se.assoc {logmult}

Standard Errors for Association Models

Description

Get standard errors for log-multiplicative association scores and intrinsic association coefficients.

Usage

se(x, ...)

## S3 method for class 'assoc'
se(x, type = c("se", "quasi.se"), ...)

## S3 method for class 'rc'
se(x, type = c("se", "quasi.se"), ...)

## S3 method for class 'hmskew'
se(x, type = c("se", "quasi.se"), ...)

## S3 method for class 'yrcskew'
se(x, type = c("se", "quasi.se"), ...)

## S3 method for class 'rcL'
se(x, type = c("se", "quasi.se"), ...)

Arguments

x

an assoc object with a non-null covmat component (for se.assoc);. or a rc, hmskew, hmskewL, yrcskew, rcL or rcL.trans object fitted with the se argument different from “none” (for other functions).

type

the type of standard errors to be computed (see “Details” below).

...

currently unused.

Details

Currently, only jackknife or bootstrap standard errors are supported, depending on the se argument passed when fitting the model. Some care is needed before using such standard errors and confidence intervals. First one must ensure all model replicates converged to a correct solution, especially for bootstrap; second, when relying on normal confidence intervals computed from these standard errors, one must ensure that the coefficients estimators follow a normal distribution. Both checks can be performed by calling plot.boot on the boot.results component of the assoc object of the models (not supported for jackknife), with the index argument identifying the coefficient of interest (call colnames on the t member of the boot.results object to find out the index you need).

If outliers are present, standard errors and confidence intervals will be artificially large; to fix this, the tolerance argument must be set to a smaller value when fitting the models (which may in turn require increasing the value of the iterMax argument if convergence is too slow). Once outliers are removed, if coefficient estimates are still not normally distributed, robust bootstrap confidence intervals can be computed using boot.ci on the same object, provided a large number of replicates (> 1000) were computed.

For each replicate, stable scores and intrinsic association coefficients are identified using an orthogonal Procrustes analysis to suppress meaningless variations due to random reflections, permutations and rotations of dimensions (Milan & Whittaker, 1995). For hmskew and hmskewL models, a rotation within each pair of dimensions and a permutation of pairs of dimensions is performed, but no reflection as it would change the sign of intrinsic association coefficients.

Quasi-standard errors are computed using qvcalc. See the help page for this function for details and references about them.

Value

An object of the same form as the assoc component of the model, but with standard errors rather than the corresponding coefficients.

Author(s)

Milan Bouchet-Valat

References

Milan, L., and J. Whittaker (1995). Application of the Parametric Bootstrap to Models that Incorporate a Singular Value Decomposition. Journal of the Royal Statistical Society. Series C (Applied Statistics) 44(1), 31-49.

See Also

assoc, rc, hmskew, hmskewL, yrcskew, rcL, rcL.trans

Examples

  # See ?rc about Wong (2010)