[R-sig-eco] Hurdle model: singular variance-covariance matrix

[Dixon, Philip M [STAT]]

Yaiza,

The problem is that two coefficients are perfectly correlated with each other.  Since this happens in the models with interactions but not additive effects, my hunch is that the cross-product variable is very highly correlated with one of the original variables.  This can happen if one variable has a large mean and large variability and the second has a large mean and small variability.  If so, this is a numerical issue, not a conceptual, statistical or ecological issue.

I suggest you center all X variables to mean 0.  That often clears up numerical issues.  If it doesn't standardizing each to unit variance can also help.  Centering the X variables also has the statistical advantage that the intercept is directly interpretable.

Best wishes,
Philip Dixon