Professional Documents
Culture Documents
Random
Random
Random sampling refers to a variety of selection techniques in which sample members are
selected by chance, but with a known probability of selection. Most social science, business,
and agricultural surveys rely on random sampling techniques for the selection of survey
participants or sample units, where the sample units may be persons, establishments, land
points, or other units for analysis. Random sampling is a critical element to the overall survey
research design.
This entry first addresses some terminological considerations. Second, it discusses two main
components of random sampling: randomness and known probabilities of selection. Third, it
briefly describes specific types of random samples, including simple random sampling (with
and without replacement), systematic sampling, and stratification, with mention of other
complex designs. The final section touches on inference, which is the reason that random
sampling is preferred in scientific surveys.
Terminological Considerations
Some authors, such as William G. Cochran, use the term random sampling to refer specifically
to simple random sampling. Other texts use the term random sampling to describe the broader
class of probability sampling. For this reason, authors such as Leslie Kish generally avoid the
term random sampling. In this entry, random sampling is used in the latter context, referring to
the broader class of probability sampling.
Critical Elements
The two critical elements of random sampling are randomness and known probabilities of
selection.
Randomness
The first critical element in random sampling is the element of randomness. Ideally, all
members in the survey's target population have a non-zero chance of selection.
In describing what random sampling is, it is helpful to highlight what it is not. The sample is not
pre-determined. Nor is a random sample selected by expert judgment. Random sampling does
not imply that the sampling is haphazard. Furthermore, random sampling is not convenience
sampling, in which the interviewers take respondents that are easiest to obtain.
The element of randomness is applied to the process scientifically. That is, there is a method,
usually mechanical, to the selection process that is rigorously followed. The precise method
may rely on random number generators or tables of random numbers. By following the
scientific process, the probabilities of selection are known and preserved.
Random number generators and tables of random numbers are not truly random, of course,
but the process needs to be random enough. This is especially important in litigious contexts.
Bruce D. McCuUough and Wendy Rotz have tested the random number generators available in
various data tools and statistical packages.
The probabilities of selection are important for the theory that enables researchers to estimate
sampling error. Because a sample is a subset of the target population and not a census
(complete enumeration), estimates derived from sample responses will rarely match the target
population values exactly. The variable difference between the sample estimate and the
population value is sampling error. (Nonsampling errors, such as inaccurate frames of the
target population and imprecise measures of the questionnaire items, affect both surveys and
censuses. Nonsampling errors are not covered in this entry.)
Having a randomly selected sample with known probabilities of selection enables the
researcher to estimate the sampling error. That is, the researcher can use the sample to make
inferences for the target population and to estimate the precision of the sample-based
estimates.
The probabilities of selection may enter into the estimation process in another way, as well.
Because the probabilities of selection for members of a random sample are known, the sample
responses can be appropriately weighted (if the probabilities are different, as in complex
sampling designs) to yield improved estimates for the target population. The weights are a
function of the probabilities of selection, which are not known precisely for purposive,
convenience, or other nonprobability samples.
Sample Designs
The most familiar type of random sampling is simple random sampling. Simple random
sampling may be with or without replacement. The drawing of names out of a hat and the
selection of official lotto numbers are examples of simple random sampling without
replacement. In simple random sampling, all members of the population have the same
probability of selection.
Simple random sampling is said to be with replacement if each selected sample member is
replaced into the available population pool for subsequent draws. In practice, sampling with
replacement is not as common as sampling without replacement.
An easy way of selecting a simple random sample of size n without replacement is to use a
random number generator to assign a random number to each member of the population in the
frame or population list, sort the frame by the random number, and select the first n units in the
randomly ordered list.
Systematic Sampling
Another random sampling design is systematic sampling, in which the population is ordered,
and every kth unit is selected. Once a random starting point is selected, the rest of the sample
is determined; however, the randomness is in the selection of the starting point. In other words,
a single sample is randomly selected from a set of k possible samples. (If k is not an integer,
more samples are possible.)
