search

ohdsi-studies / TicagrelorVsClopidogrel

OHDSI network study: Net Adverse Clinical Event between Ticagrelor and Clopidogrel in Patients with Acute Coronary Syndrome
1 stars 4 forks source link

1-on-1 matching vs. variable ratio matching #53

Closed chandryou closed 5 years ago

chandryou commented 5 years ago

1-on-1 matching using a non-conditioned outcome model might be better, because that usually gives you more power than variable-ratio matching with a conditional model (because the conditioning is very data-hungry).

In the Method Evaluation, we see evidence to support that: http://data.ohdsi.org/MethodEvalViewer/

chandryou commented 5 years ago

In the fitOutcomeModel function (https://ohdsi.github.io/CohortMethod/reference/fitOutcomeModel.html) there is the argument ‘stratified’. This MUST be set to TRUE if you’re using stratification or variable-ratio matching, but can be FALSE for 1-on-1 matching.

Conditioning (stratification) can be understood as fitting a separate model in each stratum (a matched set is a very small stratum), and then combining across strata. When you do matching and have a rare outcome, most strata don’t have an outcome and these strata do not contribute anything to the analysis.

the evidence now suggests you should 1-to-1 matching with stratified = FALSE.

chandryou commented 5 years ago

https://github.com/chandryou/TicagrelorVsClopidogrel/commit/1785a669e59930e9435696a7c13e637bc66d5286