Having performed a linear regression, I can find the confidence interval for the response conditioned on a particular x value. However, I am interested in a C.I for the mean response for a set of N new observations. That is, I need to combine the N prediction intervals.

The closest post I could find was Calculating the mean using regression data, but it only handles the univariate case.

I tried deriving the standard error of the mean response below, but I'm not sure if this correct.

$\begin{align} var(\hat{\bar{y}}) &= var \left( \frac{1}{n} \sum_i \hat{y}_i|x_1 \ldots x_n \right) \\ &= var \left( \frac{1}{n} \sum_i \hat{y}_i|x_i \right), \quad \text{where the } \hat{y_i}|x_i \text{ are independent} \\ &= \frac{1}{n^2} \sum_i var(\hat{y}_i|x_i) \\ \end{align}$

where $var(\hat{y}_i|x_i) = \sqrt{\sigma^2 x_i^T (X^TX)^{-1}x_i}$ for $x_i$ in the training data and $var(\hat{y}_i|x^*_i) = \sqrt{\sigma^2 (1+ x_i^{*T} (X^TX)^{-1}x^*_i)}$ for $x^*_i$ in the test data.

Am I on the right track here? Also, is there an R implementation somewhere, or should I do it from scratch?

Edit: I am also reading up on Bayesian regression methods which specify the predictive distribution $P(y_i|x_i^*)$, and a credible interval for the response. We face a similar problem here, namely, how to compute the predictive distribution for the mean response $P(\overline{y}|x_1^* \ldots x_n^*)$?



Your Answer

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

Browse other questions tagged or ask your own question.