Case-control

Case-control studies are one type of epidemiological study design.

While the 'gold standard' in terms of study design is the double blind prospective randomized controlled trial, in order to study very infrequent events a very large population must be tracked to see a large enough number of cases; furthermore, if the event may take a long time to develop, this large population must be tracked for many years, despite 'drop-outs' from the study. The expense involved is too large to permit this methodology to be used to investigate every suspected risk. Instead, the case-control study enables much more cost-effective study of the factors related to the effect being investigated; if the evidence found is convincing enough, then resources can be allocated to a double blind prospective study to confirm the results.

The case-control study begins with a set of several cases showing the effect of interest, whether it is negative, such as lung cancer, or positive, such as living past 100 years of age. These individuals are scored thoroughly on as many relevant variables of interest as possible; typically demographic variables such as age, sex, race, income bracket, geographic area of residence, plus anything specifically relevant to the study subject, such as smoking history and asbestos exposure for lung cancer.

Control subjects are then chosen from a similar population who do not show the effect of interest, usually but not necessarily one per case subject, to match each case subject as closely as possible on variables that are to be factored out of the study; typically such things as age, sex, etc. which are known to have large effects on many aspects of the individual's status, but are not of specific interest to the particular question being investigated. The two groups are then compared on the remainder of the variables scored, and a statistical analysis can be used to estimate the strength of association of each parameter with the effect being studied. For instance, if all the cases were found to be overweight but none of the controls, that would most likely result in an estimate of a high degree of association of overweight with the phenomenon, but if half of each group was found to be frequent coffee drinkers, that would most likely result in an estimate of little or no association of coffee intake with the effect.

Choice of variables to be scored is, obviously, critical to the success of this type of study. Omission of a variable with a real effect will cause its effect to be erroneously assigned to another variable which is related to both; for instance, omission of smoking as a variable in a lung cancer study might cause a spurious association to be seen with low weight, since smokers tend to be of lower weight than nonsmokers. After the choice of the set of all variables to be scored, it must be determined whether each variable is to be utilized to select a control with similar charactistics, or to be retained for the statistical analysis. Variables that are used to select the control group are effectively dropped out of the analysis, simplifying the mathematical work to be done and increasing the sensitivity to the other variables.

Variables may be not used to select controls and instead kept as potential 'causal variables' for the statistical analysis for a number of reasons; they may be of interest as possible alternative explanations for the effect, i.e. asbestos exposure for a study of smoking and lung cancer, to compare the relative sizes of the associations; or because they are rare enough that finding a control with the same description would be too difficult, for instance, sickle cell anemia.

After collecting the data on each case and control subject, they are analyzed, typically by a mathematical statistical technique known as multivariate regression. This works best with a limited number of variables (perhaps 10 or fewer), therefore in the frequent case of more variables than that, each is individually tested for association with the effect under study with a simple univariate regression, and only those showing some such association are then entered into the multivariate regression. The multivariate regression assigns each hypothetical 'causal variable' an estimated independent strength of association with the effect being measured, calculated on what the correlation would be if all the other 'causal variables' were identical except for the one being calculated, and an estimated confidence interval, i.e. the region within which the actual value might most likely be expected to lie. The result may be a positive association, i.e. that variable increases the chances of seeing the effect in question; negative, if the variable decreases the frequency of the effect; or zero, if the variable has no association with the effect, positive or negative. Usually, the variables used to select controls are included in the regression, to check on whether they were correctly balanced between the case and control populations.

Even more cost effective than case-control studies are cross-sectional studies, which do not require the extensive, and expensive, matching of cases to controls on several parameters, and thus frequently can be carried out on preexisting data, such as that collected by the Census Bureau or the Center for Disease Control. Such studies therefore can cover study groups as large as the entire population of the United States; the epidemiologically naive often wonder why a study of such a large population is not superior to a case-control study of a few hundred individuals or fewer. The advantage of case-control studies over cross-sectional studies of a population is the ability to determine the association between potential cause and effect on an individual basis. In the cross-sectional study, the individual variables are aggregated over the population as a whole, then an association is sought between the aggregated variables; in the case-control study, the association is determined for each individual case-control pair, then aggregated. This provides a more powerful analysis of the possible associations, and more accurately determines which possible causes are truly directly related to the effect being studied, and which are merely related by a common cause.

For instance, cross-sectional studies confirm that people who consume large amounts of alcohol also show high rates of many other diseases; but alcohol consumption is also associated with improper nutrition and hygiene, high rates of smoking and abuse of illegal drugs, and many other behaviors risky to health. Cross-sectional studies cannot differentiate between these possible causes, but case-control studies can determine that gastrointestinal bleeding, for instance, is a direct result of alcohol consumption, while memory deterioration is more a result of improper nutrition among alcoholics, therefore possibly more easily preventable.

The disadvantage of case-control studies is that they do not give any indication of the absolute risk of the factor in question. For instance, a case-control study may tell you that a certain behavior might increase the risk of death tenfold, which sounds alarming; but it would not tell you that the actual risk of death would change from one in ten million to one in one million, which is quite a bit less alarming. For that information, data from outside the case-control study must be consulted.

The great triumph of the case-control study was the establishment of a link between tobacco smoking and lung cancer, by Nobel prize winner Sir Richard Dall. Using this technique, Dall was able to show a very significant association between the two. Skeptics, largely backed by the tobacco industry, argued for many years that this type of study did not absolutely prove causation, but the eventual results of double-blind prospective studies confirmed the causal link which the case-control studies suggested, and it is now generally accepted that tobacco smoking greatly raises the risk of lung cancer.

• BMJ (formerly British Medical Journal) on Case-control and cross-sectional studies (http://bmj.bmjjournals.com/epidem/epid.8.html)