You are on page 1of 18

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

net/publication/307546330

Non-probability sampling

Chapter · January 2016

CITATIONS READS

127 73,896

3 authors:

Vasja Vehovar Vera Toepoel


University of Ljubljana Utrecht University
120 PUBLICATIONS   3,838 CITATIONS    78 PUBLICATIONS   1,958 CITATIONS   

SEE PROFILE SEE PROFILE

Stephanie Steinmetz
University of Lausanne
60 PUBLICATIONS   1,720 CITATIONS   

SEE PROFILE

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

WageIndicator View project

InGRID project & research infrastructure View project

All content following this page was uploaded by Vasja Vehovar on 01 October 2018.

The user has requested enhancement of the downloaded file.


FPO
22
Non-probability Sampling
V a s j a V e h o v a r , V e r a To e p o e l a n d
Stephanie Steinmetz

A sample is a subset of a population and we closer look at two contemporary – and per-
survey the units from the sample with the haps also the most exposed – aspects of
aim to learn about the entire population. non-­probability sampling: online panels and
However, the sampling theory was basically weighting. Finally, we summarize recom-
developed for probability sampling, where mendations for deciding on probability–non-
all units in the population have known and probability sampling dilemmas and provide
positive probabilities of inclusion. This defi- concluding remarks.
nition implicitly involves randomization,
which is a process resembling lottery draw-
ing, where the units are selected according to
their inclusion probabilities. In probability TYPOLOGY, PRACTICE AND
sampling the randomized selection is used PROBLEMS OF NON-PROBABILITY
instead of arbitrary or purposive sample SAMPLES
selection of the researcher, or, instead of vari-
ous self-selection processes run by respond- We initially defined non-probability sam-
ents. Within this context, the notion of pling as a deviation from probability sam-
non-probability sampling denotes the pling principles. This usually means that
absence of probability sampling mechanism. units are included with unknown probabili-
In this chapter we first reflect on the prac- ties, or, that some of these probabilities are
tice of non-probability samples. Second, we known to be zero. There are countless exam-
introduce probability sampling principles ples of such deviations. The most typical
and observe their approximate usage in the ones received their own labels and we pre-
non-probability setting and we also discuss sent below a non-exhaustive list of
some other strategies. Third, we provide a illustrations:

BK-SAGE-WOLF-160177-Chp22.indd 327 4/22/2016 3:44:56 PM


328 The SAGE Handbook of Survey Methodology

•• Purposive sampling (also judgemental sampling) polls and invitations to online panel, which all is
– the selection follows some judgement or arbi- typically done via banner recruitment; a consider-
trary ideas of the researchers looking for a kind able overlap exists here also with the notion of
of ‘representative’ sample, or, researchers may river sampling, where users are recruited with
even explicitly seek for diversity (deviant case various approaches while surfing on the web.
samplings); sometimes unit are added sequen- •• Network sampling – convenience sample where
tially until researchers satisfy some criteria. some units form the starting ‘seeds’, which then
•• Expert selection – subject experts pick the units, sequentially lead to additional units selected
for example when, say, only two most typical from their network; a specific subtype is snow-
settlements are to be selected from a region or ball sampling, where number of included net-
when the units with highest level of certain char- work ties at each step is further specified.
acteristics are selected (modal instance sampling). •• Quota sampling – convenience sample can be
•• Case study – an extreme example of expert selec- ‘improved’ with some socio-demographic quotas
tion where the selected unit is then profoundly (e.g. region, gender, age) in order to reflect the
studied, often with qualitative techniques. population.
•• Convenience sampling – the prevailing non-
probability approach where units at hand are In addition, we may talk about non-probability
selected; the notion roughly overlaps also with sampling also when we have probability
accidental, availability, opportunity, haphazard sampling, but:
or unrestricted sampling; most typical formats
are recruitment at events (e.g. sports, music, etc.)
a) we previously omitted some (non-negligible)
and other locations (e.g. mall intercept, where
segments of the target population, so that the
customers are approached in shopping malls,
inclusion probabilities for corresponding units are
street recruitment, where people on the street
zero (i.e. we have non-coverage);
are invited into a survey) or at virtual venues
b) we face very low response rates, which – among
(e.g. web intercept, where the web visitors are
others – occurs with almost any traditional
exposed to pop-up invitations); particularly fre-
public media recruitment, including the corre-
quent is also the recruitment of social ties (e.g.
sponding advertising, from TV and radio spots
friends, colleagues and other acquaintances).
to printed ads and out-of-home-advertising, i.e.
•• Volunteer sampling – a type of convenience sam-
OOH media, being digital (e.g. media screens in
pling, where the decision to participate strongly
public places) or non-digital (e.g. billboards).
relies on respondents due to the non-individualized
nature of invitations (e.g. general plea for par-
ticipation appears in media, posters, leaflets, The bordering line when a certain part of the
web, etc.). population is non-negligible (a) or when the
•• Mail-in surveys – a type of volunteer sampling response rate is very low (b) is a practical
with paper-and-pencil questionnaires, which are question, which is conceptually much blurred,
exposed as leaflets at public locations (e.g. hotels, disputable and depends on numerous circum-
restaurants, administration offices) or enclosed to stances. In particular, it depends on whether
magazines, journals, newspapers, etc. we know the corresponding missing data
•• Tele-voting (or SMS voting) – a type of vol- mechanism and control the probabilities of
unteer sampling where people are invited to inclusion, as well as on whether the missing
express their vote by calling-in some (toll-free or
segments substantially differ in their charac-
chargeable) number or by sending an SMS; most
typically this is related to TV shows (particularly
teristics. The opposite is also true: whenever
entertainment and reality) and various contests we have low non-coverage and high response
(music, talents, performance), but also to politi- rates – especially with relatively small popu-
cal polling. lations (e.g. hotel visitors, attendants of an
•• Self-selection in web surveys – a type of volunteer event) – some of the above types of non-
sampling, where a general invitation is posted probability sampling can turn to probability
on the web, such as invitations on web pages sampling, particularly in case of volunteer
(particularly on social media), question-of-the-day samples (e.g. mail-in, web self-selection).

