In binomial regression, I have a trial where the model predicts 2 logits (88% predicted success rate) and one data point is 10 successes out of 10 trials (100% observed success rate). What is the residual for this particular data point?

My first hunch was to logit-transform this data point (logit(10/10)) and calculate the difference in logit-space. Naturally, this is wrong since the residual is infinite when $successes = trials$. The same would be true for Bernoulli models.

I am using JAGS to write an AR(N) (autoregressive) binomial model as part of a project to infer change points in time series. The autoregressive coefficient predicts the data point $i$ from the $residual_{i-1}$, so this is where the need to compute residuals steps in. I can see the literature on Pearson residuals as well as deviance residuals, but I am unsure how to relate them to the logit scale. There is a section that looks relevant here, but it is quite involved with matrix multiplication.



Your Answer

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