1
$\begingroup$

I have a group of patients (about 100 of them) that needs to be randomized into 2 groups (treatment and control) so that these groups were as similar as possible in terms of some (about 6-10) covariates.

Some of covariates are continous. I can't cut them into intervals and use some startified randomization scheme since, with 6-10 covariates, number of resulting strata would be large (as compared do group size).

Can you suggest any method? It would be nice if it was implemented in R and also worked for more than two groups (for future work). But these are not "must haves".

So far the best method I found is covariate-constrained randomization designed for cluster randomized trials implemented in cvcrand package for R (look at package's vignette for more details). However this doesn't solve my problem because it's for clustered trials (mine is not). Or maybe this is not a problem?

Any suggestions appreciated.

Improve this question
$\endgroup$
3
  • 2
    $\begingroup$ Simple random assignment into the groups is the standard method. How would that fail to work in your case? $\endgroup$
    – whuber ♦
    Commented Jul 16, 2021 at 12:48
  • $\begingroup$ Simple random assignment does not control for any covariates. So, if I was unlucky, I can end up with groups with very different values of one (or more) of my covariates. $\endgroup$
    – Łukasz Deryło
    Commented Jul 16, 2021 at 12:51
  • $\begingroup$ On the contrary, random assignment does control for covariates in a way that permits application of standard procedures. If you insist on somehow matching covariate values, you quickly run into serious problems, of which one of the worst is having no probabilistic basis for analyzing the results. If, with 100 observations, you find a covariate split into two drastically different groups, then that covariate must have a very skewed distribution or some extreme outliers. That's a problem you ought to deal with in its own right, but not by attempting to match subgroup characteristics. $\endgroup$
    – whuber ♦
    Commented Jul 16, 2021 at 18:05

1 Answer 1

Reset to default
0
$\begingroup$

Sequential method described in

Pocock and Simon (1975), Sequential Treatment Assignment with Balancing for Prognostic Factors in the Controlled Clinical Trial. Biometrics; 103-115.

and implemented in Minirand package solves the problem of nuber of strata aproaching number of patients. So I can cut my continous covariates into some intervals.

Improve this answer
$\endgroup$

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.