GP - integration out for multivariate

Hi -

I’ve already fit a mixed linear effects model with this data and started simple, similar to Gelman and Bufami 2007.

The researchers want to see the non-linear trend of a specific covariate, so I’m doing 1 in, 1 out GP for a time series. So I’m just fitting 1,2,…,N for X, with whatever response. (As I’mmore familiar with this than any spline techniques).

So the way to do this is increment the CDF after I’ve simulated from the model WRT all my known values. However, there is no simple solution as the CDF for multivariate normal distributions are not implemented.

Is there a better way to do what I’m doing?

More generally, if I have a GP in Stan with missing data as the y variable, how am I to model the outcome?

oh wait, pretty sure I can just exclude these values when I generate the covariance matrix. That or just model it. I should think more before I ask questions.

I’ll let myself know here how it turns out.

1 Like