FOR WHAT APPLICATIONS CAN PROBABILITY AND

NON-PROBABILITY SAMPLING BE USED?

H. T. SCHREUDER1 , T. G. GREGOIRE1 and J. P. WEYER2
1 USDA Forest Service, Rocky Mountain Research Station, Fort Collins, CO, U.S.A.; 2 York
University, New Haven, Connecticut, U.S.A.

(Received 9 June 1999; accepted 28 October 1999)

Abstract. Almost any type of sample has some utility when estimating population quantities. The
focus in this paper is to indicate what type or combination of types of sampling can be used in
various situations ranging from a sample designed to establish cause-effect or legal challenge to
one involving a simple subjective judgment. Several of these methods have little or no utility in the
scientific area but even in the best of circumstances, particularly complex ones, both probabilistic and
non-probabilistic procedures have to be used because of lack of knowledge and cost. We illustrate
this with a marbled murrelet example.
Keywords: credibility, design-based inference, inductive logic, inferences, model-based inference,
non-probability samples, probability samples, sampling protocols

1. Introduction
Environmental and ecological data are important for our understanding of the structure and functions of our ecosystems. When environmental or ecological data are
obtained by sampling but the design and other aspects of the sampling protocol are
unknown, it may be difficult to extrapolate these data to some larger population
or cohort. Nonetheless they constitute information that has utility and value, at the
very least to describe and characterize the particular sample. When the sampling
strategy, that is, the combination of a sample design and an estimator of a population quantity, is known and the target population is identified, extrapolation is
possible and more credible. Taking a very broad view, sampling may be probabilistic or not (a probability sample is one for which every unit in a finite population
has a positive probability of selection, not necessarily equal to that of other units); it
may be informative or uninformative with respect to the variable of interest; it may
include response and self-selection bias and be subject to measurement error; and
it may have a temporal or longitudinal structure combined with a spatial structure.
Because all these factors affect estimation and inference, it is important to understand the constraints imposed by the sampling protocol on the interpretation of
data. In this paper we focus upon one aspect of this multifaceted problem namely
how the scope and validity of inference is impacted by sample selection. More
Environmental Monitoring and Assessment 66: 281291, 2001.
2001 Kluwer Academic Publishers. Printed in the Netherlands.

282

H. T. SCHREUDER AND T. G. GREGOIRE

specifically, the titular concern is the manner in which data from probability and
non-probability samples can be combined for various inferential purposes
The motivation for this study arose from historical practice and emerging need.
The U.S. Forest Service (USFS) has had a long tradition of conducting probabilistic surveys e.g. in Forest Inventory and Analysis (FIA) that started in 1929
(Birdsey and Schreuder, 1992). Other surveys, particularly on National Forest (NF)
lands were more often seat-of-the-pants samples, or cruises, with little statistical
validity. They were applied because of the prevailing view that valid surveys were
impractical and could not be afforded (often true) or would yield embarrassing results. Examples are James and Schreuder (1971) where the estimates generated were
embarrassing, and Schreuder et al. (1980) where the techniques developed were
unacceptable because of costs involved. Because of the threat of lawsuits coupled
with an increased awareness of the need for statistically defensible survey methods
among government, industry, and environmental groups, use of probability samples
is becoming mandatory (see for example Max et al., 1997).
Nonetheless, there remain many situations for which environmental and natural
resource data are perceived to be useful even when not collected in a statistical or
probabilistic manner. Moreover, much environmental data can not be collected in
such a way because of expense, time constraints, or other impediments. Such data
are nonetheless important and useful, providing that we can understand properly
their limitations.
Currently, inventory and monitoring issues are the most litigious in the USFS.
As yet, the quality of the natural resource data collected by the organization has not
been challenged, but this may happen soon. As noted by Meier (1986), the Supreme
Court has sanctioned formal statistical inference in a series of EEO decisions. It is
only a matter of time before the role of statisticians in the legal sector is extended
from cases of discrimination in employment to issues related to collecting and
interpreting environmental data.
With the pervasive concern over environmental quality and forest health, survey
data will be scrutinized more closely to discern dose-response and cause-effect
relationships. Documenting cause-effect is very difficult as illustrated by the most
noteworthy example of the last 50 yr, the association of smoking and lung cancer. In
the mid 1950s, a link between smoking and lung cancer was established based on
epidemiological survey data. Causality was established through verified prediction
in the next 20 yr. Only recently has the causal mechanism of how smoking causes
lung cancer been established (Peifer, 1997).
As we head into the information age, data on environmental and ecological systems will become increasingly more available, the need to understand the strength
of these data for inference likewise becomes increasingly important.