BK-SAGE-WOLF-160177-Chp22.indd 328 4/22/2016 3:44:56 PM


Non-probability Sampling 329

There are countless examples of disasters, the characteristics (e.g. attitudes) of a popu-
where results from non-probability sampling lation with millions of units could be meas-
were painfully wrong. There also exist some ured with a sample of just a few hundreds.
historical well-documented evidences, which Correspondingly, the sampling practice
have already become textbook cases. One is expanded enormously in the commercial and
related to early purposive sampling attempts in public sector.
at the end of the 19th century from official sta-
tistics in Norway (Kiaer, 1997 [1897]), which
ended in serious errors. Two cases from the Sampling Design Principles
US elections have also become popular warn-
ing against deviations from probability sam- We can select a probability sample using a
pling. The first one is the 1936 Literary Digest rich array of sampling techniques. Most text-
Poll, where around two million mail-in polls books start with a simple random sample
failed to predict the winner in the US presi- (SRS), which fully reflects lottery drawing.
dential election. The failure was basically due In SRS, all units in the population have the
to the fact that disproportionally more repub- same probability of selection, but also all
licans were in the sampling frame. The sec- samples of a certain size have the same prob-
ond example refers to the US 1948 election ability to appear. We can also use complex
prediction failure, which was – among other designs with strata, clusters, stages, phases
reasons – due to a lack of sufficient randomi- etc. The general overview on probability
zation in quota sampling. sampling designs is provided already in this
We should underline that wrong results Handbook (see Tillé and Mattei, Chapter 21).
based on non-probability samples – includ- For further reading we also point to general
ing the above three examples – are typically survey (e.g. Bethlehem, 2009; Fowler, 2014;
linked to situations where the standard statis- Groves et  al., 2009) or statistical textbooks
tical inference assuming probability sampling (e.g. Agresti and Frankin, 2013; Healey,
was used with non-probability samples. For 2011). Rich sampling literature also exist,
this reasons we first overview the probability either as general introduction (e.g. Kalton,
sampling principles and their approximate 1983; Lohr, 2010) or as advanced statistical
implementation in non-probability settings. treatment (Särndal et al., 2003).

PRINCIPLES OF PROBABILITY Statistical Inference Principles


SAMPLING1 When we infer from a probability sample to
the entire population, we face risks that due
It was only when the corresponding sampling to randomization the selected sample is by
theory was implemented in 1930s that major chance different from the population. For
breakthrough in the development of modern example, it is theoretically possible, although
surveys occurred. We should also recall that extremely unlikely, that we drew a SRS of the
only few decades before, at the end of 19th general population of a country (where men
century, the statistical profession (i.e. the have roughly a 50% share), which contains
International Statistical Institute) rejected only men, but no women. The advantage of
the practice from the Norwegian statistical probability sampling is that this risk (e.g.
office (Kiaer, 1997 [1897]) of using parts of selecting such extreme samples) can be quan-
the population to infer about the entire popu- tified, which is then often evaluated at a level
lation. One direct consequence of the changes of 5% using so-called confidence intervals. A
from 1930s was huge cost saving, because risk of 5% means that the corresponding

BK-SAGE-WOLF-160177-Chp22.indd 329 4/22/2016 3:44:56 PM


330 The SAGE Handbook of Survey Methodology

confidence interval contains the population because sampling theory was developed for
value in 19 out of 20 randomly selected sam- the probability sampling. Consequently, we
ples of certain type. The width of the confi- first discuss approaches where probability
dence interval communicates the precision of sampling principles are used in non-
the estimate, which strongly depends on the probability settings, although – due to
sample size; larger samples produce more unfulfilled assumptions (i.e. lack of known
precise estimates. Another advantage of prob- probabilities) – formally this could (and
ability samples – which is of extreme practi- should) not be done. For this reason we will
cal value – is that the confidence interval can call it as an approximate usage of probability
be estimated from the sample itself. However, sampling principles. Next, we provide an
the construction of the confidence interval is overview of some other strategies, developed
only one aspect of the process of statistical to deal specifically with non-probability
inference, which is a set of procedures for samples.
generalizing results from a sample to a popu-
lation, but it also deals with other inferential
diagnostics, as well as with statistical hypoth- Approximations of Standard
esis testing.
Probability Sampling Principles
Various inferential approaches exist. We
referred here only to the most popular fre- The approximate usage of probability sam-
quentist one, which (a) takes into account the pling in non-probability setting is sometimes
sampling design, (b) assumes that unknown understood in a sense that we first introduce
population values are fixed and (c) builds certain modelling assumptions, e.g. we
on a sampling distribution of the estimates assume that there is actually some randomi-
across all possible samples. Other inferen- zation in the non-probability sample. We will
tial approaches can differ in one or more of discuss these assumptions and corresponding
the above-mentioned aspects; however, for implementations separately for sampling
majority of practical situations they pro- design and for statistical inference part.
vide very similar results. More differences
between approaches may appear with unu- Approximations in Sampling
sual distributions and with specific modelling Design
requests. The very basic idea here is to approximate
The related inferential procedures can be and resemble, as much as possible, the prob-
very simple, as in the case of estimating the ability sampling selection of the units. This
sample mean in a SRS setting, while the esti- will then hopefully – but with no guarantee –
mation of some other parameters (e.g. ratios, also assure other features of probability sam-
various coefficients, etc.) in complex sam- pling, particularly the randomization. This
ples (e.g. with stages, strata, phases, etc.) or goal can be achieved indirectly or directly.
with complicated estimators (e.g. regression
estimator) may require additional statistical Indirect measures to approximate
resources. probability sample designs
The most essential recommendation is to
spread the non-probability sample as broadly
as possible. In practice this predominantly
APPROACHES AND STRATEGIES IN means we need to combine various recruit-
NON-PROBABILITY SAMPLING ment channels, which may appear also as
media planning strategies. A successful
We now turn to the non-probability context, example is the WageIndicator survey (see
which is, in principle, very problematic, WageIndicator.org), implemented in over

BK-SAGE-WOLF-160177-Chp22.indd 330 4/22/2016 3:44:57 PM


Non-probability Sampling 331

85 countries worldwide, which uses an elab- region, etc.), but also according to attitudes,
orated recruitment strategy by posting invita- media consumptions, purchase behaviour,
tions to their volunteer web survey across a etc. Sophisticated algorithms can be used
wide range of websites (Steinmetz et  al., to select an optimal sample for certain pur-
2013). With this strategy they approximate pose. An extreme case is individual match-
the randomization spread of recruited units. ing, which is fully effective only when we
Similarly, the Vehovar et al. (2012) study of have large databases of units available (as in
online hate speech, based on a non-probability online panels), where we first design a ‘nor-
web survey, successfully applied well-spread mal’ sample from some ideal or high-quality
recruitment procedures, ranging from off- control source. We then search for units in
line strategies (e.g. delivering leaflets with our database, using some distance measures
survey URL at public places) to systematic (e.g. nearest neighbour), which most closely
publishing of the invitation posts across resemble the ideally desired units (Rivers and
numerous online forums (covering broad Bailey, 2009).
array of topics, from cooking to cycling).
The systematic attempt to include a variety Direct incorporation of probability
of recruitment channels assured in this study sampling design principles
a considerable level of spread, which approx- Whenever possible, we can also directly
imated a good level of randomization. Of include probability sampling principles.
course, there was no guarantee in advance Sometimes components of probability
that this approach would work, but the results sample selection can be incorporated already
showed that corresponding estimates were into the early steps of the sampling process.
very close to those from a control probability For instance, in probability sampling with
sample. quotas we may first select a probability
Another indirect strategy is to shape non- sample of schools, where we then recruit
probability samples, so they would reflect, as pupils with unrestricted banner invitations on
much as possible, the structure of the survey the web according to some quotas (e.g. age,
population. If possible, of course, the match gender, year of schooling). Another example
need to be established for control variables, is a direct inclusion of randomization, which
which are related to the target ones. Quota can be incorporated into various ‘random
sampling is a typical approach here, usually walk’ selections related to the recruitment
based on controls for socio-demographics (e.g. across streets and buildings, or in randomiza-
gender, age, education, region, etc.). The quo- tion across time (e.g. days, hours). In specific
tas can be applied immediately at the recruit- types of convenience sampling (mall inter-
ment or at the screening phase of a survey. For cept, street recruitment, invitation on the
example, if we use gender quotas, this implies web) we can introduce randomization also by
that once we have reached the quota for including every ith (e.g. every tenth) passen-
women, we then reject to further survey any ger, customer or visitor.
women. Alternatively, we may also intensify Of course, none of the above-described
the recruitment channels, which attract more approaches, direct or indirect, assures with
men (e.g. advertising on sports media). some known accuracy that the corresponding
In case of a sampling frame, where we statistical inference will be equivalent to the
have additional information about the units situation with probability samples. However,
(e.g. as in online panels), we can use this it is also true that these approximations usu-
same strategy already at sample selection. ally contribute to certain improvements,
A fine selection can be then used, not only while they rarely cause any serious damage.
according to the target population structure of In the meanwhile, practitioners have devel-
socio-demographics (age, gender, education, oped elaborated skills, which often work well

