Background: Several treatment options exist for many conditions. Randomized trial evidence on the relative merits of various options may be missing or biased.
Purpose: To examine the patterns of trial evidence (network geometry) and explain their implications for the interpretation of the existing evidence on a treatment's relative effectiveness.
Data sources: PubMed and Thompson ISI Web of Knowledge (last search April 2007).
Study selection: Published networks of randomized trials that included at least 4 treatments were identified.
Data extraction: For each network, data on the number of studies per treatment comparison were extracted by one investigator and verified by a second investigator.
Data synthesis: Indices were adopted from the ecological literature that measure diversity (number of treatments and how often they were tested) and co-occurrence (whether some treatment comparisons were preferred and others avoided). Eighteen eligible treatment networks were identified for different diseases, involving 4 to 16 alternative treatments and 10 to 84 trials. Networks in which 1 option (placebo or no treatment) was the typical comparator were star-shaped, even though several treatments might have had proven effectiveness. Other networks had different shapes. Some showed important co-occurrence that avoided specific head-to-head comparisons. Comparison choices sometimes seemed justified, such as when newer treatments were not compared with older ones already shown to be inferior, whereas other choices seemed to reflect preference bias.
Limitations: Networks evolve over time as new trials accumulate, and their geometry may change. Statistical testing for co-occurrence is underpowered when few trials exist.
Conclusion: Evaluation of the geometry of a treatment network can offer valuable insights for the interpretation of total evidence when many treatment options are available.