Estimating the Average Treatment Effect of Binary and Count Outcome.

[pavan pandey]

Hi There,

I am trying to estimate the average effect of treatment using the ‘didregress’ command. However, I have a pressing concern. There are five different variables that represent the effects of treatment and each of these variables is binary e.g., whether taking the recommended diet or not; Yes== 1 and No==0.

My question is, Can I estimate the average treatment effect using the ‘didregress’ command if the outcome is binary?

I looked into the Stata 17 manual and it says “didregress estimates the average treatment effect on the treated (ATET) from observational data by difference-in-differences (DID) or difference-in-difference-in-differences (DDD). The ATET of a binary or continuous treatment on a continuous outcome is estimated by fitting a linear model with time and group fixed effects”.

I wish to calculate the average treatment effect for each of the five different binary variables individually and then combine all five variables into 1 variable which will be a count/ordinal variable having the fixed values of 0,1,2,3, 4, and 5 (adding all the five binary variables).

Can I use ‘didregress’ command for the individual binary outcome variables?

Can I use ‘didregress’ command for a count variable created by combining all five variables?

If I cannot use DiD methodology, what is the next best approach to estimate the average treatment effect in the above scenario?


Thank you very much.

[William Rossi]

Hi Pavan,
I hope I understood correctly: you have 1 treatment variable and 5 binary dependent variables. Right?

Yes, you can estimate a linear regression with a binary outcome and obtain a causal parameter.

For question 2, it depends on what type of variable is the "combined" variable. If it is a count variable, you can for sure start with a linear regression and look at what happens but I would recommed also to look at other kinds of estimators (e.g. poisson, multinomial, ordered logit/probit. !!!please note that each estimator has specific characteristics that vary with the kind of dependent variable you have).

[pavan pandey]

Hi William,

The treatment/intervention is advice given by a healthcare worker (Yes==1 and No==0).

Outcome (effect of treatment) is advised followed by the patient (Yes==1 and No==0). There are five such different outcome variables. My question is can I use DiD methodology if the outcome variable is BINARY?

[William Rossi]

As I said:

Yes, you can estimate a linear regression with a binary outcome and obtain a causal parameter.

didregress is fine since it uses a linear model to estimate the coefficients.

[pavan pandey]

Thank you William. I will get back to you if I have any more questions.

Thank you very much!

Namaste.