Comparing groups, and especially comparing outcome of different interventions, require some kind of randomisation. The main purpose of the randomization is to eliminate the possibility that there is some kind of selection bias guiding allocation to groups and that this bias is also related to the outcome. Any selection bias will distort the results and the main purpose of randomization is to eliminate this. Consequently randomized controlled trials are considered to be the best way to evaluate effects of different treatments. However, randomization is not always ethical and the only way is to use existing retrospective observations.
Traditionally retrospective observations has been analysed using multivariate regression where treatment allocation has been one of many independent variables. Propensity score matching (PSM) is a new technique allowing retrospective data to be used for group comparisons estimating effect of treatment. PSM can be done in several different ways . It has been argued that PSM is slightly better than multivariate regression for the purpose of comparing effect between different interventions .
The main downside with PSM is that it can only adjust for confounding variables that has been measured and registered. There may be other confounding variables at play that was never measured in the retrospective data set. hence, caution is required when interpreting data from studies using PSM.
Ronny Gunnarsson. Propensity Score Matcing (PSM) [in Science Network TV]. Available at: http://science-network.tv/propensity-score-matcing-psm/. Accessed February 25, 2018.