1
$\begingroup$

Suppose you have a generalised linear model (GLM) with link function $g$ so that the true expected responses in your model (with $n$ data points) are:

$$\mu_i \equiv \mathbb{E}(Y_i | \mathbf{x}_i) = g^{-1}(\mathbf{x}_i \boldsymbol{\beta}) \quad \quad \quad \text{for } i = 1,...,n.$$

Now, suppose you have a set of known weights $\omega_1,...,\omega_n$ and you want to make an inference about the weighted sum of the expected responses:

$$\mu_\omega \equiv \sum_{i=1}^n \omega_i \mu_i = \sum_{i=1}^n \omega_i \cdot g^{-1}(\mathbf{x}_i \boldsymbol{\beta}).$$

This kind of problem arises in certain applications where you want to form some kind of "standardised" or "weighted" expected response using a set of exogenous weights. In most models of interest, the link function $g$ is nonlinear, so its inverse $g^{-1}$ is also nonlinear. Thus, the inference here is (usually) about a nonlinear function of the parameter vector $\boldsymbol{\beta}$. Presumably this means that forming a confidence interval involves use of the asymptotic normal distribution for the coefficient vector, combined with the delta method to deal with the nonlinear transform. I am aware of some general ways this could be done, but I'm not sure what is "best practice" here.


Question: What is the best way to form a confidence interval for $\mu_\omega$? How is this implemented using a glm model in R? (An example for a logistic regression would be lovely.)

$\endgroup$
1
$\begingroup$

What you're asking is essentially the general form of how do I get a confidence interval for the mean of a function broadcast over a sample. This should help you out. You're already 95% of the way there.

Thus, the inference here is (usually) about a nonlinear function of the parameter vector $\beta$. Presumably this means that forming a confidence interval involves use of the asymptotic normal distribution for the coefficient vector, combined with the delta method to deal with the nonlinear transform. I am aware of some general ways this could be done, but I'm not sure what is "best practice" here.

This is precisely what you should do if you want a confidence interval. If the data are independent, then it's fairly easy to derive. After you compute your Delta method distribution, you're just multiplying by a vector. So, $$ \sqrt{n}\left(w^T g^{-1}(X\beta) - w^T\mu_\delta \right) \sim N\left(0, w^T \Sigma_\delta w \right)$$

You have to be careful about how you define the covariance matrix you use in your Delta method approximation. Do you want a confidence interval for the mean (in which case you want to use the $\Sigma = \Sigma_\beta$ from the coefficient estimates)? Or do you want a prediction interval for the quantity (in which case you'll want to use $\Sigma = \Sigma_\beta + \sigma^2 I$ combining the prediction uncertainty as well)?

You should be able to derive the necessary quantities for logistic regression with this. You seem to know the Delta method already. You know $g^{-1}$ for logistic regression, $\frac{\exp(X\beta)}{1 + \exp{X\beta}}$. You can get the necessary covariance matrix from glm. It's just a matter of plugging it in.

$\endgroup$

Your Answer

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

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