• methods

A common problem in meta-analysis of observational studies arises when the exposure variable is categorical rather than continuous. These data may be referred to as quantile or quintile data (depending on the number of categories) or dose–response data, and in this article, the term ‘categorical risk data’ will be used. These data may be reported to reflect the increase in cardiovascular risk associated with increasing weight gain, alcohol consumption or frequency of smoking. Further problems arise when studies divide the exposure variable into different numbers of categories, or the same number of categories, but using different thresholds, when data are missing, or when studies include different reference categories. These problems make it difficult to combine data in meta-analysis, but there are methods that can deal with these problems. Consider a clinical question as an example:

How is body mass index (BMI) associated with the risk of incident atrial fibrillation?

Consider three studies as examples for which this question was the focus. The first study1 provided HRs for three categories of BMI: <25 kg/m², 25≤BMI<30 and ≥30 kg/m². The second study2 reported HRs for quintiles. The third study3 reported that one unit increase of BMI was associated with a 4.3% increased risk of incident atrial fibrillation (HR 1.04, 95% confidence interval (CI) 1.02 to 1.07). This is an example of an HR reported on a continuous scale.

To carry out meta-analysis, the data from each study need to be in the same form. The same form would apply if all three studies reported HRs for incident atrial fibrillation on a continuous scale and for the same increase in BMI. For example, it may be desirable to derive a pooled HR for an increase of five units of BMI. It is possible to convert these HRs to …