I have some questions regarding the least square assumptions for causal and prediction models.

I know that in linear regressions, for the coefficient on the regressors to have causal interpretation, the conditional expectation of the error term $u_i$ on $X_i$'s has to be zero. And I always thought that if we are conducting a regression for prediction instead of causal interpretation, then the error term does not have to be zero.
However, I was studying Autoregression today and the textbook writes "The assumption that the conditional expectation of $u_t$ is zero given past values of $Y_t$––that is, $E(u_t|Y_{t-1},Y_{t-1}...)=0$––has two important implications", and I was confused.
Is this because the way the two regression models are defined are different?

  • $\begingroup$ If your autoregressive model is $Y_t=\beta_0+\beta_1Y_{t-1}+\beta_2Y_{t-2}+\cdots+u_t$ then is it very different from your linear model $Y_i=\beta_0+\beta_1X_i+u_i$? $\endgroup$ – Henry Jun 10 at 13:48
  • $\begingroup$ Oh, let me rewrite my question in a more general way. $\endgroup$ – WinnieXi Jun 10 at 14:11
  • $\begingroup$ Having conditional expectation of the error term equal to zero is not sufficient for a causal interpretation. Using the linear model you should almost always use phrases like "A w-percent increase in x is associated (not causes!) with a y-percent increase in z". $\endgroup$ – stollenm Jun 11 at 9:58

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.