What's this about?

It's called a structural break when a time series abruptly changes at a point in time. This change could involve a change in mean or a change in the other parameters of the process that produce the series.

Being able to detect when the structure of the time series changes can give us insights into the problem we are studying. Structural break tests help us to determine when and whether there is a significant change in our data.

New commands estat sbknown and estat sbsingle test for a structural break after estimation with regress or ivregress. Both are robust to unknown forms of heteroskedasticity, something that cannot be said of traditional Chow tests.

Let's see it work

We want to know whether there is a greater increase in malaria cases than would otherwise be predicted. Suppose we have data on a country where the number of cases varies over time and that variation is in general explained by the population of Anopheles mosquitoes. The model is

. regress malaria anopheles, vce(robust)

To determine whether and when there is a structural break in our data, we type

. tsset month

. estat sbsingle
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5
..................................................    50
..................................

Test for a structural break: Unknown break date

                             Number of obs =        120

Full sample:                 2005m1 - 2014m12
Trimmed sample:              2006m7 - 2013m7
Estimated break date:        2013m4
Ho: No structural break

     Test            Statistic          p-value

     swald         40966.4180           0.0000 

Exogenous variables:           anopheles
Coefficients included in test: anopheles _cons

The test rejects the null hypothesis of no structural break and detects a break in the fourth month of 2013.

We can also perform a test for more than one structural break if we have ex-ante information about when the breaks might be. It's artificial, but let's use these same data and test for a structural break, pretending that we suspect there might be one on 2013m1, which is close to 2013m4.

. estat sbknown, break(tm(2013m1))

Wald test for a structural break: Known break date

                                  Number of obs  =        120
Sample:       2005m1 - 2014m12
Break date:   2013m1
Ho: No structural break

             chi2(2)      =  209.0560
             Prob > chi2  =    0.0000

Exogenous variables:           anopheles
Coefficients included in test: anopheles _cons

The test rejects the null hypothesis of no structural break.

Tell me more

See [TS] estat sbknown and [TS] estat sbsingle for more information.