THE USE OF PROBABILITY AND NON-PROBABILITY SAMPLING

283

2. Review of Literature
The process of drawing conclusions from the analysis of observed data about unobserved parameters or underlying laws is called inductive logic, one of the most
controversial issues in philosophy. A discussion of the way in which data from
different sources may be combined necessarily involves considering the purposes
for which they are obtained. Inference, and the mechanism of inductive logic, is
not limited to the comparatively narrow field of scientific and statistical inference.
Nonetheless, the latter is an important sphere of activity, and a proper understanding of statistical inference is crucial to a discussion of the role of sampling in the
inferential process.
Scientific inference becomes statistical inference when the connection between
the unknown state of nature and the observand is expressed in probabilistic terms
(Dawid, 1984). There are two dominant paradigms for statistical inference from
sample surveys. One approach, called model-based inference, relies on a statistical
model to describe how the probability structure of the observed values depends on
uncontrollable chance variables and often on other unknown nuisance variables.
Such models may be based on a theoretical understanding of the procedure by
which the data are generated, past experience with similar processes, or experimental techniques used. But they can be ad hoc too, chosen for ease of interpretation or analysis. The other approach, called design-based inference, relies on the
probabilistic nature of the sampling. It has been the dominant paradigm since the
mid-1930s as testified to by its exclusive exposition in nearly all the prominent
texts on classical sampling theory and methods. We give below a capsule summary
of both these modes of statistical inference deferring to Srndal (1978), Srndal et
al. (1992), and Gregoire (1998) for a more comprehensive treatment.
Inherently, the design-based approach to statistical inference is tied to probability sampling, characteristics of which are that each element of the population
has a nonzero probability of being selected and the probability of each possible
sample can be deduced. The statistical behavior of estimators of a population attribute is reckoned with respect to these probabilities and the probability-weighted
distribution of all possible sample estimates. It does not rely on the distribution,
probabilistic or otherwise, of the attribute in the population being surveyed. One
critique of the design-based approach is that samples that could have been drawn,
but were not, are irrelevant: should not inference about a population parameter
rightfully be based solely upon the observed sample? (Hjek, 1981). Nonetheless,
the perceived objectivity of design-based inference has broad appeal and support
(Hansen et al., 1983). The only assumption it makes is that observational units are
selected at random so the validity of the inference only requires that the targeted
and sampled population are the same (Koch and Gillings, 1984).
One hopes that a probability sample is representative of the survey population itself, although as Kruskal and Mosteller (1979) ably demonstrate, the word
representative is subject to a wide array of interpretations. One of these is that a

284

H. T. SCHREUDER AND T. G. GREGOIRE

probabilistic sample is representative by definition. Others contend that a sample

