Loosing significance while clustering at household level

[Shafique Ahmad Khan]

I lost my significant results when I clustered the standard errors using household id. I am using one round of the data set. can i afford to run the regressions without clustering at the household level?

[Daniel Schaefer]

That depends. Why do you think the standard errors should be clustered at the household level? Could autocorrelation explain why your independent and dependent variable are related? Do you have evidence from regression diagnostics of correlated errors? How large is the number of clusters? Do you have many households with exactly one observation within? Some great advice in this thread. Take a look at the paper recommended in #4 for some more general advice.

Please don't try to "afford" anything and just focus on making sure the model is correct. If you go searching specifically for models with statistically significant results you will greatly inflate your type 1 error rate.

[Jeff Wooldridge]

If the data we sampled at the household level, in some cluster sampling scheme, there's little choice but to cluster at the household level. But you could also use a GLS ("random effects") approach to try to exploit the within-household correlation. You can trick xtreg, re to do that even though you don't have panel data. You should still use vce(robust) [equivalent, vce(cluster id)].

If the data were randomly sampled and you just happen to have repeat members in a household, you don't need to cluster (but then it wouldn't make much difference, anyway).

[Carlo Lazzaro]

Shafique:
how many clusters do you have in your dataset?