0
$\begingroup$

I am running a very simple experiment. There are two different experiment groups, participants are randomly assigned. They have to make a simple binary decision in the experiment.

I have three control variables (demographics) about the participants that I still would like to control for.

I am using a logit regression model to predict participants' choice based on the group they were assigned to (a dummy variable) and the control variables.

A colleague of mine suggested that I might have to cluster the standard errors of the logit regression on the experiment group. Is that accurate?

$\endgroup$
6
  • 1
    $\begingroup$ What does "cluster the standard errors of the logit regression on the experiment group" mean? $\endgroup$
    – dipetkov
    Commented Oct 2, 2022 at 17:11
  • $\begingroup$ Run a logit regression, clustering the standard errors at the level of the experiment group $\endgroup$ Commented Oct 2, 2022 at 18:47
  • 1
    $\begingroup$ You realize that you just wrote the same sentence as before, right? $\endgroup$
    – dipetkov
    Commented Oct 2, 2022 at 18:53
  • $\begingroup$ I did not. But similar, yes. What precisely is it that you did not understand? $\endgroup$ Commented Oct 2, 2022 at 19:08
  • 2
    $\begingroup$ This is not the most productive discussion ever. I'm suggesting that it would be helpful to explain what you mean by this statement (either version 1 or version 2). Do you know what your colleague's suggestion means? I ask because I don't actually know what your colleague means. $\endgroup$
    – dipetkov
    Commented Oct 2, 2022 at 19:23

1 Answer 1

4
$\begingroup$

No you do not need to use cluster standard errors.

Cluster standard errors are for when observations are "clustered" within other observations at a higher level of analysis. So you might have a situation where students are nested within schools, or where you have multiple observations "clustered" within the same individual ("Sarah" has two observations in the dataset, one at time 1 and another a time 2).

It might SEEM at first glance like this also applies to your situation since you COULD think about people being "clustered" into different "groups" based on whether they got the treatment or not. But if you really were thinking about it that way then there would be no way you could include "treatment" as a variable in your model, since you would be considering "treatment group" as a way of identifying observations at a higher level (they way "person ID" or "name" identifies people) and not a variable that tells you things about people.

In your situation "experimental group" is not a way of identifying other types of observations but a variable - a characteristic of observations. People are assigned to either the treatment or control group in the same way they are either male or female. So there is no reason to include cluster robust standard errors in this analysis. It would be like including cluster standard errors for "gender" or "race" or "income level," which is obviously not something we usually do.

$\endgroup$

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.