comprising the smallest n elements of a population of N > n elements is hardly
representative of the population, whether or not it is chosen probabilistically.
Other methods of sampling have been proposed with a view towards providing
a sample that is more of a microcosm of the population. Srndal et al. (1992) have
a short but illuminating discussion on several purposive sample selection methods,
i.e. expert choice sampling, quota sampling, and balanced sampling. The latter two
methods select samples from the population in ways that provide zero probabilities of selection for many units of the population. With such samples, accurate
estimation of a population attribute may be possible, but an objective measure of
precision is not possible within the design-based framework.
Schreuder and Alegria (1995) demonstrated the consequences for inference of
using a probabilistic sample from a well defined population of interest, to estimate
population totals if the probabilities of selection are unequal but unknown and,
consequently, are treated as equal. This situation arises, for example, when a prior
sample from a timber sale is used for a new purpose, e.g. ecological monitoring.
Timber sale samples are often probabilistic but focused more heavily on strata
with a greater concentration of timber than on other strata. Using the incorrect
probabilities of selection introduces a design-based bias, which can be large.
It is important to understand clearly the strengths and weaknesses of purposive
and random sampling. Often, purposive sampling can not be avoided, e.g., when
sampling wildlife for which budgets are often minimal, the sampling process is
difficult (because the animals are scattered and not stationary) and the information
collected may have considerable political value because of an Endangered Species
Act. Similarly, meteorological variables require very frequent measurement to be
useful for purposes such as compiling growth and mortality models in forestry.
At present, it is often impractical to gather such information at random sites such
as FIA plots (when this is feasible, better growth and mortality models will be
produced).
If a sample is selected in a non-probabilistic manner, inference may be made by
specifying an underlying superpopulation model for the N-dimensional distribution
of Y = (Y1 , . . ., Yk , . . ., YN ) U , with Yk a random variable tied to the kth element
(Sarndal et al., 1992). If this model is called , the actual finite population vector of
interest y = (y1 , y2 , . . ., yk , . . ., yN ) can be considered a realization of Y . Assuming
P
that we are interested in the random variable Y = N
i=1 Yi , then for a sample s,
the realized values yk of the random variable Yk are observed for ks and not
observed for kU s. Using the specification of , then for an estimator Y of Y ,
the distribution of Y Y can be derived for the specific sample s and the modelbased mean square error of Y Y can be obtained and estimated, leading to a
model-based prediction interval for Y .
In this framework, inference may extend to parameters of the superpopulation
model, and not just the particular population at hand. Hence the inference space

THE USE OF PROBABILITY AND NON-PROBABILITY SAMPLING

285

is necessarily broader than design-based inference. Sample elements need not be

chosen randomly or with known probability, so long as they are not selected on
the basis of their yk values. Because a probability structure has been assumed for
the population itself, the distribution-free (strictly speaking not true since it relies
on randomization distribution but not on an assumed distribution so that it is often
considered distribution-free) properties of design-based inference are sacrificed.
Conclusions and inferences depend on the validity of the model, which may be a
serious liability if the model is not specified correctly. Conversely, with a correctly
specified model, which can be viewed as a source of auxiliary information which
is not fully utilized in the design-based framework, an increase in precision may be
realized.
In forestry a strong linear relationship exists between tree volume (v) and diameter breast height squared times total height (d 2 h). Such models, when fitted to
data, often have R 2 values of 0.98 to 0.99, and the data exhibit heterogeneous errors with variance proportional approximately to (d 2 h)2 (Schreuder and Williams,
1997 and references therein). Yet even with such a strong relationship, there is
still considerable risk in selecting units purposively and relying on the stipulated
model for inference. Schreuder et al. (1990) showed unacceptable estimation bias
for populations estimating total volume with d 2 h as covariate.
The problems raised above are as important in epidemiology as in forestry
and an excellent example of the issues in an epidemiological context are discussed by Greenland (1990). Overton et al. (1993), Cox and Piegorsch (1996),
and Piegorsch and Cox (1996) have also examined ways to combine probability
with nonprobability samples.
T. M. F. Smith, formerly a strong advocate of model-based inference, now
seems noncommittal, viz., My view is that there is no single right method of
inference. All inferences are the product of mans imagination and there can be
no absolutely correct method of inductive reasoning. Different types of inference
are relevant for different problems and frequently the approach recommended reflects the statisticians background such as, science, industry, social sciences or
government ... I now find the case for hard-line randomization inference based on
the unconditional distribution to be acceptable ... Complete reconciliation is neither
possible nor desirable. Vive la difference (Smith, 1994, p. 17).
Brewer (1994, 1995, 1998) discusses in detail a model-based sampling procedure called balanced sampling, first suggested by Royall (1970, 1971). This procedure is totally model-dependent and accepts only samples for which the sample
moments equal the population moments for the covariates x, e.g., the sample mean
equals the population mean. This presupposes considerable information about the
population of x-values and, as Brewer (1994) indicates, the procedure is practicable
primarily for one-time surveys. Even here, stratified sampling may be a suitable
alternative. Further research is needed when balanced sampling should be used.
Schreuder and Reich (1998) discuss proposals for using models (imputation
or tree models) to update tree and plot information and suggest that while it is

