Beruflich Dokumente
Kultur Dokumente
Original citation: Skinner, Chris J. and Wakefield, Jon (2017) Introduction to the design and
analysis of complex survey data. Statistical Science, 32 (2). pp. 165-175. ISSN 0883-4237
DOI: 10.1214/17-STS614
LSE has developed LSE Research Online so that users may access research output of the
School. Copyright © and Moral Rights for the papers on this site are retained by the individual
authors and/or other copyright owners. Users may download and/or print one copy of any
article(s) in LSE Research Online to facilitate their private study or for non-commercial research.
You may not engage in further distribution of the material or use it for any profit-making activities
or any commercial gain. You may freely distribute the URL (http://eprints.lse.ac.uk) of the LSE
Research Online website.
Statistical Science
2017, Vol. 32, No. 2, 165–175
DOI: 10.1214/17-STS614
© Institute of Mathematical Statistics, 2017
165
166 C. SKINNER AND J. WAKEFIELD
designs, involving, for example, stratification and mul- population which is sampled and n denotes the sam-
tistage sampling, together with associated complica- ple size. Why would one wish to deviate from SRS
tions such as nonresponse. Although we shall provide and, in particular, from one of its properties, that each
a brief outline of complex designs at the end of this unit in the population is selected with equal probabil-
section, we focus in this paper on the methodology of ity? One reason is for efficient estimation. Tillé and
estimation and analysis and the question of how to ac- Wilhelm (2017) refers to the “false intuition that a
count for the complex sampling design, largely treated sample must be similar to a population” and explains
here as given. how more efficient estimation can often be achieved by
A further source of mystery for the many secondary sampling units with unequal probabilities. Other rea-
analysts of complex survey data is that, whereas many sons include practical constraints imposed by the na-
novel developments in the methodology of survey de- ture of frames from which the sample must be drawn
sign and estimation have been introduced by the agen- and variable costs of data collection. See Valliant, De-
cies conducting large surveys, not all the information ver and Kreuter (2013) for a detailed account of a wide
relating to these developments is made available when range of designs used in practice.
data are released for general analysis. For example, sur- We now describe some basic designs. We emphasize
vey weights and imputed values may be made avail- that in practice these often act as the building blocks of
able but some of the complex features of the sampling more complex designs, because of the characteristics
frame, design and “raw” data underlying such released of the sampling frame or the population. A common
information may remain concealed. design is stratified simple random sampling in which
Historically, survey sampling has often been seen as a group label is available for each unit in the popu-
a rather separate topic to much of the rest of statistics. lation, and SRSs are taken within each of the groups.
For example, in studies of individuals, the groups may
Not only have survey design and survey data analysis
correspond to demographic strata and geographical re-
typically been undertaken by different people but the
gions. Stratified random sampling can provide appre-
estimation methodology associated with survey design
ciable gains in efficiency if the variables defining the
has often been centered on the design-based (or ran-
strata are associated with the response. The main im-
domization) approach, quite separate from the model-
pediment to its use is the availability of strata informa-
based inference featuring in much mainstream applied
tion on all members of the population. In single-stage
statistics. While the former approach has a clear ratio-
cluster sampling, the population is again partitioned,
nale, it can be confusing to those who have received
but this time into what are called “clusters”, or primary
a conventional model-based training in statistics. The sampling units (PSUs). The PSUs are often defined ge-
slow rate of inclusion of complex survey methods in ographically.
much applied statistical software for analysis has also The key difference between cluster sampling and
contributed to this separation. Nevertheless, we have stratified sampling is that only a sample, rather than
recently sensed a greater degree of cross-fertilization all, of the clusters are selected and then informa-
of ideas between survey sampling and applied statisti- tion is obtained from all individuals within clusters.
