Table 2

Situations in which the matched-pair design is (or might be) better and those in which the stratified design is (or might be) better (shading indicates the situations prevailing in our study)

Matched pairs better
Stratified better

Large variation between pairs with respect to baseline risk
Within-stratum variation small compared to between-stratum variation
High matching correlation
Small matching correlation within strata
No individual level analysis desired
Analysis at individual level desired (2) (interactions of interventions with age, gender, medical history)
Homogeneity of effect can be assumed across pairs
Heterogeneity of effect across strata possible
No drop-outs expected
Individual hospitals may drop out
Medium number of clusters (20 to 40, so as to have 10–20 well-matched pairs)
Large number of clusters (perhaps >30 or >40 depending on the number of strata)
Calculation of ICC(1) needs special assumptions
Calculation of ICC straightforward

(1) ICC – intracluster correlation coefficient (2) Only as secondary analysis in our study

Gülmezoglu et al. BMC Medical Research Methodology 2004 4:2   doi:10.1186/1471-2288-4-2