Standard errors tend to 0 when cluster with fixed effects

[Jonathan Carr]

Hi,

I am running a difference-in-difference model and would like some help on clustering. My model is:
y(gt) = a*G1 + b*G2 + c*PolicyDummy + d*PolicyDummy * G1 +e*PolicyDummy * G2+ error

I would like to cluster by group, however when I do I find errors fall to 0.

I have read other posts and Jeff Wooldridge referred to this problem in September 2014.

Is there a practical solution to this problem, or should I not include clustered standard errors?

Thank you!

[Ariel Karlinsky]

Can you share output?

[Jonathan Carr]

Statalist.docx
This is a my model before and after I add clustering. G1 and G2 are dummy variables for group 1 and 2 respectively. Date is a time variable so DateG1 is group 1 specific trends. Post2012 is a dummy variable for when policy was introduced, 1 after 2012, 0 before.

[Ariel Karlinsky]

Please share output with [CODE] delimiters as per FAQ, and not in an attached document.

[Jonathan Carr]

Hi Ariel, sorry I had not read the FAQs.

I am running Stata 14.2, I would like to know whether a government policy in 2012 affected investments under 150,000. G1 is 1 if investments are under 150,000, G2 is 1 if 150,000-250,000 and G3 is 1 if 250,000-500,000. The groups variable is 1, 2 and 3 for G1, G2 and G3 respectively.

My problem is that when I cluster on groups my standard errors tend to 0 (t stat is e^14). This problem is mentioned in a post by Jeff Wooldridge in September 2014 but a solution is not given.

 reg Y2 G1 G2 Post2012 Post2012G1 Post2012G2 Date DateG1 DateG2

 reg Y2 G1 G2 Post2012 Post2012G1 Post2012G2 Date DateG1 DateG2 , cluster(groups)

When I use cluster robust standard errors on groups (I have 3 groups) the errors tend to 0. The same occurs if I do:

 reg Y2 i.groups , cluster (groups)

I believe running the regression without clustering / other will mean my standard errors will be underestimated. Is there anyway to make my standard errors robust?

Thanks!

[Ariel Karlinsky]

you should also share your output, not just your commands
but from what you said, clustering on 3 groups is bad practice and can do more harm than good. If you think standard SEs may be inaproriate, you can estimate robust SEs instead using vce(robust).

[Phil Bromiley]

You'll increase your chances of a useful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output (fixed spacing fonts help), and sample data using dataex.

I'm having trouble understanding the basics here. If you want to see whether government policy influenced investment, then investment should be the dv but it sounds like you're using investment categories as iv's. If you do make investment the dv, then if you have a continuous value for investment, it is usually best to keep it continuous. If your data only reports categories, than you might look at ordinal logit.

I'm curious why the cluster by group would result in extremely small standard errors, soI would want to look more closely before rejecting the possibility of different standard errors across categories. It would be nice to understand why grouping the errors changes things so much. What is the sample size? [Posting the results would help us know how well the model fits, the sample size, etc.]

[Jonathan Carr]

Thank you very much both of you. I couldn't work out how to add my results, but I've read the Abedie et al (2017) paper and worked out that I don't need to cluster. Thank you Ariel for your helpful advice about robust errors, I have used these.

Best