286

H. T. SCHREUDER AND T. G. GREGOIRE

inappropriate for public data sets, it could be useful to specific users for updating
data for their own purposes. But to quote Smith (1994, p. 34); with whom we concur: Democratic systems and good government require data bases of the utmost
integrity. Supplementing observational data may cause some in the public sector
to question the integrity of the data and inferences stemming from their analysis.
For scientific and other users of survey data it seems imperative to understand
the scope and strength of inference that can be made following surveys of various
types. The motivation and objective of this paper is to further this understanding.

3. A Suite of Sampling Scenarios

In order to satisfy the objectives of the paper we consider the following hierarchy
of inferences for which sample data are needed:
a.
b.
c.
d.
e.
f.

to draw inferences for the individual collecting the information only?

to draw inferences for an organization?
for inferences published in a scientific publication?
for inferences by the public at large, e.g. such as a public data base might be?
to identify possible cause-effect hypotheses?
to establish cause-effect relationships?

In the following we describe a series of sampling scenarios corresponding to this

hierarchy and consider valid inferences that may be drawn from each.
Assume that data sets are available relating to mortality and basal area growth
of trees in a particular forested region and each was acquired by different means,
namely:
a. An interested party reports that during his travels through the region over the
past 25 yr, he has noticed a sharp decrease in tree size and the forests seem to
be less vigorous than they used to be.
Possible conclusion: In this situation the target population is unspecified and difficult to discern and the sample is subject to sample and site selection bias. The
qualitative assessment of forest vigor is highly subjective. Therefore, no scientific
conclusions can be made. This is a common situation in that all of us draw conclusions at times based on little or no information and, one might argue, this was a key
reason why statistics impacted so profoundly on scientific inference. Basically, the
only inference to be drawn from this anecdotal evidence is that there is a hypothesis
to be examined that a change may have occurred (either in the person doing the
observing, the climate, the forest, etc.).

THE USE OF PROBABILITY AND NON-PROBABILITY SAMPLING

287

b. A data set was purposively collected by representatives of forest industry or

an environmental group to document a sharp increase in mortality and decline
in basal area growth in natural pine stands in a region. No documentation is
available of how the data were collected.
Possible conclusion: Here the population is defined to be natural pine stands. In
contrast to the scenario in a), these data are derived from a defined measurement
protocol rather than qualitative assessment. The purposive selection does not allow
design-based inference about trends in mortality and basal area growth. Without
more information about the criteria used for the purposive sampling, even modelbased inference is untenable. Therefore, only very limited scientific conclusions
can be drawn relating to an increase in mortality and decline in growth for the
cohort comprising the stands actually observed. However, these data can serve
useful descriptive purposes without any probabilistic interpretation (cf. Greenland,
1990, p. 6). One may scrutinize the data set for possible indicators of decline, which
may suggest hypotheses to be examined later in a scientific setting.
c. FIA data are selected based on screening by variables such as undisturbed plots
growing in natural stands (Gadbury et al., 1998 and references therein).
Possible conclusion: This is a probability sample of some population and any conclusions drawn are statistically valid, e.g. mortality increased and growth declined.
The problem is that one has no idea of the population represented by the sample.
d. An FIA data set comprising all the plots in the region is used to demonstrate
an increase in mortality and decline in basal area growth.
Possible conclusion: A statistically valid increase in mortality and decline in basal
area growth can reasonably be inferred for the entire region.
e. The situation as in d) but some of the possible causes are measured too. The
data may be analyzed for possible explanations for the decline (drought, beetle
infestations, etc.) but no cause-effect can be established, only hypothesized.
f. Given the situation in e), but all possible causes are measured. Even this is usually not enough to establish cause-effect but at least the data may indicate what
the most likely cause(s) is. If, in addition the analysis from the data show that
at least 2 of the 3 criteria of Mosteller and Tukey, viz. consistency and responsiveness were met (Schreuder and Thomas, 1991; Olsen and Schreuder, 1997)
than cause-effect has been established. Ideally, the third criterion: mechanism
should be established too but this is often very difficult as noted earlier in the
smoking-cancer issue. Consistency implies that the presence and magnitude
of the effect is always associated with a minimal level of the suspected causal

