cond.rsm {marg}

Approximate Conditional Inference in Regression-Scale Models

Description

Performs approximate conditional inference on a scalar parameter of interest in regression-scale models. The output is stored in an object of class marg.

Usage

## S3 method for class 'rsm'
cond(object, offset, formula = NULL, family = NULL, 
     dispersion = NULL, data = sys.frame(sys.parent()), pts = 20, 
     n = max(100, 2*pts), tms = 0.6, from = NULL, to = NULL, 
     control = glm.control(...), trace = FALSE, ...)

Arguments

object

a rsm object; for instance the result of a call to rsm.

offset

either the covariate occurring in the model formula whose coefficient represents the parameter of interest or scale if the parameter of interest is the scale parameter. In the first case, the variable may be numerical or a two-level factor. In case of a two-level factor, it must be coded by contrasts and not as two dummy variables. Can also be a call to a mathematical function (such as exp, sin, ...) or to a mathematical operator (^, /, ...) applied to a numerical variable. The call must always agree with the label used to identify the corresponding parameter in the rsm object passed through the object argument or defined by formula and family. Beware that the label includes the identity function I() if an arithmetic operator was used. Other function types (e.g. factor) and interactions are not admitted. If interest focuses on the scale parameter, it must not be fixed in object or when using the dispersion argument in case no rsm object is supplied.

formula

a formula expression (only if no rsm object is defined).

family

a family.rsm object defining the error distribution (only if no rsm object is defined). See rsm.families for the families supported.

dispersion

argument only to be used if no rsm object is defined. If NULL, the scale parameter is taken to be unknown. If known, the numerical value can be passed. The default is NULL. Huber's least favourable error distribution represents a special case. If dispersion is NULL, the maximum likelihood estimate is computed, while if TRUE the MAD estimate is calculated and the scale parameter fixed to this value in subsequent computations.

data

an optional data frame in which to interpret the variables occurring in the formula (only if no rsm object is defined).

pts

number of output points (minimum 10) that are calculated exactly; the default is 20.

n

approximate number of output points (minimum 50) produced by the spline interpolation. The default is the maximum between 100 and twice pts.

tms

defines the range MLE +/- tms * s.e. where cubic spline interpolation is replaced by polynomial interpolation. The default is 0.6.

from

starting value of the sequence that contains the values of the parameter of interest for which output points are calculated exactly. The default is MLE - 3.5 * s.e.

to

ending value of the sequence that contains the values of the parameter of interest for which output points are calculated exactly. The default is MLE + 3.5 * s.e.

control

a list of iteration and algorithmic constants that control the rsm fit. See glm.control for their names and default values.

trace

if TRUE, iteration numbers will be printed.

...

additional arguments, such as weights, subset, control etc. used by the rsm fitting routine if the rsm object is defined through formula and family. See rsm for their definition and use.

Details

This function is a method for the generic function cond for class rsm. It can be invoked by calling cond for an object of the appropriate class, or directly by calling cond.rsm regardless of the class of the object. cond.rsm has also to be used if the rsm object is not provided throught the object argument but specified by formula and family.

The function cond.rsm implements several small sample asymptotic methods for approximate conditional inference in regression-scale models. Approximations for both the modified/marginal log likelihood function and approximate conditional/marginal tail probabilities are available (see marg.object for details). Attention is restricted to a scalar parameter of interest, either a regression coefficient or the scale parameter. In the first case, the associated covariate may be either numerical or a two-level factor.

Approximate conditional (or equivalently marginal) inference is performed by either updating a fitted regression-scale model or defining the model formula and family. All approximations are calculated exactly for pts equally spaced points ranging from from to to. A spline interpolation is used to extend them over the whole interval of interest, except for the range of values defined by MLE +/- tms * s.e. where the spline interpolation is replaced by a higher order polynomial interpolation. This is done in order to avoid numerical instabilities which are likely to occur for values of the parameter of interest close to the MLE. Results are stored in an object of class marg. Method functions like print, summary and plot can be used to examine the output or represent it graphically. Components can be extracted using coef, formula and family.

Main references for the methods considered are the papers by Barndorff-Nielsen (1991), DiCiccio, Field and Fraser (1990) and DiCiccio and Field (1991). The theory and statistics used are summarized in Brazzale (2000, Chapters 2 and 3). More details of the implementation are given in Brazzale (1999; 2000, Section 6.3.1).

Value

The returned value is an object of class marg; see marg.object for details.

Note

If the parameter of interest is the scale parameter, all calculations are performed on the logarithmic scale, though most results are reported on the original scale.

In rare occasions, cond.rsm dumps because of non-convergence of the function rsm which is used to refit the model for a fixed value of the parameter of interest. This happens for instance if this value is too extreme. The arguments from and to may then be used to limit the default range of MLE +/- 3.5 * s.e. A further possibility is to fine-tuning the constants (number of iterations, convergence threshold) that control the rsm fit through the control argument.

cond.rsm may also dump if the estimate of the parameter of interest is large (tipically > 400) in absolute value. This may be avoided by reparametrizing the model.

References

Barndorff-Nielsen, O. E. (1991) Modified signed log likelihood ratio. Biometrika, 78, 557–564.

Brazzale, A. R. (1999) Approximate conditional inference for logistic and loglinear models. J. Comput. Graph. Statist., 8, 653–661.

Brazzale, A. R. (2000) Practical Small-Sample Parametric Inference. Ph.D. Thesis N. 2230, Department of Mathematics, Swiss Federal Institute of Technology Lausanne.

DiCiccio, T. J., Field, C. A. and Fraser, D. A. S. (1990) Approximations of marginal tail probabilities and inference for scalar parameters. Biometrika, 77, 77–95.

DiCiccio, T. J. and Field, C. A. (1991) An accurate method for approximate conditional and Bayesian inference about linear regression models from censored data. Biometrika, 78, 903–910.

See Also

marg.object, summary.marg, plot.marg, rsm

Examples

## Sea Level Data
data(venice)
attach(venice)
Year <- 1:51/51
c11 <- cos(2*pi*1:51/11) ; s11 <- sin(2*pi*1:51/11)
c19 <- cos(2*pi*1:51/18.62) ; s19 <- sin(2*pi*1:51/18.62)
##
## quadratic model fitted to the sea level, includes 18.62-year 
## astronomical tidal cycle and 11-year sunspot cycle
venice.rsm <- rsm(sea ~ Year + I(Year^2) + c11 + s11 + c19 + s19, 
                  family = extreme)
names(coef(venice.rsm))
## "(Intercept)"  "Year"  "I(Year^2)"  "c11"  "s11"  "c19"  "s19"      
##
## variable of interest: quadratic term
venice.marg <- cond(venice.rsm, I(Year^2))
##
detach()

## House Price Data
data(houses)
houses.rsm <- rsm(price ~ ., family = student(5), data = houses)
##
## parameter of interest: scale parameter
houses.marg <- cond(houses.rsm, scale)