Figure 6.5 Prospective cohort analytic study examining the relation between post‐partum back pain and epidural anaesthesia during labour.
Source: From Macarthur et al. (1995), © 1995, BMJ Publishing Group Ltd.
You may have noticed one prominent feature that distinguishes case–control from cohort designs: case–control studies ask the question by looking backwards (asking whether more cases than controls had been exposed to some risk factor), while cohort analytic studies ask their question by looking forwards (asking whether more people who were exposed to the risk, than people who weren't exposed, developed the outcome in question). But you need to be careful with your thinking and your terminology in relation to the words retrospective and prospective because, as we have seen, it is possible to ask the (forward‐going) question of the cohort study by starting now and looking ahead (Figure 6.5), or by starting with the presence and absence of the risk factor some time previously and following up until lately (Figure 6.4). Put another way, case–control studies are retrospective but cohort studies may be described, quite properly, as either prospective or retrospective.
VARIATIONS IN CASE–CONTROL AND COHORT DESIGNS
So‐called nested case–control studies are found in the health literature in abundance. The concept is very similar to the standard version of the case–control study, but the cases and controls are selected from a large cohort study, reducing the research work that would be needed were the whole cohort to be analysed. All the cases in the cohort will usually be selected, together with several (usually one to five) controls per case. Suppose we have decided to have five controls per case – to maximize statistical provision (more than five controls per case offer little extra precision). The controls could be selected randomly from the people in the cohort who are still at risk at the time of the index case and are time‐matched with their specific ‘case' – creating a ‘risk‐set', in this illustration consisting of one case and five controls. There are several advantages over the conventional case–control study. First, conventional case–control studies often struggle to determine control groups that truly represent the study population, and use convenience samples; here, the randomly selected controls are a representation of the people in the cohort and thereby in the study population. Second, the exposure data in the cohort design will have been collected prospectively and are thereby free of recall bias – a common problem with case–control studies (see Chapter 16). Nested studies may differ in their analysis from conventional case–control studies (see Chapter 28), but this matter is beyond the scope of this book.
Much less frequently encountered, but also derived from a large cohort study, is the case‐cohort design. In such a study, a sub‐cohort is selected from a large cohort study at the start of the follow‐up; the sub‐cohort is not time‐matched with the cases. The analysis is relatively complicated, partly because it involves the sub‐cohort and cases outside the sub‐cohort, but it offers some advantages over the conventional case–control analysis (see Chapter 28).
COMPARING AND CONTRASTING CASE–CONTROL AND COHORT STUDIES
Compared with cohort studies, case–control studies are cheap and cheerful. First, they can be completed without waiting for the outcome to develop. Second, there is no need for enormous numbers of study subjects who do not develop the outcome; because in a case–control study it is already clear who has the outcome, only a convenient number (needed for reasonable statistical precision) of these 'controls' is needed. In a cohort study, however, there is no way of knowing who will get and not get the outcome – so everyone has to be included until the study end‐point reveals who has developed the condition so far. For a great many relatively uncommon outcomes the case–control study is favoured for just these reasons. Examples of case–control studies include investigating whether deep venous thrombosis is related to taking the oral contraceptive pill, and whether depressive illness is related to adverse life events and difficulties.
On the other hand, there are serious shortcomings with case–control studies. They are prone to extra biases – in particular those concerned with recollection of past events. Suppose, for example, that a study hypothesis is that an adult condition such as motor neurone disease is predisposed to by certain kinds of childhood infections. It is likely that in an effort to understand why one has such an illness, sufferers with motor neurone disease may recall more such illnesses than controls who have no pressure to make such an effort. It is very difficult to avoid such bias.
Second, it is usually difficult to select a wholly suitable control group for a case–control study. In a case–control study concerned with smoking and lung cancer, for example, should those without lung cancer be people with other cancers, or patients with other respiratory conditions, or other patients at the hospital without chest disorder, or general practice cases (and should they exclude those with chest problems), or members of the general public who are not drawn from healthcare contacts at all? None of these possible groups is wholly suited or unsuited; in the end, the findings will often be hard to interpret whichever control group is selected (sometimes case–control studies use more than one control group for just such reasons). Put another way, the control group in a case–control study is intended by the researcher to represent, as if it were a small random sample, the great many people in the population who have not developed the outcome. Unfortunately, in practice, they are instead a specifically targeted group – easily identified and included in the research – but not actually providing the representation of the non‐cases in the population that the researcher ideally wants.
Cohort analytic studies avoid the worst aspects of the above two problems. First, because exposure is assessed at the start of the study, it is not subject to false remembering later. Second, assessment of exposure to risk often provides sensible exposed and unexposed groups, and avoids uncertainty over which comparison group to choose.
Both these kinds of analytic observational study are greatly prone to confounding. This matter is discussed in Chapters 17, 34, and 35.
‘STROBE' GUIDELINES
Readers of research want to know that the material in front of them is fairly and comprehensively described, preferably setting out what was planned, as well as what was done. We all hope that researchers will tell us in a thorough way what their question was, how the study planned to answer that question, how it was undertaken, what was found, and that they will give us a fair interpretation of what it all means – particularly in relation to the limitations of the design and its execution, and how generalizable the findings are likely to be. As a commendable attempt to render it more likely that we shall read research that attains these standards, people have been trying in recent years to set out guidance for the reporting of research. Researchers have got together with journal editors, and research