IPW with count treatment

[Luis Mijares Castaneda]

Hello, I was interested in using an IPW model to adjust for confounders on my treatment effects. My treatment effect -shooting- total number of fatal and non-fatal shootings (truncated) is a count variable. In contrast, my outcome variable, -dentist-, is a binary variable indicating a dental visit in the previous year. Is there any way to model the treatment with a ZIP or Gaussian distribution in an IPW framework? Are there alternative strategies for handling a continuous or count-valued treatment in IPW frameworks? Attached is the -dataex- of the variables.

* Example generated by -dataex-. For more info, type help dataex
clear
input byte(aid dentists shootings sex race)
 1 0 2 1 2
 2 1 1 0 0
 3 1 0 1 1
 4 0 0 0 3
 5 1 0 0 2
 6 1 1 1 0
 7 0 2 0 1
 8 1 0 1 2
 9 0 2 1 3
10 1 1 0 1
11 1 0 0 0
12 0 1 1 2
13 1 2 1 3
14 0 0 0 0
15 0 2 0 1
16 1 0 1 2
17 1 1 0 1
18 0 1 1 3
19 1 0 0 0
20 0 2 1 2
21 1 1 1 3
22 0 0 0 1
23 1 2 0 0
24 0 1 0 1
25 1 0 1 2
end
label values dentists dental
label def dental 0 "Dental visit within a last year", modify
label def dental 1 "Dental visit greater than a year", modify
label values sex gender
label def gender 0 "male", modify
label def gender 1 "Female", modify
label values race race
label def race 0 "0.White", modify
label def race 1 "1.Hispanic", modify
label def race 2 "2.Black", modify
label def race 3 "3.Other Race", modify

[Jeff Wooldridge]

A method that wouldn't account for the natural ordering of shootings is IPW where the treatment can take on more than one value. The Stata teffects command allows this. You could expand the definition of the variable to distinguish between fatal and nonfatal shootings, too. Except for some efficiency loss, which may be small, there's no real harm in modeling the different values of shooting as a multinomial outcome.

Because your outcome is binary, you might try

teffects ipwra (dental x1 ... xk) (shootings x1 ... xk)

This will first estimate a multinomial logit for the different levels of treatment and then a binary logit for whether there was a dental visit.

[Jeff Wooldridge]

I forgot to show that the model for dental should be logit:

 
 teffects ipwra (dental x1 ... xk, logit) (shootings x1 ... xk)

[Luis Mijares Castaneda]

Jeff Wooldridge, hello. Thank you for your answer. I had one more question. My data comes from a survey, so there's stratification and PSU. Do I need to incorporate this in the IPW regression adjustments?