Department of Methodology
London School of Economics
Columbia House 8th floor
London WC2A 2AE, UK
Knowledgebase
DoM Knowledgebase > Help Desk for staff only > Knowledgebase

Ask a question:


Propensity score matching

Solution Situation: Study of impact of receiving a postgraduate study scholarship on various outcomes, for Sardinian students.

Design: Recipients are matched to students who did not receive (or apply for) scholarship, using propensity score matching.

Part of data collection is an online survey, both to collect outcome measures for all, and matching variables for control group to get the same variables as for the treatment group. Final matching is to be done after the survey.

Questions discussed:

(1) Number of times for which outcome variables should be collected in the survey.

Here it seems better to collect a small number, to keep the survey short. Perhaps one before treatment (i.e. between finishing undergraduate degree and starting PG studies) and one after. For both, should define clearly which time the question refers to (either in time since graduation, or calendar time).

(2) Possible problems with lack of overlap (common) support between groups.

This remains to be seen, but does not strike me as an inevitable problem. The matching should be only one variables determined before the time of the scholarship decision. It does not seem very likely that the groups would be very different in these. If the quantity of interest is the average treatment effect of the treated and some treated students can't be matched, they should be profiled carefully. Otherwise, it's not clear to which particular subgroup of the sample you are making inferences to.

(3) What to do if some of the matching variables are missing.

This seems to be the same as the problem of missing data in explanatory variables in any regression model. In that context the main solutions are (i) omitting variables, (ii) omitting cases, and (iii) imputation. For propensity score matching, (i) and (ii) seems inappropriate, at least if it is the (fairly rare) treated individuals for whom there is missing data; while (iii) is certainly possible. An alternative is to match on the pattern of missingness. Details for this procedure can be found here: Rosenbaum, Paul R. 2009. Design of Observational Studies. Heidelberg, New York: Springer. Chapter 9.
 
Was this article helpful? yes / no
Article details
Article ID: 1
Category: Knowledgebase
Date added: 2012-01-10 14:13:53
Views: 467
Rating (Votes): Article rated 5.0/5.0 (1)

 
<< Go back

 
Powered by Help Desk Software HESK - brought to you by Help Desk Software SysAid