Sampling procedures
From InterSciWiki
Contents |
[edit] Probability sampling
If you are series about sampling, and want a complex sampling design, get the jolly green giant bible of modern sampling theory and survey design, 1965, Leslie Kish, Survey Sampling. We used this reference in designing our Irish National Sample Surveys in 1970-1972. The probabilistic designs in Kish (1965) include cluster sampling o reduce costs of surveys by sampling elements that closer together in the regions sample, and stratifiedl sampling which increases the sampling rate for rare subpopulations or those where greater accuracy is needed. WikiSysopWikiSysop 13:33, 27 February 2008 (PST)
If you have an exhaustive list of houses or people to sample in a delimited area, the advantage of some form of probability sampling from that list is that you get valid standard errors on your estimates of means and other statistics.
[edit] Opportinity sampling
Random sampling is not much good if you have only an opportunity subsample as in cross-cultural sampling, where you can pick societies randomly or on a geographic basis, but only a few cases will have ethnographic descriptions. The HRAF "probability samples" are a farce, a mere pseudoscientific veneer designed to convince unwary researchers to use or buy the HRAF collection of ethnographic texts.
Representative samples of this sort, like interviewing people off the street because the make themselves available, are not useless, because they do allow distributions to be measured and analyzed, but the conventional statistics do not apply, especially when it comes to probability theory estimates of variances (thus, standard errors and error bars). They are certainly useful for exploratory data analysis.
A more serious problem is that of sampling non-independent cases, which occurs wherever there are larger scale influences that produce greater similarities among cases that are (1) closer in space (2) in contact or communication (3) exposed to similar influences (nationality, religion, etc) (4) have common historical origin (5) share habitat (6) genetically closer, etcetera. All of these are instances of network autocorrelation which might also, for some networks, produce greater dissimilarity or differentiation.
It may be more important to include network measures for possible autocorrelation effects than to randomly sample, but the combination of probability sampling and autocorrelation tests provides the strongest research design when accurate and valid estimates are sought in later stages of researching a major problem.
[edit] Spatial sampling
[edit] SRS or Simple random sampling
Assuming you have a 'complete inventory to sample from a geographically limited or easily reachable population (the sampling universe), SRS is the best procedure to use because no corrections need be made as in the more complex sampling designs (cluster, proportional, or both together). Get your list, number them sequentially, 1 to N, pick a subset of a given size K, and use random numbers that are equiprobable between 1 and N to choose your cases. These numbers can easily be generated by computer.
[edit] Nonresponse bias
If you have lots of nonresponses you need to estimate the nonresponse bias. See survey nonresponse bias, Sampling Issues: Nonresponse, and other sources. Increasing sample size does not help.
Surveying nonrespondents is useful if this is a serious problem.
[edit] References
Bernard, HR. 2006. Ch. 6 Sampling and Ch. 7 Sampling Strategy. in Research Methods in Anthropology. 4th Edn. Lanham: Altamira Press.
Johnson, JC. 1998. Research design and research strategies, pp. 131-172 in Handbook of Methods in Cultural Anthropology. HR Bernard, Ed. Walnut Creek, CA: Altamira.
