Lucid.OsteoNorm.FitModel¶
This task type fits a statistical distribution model to a table of normative data sample measurements.
Centile curves are modelled using GAMLSS (Generalized Additive Models for Location, Scale and Shape).
Parameters¶
μ (mu): location (median for BCPE)
σ (sigma): scale (approximate coefficient of variance for BCPE)
ν (nu): skewness (transformation to symmetry for BCPE)
τ (tau): kurtosis (power of the exponential for BCPE)
mu and sigma are smoothed with cubic P-splines (cf. Rigby and Stasinopoulos). nu and tau can also be smoothed with cubic P-splines, or can be constant. (GAIC (global optimization of hyperparameters using Generalized Akaike Information Criterion) is not used.)
Distribution family¶
For now BCPE is always used.
NO: Normal (Gaussian) distribution
BCCG: Box-Cox Cole and Green distribution
BCPE: Box-Cox Power Exponential
BCT: Box-Cox t
SHASH: Sinh-Arcsinh distribution (Jones, 2005)
SHASHo: Sinh-Arcsinh distribution (Jones and Pewsey, 2009)
Scoring algorithms¶
The RS (Rigby and Stasinopoulos) algorithm is a modified version of the Fisher-scoring algorithm.
The CG (Cole and Green) algorithm is a Newton-Raphson approach suitable when parameters are highly correlated.