Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Are standard errors valid with xtivreg2 and non-nested two-way clustering?

    Hi, all,

    I am running a fixed-effects panel regression and clustering by the panel variable (administrative district) and by state-year. (There are many districts per state.) But I was a bit concerned when I noticed this in the help file for xtivreg2:

    For fixed-effects estimation with cluster, xtivreg2 makes no degrees-of-freedom adjustment for the number of fixed
    effects. This follows the formulation of a cluster-robust covariance matrix for the fixed-effects model as originally
    proposed by Arellano (1987); see, e.g., Wooldridge (2002), p. 275. Stata's official xtivreg, xtreg and areg (as of
    version 9.1, October 2005), by contrast, use the (N-N_g-K) adjustment, which is somewhat conservative in this context.
    However, the approach used by xtivreg2 requires that no panel overlaps more than one cluster. That is, either the panel
    variable is identical to the cluster variable, or panels are uniquely assigned to clusters. If any panel is contained in
    more than one cluster, the xtivreg2 approach is invalid, and it will exit with an error.
    The first cluster (district) clearly satisfies this requirement, but the second does not (as a district appears in multiple years). Does anyone know whether the standard errors in this case case can be trusted, or would we need to make a degrees of freedom adjustment for part of the variance matrix (the part corresponding to state-years and the part corresponding to district-state-years)?

    Thanks,
    Ajay
Working...
X