Skip to contents

A non-finite marginal objective is usually read as the response leaving the family's support, and the error message used to say so. That is the wrong diagnosis for a whole class of models, and a misleading one: the Laplace approximation needs the log determinant of the inner Hessian, so an indefinite Hessian produces exactly the same symptom with the data entirely inside the support.

Usage

.inner_indefinite(obj, design, method, famname)

Arguments

obj

The MakeADFun object, evaluated at its starting values.

design

The design object.

method

"REML" or "ML".

famname

The family's name, for the message.

Value

A sentence describing the negative curvature, or NULL if the Hessian is unavailable or positive definite.

Details

The Box-Cox power exponential is the case in hand. Its log-likelihood is concave in the four intercepts alone – an intercept-only fit converges in half a second – but adding a single covariate column to mu puts three negative eigenvalues into the inner Hessian, every one of them a direction mixing that column with the sigma, nu and tau intercepts. No starting value repairs it: sweeping tau from 2 to 9 and nu from 1 to 2.5 never gets below two negative directions. It is a property of the family's parameterisation, not of the start.

Under "REML" the negative curvature lands in beta, which is declared random and so passes through the inner solve carrying no prior to convexify it. The penalized blocks are not the problem – their Gaussian prior leaves them comfortably positive definite. Hence the suggestion of a basis with no null space, which is what puts those columns under a penalty.