I have some subjects from I have measured a variable (body weight). I want to do a case-control experiment where I will do an intervention on one of the groups (Experimental). I'd like for the group means to be as close as possible to begin. My questions are:

  1. Is this a legitimate way to assign subjects?
  2. Is there a statistical programming way (R) to do this?

I think this article describes what I'm trying to do, so it sounds like this is accepted practice, but I'm not sure. And it's in SAS, so not helpful to me (I'm using R).

  • $\begingroup$ It sounds reasonable, on the surface to me. You have said a philosophy, not an actual explicit method, so when you ask "legitimate way" I can't answer that. Some of your post-assignment analysis drives how appropriate your assigning method is. That would need to be specified. I assure you there is a way to do it in R. $\endgroup$ – EngrStudent Feb 2 '18 at 21:53
  • 1
    $\begingroup$ It sounds full of pitfalls to me. That's in part because groups must match on other characteristics (such as spread) to have any chance of being comparable. Another reason is that most deterministic algorithmic ways to split subjects into groups have no known (or even computable) statistical properties, making it invalid to apply most statistical testing procedures. Although there are some (weak) objections to randomization, that still is superior to and far simpler (and more defensible) than most alternatives. cc@EngrStudent $\endgroup$ – whuber Feb 2 '18 at 22:04
  • $\begingroup$ You might also consider a matched subject design. $\endgroup$ – Sal Mangiafico Feb 3 '18 at 11:56

Before revealing one option for pairing participants based on certain features, you should strongly consider @whuber's suggestion to use a totally random design. There are numerous benefits, both statistical and practical that are afforded by random assignment. For one, a random sample is easy to explain, and most people intuitively understand its value in experimental design. With random sampling any variation between the groups should be non-systematic in nature, and most statistical tests you would want to run are well-suited for these sorts of designs. In taking a random approach, you also avoid putting your finger on the scale by accident and ensure that your results are easily compared to future investigations of the same population using random designs. Any matching algorithm you use will necessarily be unique in some ways to your obtained sample, and may not be replicable in future cohorts.

That all being said, there is a package called optmatch in R. An accessible review of its features can be found here. This is but one option out there, and before you use a tool like this, I strongly recommend testing it out on some simulated data so that you can fully assess the potential consequences of specific choices available to you.

Again, I would say that whenever possible, random assignment is the ideal choice, but there are some tools out there (like optmatch) that can help in specific situations.

| cite | improve this answer | |

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.