BK-SAGE-WOLF-160177-Chp22.indd 331 4/22/2016 3:44:57 PM


332 The SAGE Handbook of Survey Methodology

for their specific circumstances, research although this is not true. Applying standard
problems and target variables. Again, there statistical inference approaches as an approx-
is no formal guarantee (in advance) that their imation in non-probability samples can thus
solutions will also work in modified circum- ‘seduce’ the users into believing that they
stances or in some other situations. have ‘real’ confidence intervals. Therefore,
it seems reasonable to consider using a dis-
Approximations in Statistical tinct terminology to formally separate a non-
Inference probability from a probability setting. For
As already mentioned, with non-probability example, as proposed by Baker et al. (2013)
samples, by definition, the inclusion proba- we might rather use the term ‘indications’
bilities are unknown or zero, so without fur- instead of ‘estimates’. This is also a direction
ther assumptions this very fact formally to which the hard-core statistical approach
prevents any statistical inference calculations against any calculation of confidence inter-
(e.g. estimates, variances, confidence inter- val in non-probability setting is slowly melt-
vals, hypothesis testing, etc.). ing in the recent AAPOR (2015) Code on
Uncontrolled inclusion probabilities can Professional Ethic and Practice.
also introduce a certain selection bias, for In practice the situation is not that bad and
which Bethlehem and Biffignandi (2012: this approximate usage often provides rea-
311) showed that it depends on (a) the varia- sonable results, which is mainly due to the
bility of inclusion probabilities and on (b) the following reasons:
correlation between target variable and inclu-
sion probabilities. For example, the Internet- a) In almost all non-probability samples there still
savvy respondents typically participate more exists a certain level of ‘natural’ randomization,
frequently in self-selected web surveys. As which, of course, varies across circumstances and
they differ from other Internet users not only from sample to sample.
b) This can be further accelerated with measures
in socio-demographics (e.g. younger) but also
researchers undertake for improving the non-
in web skills or altruism (Couper et al., 2007; probability sampling designs (see above discussion
Fricker, 2008; Malhotra and Krosnick, 2007), on spread, randomization, quotas and matching).
the results from such samples may show a c) Powerful techniques also exist in post-survey
bias towards the characteristics of these users. adjustment steps, such as imputations, weight-
Against this background, in survey practice ing, complex estimators, and statistical models
the same statistical inference procedures – (see discussion further below).
developed in probability sampling – are rou-
tinely applied also into a non-probability set- All these three streams (a, b, c) together can
ting, although there is no guarantee whether be very effective, at least according to the
and how this will work. For instance, if we perceptions of users and clients, who judge the
apply for a certain non-probability sample a results by using some external controls and
standard procedure to calculate confidence their knowledge. The above-mentioned study
intervals, we implicitly assume that prob- on online hate speech (Vehovar et al., 2012) is
ability sample selection (e.g. SRS) was used. a typical example here, because comparisons
However, at this time we are no longer con- with a parallel telephone probability sample
fident that the interval actually contains the confirmed the results from non-probability
population value with the pre-assumed, say, samples:
5% risk. It is very likely, that the actual risk
is way above 5%, but we have no possibility •• Estimated means on rating scales fully matched
to calculate it from the sample. Of course, we between the two samples.
can still pretend and use the same interpre- •• A similar result was true for the correlations,
tation as in a probability sampling context, ranks and subgroup analysis.

BK-SAGE-WOLF-160177-Chp22.indd 332 4/22/2016 3:44:57 PM


Non-probability Sampling 333

•• The estimates for the shares (percentages) were affected by deviations from probability sam-
skewed towards more intensive Internet users, pling principles. Consequently, they can be
but the ranks and relations between categories successfully dealt even with low-quality non-
were still usually preserved. probability samples; this is often the case in
marketing research. On the other hand, cer-
The above pattern is very typical for many tain variables, such as employment status,
studies with ‘well-spread’ and ‘well-weighted’ health status, poverty measures or study of
non-probability samples. Successful results rare characteristics (e.g. diseases, minori-
from non-probability samples are thus often ties, sensitive behaviour), are less robust for
reported, particularly for online panels, as deviations and require a strict probability
shown in Rivers (2007) and in Callegaro et al. sampling design and an elaborated estima-
(2014). However, the latter reference also tion treatment.
points to many exceptions. Similarly, scholars
often compare the two approaches and as a
result they typically expose advantages of the Specific Modelling Approaches
probability setting (e.g. Yeager et al., 2011).
We may add that advocates (e.g. Rivers, Being very formal and strict, we should
2010) of using probability sampling principles acknowledge that randomization – as well as
in non-probability setting also emphasise the related probability sampling – is not a
that papers in leading scholarly journals necessary precondition for valid statistical
often calculate confidence intervals for non- inference (Valliant et  al., 2000: 19). If we
probability samples, as well as that the non- have clear and valid assumptions, a specific
probability sampling approach is already modelling approach can also provide corre-
acceptable in many neighbouring fields sponding statistical inference. In such situa-
(e.g. experimental and quasi-experimental tion, we first assume that the data – or at least
studies, evaluation research and observational specific relations among the variables – are
studies). On the other hand, many statisticians generated according to some statistical
basically insist that without sound and explicit model. Next, additional external data
assumptions or models – which are usually are then particularly valuable (e.g. socio-
missing in non-probability sampling – there demographic background variables in case of
are no solid grounds to infer from any sample. online panels), so that certain model-based
In any case, whenever we use, run or dis- approaches to statistical inference can be
cuss non-probability samples, we should used to build the model and then estimate the
clearly differentiate among them, because values of interest. However, the correspond-
huge differences exist. While some carefully ing statistical work (i.e. assumption testing,
recruit and spread across the entire spectrum model development, data preparation, etc.)
of the population, using elaborated sam- can be complicated and time-consuming,
pling design procedures (e.g. strata, quotas, particular because it typically needs to be
matching), others may simply self-select done for each variable separately.
only within one narrow sub-segment. Similar Another stream of approaches is related
differences appear also with (non)using to sample matching, mentioned above in the
advanced procedures in post-survey adjust- relation to quota selection (which is a type
ment (imputation, weighting, calibration, of aggregated matching) and to individual
estimation, modelling). matching (where we may use some ‘nearest
Moreover, the specifics of the target vari- neighbour’ techniques). However, we can
ables and research problem are usually also also use some advanced models, which are
very important here. Some variables are sim- related to propensity score matching. The
ply more robust in a sense that they are less latter is closely related to causal inferences

