk B (Iu2)
y I i x TI i B I1 t Tx i B I 2 x Tk B I 2
unfavourable situation for the second variance
component arises for case (c), where the residual is
x Tk B I 2 . Thus (a) or possibly (b) has the best potential
for efficient estimation of cluster statistics.
5 Summary and discussion
The question of efficient weighting of the observed
values has always been important in survey sampling
theory. An important step was the formulation in 1952
of the HT estimator, prescribing that the weight of each
unit equals the inverse of the probability of its inclusion
in the sample. Thus, in stratified simple random
sampling (STSRS), the weight given to all units
sampled from a stratum equals the inverse of the
sampling rate in the stratum. Neymans convincing
results in 1934 on optimal estimation under STSRS laid
the foundation of what is now commonly called the
design-based theory of estimation. Another important
principle embodied in HT estimation is that the same
weight system applies to all y-variables of interest in a
multi-purpose survey. This preserves the design
unbiasedness for every y-variable. Assuming no
non-response, the sampling design alone determines
once and for all the weighting and the construction of
the point estimator.
The principle of a single weight system extends to
the calibration estimators in this paper. However, unlike
the sampling weights a k , the calibrated weights wk are
calculated only after drawing the sample. They are
1355
 2003 Joint Statistical Meetings - Section on Survey Research Methods
usually more efficient (give a smaller variance) than the
a k for every single y-variable and they produce
estimators with a negligible bias.
The literature on calibration has been based on a
model oriented construction of these estimators. Both
the model assisted and model dependent approaches to
calibration involve an explicit assumption of a linear
superpopulation model between x and y. This model is
of the form y k = xTk B + k where it is assumed that
E ( k ) = 0 and Var ( k ) = ck 2 with ck > 0 . In
practice however, this model is often invalid.
In our approach, the use of auxiliary information is
not linked to model fitting. We define a
parameterization of the calibration weights involving
the instrument vector z k and then apply the method of
automated linearization to obtain a linear approximation
of the calibration estimator. This linear approximation
is a design-based function of a set of fixed but unknown
population residuals determined implicitly without any
modelling. The wk are calculated using all or part of
the available auxiliary information. We have shown
how to do this for different designs including one-phase
and two-stage designs. It is important to note that the
construction of the point estimator has nothing to do
with the y-variables; the same weights apply to all
y-variables as is the case for the HT estimator.
However, the calibration estimator can be considerably
more efficient for some y-variables than others. This
depends on the resulting population residuals.
References
Andersson, C. (1997). Continuous labour force surveys:
performance analysis of a single weight procedure.
Internal report, Statistical Methodology Unit,
Statistics Sweden.
Binder, D. (1996). The right and wrong ways to Taylor
linearize for single phase and two phase samples: a
cookbook approach. Survey Methodology 22,
17-22.
Binder, D. and Kovaevi, M.S. (1995). Estimating
some measures of income inequality from survey
data: and application of the estimating equation
approach. Survey Methodology 21, 137-145.
Deville, J.C. and Srndal, C.E. (1992). Calibration
estimators in survey sampling. Journal of the
American Statistical Association 87, 376-382
1356
Deville, J. C. (1999). Variance estimation for complex
statistics and estimators: Linearization and residual
techniques. Survey Methodology 25, 193-203
Deville, J.C. (2002). Correction for non-response by
generalized calibration. Report, ENSAI, France.
Estevao, V. and Srndal, C.E. (2002). The ten cases of
auxiliary information for calibration in two-phase
sampling. Journal of Official Statistics 18,
233-255.
Huang, E.T. and Fuller, W.A. (1978). Nonnegative
regression estimation for sample survey data.
Proceedings Social Statistics Section, American
Statistical Association, 300-305.
Lematre, G.E. and Dufour, J. (1987). An integrated
method for weighting persons and families. Survey
Methodology 13, 199-207.
Le Guennec, J. and Sautory, O. (2002). Application of
generalized calibration to the correction of
non-response: An experiment. Report, ENSAI,
France.
Montanari, G.E. (1987). Post-sampling efficient
QR-prediction in large-sample surveys.
International Statistical Review 55, 191-202.
Nieuwenbroek, N.J. (1993). An integrated method for
weighting characteristics of persons and
households using the linear regression estimator.
Internal report, Central Bureau of Statistics, The
Netherlands.
Srndal, C.E., Swensson, B. and Wretman, J. (1992).
Model Assisted Survey Sampling. New York:
Springer-Verlag.
Thberge, A. (1999). Extensions of Calibration
Estimators in Survey Sampling. Journal of the
American Statistical Association 94, 635-644.
Woodruff, R.S. (1971). A simple method for
approximating the variance of a complicated
estimate. Journal of the American Statistical
Association 66, 411-414.