cal methods of analysis. A key purpose of this paper Cluster sampling in general reduces efficiency, be-
and special issue is to help support the sharing of such cause of within-cluster correlation, but logistically it
ideas by opening up developments in the survey sam- is very convenient, particularly in nationwide sam-
pling literature to a broader readership. pling. In two-stage cluster sampling, random samples
Before proceeding to consider inference, however, are taken within the sampled clusters (PSUs). In large
we set the scene by outlining some features of com- national surveys, stratified multistage cluster sampling
plex designs. We only consider probability sampling (in which cluster sampling and stratified SRS is exe-
in which the design is characterized via a probabil- cuted) is the norm, since it balances efficiency, logis-
ity distribution over the possible samples that may tical constraints, and the requirement for estimates of
be collected. In particular, each unit in the popula- sufficient precision to be obtained for subgroups of in-
tion of interest has a nonzero probability of being se- terest.
lected. A complex design may be viewed as one de- For the remainder of this paper, we turn to inference,
viating from the simplest design, simple random sam- but see Tillé and Wilhelm (2017) for additional proba-
pling (SRS), in which all subsets of n from N units bility sampling designs, with an emphasis on new de-
are equally likely. Here, N denotes the size of the velopments. In Section 2, we introduce and compare
COMPLEX SURVEYS 167
design-based and model-based inference for complex the randomization associated with probability sam-
survey data. Section 3 brings in the role of auxiliary pling or from a model assumed to generate the popu-
information and describes model-assisted inference in lation values yk . Inference based on models is likely to
which working models are adopted to suggest estima- be familiar to most readers and so we leave it till sec-
tors/designs, but with the design-based approach be- ond. First, we discuss design-based inference, some-
ing followed for inference. Sections 4 and 5 consid- times also called randomization-based analysis. Lohr
ers model parameters (as opposed to finite population (2010) is a popular text that is primarily concerned with
characteristics) and nonprobability sampling. The im- design-based inference. The latter is more distinctive to
portant topic of variance estimation is the subject of survey sampling, though inference based on random-
Section 6. Final remarks conclude the paper in Sec- ization is sometimes used in randomized experiments
tion 7. (Cox, 2006, Chapter 9).
2.1 Design-Based and Model-Based Approaches
2. INFERENTIAL OVERVIEW
to Inference
In this section, we provide a brief overview of in-
2.1.1 Design-based. The population values y1 , . . . ,
ferential approaches to the analysis of survey data, as-
yN , are viewed as fixed constants, with the collection
suming that probability sampling is used and there is
of units selected, S, treated as random. Assuming N
no nonresponse. First, we define some notation. Let
known, a standard weighted estimator of the popula-
yk , k = 1, . . . , N represent the values of a survey vari-
tion mean is
able of interest on all N units in a well-defined finite
population (e.g., all individuals aged 18 or greater in dk yk
(1) y HT = k∈S .
a particular administrative region); we shall also write N
k ∈ U to index this collection. A sample of these units, This will be referred to as the HT estimator, since its
denoted S ⊂ U , is taken via some probabilistic mech- numerator was proposed by Horvitz and Thompson
anism, where
p(s) denotes the probability of selecting (1952) for estimating the population total k∈U yk .
S = s with s p(s) = 1, and the values of yk are only The primary motivation for such design weighting is
observed for k ∈S. The probability of unit k being se- to remove bias, as discussed in detail in Haziza and
lected is πk = s:k∈s p(s), and the so-called design- Beaumont (2017). Bias and other moments are evalu-
weight is defined as dk = πk−1 . This weight is often ated in the design-based framework with respect to re-
interpreted loosely as the number of population units peated sampling of units from the finite population U .
