Hi,
I am estimating a causal effect of policy intervention on bounded outcome as below.
I am running 2SLS regression to estimate the effect of X1 on Y. Since my endogenous X1 is binary while exogenous Z is continuous, I first run non-linear logit regression of X1 on Z and X2, and use this predicted X1, X1_hat, as an instrument.
The problem is that my 2SLS model has very poor model fit (often resulting in negative R2) since my Y is fractional. I found "ivfprobit" newly added in Stata 18, which allows me to run IV regression when my Y is fractional.
However, this command assumes all endogenous X being continuous, which is not true in my study (X1 is binary).
Q1. Is there a way to "relax" this assumption of endogenous X being continuous, or is there a way to circumvent this assumption?
What happens to this estimator if my X1 is binary? (biased, inconsistent, etc.)
Q2. This is an extra question: is there any statistic show model fit in "ivfprobit" command?
I attached the result screenshot of "ivfprobit" and "ereturn list", but I don't see any model fit-related statistic, like R2 or information criterion.
(I excluded other exogenous variables for simplicity in this screenshot).
Thank you.
I am estimating a causal effect of policy intervention on bounded outcome as below.
- Y: Probability, fractional (continuous and bounded from 0 to 1).
- X1: Endogenous, binary
- X2: Exogenous covariates.
- Z: Exogenous, continuous.
I am running 2SLS regression to estimate the effect of X1 on Y. Since my endogenous X1 is binary while exogenous Z is continuous, I first run non-linear logit regression of X1 on Z and X2, and use this predicted X1, X1_hat, as an instrument.
The problem is that my 2SLS model has very poor model fit (often resulting in negative R2) since my Y is fractional. I found "ivfprobit" newly added in Stata 18, which allows me to run IV regression when my Y is fractional.
However, this command assumes all endogenous X being continuous, which is not true in my study (X1 is binary).
Q1. Is there a way to "relax" this assumption of endogenous X being continuous, or is there a way to circumvent this assumption?
What happens to this estimator if my X1 is binary? (biased, inconsistent, etc.)
Q2. This is an extra question: is there any statistic show model fit in "ivfprobit" command?
I attached the result screenshot of "ivfprobit" and "ereturn list", but I don't see any model fit-related statistic, like R2 or information criterion.
(I excluded other exogenous variables for simplicity in this screenshot).
Thank you.
