Ways to compare SITAR models for fit

Description

BICadj and AICadj calculate the BIC and AIC for SITAR models, adjusting the likelihood for Box-Cox transformed y variables. varexp calculates the variance explained by SITAR models, compared to the corresponding fixed effect models.

Usage

BICadj(..., pattern)
AICadj(..., k = 2, pattern)
varexp(..., pattern)

Arguments

Details

The deviance is adjusted if the y variable is power-transformed, using the formula

deviance - 2n ( (lambda-1) * log(gm) + log(abs(lambda)) )

where lambda is the power transform, and n and gm are the length and geometric mean of y.

The variance explained is given by

% explained = 100 * (1 - (sigma2/sigma1)^2)

where sigma1 is the fixed effects RSD and sigma2 the SITAR random effects RSD.

BICadj and AICadj accept non-sitar models with a logLik class. varexp ignores objects not of class sitar.

Value

For BICadj and AICadj a named vector of deviances in increasing order. For varexp a named vector of percentages in decreasing order.

Author(s)

Tim Cole tim.cole@ucl.ac.uk

See Also

BIC, AIC

Examples

data(heights)
## fit sitar model for height
m1 <- sitar(x=age, y=height, id=id, data=heights, df=5)

## update it for log(height)
m2 <- update(m1, y=sqrt(height))

## compare variance explained in the two models
varexp(m1, m2)

## compare BIC adjusting for sqrt transform
## the pattern matches names starting with "m" followed by a digit
BICadj(pattern="^m[0-9]")

Results