Sample Selection: Case-control and Cohort Studies

Sample Selection

It is always desirable for the sample in a study to be representative of the population of interest, but this is not as important in experiments as in observational studies. The sample should be chosn to be as similar as possible to the relevant population, so it is essential to be able to describe just how the sample was chosen. Another way of combating variablity is to increase the sample size. Larger samples enable us to evaluate effects of interest more precisely. In a designed experiment there may be several conditions, called factors, being controlled by the investigator. The distinction between within subject and between subject comparisons is important. It is not possible to say what the best design is in any given circumstance. The choice of factors to control, which factors are between subject and which within, and how many observations to take for each subject is difficult, and it will often take much thought to arrive at a satisfactory design. Expert statistical help is particularly valuable at this stage. Any weaknesses in the design cannot be rectified later.

Random Allocation is used to prevent bias where we want to compare treatments between groups which do not differ in any systematic way. While simple randomization removes bias from the allocation procedure, it does not guarantee, eg., that hte subjects in each group have similar age distributions. indeed in small studies it is highly likely that some chance imbalance will occur, which might complicate the interpretation of result. Even in studies with over 100 subjects there may be some substantial variations by chance, especially for characteristics that are quite rare. We can use stratified randomization to achieve approximate balance of important characteristics (passive smoker) without sacrificing the advantages of randomisation. The method is to produce a separate block randomisation list for each subgroup (stratum). It is essential that stratified treatment allocation is based on block randomisation within each stratum rather than simple randomisation, otherwise there will be no control of balance of treatments within strata, and so the object of stratification will be defeated. Stratified randomisation can be extended to two or more stratifying variables. In small studies it is not practical to stratify on more than one or pehaps two variables, as the number of strata can quickly approach the number of subjects. In some studies it is either impossible or impractical to allocate treatments to individual subjects. Suppose that we wish to evaluate the effectiveness of a health education campaign in the newspapers to increase awareness of the dangers of drugs, or indeed to change behaviour. We can target individuals at random, but rather we can randomly assign whole areas to receive different media coverage. with a large number of small areas this cluster randomisation should give reliable results, but with a small number of very large areas, there are problems in ensuring the comparability of the areas. Here, it is valuable to obtain baseline data before the study starts so that changes within areas over the time of the study can be compared. Other clusters sometimes used in experimental research are schools, hospitals and families.

Observational Studies


Many studies are carried out to investigate possible associations between various factors and the development of a particular disease or condition. There is no logical difference between comparing the outcome of two groups of patients given alternative treatments and comparing the outcome of groups receiving different exposures. In general however, areas of epidemiological research are not amenable to being investigated by randomized trials. We cannot randomise individuals to smoke or not to smoke nor to work in particular jobs, and other factors such as age and race are not controllable by the individual. We must use observational studies, therefore, to study factors or exposures which cannot be controlled by the investigators. Nevertheless, the goal of an observational study should be to arrive at the same conclusions that would have been obtained by an experimental trial.
There are two main types of observational study that are used to investigate the causal factors - the case-control study and the cohort study. In a retrospective case-control study a number of subjects with the disease in question (the cases) are identified along with some unaffected subjects (controls). The past history of these groups in relation to the exposures of interest is then compared. In contrast, in a prospective cohort study a group of subjects is identified and followed prospectively, perhaps for many years, and their subsequent medical history recorded. The cohort may be subdivided at the outset into groups with different characteristics, or the study may be used to investigate which subjects go on to develop a particular disease.

In the case control study we identify a group of subjects (cases) with the disease or condition of interest, say lung cancer, and an unaffected group (controls), and compare their past exposure to one or more factors of interest, such as consumption of carrots. If the cases report greater exposure than the controls we may infer that exposure is causally related to the disease of interest, for example that consumption of carrots affects the risk of developing lung cancer.

Disadvantages of case control studies:
- They are inefficient for the evaluation of rare exposures: it can be time consuming and expensive to find people with the relevant condition.
- We cannot compute incidence rates of disease in exposed/non-exposed individuals unless the study is population based, because we are deliberately selecting a subset of people with the condition of interest rather than studying the population at large.
- The temporal relationship between exposure and disease may be difficult to establish. This is a particular problem in retrospective studies when people are asked to recall periods of time and sequences of events.
- They may be prone to bias arising from difficulties in ensuring that cases and controls are similar.

Advantages of case control studies
Despite their weaknesses, case control designs are widely used due to a number of advantages over longitudinal designs:

- Rare diseases can be studied, as the specific recruitment of the individuals affected can ensure that a sufficiently large number of cases will be included in the subject for meaningful analysis
- Diseases with long latency can be studied efficiently since recruitment can focus on subjects that already have the medical condition of interest.
- They allow the effects of multiple exposures on the development of the disease to be studied at the same time.
- They are relatively quick and cheap to undertake due to their retrospective nature.



Altman, Medical Statistics, Oxford Univ.