Selecting a Sampling Design
If you are... | consider using ... |
---|---|
in an emergency situation. . . | judgmental sampling |
a screening situation. . . | judgmental sampling for small-scale problems with small budgets and limited schedule |
systematic/grid sampling | |
simple random sampling | |
searching for rare characteristics (including hot spots). . . | adaptive cluster sampling |
systematic/grid sampling | |
identifying areas of contamination. . . | adaptive cluster sampling |
stratified sampling ^{A, B} | |
systematic/grid sampling or simple random sampling when no professional knowledge or prior information exists ^{A, B} | |
estimating the prevalence of a rare trait. . . | simple random sampling ^{A, B} |
stratified sampling ^{A, B} | |
estimating/testing an area/process mean
or proportion. . . |
simple random sampling or systematic/grid sampling when no professional knowledge or prior information exists ^{A, B} |
ranked set sampling (for means only) | |
stratified sampling ^{A, B} | |
comparing parameters of two areas/processes.
. . |
simple random sampling or systematic/grid sampling ^{A, B} |
ranked set sampling | |
stratified sampling ^{A, B} | |
^{A}Consider using compositing
in conjunction with this design if analytical costs are much higher
than sampling costs and samples can be homogenized.
^{B}Use compositing only when interested in means. |
Simple Random Sampling - A random number generator (or equivalent process) is used to select all sampling locations.
Can be used for any objective - estimating/testing means, proportions, etc., comparing means, proportions, etc., of two or more areas/processes, delineating boundaries, etc., but is one of the least efficient (though easiest) designs since it doesn't use any prior information or professional knowledge. It is primarily used in conjunction with other sampling designs, as the last stage of sampling in multi-stage projects (i.e., a sample of units is selected at the first stage and then subunits are selected from each unit), and for assigning units in experimental (e.g., intra-laboratory studies). Use when:
- the area/process to sample is relatively homogeneous (i.e., no major patterns of contamination or "hot spots" expected) and there is no prior information or professional knowledge available;
- there is little to no prior information or professional judgment available;
- there is a need to protect against any type of selection bias (for example, when any professional judgment used to define 'areas' may be challenged); or
- it is not possible to do more than the simplest computations on the resulting data.
Stratified Random Sampling - Prior information about the area/process is used to create groups that are sampled independently using a random process. These groups can be based on spatial or temporal proximity, or on preexisting information or professional judgment.
Can be used for any objective - estimating means, proportions, etc., delineating boundaries, etc. Use when:
- the area/process can be divided based on prior knowledge, professional judgment, or using a surrogate that is highly correlated with the item of interest;
- the target area/process is heterogeneous;
- you need to ensure representativeness by distributing the samples throughout the spatial and/or temporal dimensions of the area/process;
- you need to ensure that rare groups (e.g., shrimps clustering in large but scattered schools, unevenly distributed contamination, rare and endangered species) of the area/process are sampled sufficiently (i.e., you take enough samples to draw conclusions about these groups);
- costs and/or methods of sampling differ within the area/process, or
- you need information about the entire area/process and specific subgroups.
Systematic and Grid Sampling - A random number generator (or equivalent process) is used to select an initial sampling point (either spatial or temporal) and the remaining points are based on a specific pattern (weekly, rectangular, square, triangular, etc.)
Can be used for any objective - estimating means/testing, proportions, etc.; delineating boundaries; finding hot spots; and estimating spatial or temporal patterns or correlations. It is primarily used for pilot studies, scoping studies, and exploratory studies. Use when:
- the item of interest could not possibly be aligned with the sampling pattern (e.g., don't take air samples every Monday morning if a nearby plant always pressure-cleans the duct work on Monday morning);
- little to no prior information is available;
- regular spacing makes it easy for field teams to locate the sampling points; or
- when uniform coverage of an area/process is necessary.
Ranked Set Sampling - Screening (inexpensive) measurements are used on an initial random sample. The results are ranked into groups based on relative magnitude (high, medium, low), then one location from each group is sampled.
Used primarily for estimating/testing means or comparing two means. Use when:
- inexpensive measurement techniques are available;
- professional judgment, on-site measurements, or an inexpensive auxiliary variable is available to rank samples based on magnitude; and
- the cost of locating samples and using screening measurements or professional judgment to rank these samples is low compared to laboratory measurements.
Adaptive Cluster Sampling - Take random samples. If result shows characteristic of interest (i.e., "hit"), take additional samples adjacent to the original.
Can be used for estimating or searching for rare characteristics, delineating hot spots, estimating means, and determining extent of contamination. Use when:
- inexpensive, rapid measurements techniques, or quick turnaround of analytical results are available (for example, field measurement technologies);
- the item of interest is sparsely distributed but highly aggregated (e.g., shrimps clustering in large but scattered schools, unevenly distributed contamination, rare and endangered species); or
- you wish to concentrate resources in areas of greater interest (for example, take samples close to any "hits" to determine extent).
Composite Sampling - First, another sampling design is used to select sample locations. Then composite samples are created by physically combining and homogenizing these samples based on a fixed compositing scheme.
Can be used to estimate/test means, compare two or more means, estimate the prevalence of a trait (or the proportion of an area/process that has a particular trait), or to identify samples with a specific trait. Use
- when analysis costs are large relative to sampling costs;
- the individual samples are similar enough to homogenize, the mixing process (weighing, homogenizing, etc) will not create large errors, and there are no safety hazards or potential biases (i.e., loss of volatile organic components) resulting from compositing process;
- you wish to increase sample coverage of the area/process without increasing laboratory analysis costs;
- information on individual samples, variance, and any potential associations (for example, correlations between concentration levels of two contaminants) is not important;
- concentrations of relevance are much larger than detection limits; or
- when individual samples may not have an adequate mass for analysis (for example, for dust samples or tissue samples).