I estimated a linear probability model (LPM) $P(y=1|x_1) = b_0 +b_1x_1 + u $ and a probit model $P(y=1|x_1) = \Phi(b_0 +b_1x_1 + u) $, where $\Phi()$ denotes the cumulative normal distribution. The regressor $x_1$ is the same binary variable.

I observed that the predicted probabilities for both models $\hat{y} =\hat{b_0} +\hat{b_1}x_1 $ and $\hat{y} =\Phi(\hat{b_0} + \hat{b_1}x_1)$ are identical across both models. Why is this the case from a theoretical point of view?

  • 2
    $\begingroup$ Assuming $y$ is binary, $y = b_0 +b_1x_1 + u$ is not a linear probability model (LPM), just a usual linear regression model. $P(y=1\mid x_1) = b_0 +b_1x_1 + u$ is the specification for the LPM. Also $y = \Phi(b_0 +b_1x_1 + u)$ is not a probit model - $P(y=1\mid x_1) = \Phi_u(b_0 +b_1x_1)$ is. Are these just typos of the question, or you actually estimated the specifications that you write? $\endgroup$ Dec 18, 2013 at 19:52

1 Answer 1


Because the model's saturated: with a binary response & a single binary predictor you get a perfect fit no matter what link function you use, with the estimated probability being the observed probability for each value of $x_1$. Think of fitting a line to two points: it doesn't matter whether it's straight or curvy; if it has two parameters it'll go through them.

  • $\begingroup$ thanks a lot! if i add more explanatory variables (continuous and binary) the fitted probabilities are not identical anymore, but still highly correlated (~98% of correlation). Why is this the case? $\endgroup$ Dec 18, 2013 at 18:19
  • $\begingroup$ well, it's to be hoped that the fitted probabilities aren't wildly different for sensible link functions within the range of predictors in the sample. $\endgroup$ Dec 18, 2013 at 18:27

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.