Cluster Randomized Controlled Trial

[Joe Malaco]

I have a question about the Cluster Randomized Controlled Trial. Is it recommended to perform the svyset command when doing the Cluster Randomized Trial? Another question is that what command can we use if we want to adjust for clustering?

Thanks!

[Clyde Schechter]

Assuming that the selection of the clusters themselves and the selection of individual units within the clusters does not involve any departures from simple random sampling, there is no reason to use -svyset- in the analysis of these trials.

Some key elements in the analysis of cluster randomized trials is the inclusion of cluster-level fixed or random effects. You might also use cluster robust standard errors if the number of clusters is sufficiently large.

[wbuchanan]

If you are setting up the experiment yourself, I’d suggest looking at some of the work that was presented at the Baltimore Stata conference a couple of years ago on covariate constrained/balanced clustered randomization. Depending on the size of the sample and dimensions that potentially affect the outcome you could end up with randomization that assigns clusters with the same value on a covariate to the same condition which would end up creating incomparable groups.

[wbuchanan]

If you are setting up the experiment yourself, I’d suggest looking at some of the work that was presented at the Baltimore Stata conference a couple of years ago on covariate constrained/balanced clustered randomization. Depending on the size of the sample and dimensions that potentially affect the outcome you could end up with randomization that assigns clusters with the same value on a covariate to the same condition which would end up creating incomparable groups.

[Joe Malaco]

Thank you both for the input! Here is the command that I used to tell Stata to run the Wilcoxon signed rank test on the post-test and pre-test. But I wanted to adjust for the school as my cluster within the school district. How can do that! Command: signrankstrsuc_pro_2 = strsuc_pro_1

[lourdes gastearena]

Hello statalist,
I have a question about the command "misstable tree, frequency" .Does anyone know how to interpret the output?
Thank you!

[Joe Malaco]

Thank you both for the input! Here is the command that I used to tell Stata to run the Wilcoxon signed rank test on the post-test and pre-test. But I wanted to adjust for the school as my cluster within the school district. How can do that! Command: signrank strsuc_pro_2 = strsuc_pro_1. I guess my overall goal is to compare the pre- and post-tests of the intervention vs control groups adjusting for the cluster (i.e., schools)

[wbuchanan]

This depends largely on many specifics of your data and theoretical framework (e.g., # clusters, do schools affect the outcome or are you only interested in adjusting standard errors due to a lack of independence). Is there a specific reason why you are using a rank sum test instead of any other models?

[Joe Malaco]

Yes, the reason I'm using the Wilcoxon signed rank test is because the data is not normally distributed. I'm interested in seeing how schools affect the outcome.

Thanking you again!

[Clyde Schechter]

Lack of normality in the outcome distribution is not a good reason to resort to non-parametric tests. Normality of the outcome distribution as a whole is, in fact, not a reason at all--nothing in ttests or linear regression calls for that. Normality of the outcome within treatment groups is a classical condition for the use of ttests or linear regression. But, it is a sufficient condition, not a necessary one. In particular, if your sample size is sufficiently large, the central limit theorem will kick in and the test statistics will have the desired t- and F- distributions to make inference in the usual way just fine. So lack of normality of outcome within each treatment group is only a problem if the sample is also small.

[Joe Malaco]

Thank you so much, Clyde!