“represented” by the kth sampled unit. We write ES [y HT ], with the subscript S on the expecta-
We suppose here that finite population characteris- tion emphasizing that we are averaging over possible
tics are the targets of inference. Other targets are con- subsets that could have been selected. Similarly, the
sidered in Section 4. It is common in the survey sam- variance of the estimator will be written as varS (y HT ).
pling literature to emphasize first the estimation of pop- We informally define two particular criteria: an estima-
ulation totals, before moving on to targets which are tor is design unbiased if its expectation (over all possi-
functions of totals. There are a variety of reasons for ble samples) is equal to the true value, and an estimator
this, beyond the obvious one that totals may indeed be is design consistent if both the design bias and the vari-
targets of inference. One reason is that issues of bias ance go to zero as the sample size increases. For the lat-
can be dealt with more simply with totals than, say, ter, one must consider a sequence of populations, with
means, especially when the population size N is un- the finite population size and the sample size tending
known, as it often is. However, to present basic ideas to infinity.
in this paper
we shall treat the finite population mean An alternative estimator of the mean, defined whether
y U = N1 k∈U yk as our target of inference, since it is a N is known or not, is the Hájek estimator (Hájek,
more natural “unit-level” parameter of interest in much 1971):
survey data analysis and includes, for example, a pro-
k∈S dk yk
portion as a special case. (2) y HJ =
,
We next contrast two broad paradigms under which N
=
inference based on survey data may be performed: the where N k∈S dk , vindicating dk ’s interpretation
design-based and model-based approaches. These re- earlier as the number of population units represented
fer to two different sources of randomness, either from by the kth sampled unit. The estimator (2) is biased
168 C. SKINNER AND J. WAKEFIELD
so that we recover a familiar form for the variance of The prediction variance is
the estimator, albeit with a finite population correction. 2
2 1 n
1 N
so that the Hájek estimator is always unbiased under corresponds to y HT , the HT estimator. Model (7) is
this model but the HT estimator is only unbiased under not required to be correct for the properties of the
= n dk = N . This condition often
the condition N HT estimator to be valid, but it does suggest situa-
k=1
does hold, for example, under SRS when dk = N/n, tions in which we would expect the estimator to per-
but there may be problems in the performance of the form well (or not). This could be viewed as a model-
HT estimator for sampling designs where it does not. assisted approach, which we discuss in more detail
An example is a design in which the sample size n is in Section 3. The widely cited Basu elephant exam-
random and the weights are constant (i.e., dk = d). ple (Basu, 1971, Hájek, 1971) provides an extreme ex-
We switch now to consider the design bias of the ample in which the HT estimator performs poorly in a
situation in which the responses yk are not related to
model-unbiased estimator Y n , where Yk is replaced by
yk . We have the sampling probabilities πk . Briefly, a fictional cir-
cus owner would like to estimate the weight of his 50
1 N
1 N strong herd of elephants, based on measuring a sin-
ES [Y n ] = ES Ik yk = πk yk gle elephant. Drawing on 3-year old records, he pro-
n k=1 n k=1
poses to measure the weight of Sambo, an elephant
and so this estimator will generally be design biased. who was previously of average weight, and multiply
Haziza and Beaumont (2017), Section 3, show that the this weight by 50. This purposive design traumatizes
design bias will only disappear if the πk are uncor- the circus statistician, who is obsessed with using a
related with the yk , which corresponds to the notion design-unbiased estimator. To this end, he proposes an
of noninformative sampling discussed in Section 2.1.2. alternative plan in which Sambo is selected with prob-
This illustrates the (design) bias impact of model mis- ability πk = 10099
, and one of the remaining 49 crea-
tures with probability πk = 100 1
× 49
1
. Letting y de-
specification. The unweighted estimator Y n is unbiased
for Y U under the model μ = EM [Yk ] (and the assump- note the weight of the selected elephant, the HT esti-
tion of noninformative sampling) but does not protect mator is dy where d = 100 99 if Sambo is selected, and
against bias if these assumptions fail, unlike the design d = 100 × 49 if any of the other elephants is selected.
weighted (HT and Hájek) estimators. Clearly, whichever elephant is selected, this estimator
So far as the variance is concerned, if we assume is unsatisfactory. Putting aside the wisdom of an n = 1
that varM (Yk ) = σk2 and that Yk and Yl are independent design, is it clear here that, by construction, yk is not
under the model, then the variance under the model of, proportional to πk . For further discussion, see Brewer
for example, the Hájek estimator may be expressed as (2002) and Lumley (2010), page 149.
