Essential Statistical Concepts, Definitions, and Terminology - Sampling Spatial Units for Agricultural Surveys - page 2

include population, housing, industry, service, and agricultural censuses. The main
disadvantages of censuses are the high cost, the time consuming process, and the
significant non-sampling error caused by the management of a large number of
population units.
Some margin of error is generally acceptable for practical purposes, so a sample
survey is an effective alternative to a complete enumeration survey. In a sample
survey, some units are selected in a suitable manner. Observations from these units
are used to infer information about the entire population.
A sample selected using appropriate criteria can give information that can be
extended to the population from which it came, with a degree of uncertainty
determined using probabilistic methods. A sample survey can even lead to more
reliable results than a census survey. In fact, in census surveys, data collection may
take place under very different conditions if the population is very large. This can
result in major measurement errors (see Sect. 9.5 ).
The sample should be representative of the population under investigation.
Improperly conducted sample surveys may only collect data from particular groups
of the population. Typically, these groups are chosen because their information is
easily obtained.
The main sampling issues are: how best to select the sample, and how best to
extrapolate the information to estimate the variables of the whole population. Other
important aspects to be considered are the sample size and particular recorded
measurements.
There are many approaches to sampling. The main practical distinction is
between sampling from finite and infinite populations. In other words, we distin-
guish between the two methods by considering if we are able to enumerate all the
statistical units representing the target population with codes from 1 to N . In this
topic, for the sake of simplicity, we only consider finite populations of size N (see
Chap. 6 for details).
We can also distinguish between sampling approaches using stochastic assump-
tions of the observed data. In particular, we can suppose that the data are affected by
a random measurement error, which implies that a random variable should be used
to model each variable of interest for each unit. In this case, we often refer to the
superpopulation approach.
Another consideration when defining a sampling strategy is the efficiency
concept: the design-based , model-assisted , and model-based approaches. The
design-based approach has its origin in Neyman

s seminal paper (Neyman 1934 ).
It represents the basic underlying philosophy of most traditional sampling theory
texts (Cochran 1977 ; Kish 1965 ; Lohr 2010 ). Under this framework, the properties
of an estimator are measured by averaging over all the possible samples, s (see
Chap. 6 ). The model-assisted approach is also based on this same feature, but it
explicitly makes use of models that link the survey variable y to the covariates
X (see Chap. 10 ) . The model-based approach uses these methods, and also evalu-
ates the performance of an estimator using its expected behavior over the possible
realizations of the distribution of the observed data, conditioned to the selected
sample s (see Chaps. 11 and 12 ) .

'