The subjects selected must be appropriate to the hypothesis being tested. For example, you may wish to know how well diabetic patients adhere to the dietary advice they have been given. Those patients who attend a diabetic clinic provide an accessible group of patients to study, but the findings will not necessarily apply to other diabetics, since those who attend a clinic are a select group who probably differ from those who do not. Therefore, you need to decide in advance which group you are interested in and choose your population accordingly. If you are specifically interested in a certain subgroup, you need to ensure that it is adequately represented among those actually seen. This can be done by defining the population more narrowly, by enlarging its size, or by some selective sampling procedure, otherwise the subgroup may be too diluted for an effect within it to be detected in the whole population, and too small to be studied separately. On the other hand, if you want the results to be helpful to say a general practitioner who sees a wide range of cases, the population should encompass a similar range. Otherwise the results, however interesting to you, will seem largely irrelevant to the practitioner.

If you wish to study adults from the general population, the electoral roll is a useful sampling frame for a defined geographical area.

When selecting subjects for a research study it is essential to avoid bias due to selection. The results of a study can only be extrapolated to a larger group of subjects with confidence if the group studied can reasonably be assumed to be representative of the larger group. In order to ensure that subjects are representative of the population from which they are chosen, a random sample should be selected. Random is not the same as haphazard. Random sampling of subjects can best be achieved by use of a table of random numbers. For example, if a sample of say 1 in 2 is to be taken, each potential subject is allocated the next number from the table and included in the study sample if the number is even and excluded if the number is odd. If a 1 in 3 sample is required, subjects are allocated consecutive numbers from the table, ignoring zero, and those with numbers ending in say 3, 6, & 9 are selected for the study.

Sometimes a more complex procedure than straight random sampling is required, e.g. if you wish to compare subgroups of patients. Within a random sample, it is possible you might obtain too few subjects of one type. You might therefore decide to obtain a balanced sample, with equal numbers in each group. The way this is often achieved in a study based on patients attending an out-patient clinic is to take say, the first 20 patients within each group who present at the clinic. This is hazardous, however, as the most numerous group will be recruited long before the least numerous group. During the extended time period required to recruit the latter group, seasonal effects, drift in the laboratory measurements, changes in the methods of measurement etc. may all occur. It is therefore advisable for sampling ratios for each subgroup to be chosen based on the likely frequency of patients in the various subgroups. For example, you might decide to recruit a random sample of 1 in 10 for the most numerous subgroup and recruit all of those in the least numerous sample of 1 in 10 for the most numerous subgroup and recruit all of those in the least numerous subgroup. In this way, the required number for each subgroup will be obtained in roughly the same period of time. This method is known as stratified sampling, becausethe different groups (or strata) are sampled separately.

In studies covering a wide geographical area, the efficiency of resource utilisation can be maximised by multi-stage random sampling. For example, if you wish to study a random sample of schoolchildren, you could define a number of levels, and then draw a random sample within each level. The levels might be defined as Local Education Authorities (LEA), schools within each LEA, classes within schools and children within classes. This method of sampling enhances efficiency as the children to be studied are clustered in a relatively small number of sites.

Response rate is very important. If the response rate is low, those seen cannot be assumed to be representative of the total group. Those who refuse to take part in a research study are likely to differ in many ways from those who agree to take part. A low response rate is therefore a source of bias and makes it impossible to draw valid conclusions from the study.