From a modeling perspective, the Hájek estima-
n
tor (2) arises from the model with EM [Yk ] = θ and
EM (y HJ − Y U )2 = −1 dk − N −1 2 σ 2
N k varM (Yk ) = πk2 σ 2 , k = 1, . . . , n, and so one would ex-
k=1
pect it to outperform the HT estimator when the re-
N sponse is approximately constant (as opposed to be-
+ N −2 σk2 ing proportional to the sampling probabilities), which
k=n+1 argues for its use in many instances, regardless of
(the same expression holds for the HT estimator with whether N is known,
N replaced by N ). This expression will almost cer- To summarize, it is informative to view model-based
tainly not equal (3). estimators from a design-based perspective and vice
Studying the model variance of design-based estima- versa, since it gives insight into situations in which the
tors may help in assessing their efficiency. Consider, respective estimators will perform well. Conclusions
for example, the model we have drawn here include that the HT estimator may
be design-unbiased, but biased with respect to particu-
(7) Yk = θ πk + εk , lar models, and the model variance may not correspond
to the design variance.
where the error terms εk are independent with
EM [εk ] = 0 and varM (εk ) = πk2 σ 2 , k = 1, . . . , n. The
3. AUXILIARY VARIABLES AND
weighted least squares estimator θ , that minimizes
MODEL-ASSISTED ESTIMATION
n
(yk − θ πk )2
, In most survey settings, auxiliary information about
k=1 πk2 σ 2 the population units will be available to assist both de-
COMPLEX SURVEYS 171
sign and inference. From a predictive model-based per- design consistent for y U and x U , respectively. The ra-
spective, it is very natural and commonplace to include tionale for using the regression estimator rather than
auxiliary variables as covariates in regression models. the simple estimator y S is that it improves precision.
This enables more efficient predictions to be achieved. It does this because, under (8) with B1 = 0, the er-
Conditioning on auxiliary variables used in the design ror y S − y U is correlated with x S − x U . Under SRS,
also helps to ensure that sampling is noninformative, the large-sample variance of the regression estimator is
that is, that Ik is independent of Yk conditional on these given by
covariates; hence it helps avoid the kinds of selection
bias mentioned in Section 2.1.2. n Se2
1− ,
From a design-based perspective, it is also possible N n
to include auxiliary variables to improve inference, in
particular, inference may be “assisted” by considera- where Se2 = N−1 1
k∈U ek with ek = yk − (B0 + B1 xk )
2
tion of suitable covariates in a regression model for Yk . denoting the residual. By comparison with the expres-
The definitive text on model-assisted inference in sur- sion before (5) we see that the use of auxiliary infor-
veys is Särndal, Swensson and Wretman (1992). For mation has reduced the variance by a factor of approx-
simplicity, suppose we know the mean x U for a sin- imately R 2 , the squared correlation between the yk and
gle variable xk that we believe is associated with yk . xk .
Consider the “working model”, For general complex designs, we may use design
weights in the estimators B 0 and B1 and obtain, as dis-
(8) Yk = B0 + B1 xk + εk ,
cussed by Breidt and Opsomer (2017), the generalized
where the error terms εk are independent with EM [εk ] = regression (GREG) model assisted estimator of y U as
0, varM (εk ) = σ 2 , k = 1, . . . , N and the intercept and
0 + B
y GREG = B 1 x U = y + B
1 (x U − x HT ),
slope are defined with respect to the finite population: HT
N
k=1 (xk − x U )(yk − y U )
so that the HT estimator of the mean of the yk is ad-
B1 = N justed via the difference between the population mean
k=1 (xk − x U )
2
(9) of the xk and its sample estimator. The GREG estima-
cov(x, y) RSy2 tor is design consistent with finite sample design bias,
= = 2 ,
var(x) Sx but for large samples its precision will be greater than
that of the HT estimator. The estimators B0 and B
1 can
(10) B0 = y U − B1 x U ,
also include weighting for heteroskedasticity in model
where R, Sx and Sy are the correlation, standard devia- (8) giving, for example, when B0 is taken as zero, the
tion of x and standard deviation of y, in the population. widely used ratio estimator, as a special case (Breidt
This model motivates B0 + B1 x U as an estimator of and Opsomer, 2017).
