Power analysis for difference-in-difference

[Gustav Egede Hansen]

Hi everybody

I would like to calculate the power of my difference-in-difference estimation. So far, I have only managed to calculate the power by comparing the posttreatment difference in means using this code:

power twomeans 43183.2 44557.72, sd1(19151.54) sd2(16492.9) n1(2926) n2(279) graph onesided
*the first mean is the controlgroup and the second mean is the treatmentgroup

I have looked at the different stata options for calculating power, but I cannot find anything about diff-in-diff. If anyone could help me out, I would be very grateful.

The pretreatment numbers are:

controlgroup:
mean: 41878.44, sd: 18931.28

treatmentgroup:
mean: 40983.5, sd: 19267.45

There are 12 observations per individual in eachtime period.


Best,
Gustav

[Fei Wang]

Gustav, a difference-in-differences estimation can be implemented with a multiple linear regression. Assume y is the dependent variable, d is the treatment group dummy, p is the post-treatment dummy, and w = d*p. Then the DiD estimate would be the coefficient of w in the following regression.

regress y d p w

For power calculation, you may use command -power pcorr-.

power pcorr r2_p, ntested(1) ncontrol(2) n(sample_size)

where "sample_size" is the sample size of the linear regression, "r2_p" is the squared partial correlation between y and w and can be obtained from -pcorr-, as below.

pcorr y d p w

[Gustav Egede Hansen]

Hi Fei

Thanks for the reply. I have two questions. Regarding:

power pcorr r2_p, ntested(1) ncontrol(2) n(sample_size)

Do I have to specify that indviduals are clustered? In my study, n is 76.920 with 3205 clusters.

And also could you explain why you specify:

ncontrol(2)

Is it because I need to specify that d and p is in the model?

Best, Gustav

[Fei Wang]

Do I have to specify that indviduals are clustered? In my study, n is 76.920 with 3205 clusters.

The sample size is the total number of individual observations.

Is it because I need to specify that d and p is in the model?

Yes.

[Gustav Egede Hansen]

Alright thank you Fei. I really appreciate you taking the time.


Best,
Gustav