288

H. T. SCHREUDER AND T. G. GREGOIRE

agent. Responsiveness is established if the symptoms are reproduced by experimentally exposing the population to the suspected causal agent. Mechanism
establishes the cause-effect linkage by a step-by step approach, indicating clear
understanding of what occurs (Mosteller and Tukey, 1977).

Possible conclusion: This is the ideal situation rarely met in practice. It is a longterm process and usually requires large scale surveys such as FIA, preceded or
followed by careful experimentation (Smith and Sugden, 1988).
An example: We want to determine the maximum (Dmax ) and preferred distance
(Dopt ). Marbled Murrelets (MM) will fly inland to establish a nest and we want
to know for management purposes why these are the maximum and preferred
distance. Clearly this assumes that there is a maximum distance Dmax (very likely)
and a preferred distance (Dopt ) less likely) which may have to be estimated by
assuming a model or models containing such parameters. The murrelet is a small
sea bird that is listed as an endangered species. Considerable time and money is
being spent on recovery efforts to increase the population. An important part of the
recovery is knowing where exactly they nest and why.
MM nests are very difficult to find. A random selection approach to determine
their location is not feasible. This is clearly a situation where, using volunteers and
wildlife specialists in various federal and state agencies, a series of nests has to be
found and their distance from the coast measured. Assume that we find n nests that
are occupied or show evidence of having been occupied and their distances to the
ocean.
Since this is not a probabilistic sample, we can not fit a statistical distribution
such as the Sb distribution (Schreuder et al., 1993) to the data although we can use
similar mathematical models to estimate the maximum (Dmax ) and the preferred
distance (Dopt ). Clearly, our estimates of both parameters will be biased because
the sampling frame does not include all the target population (the population of all
MM that nest during the season sampled). For example, there may be nests at some
distances from the ocean that are not detected because of thorny or impenetrable
vegetation that deters observers but (perhaps even for the same reasons) may be
attractive to the MM or perhaps have been vacated due to predator kill. If such
nests tended to be very far from the ocean, their omission from the sample would
result in an underestimate of the maximum nesting distance.
Despite the limitations of the sample described above, it can still be useful to
help formulate hypotheses on cause-effect to explain the distances estimated. To
illustrate this, assume that we formulate three hypotheses about factors determining the maximum distance for nesting (other equally valid hypotheses can be
formulated, which is why establishing cause-effect relationships is so difficult),
namely:

THE USE OF PROBABILITY AND NON-PROBABILITY SAMPLING

289

1. The MM can only fly this maximum distance and still supply its brood with
food. This can best be established through experimentation.
2. The nest needs to be in a safe place, relatively protected from predators. Evidence of this, certainly initially, can only be established through observational
studies on all n observed breeding bird sites (both those occupied and those
showing evidence of having been occupied) to characterize the nest environments including freedom and protection from predators. This may also require
observation at apparently suitable nesting sites that are not used. Considerable
data would need to be collected and analyses done before it would be possible
to formulate a well-defined hypothesis such as a model testable by experimentation. Obviously, this hypothesis could be seriously affected by differences
between the target and sampled populations. There may be characteristics of
the non-sampled part of the target population that are essentially not observed
in the sampled part.
3. A combination of 1) and 2) above. This would most likely result in one or more
models including (Dmax ) and (Dopt ) to describe the behavior of the MM in
locating nesting sites in terms of distance from the sea. To study this hypothesis
properly is likely to require both experimentation and observational studies.
To establish whether these hypotheses are true requires considerable time and
money. The political will to support this work over the long period of time required
may not exist.

