6
$\begingroup$

I want to look into the causal effects of an education program on a binary employment outcome (positive vs negative). I have two groups of students- one that took the education program (the treatment group) and one that did not (the control group).

The groups are highly unbalanced proportionally on background demographics. I have implemented 1:1 nearest neighbour propensity score matching using the background demographics such as sex, ethnicity, socioeconomic class etc in the MatchIt package in R. After assessing my balance statistics, I am now confident that the sample is well matched on these background demographics.

I have spent a long time looking through papers such as Stuart (2010) and Austin (2007) with regards to advice on how to proceed with post-matching analysis and this is where my confusion begins.

I am trying to follow Stuart's paper. and it suggests that it is not necessary to take into account the matched pairs after k:1 matching but rather 'conditioning on the variables that were used in the matching process (such as through a regression model) is sufficient.' Should I interpret this as to include the background demographics, alongside the educational program indicator as independent variables in the logistic regression model with employment outcome being my dependent variable (e.g., using glm in R)? Or is it suggesting that I should be using conditional logistic regression? From my understanding, conditional logistic regression does take into account the matched aspect of the data, is that right?

I am also not sure if it is appropriate to incorporate the background demographics in the modelling of the post-propensity score matching analysis? I have seen some studies do this as they say although there is no difference between the treatment/control groups on these variables they may have an impact on the outcome of interest, e.g., although the socioeconomic classes of those students in / not in the educational program are balanced, this variable may have an impact on employment outcome. I have also seen some studies completely ignore them.

I don't have access to many books but am happy to be directed to available papers on the web.

$\endgroup$
  • $\begingroup$ How many patients in the treatment group were removed in the matching process? How many in the control group? Note that it is not clearly correct to do 1:1 matching. $\endgroup$ – Frank Harrell Aug 6 '15 at 22:23
2
$\begingroup$

Stuart et al. mention the R package Zelig, which seemlessly works for post-matching analysis after matching with MatchIt. It is mentioned quite often that you should NOT simply compare the means after matching, although this is quite common practice. You can make optimal use of the matching process by using regression models. When matching has been performed, further (parametric) statistical analysis need to be performed. Matching is just a first step. It can be seen as a non-parametric method for pre-processing the data in order to create a quasi-randomized study and thus decrease or eliminate the dependence of the outcome variables on the confounding covariates.

$\endgroup$
  • $\begingroup$ Why does the further statistical analysis have to be parametric? $\endgroup$ – Brash Equilibrium Aug 25 '16 at 4:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.