The self controlled case series method

The study by Douglas and Smeeth (doi: 10.1136/bmj.a1227) uses the self controlled case series method to study the association between exposure to antipsychotics and the risk of stroke.1 The study found that use of any antipsychotic agent significantly increased the risk of stroke (relative risk 1.73, 95% confidence interval 1.60 to 1.87). The risk of stroke in people with dementia taking any antipsychotic was higher (3.50, 2.97 to 4.12) than in people without dementia taking similar medication (1.41, 1.29 to 1.55).

The self controlled case series method, or case series method for short, can be used to study the association between an acute event and a transient exposure using data only on cases; no separate controls are needed.2

The method uses exposure histories that are retrospectively ascertained in cases to estimate the relative incidence. That is, the incidences of events within risk periods—windows of time during or after experiencing the exposure when people are hypothesised to be at greater risk—relative to the incidences of events within control periods, which includes all time before the case experienced the exposure and after the risk has returned to the baseline value.

This method has two major advantages. Firstly, no controls are needed, which reduces the time needed to carry out a study, as well as the costs. Secondly, all fixed multiplicative confounders that do not vary over the study periods and that act proportionally on the baseline risk are controlled for implicitly. Another advantage is that in certain circumstances, when the risk periods are short compared with the total observation time, the case series method is almost as efficient as the cohort method with the same number of cases. Such a situation often arises in vaccine safety studies, for which this method has been used most often to date.

For many researchers, the main appeal of the self controlled case series method is the implicit control of fixed confounders. This was the case for Douglas and Smeeth, who stated that the underlying cardiovascular risk for people prescribed and not prescribed antipsychotic drugs differs in ways that are difficult to quantify and control for. Databases often do not include as much information on potential confounders as researchers would like, making the case series design attractive for studies using database data, as was the case in the present study.3

Whereas fixed confounders are controlled for, confounders that vary with time are not, although it is possible to allow for them explicitly. Age effects are almost always included in a case series analysis; the increasing incidence of stroke with age was allowed for by including an age effect in five year age bands. Results can sometimes be sensitive to the choice of age groups, and a semi-parametric version of the method—which avoids the need to specify the age groupings at the cost of increased computational time— is available.4 For large data sets, a more practical approach is to undertake sensitivity analyses with different age groupings.

The largest limitation of the self controlled case series method is that the probability of exposure must not be altered by a previous event. If stroke were a contraindication to antipsychotic drugs the relative incidence would be biased upwards. Is it possible that the occurrence of a stroke would lead to people being more or less likely to be prescribed antipsychotic drugs? This study looked at patients only in the time period before the end of 2002 because this was when concerns about the possible effects of antipsychotic drugs first emerged, hence minimising the chance that any such bias would be present. Events that can result in death, as is the case with stroke, can also introduce bias—patients must be alive to receive a prescription. It is unclear what effect this would have in the present study, although the bias was shown to be small in a study of myocardial infarction.4

Several studies have carried out both a case-control analysis and a case series analysis.5 The sources of bias are different in the two types of studies, and comparing results from each might provide insight into how these biases affect the results or might increase confidence in study conclusions.

Case series studies are relatively straightforward to perform. Observation periods are defined for each case by fixing age and time boundaries for the study, exposure histories are ascertained, and age groups and risk periods are defined. Perhaps the most difficult part is choosing how to define the risk periods. These should be defined a priori, on the basis of the study hypotheses, previous studies, and biological mechanisms. Where uncertainty exists, several contiguous periods can be used. For example, Douglas and Smeeth defined one set of risk periods while people were taking antipsychotic drugs using information on prescriptions as a guideline. A second set of risk periods for washout periods was included because the incidence may not have returned immediately to the baseline value. Instructions on the basic theory, data analysis, and interpretation are available in a tutorial,6 and example files for use in several statistical packages are available on the Open University website (http://statistics.open.ac.uk/sccs).