backward {CoCoRaw}R Documentation

Stepwise model selection by backward elimination and forward selection of edges or interaction terms

Description

This function perform (depth first) search for a single model with all edges or interaction terms significant (backward), or all terms not in the model not significant (forward). The function can also be used to compute tests for all terms in the model or all terms not in the model.

Usage

backward(sorted = FALSE, reversed = FALSE, only = FALSE, short = FALSE, 
         p.accepted = FALSE, p.rejected = FALSE, decomposable.mode = NULL, 
         coherent = FALSE, headlong = FALSE, recursive = FALSE, 
         follow = FALSE, 
         least.significant = TRUE, 
         components = NULL, p.components = FALSE, 
         separators = FALSE, p.separators = FALSE, 
         edges = TRUE, model = FALSE, fix.edges = NULL, 
         return.tests = FALSE, data = NULL, 
         object = .object.of.model(model, data = data, ...), ...)

forward(sorted = FALSE, reversed = FALSE, only = FALSE, short = FALSE, 
        p.accepted = FALSE, p.rejected = FALSE, decomposable.mode = NULL, 
        coherent = FALSE, headlong = FALSE, recursive = FALSE, 
        all.significant = TRUE, 
        components = NULL, p.components = FALSE, 
        separators = FALSE, p.separators = FALSE, 
        edges = TRUE, model = FALSE, fix.edges = NULL, 
        return.tests = FALSE, data = NULL, 
        object = .object.of.model(model, data = data, ...), ...)

Arguments

sorted Logical, if sorted is TRUE then a list of tests is printed sorted according to the selected p-value (statistic, if IC is selected). For backward the p-values are ordered in increasing order (IC decreasing), and for forward the p-values are ordered in decreasing order (IC increasing). The least significant edge is then last for backward, and the most significant edge is then last for forward.
reversed Logical, if reversed is TRUE then the sorted list is printed in reverse order.
only Logical, if only is TRUE then only the sorted list is printed.
short Logical, if short is TRUE then only a short report is printed: The test of the edge removed (entered) at each cycle.
If short is TRUE and dump (optionsCoCo) also is TRUE then for each completed set of backward (e.i. each removal of an edge or interaction terms), the dump file is rewinded, and a report of rejected, accepted, eligible terms and the model resulting of the step is printed on the file. In the following step of the backward elimination each tested term and the selected test statistic for the test of the term is printed on the dump file. Similar for forward.
p.accepted Numeric, edges and interaction terms with tests with p-value (or minus the IC value) greater than p.accepted are accepted for removal in backward.
If coherent is TRUE for forward then terms with p-value (or minus the IC value) greater than p.accepted are not eligible for entering in sub sequential cycles of the forward selection.
p.rejected Numeric, and interaction terms with tests with p-value (or minus the IC value) smaller than p.rejected are rejected, and thus entered into the model in forward (according to all.significant).
If coherent is TRUE for backward then terms with p-value (or minus the IC value) smaller than p.rejected are not eligible for removal in sub sequential cycles of the backward elimination.
decomposable.mode Logical, if decomposable.mode is TRUE then only decomposable models are visited.
coherent Logical, if coherent then the principle of coherence is applied: Once an edge in backward (forward) is rejected (accepted) it is no more tested for removal (entering).
headlong Logical, if headlong is TRUE then in each cycle, e.i. visiting of the edges or interaction terms of the model, terms are only visited until a term is found to remove or enter. In backward edges are visited until an edge with a p-value (or minus the IC value) greater than p.accepted is found, and in forward edges are visited until an edge with a p-value (or minus the IC value) smaller than p.rejected is found.
recursive Logical, if recursive is TRUE then cycles of visiting edges or interaction terms are repeated until no more edges are removed in backward and until no more edges are added in forward.
follow Logical, if follow is TRUE then tests in backward are performed against the model selected in the previous cycle, else tests are performed against the BASE model. The option is not available for forward.
least.significant Logical, if least.significant is TRUE then the least significant edge is removed in backward, else all non significant edges are removed. The option is not available for forward.
all.significant Logical, if all.significant is TRUE then all significant edges are entered in forward, else only the most significant edge is entered. The option is not available for backward.
components Logical, if components is TRUE, then common decompositions of the two models of each tests are found. If any p-value (or minus the IC value) of the tests on the resulting components is less than p.components then the test is rejected.
p.components Numeric, edges with a p-value (or minus the IC value) less than the value p.components for any component is rejected, if components is TRUE.
separators Logical, if separators is TRUE, then for each edge all separators of the two vertices of the edge is found, and the edge is rejected, if for any of these separators the two variables not are conditional independent, determined by p.separators.
p.separators Numeric, edges with a p-value (or minus the IC value) less than the value p.separators for any separator is rejected, if separators is TRUE.
edges Logical, if edges is TRUE, then in each cycle edges of the model are visited, else the maximal interaction terms are visited.
model See returnModel.
fix.edges A character string with a generating class. Edges in fix.edges are not eligible for removal in backward, and are not eligible for entering in forward. if fix.edges is NULL, default, the fixing of the previous backward or forward is retained. Use fix.edges = "" to clear fixing of edges.
return.tests Logical, if return.tests is TRUE then the test values are returned in a matrix. (recursive should be FALSE.)
data See exportCoCo.
object See exportCoCo.
... Additional arguments to generate the CoCo object from the data argument. See propertyModel.

Value

TRUE

Note

Note also the options of the sections "tests" (optionsCoCo).

p.accepted, p.rejected, p.components, p.separators when set to a value not FALSE will set these options for the CoCo object. If the arguments are not given to backward and forward the the values of the CoCo objects will be used.

The arguments decomposable.mode and separators when different from NULL sets respectively the options decomposable.mode and partitioning.

Note that to use BIC both IC and BIC has to be set to TRUE, since IC controls whether to use p-values or IC-values, and BIC controls which IC-value to use.

Author(s)

Jens Henrik Badsberg

References

Badsberg, J.H.: A guide to CoCo, JSS, 2001 ( http://www.jstatsoft.org/v06/i04/ ) and Badsberg, J.H.: Xlisp+CoCo, Aalborg University, 1996.

See Also

editModel and eh.


[Package CoCoRaw version 0.1.7.4 Index]