BK-SAGE-WOLF-160177-Chp22.indd 333 4/22/2016 3:44:57 PM


334 The SAGE Handbook of Survey Methodology

techniques (Rubin, 1974), developed for perhaps the most frequent and most advanced
quasi-randomized and non-randomized set- example of contemporary non-probability
tings, and in particular to propensity score sampling, and weighting, because it is subject
procedures developed for causal inference for of high expectations about solving the prob-
observational studies (Rosenbaum and Rubin, lems arising from non-probability samples.
1983). An extensive overview of sample
matching implementation for non-probability
survey settings is provided in Baker et  al. Non-probability Online Panels
(2013: 34).
The achievements of some modelling When the same units are repeatedly surveyed
approaches are in fact admirable. For exam- at various points in time, we talk about longi-
ple, the successes of various voting pre- tudinal survey designs. Important advantages
dictions, which are based on results from over independent cross-sectional studies are
non-probability samples, are very notable. efficiency gains in recruiting and reduced
Sometimes they are perhaps even shocking sampling variation, because we can observe
for many traditional statisticians, who believe change at the individual level (Toepoel et al.,
that no statistical inference exists outside 2009). In traditional longitudinal designs, all
probability sampling. It is true, however, that units in the sample are repeatedly surveyed
weak documentation of the corresponding with questions belonging to a certain topic
procedures (particularly in commercial sec- (e.g. labour market). However, we will use
tor) might possibly violate the principles of here the notion of panel somehow differently
research transparency (AAPOR, 2014). and according to the understanding, which
Nevertheless, due to their specific and nar- actually prevails within non-probability sam-
row targeting, these approaches are often too pling context: panels are large databases of
demanding in skills and resources to be gen- units, which initially agreed to cooperate,
erally used. As a consequence, whenever we provided background information and were
need a straightforward practical implementa- recruited to serve for occasional or regular
tion of statistical inference, such as running selection into specific sample surveys
descriptive statistics of all variables in the addressing related or unrelated topics.
sample, we usually still run standard statis- With the rise of the Internet, the so-called
tical inference procedures (which assume online panels of general populations –
probability-based samples). Of course, as sometimes also labelled as access panels –
indicated above, this is formally correct only emerged in developed countries and they are
if we can prove that the necessary assump- rapidly becoming the prevailing survey data
tions hold, which is usually not the case. collection method, particularly in marketing
Otherwise, we run them as approximations research (Macer and Wilson, 2014). Our focus
or indications, together with the related here is on non-probability online panels,
risks, which then need to be explicitly stated, which are the dominant type of online panels,
warned and documented. while probability online panels are relatively
rare. We discuss general population panels,
but the same principles can be extended to
other populations, e.g. customers, profession-
SELECTED TOPICS IN NON- als, communities, etc.
PROBABILITY SAMPLING In general, non-probability online panels
comply with issues, challenges and solu-
When discussing non-probability sampling, tions (i.e. approximations), which have been
two issues deserve special attention: non- discussed in previous sections. Therefore,
probability online panels, because they are conceptually, there is not much to add. On the

BK-SAGE-WOLF-160177-Chp22.indd 334 4/22/2016 3:44:57 PM


Non-probability Sampling 335

other hand, the practices of online panels than others, but sometimes serious discrepan-
are very interesting and illustrative for non- cies to benchmark values were found for all
probability sampling context. One specific panels (e.g. political party preference).
is that recruitment (i.e. sample selection) A similar study, where 17 online panels
actually occurs at two points: first, with the in the US contributed a sample of respond-
recruitment of units into the panel from, say, ents (Thomas and Barlas, 2014), showed that
the general population, where we need to online panels actually provided consistent
focus on a good approximation of probability estimates on a range of general topics, such
sampling (i.e. spread and coverage). Second, as life satisfaction and purchasing behaviour.
when we recruit units from our panel into a The corresponding variations often resem-
specific sample, we need to focus on select- bled variations in estimates from probability
ing an optimal structure by using quotas, samples. However, there were also specific
stratification, matching, etc. topics, particularly those related to small
Careful management is required to prove population shares, where larger variations
the representation of essential subgroups in were found, such as occurrence of binge
each sample, but also to control the work- drinking (i.e. drinking more than 4–5 glasses
load for each unit. A panel management sys- of alcohol in one evening).
tem thus needs to record and optimize when An exhaustive overview of studies evaluat-
and how units are recruited, how often they ing results from non-probability samples can
participate in surveys, and how much time it be found in Callegaro et al. (2014). The over-
takes them to complete a survey. It also has view reveals that professional online panels
to detect various types of ‘cheaters’, which often provide results, which rarely differ
can be among others recruited from ‘profes- significantly from the corresponding bench-
sional respondents’, who may maintain fake marks. On the other hand, it also showed that
profiles to participate in online panels. After exceptions do exist. Similarly, this overview
the data are collected, another step of opti- demonstrates that the estimates from prob-
mization occurs by editing, imputing and ability samples consistently outperform the
weighting, which can be especially tailored ones from non-probability online panels.
to certain circumstances. Nevertheless, in practice, the dramatic differ-
Results from the NOPVO-study (Van ences in costs between the two alternatives
Ossenbruggen et al., 2006), in which 19 non- usually (but not always) outbalance the dif-
probability online panels in the Netherlands ferences in the estimates.
were compared, showed that panels use a We may add here that cost-error optimi-
broad array of recruitment methods, with the zation is a very complex and much under-
main approach relying on online recruitment researched issue even for probability-based
and various network procedures. Sometimes samples, where we can minimize the product
a subset – particularly hard-to-reach sub- of total survey error (usually approximated
groups – of panel members is also recruited with means squared error) and the costs (see
via probability-based sampling using tradi- Vehovar et  al., 2010; Vannieuwenhuyze,
tional modes (telephone, face-to-face). The 2014). Also in this case, the cost-error opti-
incentives also vary, from lotteries, charity mization in non-probability samples could,
donations, vouchers or coupons for books in principle, approximate the optimization in
or other goods, to prevailing method of col- probability-based approach, although in real-
lecting points, which can be then monetized. ity these judgements are often much more
Due to these variations, it is not surprising arbitrary.
that considerable differences were found also With non-probability online panels one
with respect to the estimates. Some panels has to be careful with statistical inferences,
in the NOPVO study performed much better as noted previously. Different to probability

