Structural Break Test (CUSUM or Chow) in SAS

[AxelRoses]

Hello,



First let me tell you that I'm not so experienced in SAS.

I was given a task to find a structural break in labor productivity using the following regression model:

Labor_Productivity = Unemployment Oil_Prices Real_Interest_Rates.

When I use growth rates, which I was advised to do, and run the regression, the R-square shows only 4% and t-statistics for the variables are not so good.

However, when I use logs of the variables, the R-square is around 82% and t-statistics are much better, but Durbin-Watson statistic shows a positive serial correlation.

Why is it such a difference? I think I need to use growth rates in my regression, but due to low statistics, the Chow test gives wrong output (I think) or maybe I'm doing it wrong.

Because I don't know when the break (breaks) occurred, I tried CUSUM test. Here is the code I used:



proc autoreg data=AJ.WD1 ;

model prod_change = unem_change oilprice_change realintrate_change;

*model ln_prod = ln_unem ln_oilprice realintrate;

output out=cusum cusum=cusum cusumub=ub cusumlb=lb;

run;



proc print data=cusum;

var lb cusum ub;

run;



However SAS gives CUSUM ERROR and dots for lb cusum ub. What am I doing wrong?

Is there any other way to find unknown structural breaks in panel data?

Please see the attached file with the data.

Please, any advice or suggestion, or help would be greatly appreciated.

 

[mmwave]

Hello,

Thank you for reaching out to us with your question. It sounds like you are on the right track with using growth rates in your regression model, as this is a common approach when analyzing panel data.

The difference in R-square and t-statistics between using growth rates and logs of the variables is likely due to the fact that taking the log of a variable can change its interpretation and relationship with other variables in the model. For example, taking the log of a variable can make it more linear in relation to other variables, which can improve the model's fit. However, it is important to consider the assumptions and potential biases that can arise when using log transformations in regression models.

As for the CUSUM test, it is possible that the error you are receiving is due to an issue with the code or data formatting. I would recommend checking the SAS documentation or reaching out to SAS support for assistance with this specific test. Alternatively, you may want to explore other statistical tests or methods for detecting structural breaks in panel data, such as the Chow test or the Bai-Perron test.

Overall, my advice would be to carefully consider the assumptions and potential biases of any statistical tests or methods you are using, and to seek out additional resources or support if needed. Best of luck with your research!