Generate dataset for simulation (non-i.i.d panel data)

[Fahad Mirza]

Hello everyone,

Hope you all are well and keeping safe.

I have a question for the community here and I have tried looking for a method online but could not find a solution which is why I am posting my request here.

I have to construct a Hausman test for non-i.i.d panel data and check it's efficacy against traditional one. For that I need to generate a dataset having small T and large N.

At the moment I am thinking 20 time periods and 50 observations per time period (total 1000 observations).

I am unable to figure out how to generate such a dataset. I looked at this presentation by Christopher F. Baum (which had useful information on simulations) however, I am still stuck at square one.

I will really appreciate the help on this.

Thank you!

[Ken Chui]

Hello, my understanding is that you're not sure how to set up that scaffold? Try this:

clear
set obs 50
gen id = _n
expand 20
bysort id: gen timeunit = _n

[Fahad Mirza]

Ken Chui thank you for the reply. I feel my post may have caused this confusion. I am unable to understand how I can create a Non-identical and Non-independent dataset. The scaffold is something I am familiar with but unable to figure out what makes a dataset non-iid. I hope I am able to convey this

[Joseph Coveney]

I am unable to understand how I can create a Non-identical and Non-independent dataset. The scaffold is something I am familiar with but unable to figure out what makes a dataset non-iid.

You could induce dependence with a shared random effect. You could render the distributions nonidentical by varying the residual variance parameter. Maybe consider something along the lines of the following.

version 16.1

clear *

set seed `=strreverse("1602717")'

quietly set obs 50
generate byte pid = _n
generate double pid_u = rnormal()
generate double var_e = runiform(1, 4)

quietly expand 20
bysort pid: generate byte tim = _n

generate double out = 0 + 0 * tim + ///
    pid_u + /// <- nonindependent
        rnormal(0, sqrt(var_e)) // <- not identically distributed residuals

exit

I think that most would try to induce some form of autocorrelation between the residuals, too.

I hesitate to consider N = 50 and T = 20 "a dataset having small T and large N".