[R-sig-ME] multilevel ZINB mixture model (interpreting interactions with time)

[MARIA OLGA VIEDMA SILLERO]

Dear all:

I am working with a multilevel Zero-inflated mixture model for longitudinal data. Concretely, I am predicting fire frequency and excess of zeros over 445 spatial units across 35 years. I am running the models using MCMCglmm package in R.

I used the ID of the units and Time (as categorical variable) as random factors, and time (in continuous) and other explanatory variables as fixed effects. Moreover, I tested the interaction between time (continuous: 0-34 years) and other explanatory covariates (both, time-no variant and time-varying) in both parts of the model (the count part: negative binomial and the zero-part: logistic).

My problem resides on how interpret the interactions with time. If I understood well, the coefficient of the fixed effect in longitudinal models is the averaged effect of any co-variable on the outcome across time, and the interaction of any co-variable with time indicates the change of that averaged effect at each time step.

However, I observed that in almost all co-variables when the fixed effect was negative, the interaction with time resulted positive, while the opposite was observed when the mean fixed effect was positive. Actually, I don´t know how interpret this results. Will it be that the averaged effect (for example negative) is stronger (more negative) at each time step if the sign of the interaction is positive? Or that the averaged negative effect is turning more positive across time?

Thanks in advance,

Olga Viedma

Dª. Olga Viedma Sillero
Profesora Ordenación del Territorio
Facultad de Ciencias del Medio Ambiente y Bioquímica
Universidad de Castilla-La Mancha
Avd/ Carlos III, s/n. 45071 Toledo
Tel: 925 268800 (ext. 5780)
Email: olga.viedma at uclm.es
http://blog.uclm.es/grupofuego


	[[alternative HTML version deleted]]