How does the INLA model handle 0 values in continuous variables?

[Wen FU]

Dear all.

As I checked inla.list.models() and found some zero-inflated models for counting data, however, if I have a continuous data set including many 0 values. In addition to the exponential model, is the other choice provided by Inla?

Would you have some likelihoods such as zero-inflated gamma likelihood or zero-inflated lognormal likelihood? Or we can modify some parts of the codes by ourselves?

Thank you in advance for your help!

Wen

[Denis Rustand]

Hi Wen,

For zero-inflated continuous distributions, an additional option that you can fit with R-INLA is the two-part model. It consists in a decomposition of the distribution of your zero-inflated continuous outcome into a binary outcome (probability of zero) and a positive continuous outcome (where you can specify a lognormal or gamma likelihood). It corresponds to the hurdle model for zero-inflated count data. We made a paper where we fit this model to a longitudinal zero-inflated continuous outcome jointly with survival times, see https://arxiv.org/pdf/2010.13704.pdf

And you can take a look at the associated R code available here: https://github.com/DenisRustand/TPJM_sim/blob/master/TPJM_INLA.R

Best,
Denis Rustand
he...@r-inla.org

[Wen FU]

Thank you so much, Denis!  It's very helpful