Adding correlated errors to Insurance Loss Payments example


In the Advanced Bayesian Multilevel Modeling with the R Package brms vignette,

Example 3 covers a model for Insurance Loss Payments using a non-linear growth model. Since this is modeling cumulative payments over time, it makes sense to try adding in some error correlation between development ages dev for a given year AY.

Looking through the documentation, it seems like this could be done by adding

autocor = cor_ma(formula = ~ dev|AY, q = 1, cov = TRUE)

to the brm function. The warning says I have to use cov = TRUE for a non-linear model.

Is this the best way to implement the correlation?


I am hesitant to talk about a “best” approach. The specification of the MA correlation looks correct to me.



I understand the hesitation to say what is “best.” When making the model more complex by adding something like a correlation structure, I’d expect the convergence to take longer. But helps to advice that I’m generally on the right track.