Overarching questions about Gaussian models

[psolymos]

Non-centered covariates in Gaussian models can lead to negative estimates. This is undesirable from various reasons:

• Lorenz-curve based thresholding will not work, and in the absence of that, we cannot assign split labels to multicut1 objects.
• mu1/mu0 type prior indicator potential measures will be incorrect (even x + abs(mu0) will be NaN due to division by 0). mu0=0 (or -10^-14) is quite likely when data has lots of 0.
• Lognormal, Gamma etc models have 0 mass at 0, thus require ZI component for almost all practical purposes

Having negative values, however, might be justified when one used non-abundance data, e.g. instrument measurements that can take negative values (and do not want to transform due to variances, 0s, etc)

Possible options:

• Do not care (no error/warning), let the users do what they want.
• Warning: give warning and a workaround (e.g. centering, fix_fitted option) for methods that otherwise fail (like bestpart, summary, plot). But the estimates can be processed further.
• Error: constrain the analysis workflow to abundance type data.

Right now the implementation uses warnings. This is maybe as good as it gets.

[psolymos]

Keep this note for the record, and maintain current implementation (warning + workarounds).