# Chow-test and multi-significant breakpoints

CHOW.xlsHi all, Background: In connection to time series analysis you often experience that the series has been exposed to structural changes and potentially can have at least one breakpoint. Let’s say you estimate a parsimonious model and evaluate potential breaks by using Shazam’s sequential Chow-test option under Diagnose. Let’s say the output shows multiple significant breakpoints, i.e. with p-values less than 0.05. Question: How do we interpret the result when we have multiple significant breakpoints? I have attached an output which illustrates the “problem” (This is an output from an AR(1)-model). The output indicates multiple significant breakpoints, for example about N1 = 38, N2 = 38, about N1 = 50 N2 = 24, and about N1 = 62 , N2 = 14. I hope Shazam staff or other in the field can comment the “case” (please, see attached file).

edit retag close merge delete

Sort by » oldest newest most voted

The DIAGNOSE command just splits into two groups, trying various split points. However, you always calculate a multipoint chow test manually but simply splitting the samples in as many points as you like, run regressions on each portion separately and compare the combined sum of squared residuals (which would be an unrestricted model) to the sum of squared residuals from the regression with the whole sample. This is a simple extension of the two sample chow test.

more

Hi, Thank for the answer. I agree on your recommendation. But what I find interesting is that the single run of the sequential Chow test is probably not consistent with regard to identifying significant multiple breakpoints. The argument is that the single run is a sum of sequential tests which are difficult to compare against each other because they are mutually exclusive. Maybe it is consistent to: Step 1: Identify the first significant breakpoint (lowest p-value) by running the first sequential chow-test. Step 2: Incorporate the breakpoint (including the dummy-variables). Step 3: run a sequential Chow test with the included dummy-variables and check for a (new) significant breakpoint. Then repeat the process. I imagine that this is a time consuming process – and maybe it is not correct either? Comments are welcome.

( 2013-07-19 07:24:51 +0000 )edit