BK-SAGE-WOLF-160177-Chp22.indd 335 4/22/2016 3:44:57 PM


336 The SAGE Handbook of Survey Methodology

samples, in non-probability online panels – as found in Bethlehem (2009), Biffignandi and


in non-probability samples in general – the Bethlehem (2012) and Valliant et al. (2013).
corresponding quality relies on a successful A core requirement for weights is that the
mix of approximations. The quality of a cer- variables used have to be measured in the
tain panel can be checked by monitoring their sample and also in the population. Moreover,
documentation and past results. Both ISO for weighting to be effective, they also need
20252 (Market, Opinion and Social Research) to be correlated with target variables.
and ISO 26362 (Access Panels in Market, Different weighting approaches exist;
Opinion and Social Research) standards for one is based on generalized regression esti-
instance specify key requirements to be met mation (linear weighting), where linear
and reported. ESOMAR (2015), the global regression models are used (Kalton and
organization for market, opinion and social Flores-Cervantes, 2003). Poststratification is
research, also regularly issues practical guide- a special and also most frequently used exam-
lines to help researchers make fit-for-purpose ple of this approach, where auxiliary varia-
decisions about selecting online panels. bles simply divide the population into several
We can summarize that contemporary strata (e.g. regions). All units in a certain
non-probability online panels successfully strata are then assigned the same weight. The
serve for studying numerous research top- problems appear, if strata are too small and
ics. However, here too, we cannot escape we have insufficient or no observations in a
from the usual limitations of non-­probability strata, which consequently hinder the compu-
samples. Correspondingly, for certain tation of weights for such strata. We may also
research problems these panels may not be lack corresponding population control data.
the right approach. Still, they are perhaps the The second stream of approaches is based
best example, how the limitations of non-­ on raking ratio estimation. This approach
probability setting can be overcome or at is closely related with iterative proportional
least minimized through carefully applied fitting (Bethlehem and Biffignandi, 2012),
methodological approaches. where we sequentially adjust for each aux-
iliary variable separately (e.g. age, gender,
etc.) and then repeat the process until con-
vergence is reached. This saves the above-
Weighting
described problem of small number of units
In probability samples so-called base weights in stratification and lack of population con-
are used first to compensate for the inclusion trols, and is thus very convenient when more
probabilities. In a second step, specific auxiliary variables are considered. However,
non-response weights are applied to reduce we are limited here to categorical auxiliary
the bias – a difference between the estimate variables; problems also appear with vari-
and the true value – resulting from non- ance estimation. In most situations the linear
response. In a similar way, specific weights weighting and raking provide asymptotically
can be developed for other methodological similar results. Still, differences may appear,
problems (e.g. non-coverage). In addition, depending on whether the underlying rela-
with so-called population weighting we usu- tion between auxiliary (control) and target
ally correct for any discrepancies in available variables is linear or loglinear (as assumed in
auxiliary variables. Typically, these are raking).
socio-demographic variables (e.g. age, We should also mention general calibra-
gender, region), but when available, other tion approaches, which enable to simulta-
variables are also considered (e.g. media neously include auxiliary information, as
consumption). A good overview regarding well as potential restrictions for the range
the general treatment of weights can be of weights. In addition, they produce the

BK-SAGE-WOLF-160177-Chp22.indd 336 4/22/2016 3:44:57 PM


Non-probability Sampling 337

estimates together with corresponding vari- than 1, while the opposite happens to over-
ance calculations. Linear weighting and rak- represented ones, which is similar as in popu-
ing can be then treated as special cases of lation weighing. For example, if the share of
these general calibration approaches. units from a certain segment in our sample is
We should be aware that weights may (or 25%, while it is 20% in the population, these
may not) remove the biases in target vari- units receive a weight proportional to 20/25
ables, but they for sure increase the sampling = 0.8.
variance. However, the underlying expecta- Within this framework, the propensity
tion is of course, that the gains in reducing score adjustment (PSA) has been proposed
the bias outweigh the corresponding loss due as a specific approach to overcome problems
to increased variance. This is not necessary of non-probability samples (e.g. Lee, 2006;
true, particularly in case of weak correla- Steinmetz et al., 2014; Terhanian et al., 2000;
tions between auxiliary and target variables, Valliant and Dever, 2011), particularly in
where weights may have no impact on bias situations when weighting based on stand-
removal. The bias–variance trade-off is usu- ard auxiliary variables (socio-demographics)
ally observed in the context of the so-called failed to improve the estimates. One reason
mean squared error (MSE), which is typi- for this is the fact that the attitudes of Internet
cally reduced to the sum of sampling variance users might differ substantially from those
and squared bias. In addition, the sampling of non-users (Bandilla et al., 2003; Schonlau
variance and potentially also the bias might et  al., 2007). Therefore, additional behav-
further increase when the auxiliary informa- ioural, attitudinal or lifestyle characteristics
tion cannot be obtained from the population, are considered. Observational and process
but only from some reference survey. data (so-called paradata) can also be use-
With probability samples the weighting ful (Kreuter, et al., 2011). PSA then adjusts
typically reduces only a minor part of the bias for the combined effects of selection prob-
(Tourangeau et  al., 2013), while the differ- abilities, coverage errors and non-response
ences between weighting approaches them- (Bethlehem and Biffignandi, 2012; Schonlau
selves are usually rather small (of course, as et al., 2009; Steinmetz et al., 2014). To esti-
long as they use the same amount of auxiliary mate the conditional probability of response
information). in the non-probability sample, a small
The above-described procedures (and probability-based reference survey is required,
their characteristic) developed for prob- which has been ideally conducted in the same
ability sampling can be used also in a non- mode using the same questionnaire. After
probability setting. Again, as we do not know combing both samples, a logistic (or probit)
the probabilities of inclusion, this usage is model estimates the probability of participat-
conditional and should be treated only as an ing in the non-probability sample using the
approximation. Correspondingly, the potentials selected set of available covariates. There
of weighting in the case of non-probability are several ways of using propensity scores
sampling hardly surpass the benefits in prob- in estimations (Lee, 2006). For instance,
ability sampling. Still, in survey practice, while Harris Interactive uses a kind of post-
basically the same procedures are applied in stratified estimator based on propensity score
non-probability setting, hoping they will some- (Terhanian et  al., 2000), YouGovPolimetrix
how work. With non-probability samples we uses a variation of matching (Valliant and
often have only one step, because we cannot Dever, 2011). Again, we should repeat that
separate sampling and non-response mecha- differences between weighting approaches
nisms. The weighting process thus simply are usual very small, which also means that
assign weights to units in the sample, so that PSA rarely significantly outperform alterna-
the underrepresented ones get a weight larger tive weighting procedures, given that both

