Clustered standard errors with experimental data

[Danielle Standish]

I'm working on regressions using experimental data that I collected for my MSc dissertation. The sample size is 105 and the data was collected over 8 sessions i.e. 8 'groups'.

When I do the regressions using clustered standard errors wipes out a good amount of the significance in comparison to standard robust. Am I correct that using clustered s.e.'s when there are only 8 clusters is probably over-stringent? Discussion in Angrist & Pischke suggests that 8 clusters is too small.

Thanks for any advice!!

[Carlo Lazzaro]

Danielle.
welcome to the list.
Yes, 8 clusters are probably too few (by the way, please provide full reference of the contribution you mentioned, as per FAQ. Thanks).
In order to check the robustness of your base case findings, you may want to opt for bootstrapped standard errors.

[Danielle Standish]

Thank you Carlo - I will check out bootsrapped standard errors. Reference; Angrist & Pischke, (2008), Mostly harmless econometrics, Chapter 8.

[Carlo Lazzaro]

Danielle:
Thanks for the full reference.
I owe that textbook and remember the discussion about the issue you're interested in.
However, others on this multidisciplinary list may benefit from that textbook too, but they probably ignore its existence.
That's why that seemingy pedantic FAQ remind us: "full reference, please!".