y U , where B0 and B1 are design-based estimators of The overall model-assisted approach has a similar
B0 and B1 . For simplicity, consider SRS and the esti- flavor to robust estimation using sandwich variance es-
mators: timators, where a working model is specified, but the
1 = k∈S (xk − x S )(yk − y S ) rsy2 consistency of the estimator is guaranteed, under very
B = 2, weak assumptions, and in particular consistency does
k∈S (xk − x S )
2 sx
not depend on strong modeling assumptions. Breidt
0 = y − B
B 1 x S ,
S and Opsomer (2017) provide a much fuller account of
where r is the sample correlation, and y S , x S and sx , sy model-assisted inference, including a wide range of ex-
are, respectively, the means and standard deviations of tensions of the GREG approach.
x and y in the sample. Since ratios are involved, these We note that GREG and related estimators can be
estimators are not design unbiased for B0 and B1 but represented as weighted estimators, where the weights
they are design consistent. The resulting estimator of extend the simple idea of design weights introduced
y U is earlier by incorporating auxiliary population informa-
0 + B
1 x U = y + B
1 (x U − x S ). tion. Various adjustments can be made and the con-
B S struction of weights can be complex; the relevant is-
This is the traditional regression estimator and is de- sues are discussed in this issue by Haziza and Beau-
sign consistent for y U under SRS since y S and x S are mont (2017) and Chen et al. (2017).
172 C. SKINNER AND J. WAKEFIELD
design, just the strata identifiers, PSU identifiers and and BRR techniques are discussed more fully in Rust
weights provide sufficient design information for con- and Rao (1996).
structing consistent variance estimators. Valliant, De- Another approach to variance estimation in com-
ver and Kreuter (2013), Chapter 15, describe how this plex designs is to adopt a model-based approach
may be done for linear estimators such as the HT es- and to accommodate the population complexity in
timator. For more complex nonlinear estimators, the the model. For example, the induced dependence be-
method of linearization (more commonly referred to tween units in a clustered population may be acknowl-
as the delta method in mainstream statistics) may be edged in a model-based approach using mixed mod-
employed. This approach depends on the nature of the els, an approach that was championed by Scott and
estimator, but is implemented in most statistical survey Smith (1969). In general, sandwich estimation is of-
software. Other approximations, free of joint selection ten utilized when adopting a model-based approach
probabilities, and usable for design-based variance es- (Pfeffermann et al., 1998, Rabe-Hesketh and Skrondal,
timation, are reviewed by Berger and Tillé (2009). 2006). The underlying idea behind sandwich estima-
Another broad approach is replication variance esti- tion is the empirical construction of variances under
mation. The bootstrap and jackknife methods are per- a (usually) simple working model. Sandwich estima-
haps the most well-known examples; each of these tion produces consistent standard error estimates under
are standard techniques in statistics; see Shao and reduced assumptions when compared with a model-
Tu (2012) for a book length treatment. Suppose that based approach, and is robust to the assumed variance
model.
survey weights, denoted wk , k ∈ S, are used to es-
timate a parameter, y U , say, via a weighted estima-
7. CONCLUDING REMARKS
tor y = k∈S wk yk . These weights may be simple
design weights, or include post-stratification, nonre- In this short paper, we have given an overview of the
sponse, etc., adjustments. Sets of L replicate weights design- and model-based approaches to inference for
wk(l) , k ∈ S for l = 1, . . . , L, are constructed and the complex survey data. There are many important and
variance estimator takes the form: emerging topics that we have not touched upon. For
example, combining different sources of data is being
L
(l) increasingly carried out, and Lohr and Raghunathan
cl
y − y 2,
(2017) review this endeavor.
