[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: RE: test for structural break in panel

From   "Nick Cox" <[email protected]>
To   <[email protected]>
Subject   st: RE: test for structural break in panel
Date   Thu, 18 Sep 2008 17:41:18 +0100

What does this mean, precisely? That all members of a panel jump at the
same date? If that's true it should be obvious from a plot of the data
or of some appropriate summary across panels, e.g. some kind of mean or
quantile or smooth. 

Alternatively, I fear that you want to appease your disciplinary devils
with some sanctified test leading to a P-value. But in turn that should
surely depend on specification of a suitable generating process. The
idea that there is a totally general way of doing this formally without
specification of a generating process seems fairly far-fetched. 

This problem is controversial enough for non-panel data, where only a
smidgen seems to separate good statistics from snake oil. Inferentially,
panels seem even more difficult. Practically, panels seem easier in that
genuine structural breaks should presumably be confirmed multiply by a
majority of panels. (If not, how can they be genuine or structural or

As part of my Ph.D. thesis I explored this question for hillslope
profiles, which are traces of hillside from valley bottom to ridge top.
Although the matter was not put in these terms in my thesis, my general
conclusion was that clear breaks are detectable with very simple
graphical methods; otherwise you have the usual multiplicity of cluster
analysis, for that is what this is, in a different guise: namely an
embarrassing confusion of numerous equally objectionable and
unobjectionable ways of doing this almost always producing moderately
inconsistent results. 

-group1d- on SSC lies in a direct line from that thesis work. It does
not apply to panel data, but it offers yet another tool for structural
break enthusiasts. 

[email protected] 


is there a way to search for a structural break in a panel data set?

I found the following:

where I could re-run the analysis for all my years. But I am wondering 
whether there is a
rolling option along the lines of:

But I do not to get meaningful results when combining /rolling/ or 
/rolling2/ with /xtreg/.

Where am I wrong?

*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index