BK-SAGE-WOLF-160177-Chp22.indd 337 4/22/2016 3:44:57 PM


338 The SAGE Handbook of Survey Methodology

approaches use the same auxiliary informa- repeatedly work well, while in others they
tion (Tourangeau et al., 2013). do not. Similarly, sometimes even a very
To summarize, weighting is by no mean deficient probability sample (e.g. with a very
a magic wand to solve the non-probability low response rate) outperforms high-quality
sampling problems. Still, it usually does pro- non-probability samples, but it can also be the
vide a certain improvement; at least it assures opposite. In any case, only repeated empiri-
cosmetic resemblance to population controls. cal evidence within specific circumstances
In practice it is thus routinely implemented, may gradually justify a certain non-probability
although it might sometimes also cause cer- sampling approach. Within this framework,
tain damage for some estimates. the justification of New York Times (NYT)
about why they switched from probability to
non-­probability sampling is very informative
(Desilver, 2014), i.e. they simply observed the
DECIDING ABOUT PROBABILITY repeated empirical evidence of the new supplier
SAMPLING APPROXIMATIONS (which relied on non-probability sampling) and
were persuaded by the corresponding success
When we implement approximations from in the past (and, perhaps, also by lower costs).
probability samples into non-probability set- Of course, with this decision they also took the
ting, there is by the very definition little theo- above-mentioned risk that past success might
retical basis for running a sound statistical not guarantee the success also in the future.
inference. Therefore, the justification and When deciding between probability and
validation of these procedures rely solely on non-probability sampling, the first stream of
the accumulation of anecdotal practical evi- considerations is thus to profoundly check all
dence, which is sometimes ironically called past evidence of approaches suitable for cer-
‘faith-based’ sampling. Only when a certain tain type of survey research. Whenever pos-
non-probability approach has repeatedly sible, for available non-probability samples
worked sufficiently well in specific circum- in the past we should inspect the distributions
stances, can this then become a basis for its of relevant target variables, together with
justifications. the effects of accompanying weights, and
We may reflect here on two comments on also compare them with external controls,
the initially mentioned Literary Digest poll e.g. administrative data or data from official
failure. First, despite long series of successful statistics. This will give us an impression of
evidence on previous election forecasts (from how the non-probability sample resembled
1916 onwards), the applied non-probability the probability ones in terms of randomiza-
sampling approach unexpectedly failed in tion and deviations from population controls
1936. This is an excellent illustration about (i.e. ‘representativeness’).
the risk when relying on successful anec- Another stream of considerations is to
dotal evidence from past success. Second, systematically evaluate broader survey
with today’s knowledge and approaches it is settings. Namely, we are rarely in the situation
very likely that this failure could have been to decide only about the probability vs non-
prevented. In other words, an elaborated probability dilemma, but rather between
non-probability sampling strategy would whole alternative packages, each containing
have been less likely to fail so unexpectedly, an array of specifics related also to non-
although it perhaps still could not provide the coverage, non-response, mode effects and
same assurance as a probability sample. other measurement errors, as well as those
In this context, we should also point out related to costs, timing and comparability.
again that in some circumstances, certain Extreme care is often required to separate the
non-probability sampling approaches may genuine problem of non-probability sampling

BK-SAGE-WOLF-160177-Chp22.indd 338 4/22/2016 3:44:57 PM


Non-probability Sampling 339

(e.g. selection bias, lack of randomization) from breakthrough from 1930s, which brought
other components. For example, it would be of spectacular changes into survey practice.
little help to reject a non-probability sample, if After the elaboration of non-probability sam-
problems actually stem from mode effect. In pling approaches we can perhaps paraphrase
such case replacing the non-probability web the same also for the gains arising from
survey with a probability web survey brings no modern non-probability sampling. The
improvement, but only lots of costs. advent of non-probability web surveys, in
We thus rarely decide separately about particular, revolutionized the survey land-
the probability–non-probability component scape. Numerous survey projects were ena-
alone, but rather simultaneously judge over bled, which otherwise could not have been
a broad spectrum of features. For instance, done, and many others were facilitated to run
in a survey on youth tourism we may choose dramatically faster, simpler and less
between a probability-based telephone, mail expensive.
or web survey, and a non-probability online We continuously indicated in this chapter –
panel or a sample recruited through online explicitly or implicitly – that in a non-
social media. The decision involves simulta- probability setting the statistical inference
neous considerations of many criteria, where based on probability sampling principles
all above-mentioned components (non- formally cannot be applied. Still, the
response, non-coverage, selection bias, mode practitioners routinely use the corresponding
effect, costs, timing, comparability, accuracy, ‘estimates’, which should be rather labelled
etc.) are often inseparably embedded into as ‘indications’ or even ‘approximations’.
each alternative (e.g. the social media option Of course the corresponding evaluation of
perhaps comes with high non-coverage and such ‘estimates’ cannot be done in advance
selection bias, but also with low costs and (ex-ante), as in probability settings, but only
convenient timing, etc.). Within this context, afterwards (ex-post). Still, such evaluations
the selection bias itself, arising from the use are very precious, as we can improve the
of a non-probability setting, is only one spe- related procedures, or, at least recognize their
cific element in the broad spectrum of factors limitations and restrict their future usage only to
and cannot be isolated and treated separately. a narrow setting for which it works. After many
Sometimes, combining the approaches can successful repetitions, with corresponding
be a good solution. We may thus obtain popu- trials and errors, the practitioners often
lation estimates for key target variables with develop very useful practical solutions. Further
some (expensive) probability survey, while in- discussion on these so-called ‘fit-for-purpose’
depth analysis is done via some (inexpensive) and ‘fitness-to-use’ strategies can be found in
non-probability surveys. Again, the study by the AAPOR report on non-probability samples
Vehovar et  al. (2012) on online hate speech (Baker et al., 2013), which systematically and
might serve as a good example, where key cautiously addresses selection, weighting,
questions were put in a probability-based omni- estimation and quality measures for non-
bus telephone survey, while extensive in-depth probability sampling
questioning was implemented within a broadly We should be thus clearly aware that appli-
promoted non-probability web survey. cation of standard statistical inference proce-
dures in a non-probability setting implicitly
assumes that inclusion probabilities actually
do exist (which is of course not true). For
CONCLUDING REMARKS example, calculating the mean for a target
variable in a non-probability sample of size n
When introducing sampling techniques for automatically assumes that probabilities are all
probability samples we recalled a major equal (e.g. SRS). Consequently, the inclusion

