I am conducting a retrospective cohort analysis looking at major psych diagnosis (exposure) on outcomes following traumatic injury using a large registry. n exposed = 36,000 ; n unexposed = 3.5 million. I want to use matching so that my computer doesn't blow up running logistic regressions in SAS (I'd prefer to control for covariates in regression...). A recent paper used the same registry that I will be using and used propensity score matching (link http://www.ncbi.nlm.nih.gov/pubmed/24368375) but I think their use of propensity score matching was inappropriate because their outcome was not a treatment and some of the variables they used for matching (i.e. Injury severity score) has absolutely nothing to do with the risk of being obese... am I wrong or am I missing something? Should I use case-control matching?
-
$\begingroup$ This paper covers all the essentials. Elizabeth A. Stuart. "Matching Methods for Causal Inference: A Review and a Look Forward." citeseerx.ist.psu.edu/viewdoc/… $\endgroup$ – Dr. Beeblebrox May 30 '15 at 12:18
Case-control matching is commonly done with medical studies. The idea of propensity score matching is you match on several covariates that one wants to control for. So in the example you gave, they matched with injury severity score maybe due to the higher the severity score the a person may be less active and thus more likely to be obese? So they would want to control for that if they are trying to tease out the effects of the treatment on obesity. Case-control matching does the same thing-- you match by several covariates (age, sex, etc) that you want to control for. Propensity score matching follows the same idea, however, a score is assigned to each unit and one matches based on the scores.