4. Recommendations
As discussed above, both purposive and design-based sampling have to play a role
in developing an understanding of the status and function of our ecosystems. Cost
and practicality dictates that. We suggest that the following recommendations be
kept in mind to develop such understanding.
1. Use design based sampling when the target and actual population sampled are
the same.
2. Consider purposive sampling when a decision is to be made quickly. The
method is generally a lot cheaper and offers more protection against small
sample sizes.
3. Use either design based or MB sampling for formulating hypotheses to be
tested. Purposive sampling is usually more efficient for that purpose but can
be more misleading.
4. Avoid any pretense that a purposive sample is a probabilistic, and hence representative, sample. Clearly state the assumptions made, what the sampled information can and cannot be used for, and draw inferences on that basis.

290

H. T. SCHREUDER AND T. G. GREGOIRE

5. Use design-based sampling to identify cause-effect and then establish causeeffect through experimentation or perhaps more efficiently use experimentation
to document cause-effect under controlled conditions and then attempt to infer
generality through design-based sampling.
With further advances in technology, it may be possible to solve the problems more
efficiently. If the MM and also its predators can be captured (at sea in the case of
the MM and wherever with predators) and fitted with miniature radio transmitters,
a larger sample of MM (and the predators too) or even a census from which a probabilistic sample could be drawn, could result. The sample could be probabilistic,
or at least more representative for estimating the maximum and optimal distances
for nesting. The information derived from the studies undertaken to establish the
hypotheses could be used to plan the radio-tagging studies better and would not be
wasted in waiting for the necessary technologies to become available.

Acknowledgements
We greatly appreciate reviews by Anthony R. Olsen and Geoffey B. Wood that
made this a considerably improved manuscript.

References
Birdsey, R. A. and Schreuder, H. T.: 1992, An overview of forest inventory and analysis estimation
procedures in the eastern U.S. with an emphasis on components of change. USDA Forest
Service RM Tech. Rep. RM-214, 11 p.
Brewer, K. R. W.: 1994, Survey Sampling Inference: some past perspectives and present prospects,
Pak. J. Stat. 10(1) A: 213233.
Brewer, K. R. W.: 1995, Combining Design-Based and Model-Based Inference, Ch. 30, in: B.
G. Cox, D. A. Binder, B. N. Chinnappa, A. Christianson, M. J. Colledge and P. S. Kott (eds),
Business Survey Methods, J. Wiley & Sons, Inc. N.Y., pp. 589606.
Cox, L. H. and Piegorsch, W. W.: 1996, Combining environmental information. I. Environmental
monitoring, measurement, and assessment. Environmetrics 7, 299308.
Dawid, A. P.: 1984, Inference, Statistical: I, in S. Kotz and N. L. Johnson (eds), Encyclopedia of
Statistical Sciences 4, pp. 89105.
Fisher, R. A.: 1922, On the mathematical foundations of theoretical statistics, Philos. Trans. Soc.
Lond. A 222, 309368.
Gadbury, G. L., Iyer, H. K., Schreuder, H. T. and Ueng, C. Y.: 1998, A nonparametric analysis of
plot basal area growth using tree based models. USDA For. Serv. RM For. And Range Exp. Sta.
RMRS-RP-2, 14 p.
Greenland, S.: 1990, Randomization, statistics, and causal Inference, Epidemiology 1, 421429.
Gregoire, T. G.: 1998, Design-based and model-based inference in survey sampling: Appreciating
the difference, Canadian J. For. Res 28, 14291447.
Hjek, J.: 1981, Sampling From A Finite Population, Marcel Dekker, Inc.

