Structural break analysis with panel data

[Michael Joseph]

I am using stata 14. I have panel data of form country-year data (1 observation is a particular year in a particular country) I am interested in the effect of housing prices and conflict. I want to test for a structural break in the relationship in a particular year. So I set my panels using xtset then run a logistic regression. However, I cannot use the structural break command sbsingle. I get an error that I haven't set it as time time series. However, I cannot set the data as time series (tsset) because I have multiple observations per year. Is there a way I can test for a structural break in the relationship between two variables that are of panel data form?

[Ashish Rajbhandari (StataCorp)]

Michael Joseph asked about using the command -estat sbsingle- for testing a
single unknown structural break for panel data.

"I have panel data of form country-year data ... I want to test for a
structural break in the relationship in a particular year ..."

The command -estat sbsingle- was implemented specifically for time-series
data with large T and N=1. Inference based on the test statistics of -estat
sbsingle- does not apply to a panel-data setting with N>1.

If Michael is willing to suppose that the break point is known, he could
implement a Wald test for structural break. Here is a simple example to test
for a structural break with a known break date for a random effects logistic model using -xtlogit-.

* Generate data for a random effect logistic model with a structural break in the intercept
* and the slope term
clear
set seed 2015
set obs  200                   
generate id = _n
generate ui = rnormal()        
expand 100                    
bysort id: generate time = _n
generate x  = rchi2(1)
generate e  = rlogistic()      
generate y = cond(time<=35, (-.25 + .3*x) + ui + e, (-.40 + .5*x) + ui + e)
replace y = (y>0)

* Estimate the model after creating an indicator variable for a known break date at t = 36
xtset id time
generate break = (time>35)     
xtlogit y ibn.break ibn.break#c.x, nocons coeflegend

* Test for a known break date
test (_b[0bn.break] = _b[1.break]) (_b[0bn.break#c.x] = _b[1.break#c.x])

- Ashish Rajbhandari

[Michael Joseph]

Ashish write: "If Michael is willing to suppose that the break point is known, he could implement a Wald test for structural break."

Thanks Ashish, but what if I don't want to assume a particular date ?

Basically, I have a theoretical reason that a break exists at a particular time, and I have used the Wald and Chow tests to show it is in fact real. But the criticism I get in review is that the break may be in another year nearby. Is there a non-parametric method to test that given I have panel data?

Thanks

Michael

[Ashish Rajbhandari (StataCorp)]

Michael Joseph asked if there was a non-parametric method to test for an unknown structural break in panel data. As of now we do not have this feature in Stata 14.

- Ashish

[Daisy]

Hi Ashish, reading your answer, I guessed you perform a Chow test for a known break. As Michael also concerned about unknown break, can we do the same for all year, and see if there is a break. Or if policy reform hapenned, we can expect the change 1,2 years after?

[Aaditya Dar]

Hi Ashish, one follow up question: what is the restricted model in your example? If it is given by:

xtlogit y x, nocons

then shouldn't the unrestricted model be:

 xtlogit y ibn.break##c.x, nocons

[Jahangir Alam]

Hi Ashish, how should we use this break if we have continuous dependent variable instead of dummy variable? I would like to check only the intercept. Could I find how much does the intercept differ due to this natural break?