l=1
A particular reason for the growing interest in com-
where
(l) (l)
y = k∈S wk yk . For suitable constants cl , bining data sources is the increased availability of “big
depending on the replication method, such replication data” sources, as noted in Section 1. There is also
weights can be constructed for a range of designs to huge current interest in a landslide of associated new
achieve consistent variance estimation. The data file, data analysis techniques, which often have their origins
released by the survey agency, now contains an addi- in machine learning. However, while many of these
tional L fields corresponding to the replicate weights, methods appear intoxicating, they must be carefully as-
alongside the basic weights wk . sessed to see whether they will provide valid inferences
For both the bootstrap and the jackknife, the replica- in the face of multiple sources of noncoverage and se-
lection. If such aspects are ignored, there is a genuine
tion weights wk(l) , k ∈ S contain zeros, either from sys-
possibility that big data analyses will produce really
tematic deletion (the jackknife) or as a result of random
big inferential train wrecks.
subsampling (the bootstrap). The implementation of
each of these techniques requires care since one must ACKNOWLEDGEMENTS
acknowledge the complex design. For example, under
multistage sampling one may remove a complete PSU, Jon Wakefield was supported by award
which preserves the dependence structure of responses R01CA095994, from the National Institute of Health.
in the same cluster, and the weights are adjusted so as
to preserve the sum of the weights. A further replica- REFERENCES
tion technique is balanced repeated replication (BRR).
BASU , D. (1971). An essay on the logical foundations of survey
Under BRR, symmetries within the design are ex- sampling, part I. In Foundations of Statistical Inference (Proc.
ploited to produce variance estimates from partially in- Sympos., Univ. Waterloo, Waterloo, Ont., 1970) 203–242. Holt,
dependent splits of the data. The bootstrap, jackknife Rinehart and Winston, Toronto. MR0423625
COMPLEX SURVEYS 175
B ERGER , Y. G. and T ILLÉ , Y. (2009). Sampling with un- L OHR , S. L. (2010). Sampling: Design and Analysis, 2nd ed.
equal probabilities. In Handbook of Statistics, Vol. 29A, Sam- Brooks/Cole Cengage Learning, Boston, MA.
ple Surveys: Design, Methods and Applications (D. Pfeffer- L OHR , S. and R AGHUNATHAN , T. (2017). Combining survey data
mann and C. R. Rao, eds.) 39–54. North-Holland, Amsterdam. with other data sources. Statist. Sci. 32 293–312.
MR2654632 L UMLEY, T. (2010). Complex Surveys: A Guide to Analysis Us-
B INDER , D. A. (1983). On the variances of asymptotically normal ing R. Wiley, Hoboken, NJ.
estimators from complex surveys. Int. Stat. Rev. 51 279–292. L UMLEY, T. and S COTT, A. (2017). Fitting regression models to
MR0731144 survey data. Statist. Sci. 32 265–278.
B REIDT, J. and O PSOMER , J. (2017). Model-assisted survey esti- P FEFFERMANN , D., S KINNER , C. J., H OLMES , D. J., G OLD -
mation with modern prediction techniques. Statist. Sci. 32 190– STEIN , H. and R ASBASH , J. (1998). Weighting for unequal se-
205. lection probabilities in multilevel models. J. R. Stat. Soc. Ser. B
B REWER , K. (2002). Combined Survey Sampling Inference: 60 23–40. MR1625668
Weighing Basu’s Elephants. Arnold, London. R ABE -H ESKETH , S. and S KRONDAL , A. (2006). Multilevel mod-
C HAMBERS , R. L. and C LARK , R. G. (2012). An Introduction to elling of complex survey data. J. Roy. Statist. Soc. Ser. A 169
Model-Based Survey Sampling with Applications. Oxford Univ. 805–827. MR2291345
Press, Oxford. MR3186498 RUST, K. F. and R AO , J. N. K. (1996). Variance estimation for
C HEN , Q., E LLIOTT, M. R., H AZIZA , D., YANG , Y., G HOSH , M., complex surveys using replication techniques. Stat. Methods
L ITTLE , R., S EDRANSK , J. and T HOMPSON , M. (2017). Ap- Med. Res. 5 283–310.
