To clarify, I have several time series of roughly 30 years with monthly datapoints. Given the fact that the time series are quite volatile, I expected more than 2 structural breaks and conducted the Bai-Perron multiple structural breaks test. While doing so, the test provided for each of the time series 3 to 4 structural breaks. As no unit root test or cointegration test available in EViews is capable of taking into account of so many structural breaks, it is best to divide the time series into subsamples. However, given the fact that the time series consist out of monthly datapoints, dividing the sample into 4 subsamples provides for subsamples with relatively few observations (E.g. 56). Do you think this has an impact on the accuracy of the unit root test/cointegration test and if so: would it better to ignore a structural break based on its "size"? Is that a possibility? Many thanks!
Given the number of experts used to measure the diversity of criteria, I think this gives you an advantage in addition to having enough observations to issue or make correct decisions. The Experton Method and Fuzzy Delphy will help you in any decision-making situation.
I faced a similar (if not more difficult) situation a lot of years ago, with annual series with a lot of structural breaks present. I had only 51 observations so splitting the sample was not an option (see below for the paper). At the time, the unit root tests (Perron and Zivot-Andrews) allowed incorporating only one structural break at the time.
What I did? I performed, for each variable, as many unit root tests as the structural breaks present in the series. That provided feedback regarding the 'true' nature of each variable, that is stationary or not. Then I used this in order to select the right model, without risking spurious regressions and invalid results.
Therefore, my suggestion is to either use a way to make robust conclusions in your analysis, despite the limitations of the unit root tests, or split the sample (if you have enough degrees of freedom, that's fine).
Mitsis, P. (2015), “Effects of Minimum Wages on Total Employment: Evidence from Cyprus”, Journal of Labor Research, 36(3), 318-345.
I think splitting the data sample as per the number of structural breaks may create problem . As you mentioned there are 4 breaks in your data set which seems little bit unusual, in general there may be 2 to 3 breaks but in your case it is 4 which is extreme. if you will break the data in to 4 points and run a model it will reduce the credibility of your analysis, additional there might be problem of degree of freedom. So I would finally suggest that you can apply other advance test of break point and see is result identical and if it is then you can run the model on whole data set and make the inference. you can also show the breaks a limitation of your analysis
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series