THE USE OF PROBABILITY AND NON-PROBABILITY SAMPLING

291

Hansen, M. H., Madow, W. G. and Tepping, B. J.: 1983, An evaluation of model-dependent and
probability-sampling inferences in sample surveys, Journal of American Statistical Association
78(384), 77693.
James, G. A. and Schreuder, H. T.: 1971, Estimating recreation use of the San Gorgonio Wilderness,
J. For. 68(8), 490493.
Koch, G. G. and Gillings, D. B.: 1984, Inference, Design Based vs Model Based, in S. Kotz and N.
L. Johnson (eds), Encyclopedia of Statistical Sciences 4, 8488.
Kruskal, W. and Mosteller, F.: 1979, Representative sampling, III: The Current statistical literature,
Intern. Stat. Rev. 47, 245265.
Max, T. A., Schreuder, H. T., Hazard, J. W., Teply, J. and Alegria, J.: 1997, The Region 6 Vegetation
inventory and monitoring system, USDA For. Serv. PNW Res. Paper PNW-RP-493, 22 p.
Meier, P.: 1986, Damned liars and expert witnesses, J. Amer. Stat. Assoc. 81, 269275.
Mosteller, F. and Tukey, J. W.: 1977, Data Analysis and Regression, Addison-Wesley, Reading, MA.
Olsen, A. R. and Schreuder, H. T.: 1997, Perspectives on large-scale resource surveys when causeeffect is a potential issue, Env. and Ecol. Stat. 4, 167180.
Overton, J. M. C., Young, T. C. and Overton, W. S.: 1993, Using Found data to augment a
probabilistic sample: procedure and case study, Env. Monitoring and Assessment 26, 6583.
Peifer, M.: 1997, Cancer-beta-catenin as oncogene: The smoking gun, Science 275, 17521753.
Piegorsch, W. W. and Cox, L. H.: 1996, Combining environmental information II: Environmental
epidemiology and toxicology, Environmetrics 7, 309324.
Royall, R. M.: 1970, On finite population sampling theory under certain linear regression models,
Biometrika 57, 377387.
Royall, R. M.: 1971, Linear Regression Models in Finite, Population Sampling Theory, in V. P.
Godambe and D. A. Godambe (eds), Foundations of Statistical Inference, Holt, Rinehart and
Winston, Toronto, pp. 259279.
Srndal, C.-E.: 1978, Design-based and model-based inference in survey sampling, Scandinavian J.
Stat. 5, 2752.
Srndal, C.-E., Swensson, B. and Wretman, J.: 1992, Model Assisted Survey Sampling, SpringerVerlag, NY, 694 p.
Schreuder, H. T. and Alegria, J.: 1995, Stratification and plot selection rules, misuses, and consequences. USDA Forest Service RM Forest and Range Exp. Sta. Res Note RM-RN-536,
4 p.
Schreuder, H. T. and Reich, R.: Data Estimation and Prediction for Natural Resources Public Data
Sets, USDA For. Serv. RM Res Note (in press).
Schreuder, H. T. and Thomas, C. E.: 1991, Establishing cause-effect relationships using forest survey
data, For Sci 37, 14971525.
Schreuder, H. T. and Williams, M.S.: 1997, Weighted linear regression for tree volume using D2H
and D2 as the dependent variables, USDA For. Serv. RM For and Range Exp Sta Res Paper
RMRS-RP-6, 11 p.
Schreuder, H. T., Gregoire, T. G. and Wood, G. B.: 1993, Sampling Methods for Multiresource
Inventory, J. Wiley & Sons, Inc., N.Y., 446 p.
Schreuder, H. T., Li, H. G. and Wood, G. B.: 1990, Model-dependent and design-dependent sampling
procedures A simulation study, USDA For. Serv RM For and Range Exp Sta Res Paper RM291, 19 p.
Schreuder, H. T., Clerke, W. H., Barry, P., and Holland, D.: 1980, Two stage stratified sampling with
regression to assess southern pine beetle damage, USDA For. Serv. SE For. Exp. Sta. Res. Paper
SE-212, 6 p.
Smith, T. M. F.: 1994: Sample Surveys: 19751990; An age of Reconciliation? Intern. Stat. Rev.
62, 534.
Smith, T. M. F. and Sugden, R. A.: 1988, Sampling and assignment mechanisms in experiments,
surveys, and observational studies, Intern. Stat. Rev. 56, 165180.