Sampling
Needless to say the most accurate information about the incidence of adverse outcomes in pregnancy caused by working with VDTs would be gained if we could gather data from all women who had ever worked in front of these terminals at some point during their pregnancy. Just as obviously, however, this would be impractical; there may be hundreds of thousands of such women scattered over most of the globe. Practical considerations dictate that we could follow up only a small proportion of these women, and if we select them appropriately, our estimates won’t be off too far. (However, the famous prediction in 1936 that Alf Landon would decisively beat Franklin Roosevelt must serve as a constant reminder that “appropriately“ isn’t all that easy to define—much to Roosevelt’s relief.) In this section we discuss various ways in which we could go about choosing the group or groups we will include in our study.
Basic Terminology PopulationAll of the people to whom the results should be applicable constitute the populations. In this example the population would consist of all women who worked at VDTs at some time during their pregnancy ( Fig.1218). (Note that “population” does not refer to all the people in the world, just to those who have a specific disorder, were exposed to some agent, or underwent some procedure.)
SampleIn most cases the population is large, and it is impractical to study all people. We limit our study to a subset of the population; this smaller group is called the sample ( Fig.1221).
Figure 1218 – (Figure 3-3) Example of a population
Streiner DL, Norman GR. PDQ Epidemiology-Second Edition, 1996, BC Decker Inc., Hamilton, Ontario.
Some figures may not display clearly when rendered as a PDF or printed.
Figure 1221 – (Figure 3-4) The sample is a subset of the population
Streiner DL, Norman GR. PDQ Epidemiology-Second Edition, 1996, BC Decker Inc., Hamilton, Ontario.
Some figures may not display clearly when rendered as a PDF or printed.
CohortOriginally, cohort referred to a group of people born in the same year. Nowadays it has the broader, if less precise, meaning of a group of people who share some attribute. For instance, all people who began working at a specific job within a given time period can be referred to as a cohort, as can all people who entered the study at a certain time.
Probability SamplingProbability sampling refers to a number of different strategies used to choose a sample. The term comes from the procedure used; every person in the population has a fixed and known probability of being selected to be part of the sample. For a number of reasons, most studies try to use one or more of these strategies if at all possible.
The primary reason is that this method allows the investigator to generalize the results from the sample to the population, which is usually the major reason for doing a study. Second, it can tell the researcher the margin of error that could be expected from these estimates, that is, how far off the estimates can be. We see this in the reporting of polls, which often have a line stating that the results are accurate to within plus or minus 4 percent. In a related vein, most statistical tests are based on the assumption of some sort of random sampling. When probability sampling is not used, we shouldn’t use these tests (although that has never stopped people from doing so), and the ability to generalize the results from the sample to the population is questionable. (This is in contrast to the view of one politician who trusted letters he received more than polls and complained that the latter were “only’’ random.)
Random SamplingIn random sampling (sometimes called “strictly random sampling” to differentiate it from the other varieties), each subject in the population has an equal chance of being chosen for the study. As we’ve mentioned, this approach maximizes the likelihood that the results of the study can be generalized to the entire population.
Random sampling is most often used in survey research ( Fig.1224). Nearly all towns and cities have lists of taxpayers (for obvious reasons) or of street and house addresses. This makes it relatively simple for the researcher to select people, or at least dwellings, at random. These days approximately 98 percent of people have telephones so it is also easy to draw a random sample from municipal or telephone lists or from dialing digits at random.
Once we move out of the realm of surveys of the general population, however, it often becomes impossible to draw a pure random sample. We would have to how, for example, every company that used VDTs and all of the pregnant women at each business who had ever worked with VDTs to select people randomly for the study.
Figure 1224 – (Figure 3-5) Random sampling
Streiner DL, Norman GR. PDQ Epidemiology-Second Edition, 1996, BC Decker Inc., Hamilton, Ontario.
Some figures may not display clearly when rendered as a PDF or printed.
More often we choose one or a number of businesses and hope that the use of VDTs within them is representative of companies in general and that the women who work there are representative of female workers in other companies. We would then randomly select people within those companies for our study.
The same situation exists even for experimentally based studies. The hospital where a new treatment is tried out is not really chosen at random; it is most likely selected on the basis of convenience (e.g., the investigator works there or knows someone there who owes him or her a favor). The assumption is made that it is representative of hospitals in general and that the randomly selected patients from that hospital are representative of the general population of patients with that condition. Unfortunately, these assumptions are not always correct and result in many of the various types of selection biases, which we discuss starting on p. 48 in Threats to Validity.
