I know that textbooks tell us matching is an alternative to random assignment when it comes to quasi-experimental research. It allows us to make treatment and control group similar to each other.

However, another question arises: Can't I use matching not just for that but also to actually achieve quasi-independent(or non-manipulated if you will) variable?

For example, I want to see the effect of party membership on survey score. However, it's impossible to manipulate party membership variable in reality because of numerous reasons, so it should be regarded as a natural trait. If I wanna acquire treatment and control group which are similar to each other except for the party membership only, I randomly assign samples to treatment and control group, and the ask them their party membership, and then finally leave the same number of sample of each party membership category for both treatment and control group.

I tried to support or find a rationale for my idea, but I couldn't. Anybody has some knowledge or opinion on this? Please help.


1 Answer 1


Using a randomization unrelated to the content of your study is called the instrumental variable approach. Try looking into that.

Your example won't work as you've described it. Let's say Party A has a higher average net worth than Party B. You want to create two groups that are equal on net worth but differ in party affiliation. When you randomize into 2 conditions, 1 and 2, it is true that what you are left with is two conditions, 1 and 2, that should have equal average net worth. Now you want to sample one party from each condition, so you leave all Party A people in condition 1 and all Party B people in condition 2, and throw out the rest.

The problem is that Party A members (even just those in condition 1) still have a higher average net worth than Party B members. Essentially what you doing is randomly throwing away some members of Party A and some members of Party B; doing so doesn't change the distribution of net worth in either party. You need to use strategic matching (e.g., propensity score) to throw out the right members of each party. An unrelated randomization wont help using the matching approach.

  • $\begingroup$ Thanks for the comment. However, I can't understand the point that group A and B need to be equal on net worth. Group A and B are randomized except for trait 'party membership', because that is the independent variable I'm concerned with but at the same time cannot be manipulated because it's kind of a natural trait. So group A and B should have different net worth, because hypothetically they are already treated. Could you add some explanation on this point? $\endgroup$
    – Kang Inkyu
    Commented Oct 9, 2016 at 4:23
  • $\begingroup$ I made some errors, which I fixed. Hopefully that clears it up. Net worth is a confoudner; it causes both survey response and party membership. So if in yoru sample Party A and Party B are not equal on net worth, you will have a biased effect estimate of Party affiliation on survey response if you simply compare them. Matching and randomization seek to alleviate this imbalance. But a randomization unrelated to party membership will not suddenly create two parties with equal net worth. $\endgroup$
    – Noah
    Commented Oct 11, 2016 at 22:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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