BK-SAGE-WOLF-160177-Chp22.indd 339 4/22/2016 3:44:57 PM


340 The SAGE Handbook of Survey Methodology

probabilities n/N are implicitly assigned to all not a scientific method with precise definition. It is
units in the sample, where N is the population more of an art practiced widely with different skills
and diverse success by many people in different
size. It then depends on corresponding ex-
places. There exists no textbook on the subject to
post evaluation procedures to judge the qual- which we can refer to base our discussion. This
ity of such an ‘estimate’, which as mentioned alone should be a warning signal against its use.
should be perhaps better called ‘indication’ or
‘approximation’. On the other (practical) side, it is also true
By abandoning probability sampling prin- that it is hard to deny advances in modern
ciples we usually also abandon the science of approaches for dealing with non-probability
statistical inference and enter instead into the samples. It is also hard to object to countless
art and craft of shaping optimal practical pro- examples of successful and cost-effective
cedures. The experience based on trials and implementations in research practice.
errors is thus essential here, as well as the We thus recommend to more openly accept-
intuition of the researcher. Correspondingly, ing the reality of using a standard statistical
with complex problems an expert advice or a inference approach as an approximation in
panel of experts with rich experience can be non-probability settings. However, this comes
extremely valuable. with two warnings: first, the sample selection
Of course, hard core statisticians may feel procedure should be clearly described, docu-
doubts or even strong professional opposition mented, presented and critically evaluated.
towards procedures, where inferential risks We thus join the AAPOR recommendations
cannot be quantified in advance. On the other that the methods used to draw the sample, to
hand, is this not also the very essential char- collect the data, to adjust it and to make infer-
acteristic of the applied statisticians – which ence should be even more detailed compared
separates them from ‘out-of-reality’ math- to probability samples (Baker et  al., 2013).
ematical statisticians – to stop and accept the The same should also apply for reporting
procedures, which bring them close enough on standard data quality indicators. Within
for all practical purposes? this context, the AAPOR statement related
We should also add that the above-­discussed to NYT introduction of non-probability sam-
issues are nothing really new. More than a pling is very illustrative (AAPOR, 2014).
hundred years ago researchers were already Second, we should also elaborate on the
trying various inexpensive alternatives to underlying assumptions (e.g. models used)
replace studying entire population or to replace and provide explanations about conceptual
expensive probability sampling procedures. divergences, dangers, risks and limitations
Wherever these alternatives worked suffi- of the interpretations (AAPOR, 2015). When
ciently well, they were simply preserved. standard statistical inference is applied to
Therefore, on one (conceptual) side we any non-probability sample, the minimum
have serious objections to non-­ probability should be thus to clearly acknowledge that
samples, summarized in the profound estimates, confidence intervals, model fitting
remarks from Leslie Kish (1998) in his reply and hypothesis testing may not work properly
to the repeated requests from the marketing or may not work at all. Of course, the entire
research industry for a review of his attitudes approach should be accompanied with corre-
to quota samples: sponding survey profession integrity, which
means exhaustively and transparently.
My section 13.7 in Survey Sampling expresses Even though these concluding remarks pre-
what I thought then (1965) and I still think. I dominantly relate to the prevailing situation,
cannot add much, because quota methods have
not changed much, except that it is done by many
where we use probability sampling principles
more people and also by telephone. Quota sam- as approximations in non-probability setting,
pling is difficult to discuss precisely, because it is they are also highly relevant to dedicated

BK-SAGE-WOLF-160177-Chp22.indd 340 4/22/2016 3:44:57 PM


Non-probability Sampling 341

statistical modelling approaches, which have MainSiteFiles/ NPS_TF_Report_Final_7_


particular potential for future developments revised_FNL_6_22_13.pdf
in this area. Bandilla, W., Bosnjak, M. and Altdorfer, P.
(2003). Survey administration effects? A
comparison of web-based and traditional
written self-administered surveys using the
Note ISSP environment module. Social Science
Computer Review, 21, 235–243.
 1  Some parts of this chapter rely on Section 2.2, which Bethlehem, J. (2009). Applied Survey Methods.
Vasja Vehovar wrote for Callegaro et al. (2015). A Statistical Perspective. (Wiley Series in
Survey Methodology). Hoboken, NJ: Wiley.
Bethlehem, J. and Biffignandi, S. (2012).
Handbook of Web Surveys. Hoboken, NJ:
RECOMMENDED READING Wiley.
Callegaro, M., Baker, R., Bethlehem, J., Göritz,
Report of the AAPOR Taskforce on Non-Proba- A., Krosnick, J. and Lavrakas, P. (eds). (2014).
bility Sampling by Baker et al. (2013) provides Online Panel research: A Data Quality
an exhaustive and systematic insight into Perspective. Chichester: Wiley.
issues related to non-probability sampling. Callegaro, M., Lozar Manfreda, K. and Vehovar,
Web Survey Methodology by Callegaro et  al. V. (2015). Web Survey Methodology.
(2015) provides extensive treatment of non- London: Sage.
probability sampling approaches in web Couper, M., Kapteyn, A., Schonlau, M. and
surveys. Winter, J. (2007). Noncoverage and
Online Panel Research: A Data Quality Perspec- nonresponse in an Internet survey. Social
tive by Callegaro et  al. (2014) provides a Science Research, 36, 131–148.
state-of-the-art insight into the non-proba- Desilver, D. (2014). Q/A: What the New York
bility nature of contemporary online panels. Times’ polling decision means. Retrieved
July 1, 2015, from http://www.pewresearch.
org/fact-tank/2014/07/28/
qa-what-the-new-york-times-polling-
References decision-means/
ESOMAR (2015). ESOMAR/GRBN Guideline for
AAPOR (2014). AAPOR Response to New York Online Research. Retrieved June 25, 2015,
Times /CBS News poll: The Critical Role of from: https://www.esomar.org/knowledge-
Transparency and Standards in Today’s World and-standards/codes-and-guidelines/esomar-
of Polling and Opinion Research. Retrieved grbn-online-research-guideline-draft.php
July 1, 2015, from http://www.aapor.org/ Fowler, F. J. (2014). Survey Research Methods
AAPORKentico/Communications/Public- (5th edn). Thousand Oaks, CA: Sage.
Statements.aspx Fricker, R. (2008). Sampling methods for web
AAPOR (2015). The Code of Professional Ethics and E-mail surveys. In N. Fielding; R. Lee
and Practices. Retrieved July 1, 2015, from and G. Blank (eds), The SAGE Handbook of
http://www.aapor.org/AAPORKentico/ Online Research Method (pp. 195–216).
Standards-Ethics/AAPOR-Code-of-Ethics.aspx London: Sage.
Agresti, A. and Franklin, C. (2013). Statistics: Groves, R., Fowler, F., Couper, M., Lepkowski,
The Art and Science of Learning from Data J., Singer, E. and Tourangeau, R. (2009).
(3rd edn). Oak View, CA: Pearson. Survey Methodology. Hoboken, NJ: Wiley.
Baker, R. P., Brick, J. M., Bates, N., Battaglia, M. Healey, J. F. (2011). Statistics: A Tool for Social
P., Couper, M. P., Dever, J. A., Gile, K. J. and Research. Belmont, CA: Wadsworth
Tourangeau, R. (2013). Report of the AAPOR Publishing.
Taskforce on Non-Probability Sampling. Kalton, G. (1983). Introduction to Survey
Retrieved from https://www.aapor.org/ Sampling. SAGE University Paper series on
AAPORKentico/AAPOR_Main/media/ Quantitative Applications in the Social