Stratified Random SamplingThere are some circumstances in which we may wish to deviate from strictly random sampling. One major reason is that with random sampling, we may end up with too few people in one subgroup or another. For instance, if we thought that the teratogenicity of VDTs was related to the number of previous pregnancies, random sampling might result in few women who had three or more children before working on the terminals; the sample would be too small to allow us to analyze the effects of parity. Similarly, we may want to have equal numbers of women in each age category to maximize the power of our statistical tests.
Conversely, we may want to ensure that our sample is equivalent to the general population in terms of a few key variables, such as age at first pregnancy or number of children (it’s obviously not necessary to match for sex). Random sampling ensures this matching in the long run with large enough samples hut not necessarily in our particular study, especially if there are fewer than 1000 subjects. By chance, we could oversample or undersample people from a particular age or parity group.
To achieve these goals, we divide the key variables into various levels, or strata. For instance, we can divide age into 10-year increments or parity into one kid, two kids, and three or more ( Fig.1226). Then subjects are selected randomly from the stratum into which they fall. If toward the end of the study we have enough women who have had one or two children but not three or more, we would restrict entry into the study to only this latter group. Because we know how our strata deviate from a strictly random sample, we can correct for this during the analyses when we derive the prevalence figures.
Cluster Sampling
In some designs, it is impractical to assign individual subjects to the various groups. For example, in the Burlington Randomized Trial, nurse practitioners were placed in the offices of some family physicians to see whether they could reduce the cost of primary care without adversely affecting its quality. Outcome was measured at the level of the individual patient. However, because most families tend to use the same family physician, it would have been infeasible to allocate random members of the same family to different practices. In this case, each family was considered to be a cluster, and the unit of randomization was the family rather than the individual ( Fig.1227).
Figure 1226 – (Figure 3-6) Stratified random sampling
Streiner DL, Norman GR. PDQ Epidemiology-Second Edition, 1996, BC Decker Inc., Hamilton, Ontario.
Some figures may not display clearly when rendered as a PDF or printed.
Figure 1227 – (Figure 3-7) Cluster sampling
Streiner DL, Norman GR. PDQ Epidemiology-Second Edition, 1996, BC Decker Inc., Hamilton, Ontario.
Some figures may not display clearly when rendered as a PDF or printed.
However, the two, three, or more people in the same household cannot he considered to be independent of one another in terms of health status; they share the same diet, environment, and likely have similar attitudes toward exercise or other behaviors that affect health. Consequently the husband’s health is probably more correlated with his wife’s than it is with that of another randomly chosen person.
Because the outcomes are correlated to some degree across people (who are usually considered to be independent in the usual statistical tests), studies that use cluster sampling usually need larger sample sizes than investigations in which the subjects are truly independent. How much larger the sample size has to be depends on the average number of people in the cluster and on how strongly the variables are correlated within members of the cluster.
Haphazard SamplingIn a haphazard sample, which is also called a “sample of convenience,” subjects are selected on the basis of their availability or in any other nonrandom way. For example, a researcher can interview people who pass a certain street corner or take blood samples from the research assistants who work in his or her laboratory. There is always the real danger that this is a biased, nonrepresentative sample. During the day, housewives, shift workers, or the unemployed are more likely to be walking around outside than are people who work 9 to 5, and the location of the specific corner may differentially favor people from one social class over another. (On Wall Street in New York and Bay Street in Toronto, you were more likely to find yuppies in 1989 and the unemployed in 1991.) Similarly, those working in a laboratory may be healthier, brighter, or disproportionately female compared with the population of interest.
Unfortunately, newscasters rely on just this sort of haphazard “person in the street” interview to find out (often erroneously) what the people “really“ think about some issue. Politicians who rely on letters they receive fall prey to the same trap; those who are concerned enough to write are not representative of the electorate in general. Lest we as researchers develop undue pride about our avoidance of such egregious errors as are committed by those who are untrained in the strict disciplines of science, two examples may suffice to remind us of our fallibility. Mueller and his colleagues developed a test for plasma unesterified fatty acid to be used for patients with neoplastic disease. Their 30 normal subjects were “members of the professional staff...or hospitalized normal volunteers.” The sampling for this test may have been a marked improvement over another test, which studied hemolysate prothrombin consumption time; the authors gave no indication at all regarding how many normal blood samples were used, much less from where they came. To assume that these samples were randomly selected, and hence representative of normal people, requires a leap of faith that we at least cannot make.
Content on this page was last changed on March 19, 2009.
© 2002 BC Decker Inc. Show Disclaimer
| 5476. | Streiner DL, Norman GR. PDQ Epidemiology. 2nd ed. Hamilton, Ontario: BC Decker Inc.; 1996. |
Next Page: Subject Allocation »