Dear Navin,

As I understand from your discussion I think you have annually data. therefore, I should say two dummies is better in this case (otherwise your dummy variables is nothing but constant).

Regarding to point 2, you are correct there is no direct effect for the dummy if it is not sig. However, you still have to check what happened of you ignore such structural break (just use the traditional test like omitted variable test)

@Navin

you could code 2004 as -1 and 2006 as 1 and the same for 2008 and 2017. You can also experiment with coding 2008 through 2017 putting 1 for 2008 and 2009 1 or -1 , and playing with it until you capture the data generating process. There are no magics but researcher' s pain. In the end know that that your dummy captures changes in the constant term a la intercept effect.

To test for a statistical difference between two periods of time by using a dummy variable, this dummy variable should acquire values 1 for one period and 0 for the other.

If the DV is not statistically significant, you can remove it from your model.

If you DV is statistically significant, you can hold it in your model or you can create two models, one for each period. Also, it is advisable to backup the found structural break with some historical/economical event.

Dear Huthaifa Alqaralleh, Chuck A Arize and Nip Nip ,

Many thanks for your responses.

@ Huthaifa, I am dealing with weekly data, and what i did was mark those weeks in years 2006-2008 as 1, and the rest as 0.

@Chuck and @Nip Nip, the Dummy variable turned out to be insignificant, which means I could explain that i have tested for structural breaks, but it turned out to be insignificant - hence dropped from the model.

Thanks again for the timely feedback - and i'll have more queries going forward. :)

It make sense to have two dummies now...

However, you still have to check the importance of your dummies(even they are insignificant)

Dear Navin,

As Huthaifa said, it's a matter of creating dummy variables: one for each regime. Moreover, I would suggest to do those structural break test using Bai and Perron (2003) code. In fact, Perron's website is a good starting point to start studying structural breaks since he publishes his codes:

http://people.bu.edu/perron/

Best,

Were there any policy changes or changes in methods of calculation of your series that might have caused the breaks? Are there other missing variables that might explain the breaks in your series? I would always prefer an economic or common sense explanation than one imposed to force the fit of a model. Do you pass other specification tests?

While I agree with John Frain's point that instability may be the result of an omitted variable, in this particular case, the sample period is small and a dummy variable seems much more preferable.

Thanks John and Chuck for your replies.

@John, there is a great deal of volatility (as i'm dealing with interest rates) during/ leading to the GFC period (data for Sri Lanka, an EME). I hope that there is no omitted variable as I've included almost all the explanatory variables (that I could think of) and even the simplest (and parsimonious) model drew an R-sqd of 0.75. Residuals turned out to be heteroscedastic - so trying out robust least squares.

@Chuck, my data are on a weekly basis - about 700+ observations. But i agree with you on using DVs.

I have no problem with using dummy variables. I simply like to have some explanation as to why a dummy variable is required. With 50+ observations for each year you could include 14 dummy variables ( d2004, d2005, d2006,,,,d2016, d2017) where, for example d2004 = 1 in 2004 and zero elsewhere, d2005 = 1 in 2005 and zero elsewhere etc. (no constant in regression) Looking at the value of the estimates of the coefficients should give you an idea as to what problems are in your original model and when the problems occur. I would then go and talk to the people who were involved in open market operations, banking supervision, changes in reserve requirements, exchange rate policies and any other actions that may have impacted on interest rates. This may involve a lot of extra research but determining a model which has economic meaning will be very worthwhile

Good commeets by John so I will recommend his last 6 lines.

Dear Mr. Perera,

If you are doing structural modelling you would want to identify structural breaks. If you are trying to estimate parameters per period or the long run mean you would try to put zero probability on idiosyncratic events including structural events. I think that's the logic for the programme. You can check (Mallick, Sarka & Roy (2003)) on www.researchgate.net/profile/Soumitra_Mallick.

Soumitra Kumar Mallick

Following

Dear All,

Many thanks for your answers and suggestions. Appreciate the time taken by you to help me out on this.

Warm regards,

-Navin

I am stuck with same issue. Thanks for that guys. But I saw some papers that added a dummy 0 before the breaks and 1 after, also they multiply the independent variable by the dummy? What is the best way and what is difference?

Like y= c+ b x + b d + b dx +u

I keep going back to the determination of the break.

I want to ask something also.

Rather than to determine a structural break by nature of economic volatility (or test not to determine itself) i want to evaluate the results of a sudden change ( a break) in an explanatory variable on dependent variable that how long this effect sustain? I study permanent income hypothesis and i need a specific structural break that sudden changes in the average expected permanent income causes long term effects on household´s consumption.How i can specify this break in the model without general tests that shows stems of the break is reflected on constant or trend ?

Defne Alpdogan Navin Perera I have the same question, do I use a insignificant dummy or do leave it out of my model?

Awad Asiri Hi,

There are primarily 2 ways to introduce structural breaks in a model:

1) Put only dummy variable: This helps in correcting the model intercept only but is unable to handle the change the slope.

2) Put dummy variable and its interaction with other independent variables (by multiplication): This helps in correcting both i.e. model intercept and slope.

Ideally, one should introduce break in model by having dummy along with its interaction term. However, the loss of degree of freedom (due to increased model parameters) is the price for it. So, if you have large datasets, pl go ahead.