Implementing Propensity Score Weights

[MB Ross]

Hello all,

Quick question about implementing propensity score weighting ala Hirano and Imbens (2001)

In Hirano and Imbens (2001) the weights are calculated such that w(t,z)= t + (1-t)[e(z)/(1-e(z))] where the weight to the treated group is equal to 1 and the weight for control is e(z)/(1-e(z))

My question is about how I use the pweight command in my estimation after I’ve calculated the propensity score… would I simply specific the weight variable as (1-e(z))/e(z) and point pweight to that variable…. or am I misunderstanding something about that command???

Thanks in advance!

[wbuchanan]

Please use full references, which is asked of all members of the community in the FAQ. I'm not sure about this particular use of propensity score weights, but if you look up the annotated output for the book Methods Matter on the UCLA stats website the example they show suggests the weights should be specified as analytic weights [aw = wgt] rather than p-weights.

[Jeff Wooldridge]

You should use the Stata command teffects with the ipw option. If you specify atet as the parameter of interest, it will properly compute the Hirano/Imbens weights. If I remember the reference correctly, Hirano and Imbens actually combine linear regression adjustment with IPW, and so it uses a weighted least squares procedure.

The generic command would be

teffects ipwra (y x1 ... xk, linear) (d x1 ... xk, logit), atet vce(robust)

[Melkamu Tadesse]

I appreciate if you extend the above fruitful discussion to my case.
I am trying to apply the DiD Method Combined with Propensity Score Matching.
Given the recommended weighted least square regression framework like this:
Δ𝑌𝑖𝑡 = 𝛼 + 𝛽𝑇𝑖 + 𝛾ΔΧ𝑖𝑖 + 𝜀𝑖𝑡 ,
(Where 𝛽 is the DiD estimator and T is treatment dummy)
And the weights in the above regression equation are equal to 1 for treated units and to Pˆ(X )/(1 – Pˆ(X )) for comparison units(both treatment group and control groups have their own weights/from Propensities). I got the propensities with pscore.(teffects then?)

How can I execute with STATA 13 then? I mean what could be the syntax in STATA 13? The weights are associated with treatment variable (which is dummy and 1 for those with treatment and 0 for those control groups).
Many thanks!

[Steve Samuels]

Welome to Statalist, Melkamu! This is a thread that ended over eight months ago. Start a new thread.

[Melkamu Tadesse]

Ok Steve, hope you will be the first on the new thread.