Complex Designs
In practice, simple random sampling and systematic sampling are rarely used alone for large
surveys, but are often used in combination with other design features that make for a more
complex design. Complex random sampling designs tend to have smaller sampling error or
lower cost, or both. The complex random sampling designs may incorporate elements that
resemble purposive sampling, such as stratification. Double sampling, cluster sampling, multi-
stage sampling, and sampling with probability proportional to size are other examples of
complex probability or random sampling.
involves dividing the target population into groupings, or strata, of relatively homogeneous
members and selecting a random sample independently within each stratum. For example, a
list of individuals may be divided into strata based on age and gender, or the population of
businesses may be divided into strata based on geography, size, and industry. The variables for
stratification need to be available for all population members and presumably related to the
survey response variables or the propensity to respond. Because members within a stratum
tend to be more alike, the emphasis is on selecting a broad sample across strata rather than
extensively within each stratum. By selecting members from each stratum, the sample will
capture much of the diversity in the population more efficiently than if the sample were selected
purely randomly. Stratification also enables the researcher to sample from the strata at different
rates if, for example, the researcher wants estimates for individual strata as well as for the
population as a whole.
Inference
The survey design includes both the sampling technique and the corresponding estimators for
inferences. With scientific random sampling and known probabilities, mathematical formulas
exist for making inferences about the target population and for estimating the sampling error
attributed to the inferences. Confidence intervals, tests of hypotheses, and other statistics are
possible with random sampling and estimates of sampling error. While an expert may
judiciously select a sample that is a good representation of the target population on some
measure, a purposive sample of this sort cannot, by itself, be used to estimate the precision of
the sample-based estimates because no such mathematical formulas are possible. Neither can
the sampling error be estimated from a convenience sample.
Under simple random sampling, the distribution of the sample mean often approximates the
normal distribution, where the variance decreases with sample size. That is, the sample mean
is a good estimator for the population mean, and the error associated with the sample-based
estimate is smaller for larger samples. This result is based on the central limit theorem.
Alternatively, in some circumstances, especially when sample sizes are small, the distribution of
sample statistics may approximate Poisson, hyper-geometric, or other distributions. The
approximate distribution is what enables the researcher to make inferences about the
population based on sample estimates.
Complex probability designs may have more complex mathematical forms for statistical
inference and may require specialized software to properly handle the estimation. Some
methods of variance estimation developed for complex designs may be less stable in some
circumstances, an issue to be aware of as the field moves toward smaller samples and
estimation using subsets of the sample. Nevertheless, inference is possible because complex
designs share the underlying concepts and theory of random selection with known
probabilities.
RachelHarter
http://dx.doi.org/10.4135/9781412963947.n440
See also
Cluster Sample
Cochran, W. G.
Complex Sample Surveys
Convenience Sampling
Kish, Leslie
Multistage Sampling
Nonprobability Sample
Nonsampling Error
Probability of Selection
Probability Proportional to Size (PPS) Sampling
Probability Sample
Purposive Sampling
Random
Sampling Error
Sampling Without Replacement
Simple Random Sample
Strata
Stratified Sampling
Systematic Sampling
Further Readings
Liu, Y., Batcher, M., & Rotz, W.(2001, August). Application of the hypergeometric distribution
in a special case of rare events. Proceedings of the Section on Survey Research Methods (pp.
941–954). Alexandria, VA: American Statistical Association.
Marsaglia, G.(1995). The Marsaglia Random Number CDROM including the Diehard Battery of
Tests of Randomness. Retrieved May 1, 2008, from http://w w w. stat. fsu. edu/pub/diehard/
National Institute of Standards and Technology.(1980). NMS Minimal BASIC Test Program for
random generators (Special Publication No. 500–70/1 and No. 500–70/2). Gaithersburg, MD:
Author.
National Institute of Standards and Technology.(2001). A statistical test suite for random and
pseudorandom number generators for cryptographic applications (Special Publication No. 800–
22). Gaithersburg, MD: Author.
Rotz, W., Falk, E., & Joshee, A.(2004, August). A comparison of random number generators
used in business—2004 update. Proceedings of the Section on Business and Economic
Statistics (pp. 1316–1319). Alexandria, VA: American Statistical Association.
Rotz, W., Falk, E., Wood, D., & Mulrow, J.(2001, August). A comparison of random number
generators used in business. Proceedings of the Annual Meeting of the American Statistical
Association, Section on Computational Statistics. Alexandria, VA: American Statistical
Association.