proaches to improving survey-weighted estimates. Statist. Sci. S ÄRNDAL , C.-E., S WENSSON , B. and W RETMAN , J. (1992).
32 227–248. Model Assisted Survey Sampling. Springer, New York.
C OX , D. R. (2006). Principles of Statistical Inference. Cambridge MR1140409
Univ. Press, Cambridge. MR2278763 S CHONLAU , M. and C OUPER , M. (2017). Options for conducting
E LLIOTT, M. and VALLIANT, R. (2017). Inference for non- web surveys. Statist. Sci. 32 279–292.
probability samples. Statist. Sci. 32 249–264. S COTT, A. and S MITH , T. M. F. (1969). Estimation in multi-stage
G ELMAN , A. (2007). Struggles with survey weighting and regres-
surveys. J. Amer. Statist. Assoc. 64 830–840.
sion modeling. Statist. Sci. 22 153–164. MR2408951
S EAMAN , S. R. and W HITE , I. R. (2013). Review of inverse prob-
G RAUBARD , B. I. and KORN , E. L. (2002). Inference for super-
ability weighting for dealing with missing data. Stat. Methods
population parameters using sample surveys. Statist. Sci. 17 73–
Med. Res. 22 278–295.
96. MR1910075
S HAO , J. and T U , D. (2012). The Jackknife and Bootstrap.
H ÁJEK , J. (1971). Discussion of ‘An essay on the logical founda-
Springer, New York. MR1351010
tions of survey sampling, part I’, by D. Basu. In Foundations of
S KINNER , C. J. (2003). Introduction to part b. In Analysis of Sur-
Statistical Inference (Proc. Sympos., Univ. Waterloo, Waterloo,
vey Data (R. L. Chamber and C. J. Skinner, eds.) 75–84. Wiley,
Ont., 1970). Holt, Rinehart and Winston, Toronto.
H AZIZA , D. and B EAUMONT, J.-F. (2017). Construction of Chichester. MR1978845
weights in surveys: A review. Statist. Sci. 32 206–226. S KINNER , C. J., H OLT, D. and S MITH , T. M. F., eds. (1989).
H ORVITZ , D. G. and T HOMPSON , D. J. (1952). A generalization Analysis of Complex Surveys. Wiley, Chichester. MR1049386
of sampling without replacement from a finite universe. J. Amer. T ILLÉ , Y. and W ILHELM , M. (2017). Probability sampling de-
Statist. Assoc. 47 663–685. MR0053460 signs; principles for the choice of design and balancing. Statist.
JAPEC , L., K REUTER , F., B ERG , M., B IEMER , P., D ECKER , P., Sci. 32 176–189.
L AMPE , C., L ANE , J., O’N EIL , C. and A SHER , A. (2015). Big VALLIANT, R., D EVER , J. A. and K REUTER , F. (2013). Practical
data in survey research: Aapor task force report. Public Opin. Q. Tools for Designing and Weighting Survey Samples. Springer,
79 839–880. Berlin. MR3088726
L ITTLE , R. J. (2013). Calibrated Bayes, an alternative inferential VALLIANT, R., D ORFMAN , A. H. and ROYALL , R. M. (2000).
paradigm for official statistics (with discussion). J. Off. Stat. 28 Finite Population Sampling and Inference: A Prediction Ap-
309–372. proach. Wiley-Interscience, New York. MR1784794