BK-SAGE-WOLF-160177-Chp22.indd 341 4/22/2016 3:44:57 PM


342 The SAGE Handbook of Survey Methodology

Sciences, series no. 07–035. Beverly Hills and Rivers, D. and Bailey, D. (2009). Inference from
London: Sage. matched samples in the 2008 U.S. national
Kalton, G. and Flores-Cervantes, I. (2003). elections. Presented at the American
Weighting methods. Journal of Official Association for Public Opinion Research
Statistics, 19(2), 81–97. (AAPOR) 64th Annual Conference,
Kiaer, A. N. (1997). The Representative Method of Hollywood, Florida.
Statistical Surveys (English translation by Rosenbaum, P. and Rubin, D. (1983). The
Christiania Videnskabsselskabets Skrifter), Oslo, central role of the propensity score in
Norway: Statistics Norway, (Reprinted from observational studies for causal effects.
Den repräasentative undersökelsesmetode, II. Biometrika, 70, 41–55.
Historiskfilosofiske klasse, No. 4, 1897). Rubin, D. (1974). Estimating causal effects of
Kish, L. (1965). Survey Sampling. New York: treatments in randomized and
Wiley. nonrandomized studies. Journal of
Kish, L. (1998). Quota sampling: Old Plus Educational Psychology, 66(5), 688–701.
New Thought. Personal communication. Särndal, C.-E., Swensson, B. and Wretman, J.
Retrieved June 25, 2015, from http://www. (2003). Model Assisted Survey Sampling.
websm.org/db/12/11504/Web%20 (Series in Statistics). New York, Heidelberg,
S u r v e y % 2 0 B i b l i o g r a p h y / Dordrecht, London: Springer.
Quota_sampling_Old_Plus_New_Thought/ Schonlau, M., van Soest, A. and Kapteyn, A.
Kreuter, F., Olson, K., Wagner, J., Yan, T., Ezzati- (2007). Are ‘Webographic’ or attitudinal
Rice, T., Casas-Cordero, C., Lemay, M., questions useful for adjusting estimates from
Peytchev, A., Groves, R. and Raghunathan, T. web surveys using propensity scoring?
(2011). Using proxy measures and other Survey Research Methods, 1, 155–163.
correlates of survey outcomes to adjust for Schonlau, M., van Soest, A., Kapteyn, A. and
non-response: examples from multiple Couper, M. (2009). Selection bias in web
surveys. Journal of the Royal Statistical surveys and the use of propensity scores.
Society Series, A 173(2), 389–407. Sociological Methods Research, 37, 291–318.
Lee, S. (2006). Propensity Score Adjustment as Steinmetz, S., Bianchi, A., Biffignandi, S. and
a weighting scheme for volunteer panel web Tijdens, K. (2014). Improving web survey
surveys. Journal of Official Statistics, 22, quality – potentials and constraints of
329–349. propensity score weighting. In: Callegaro,
Lohr, S. L. (2010). Sampling: Design and M., R. Baker, J. Bethlehem, A. Göritz,
Analysis. Pacific Grove, CA: Brooks/Cole J. Krosnick and P. Lavrakas (eds), Online
Publishing Company. Panel research: A Data Quality Perspective
Macer, T. and Wilson, S. (2014). The 2013 (chapter 12, pp. 273–298). (Series in Survey
Market Research Technology Survey. London: Methodology), New York: Wiley.
Meaning ltd. Steinmetz, S., Raess, D., Tijdens, K. and de
Malhotra, N. and Krosnick, J. (2007). The effect Pedraza, P. (2013). Measuring wages
of survey mode and sampling on inferences worldwide – exploring the potentials and
about political attitudes and behaviour: constraints of volunteer web surveys. In:
comparing the 2000 and 2004 ANES to Sappleton, N. (Ed.): Advancing Social and
Internet surveys with nonprobability samples. Business Research Methods with New Media
Political Analysis, 15, 286–324. Technologies. Hershey, PA: IGI Global.
Rivers, D. (2007). Sampling for web surveys. In: Terhanian, G., Bremer, J., Smith, R. and Thomas,
Proceedings of the Joint Statistical Meeting, R. (2000). Correcting data from Online Survey
Survey Research Methods Section. Salt Lake for the Effects of Nonrandom Selection and
City, UT: AMSTAT. Nonrandom Assignment. Research paper,
Rivers, D. (2010). AAPOR report on online panels. Harris Interactive, Rochester, NY.
Presented at the American Association for Thomas, R. and Barlas, F. M. (2014).
Public Opinion Research (AAPOR) 68th Annual Respondents playing fast and loose?
Conference, Chicago, Illinois and at the PAPOR Antecedents and Consequences of
2010 Annual Conference, San Francisco. Respondent Speed of Completion. Presented

BK-SAGE-WOLF-160177-Chp22.indd 342 4/22/2016 3:44:57 PM


Non-probability Sampling 343

at The American Association for Public Van Ossenbruggen, R., Vonk, T. and Willems, P.
Opinion Research (AAPOR) 69th Annual (2006). Dutch Online Panel Comparison
Conference, Anaheim, California. Study (NOPVO). Retrieved July 1, 2015, from
Toepoel, V., Das, M. and van Soest, A. (2009). http://www.nopvo.nl/
Relating question type to panel conditioning: Vannieuwenhuyze, J. (2014). On the relative
a comparison between trained and fresh advantage of mixed-mode versus single-
respondents. Survey Research Methods, mode surveys. Survey Research Methods,
3(2), 73–80. 8(1), 31–42.
Tourangeau, R., Conrad, F. and Couper, M. Vehovar, V., Berzelak, N. and Lozar Manfreda,
(2013). The Science of Web Surveys. New K. (2010). Mobile phones in an environment
York: Oxford University Press. of competing survey modes: applying metric
Valliant, R. and Dever, J. (2011). Estimating for evaluation of costs and errors. Social
propensity adjustments for volunteer web Science Computer Review, 28(3), 303–318.
surveys. Sociological Methods, 40(1), Vehovar, V., Motl, A., Mihelič, L., Berčič, B. and
105–137. Petrovčič, A. (2012). Zaznava sovražnega
Valliant, R., Dever, J., and Kreuter, F. (2013). govora na slovenskem spletu. Teorija in
Practical Tools for Designing and Weighting Praksa, 49(1), 171–189.
Survey Samples. (Statistics for Social and Yeager, D., Krosnick, J., Chang, L., Javitz, H.,
Behavioral Sciences Series): New York Levendusky, M., Simpser, A. and Wang, R.
Heidelberg Dordrecht London: Springer. (2011). Comparing the accuracy of RDD
Valliant, R., Dorfman, A. H. and Royall, R. M. telephone surveys and Internet surveys
(2000). Finite Population Sampling and conducted with probability and non-
Inference: A Prediction Approach. New York: probability samples. Public Opinion Quarterly,
John Wiley. 75(4), 709–747.

BK-SAGE-WOLF-160177-Chp22.indd
View publication stats 343 4/22/2016 3:44:57 PM

You might also like