Log-Multiplicative Models, Including Association Models
Functions to fit log-multiplicative models using 'gnm', with
support for convenient printing, plots, and jackknife/bootstrap
standard errors. For complex survey data, models can be fitted from
design objects from the 'survey' package. Currently supported models
include UNIDIFF (Erikson & Goldthorpe), a.k.a. log-multiplicative
layer effect model (Xie), and several association models: Goodman's
row-column association models of the RC(M) and RC(M)-L families
with one or several dimensions; two skew-symmetric association
models proposed by Yamaguchi and by van der Heijden & Mooijaart.
Functions allow computing the intrinsic association coefficient
(and therefore the Altham index), including via the Bayes shrinkage
estimator proposed by Zhou; and the RAS/IPF/Deming-Stephan algorithm.
2018-10-18 - Version 0.7.0
- Change scaling of weights in svygnm() to reflect a change in survey 3.34.
2017-11-15 - Version 0.6.5
- Add iac() function to compute intrinsic association coefficient, either
via a non-parametric estimator or via the semi-parametric Bayesian shrikage
estimator proposed by Zhou (2015).
- Remove norm argument to maor(), as the 1-norm variant isn't used.
2017-03-20 - Version 0.6.4
- Add vignette.
- No longer eliminate parameters by default in unidiff(), rcL() and anoasL().
Parameters can still be eliminated (for faster estimation) by passing
- Fix failures in plot.assoc() in corner cases.
- Improve plotting of RC(1) models.
- Rename conf.ellipses argument to plot.assoc() to conf.int.
- Change 'phi' and 'maor' components of 'unidiff' objects to vectors, with
one value for each layer. The value which was previously stored in these
components is the first value of the vector.
2016-01-15 - Version 0.6.3
- Fix computing standard errors when ncpus > 1.
- Add svygnm() function, equivalent for generalized nonlinear models
to svyglm() from package survey
- Add ras() function to run RAS/Demming-Stephan algorithm
- Allow regression-type (Goodman-Hout) symmetric association in hmskew models
- Allow computing symmetric and antisymmetric components of MAOR
- Change (undocumented) definition of unweighted MAOR
- Fix plot.assoc() with layers and diagonal when keeping default arguments
- Plot variable names next to arrow in hmskew models
- Fix calling anova() on logmult objects (noticed by Heather Turner)
- Accept type="n" and add=TRUE arguments in plot.assoc()
- Fix references in docs
2015-04-22 - Version 0.6.2
- Fix yrskew to be compatible with gnm version 1.0-8 (thanks to Heather Turner).
2015-01-13 - Version 0.6.1
- Support 1D association plots using dotchart().
- Add col.ellipses argument to plot.assoc() functions (based on work by Michael Friendly).
- Avoid plotting confidence bars outside of box in plot.assoc().
- Plot first layer instead of failing by default for heterogeneous RC-L.
- Return coordinates of plotted points from plot.assoc().
- Support three-way tables and NA cells in maor().
- Improve documentation for maor().
- Do not fail when dimension names are empty or not unique.
- Fix comments in documentation (spotted by Michael Friendly).
- Various small fixes.
2014-03-10 - Version 0.6
- COMPATIBILITY BREAK: store marginal weights for each layer instead
of only the global average. Saved models will have to be fit again
in order to be plotted.
- Report BIC/AIC variant equal to 0 for the saturated model.
- Allow fitting association models with complex survey designs
('survey' package) via svy* functions.
- Fix direction of skewness in hmskew and hmskewL models, which could
be wrong in some special cases.
- Added support for Mean Absolute Odds Ratio (MAOR) in new maor()
function, and with association and UNIDIFF models.
- Added support for supplementary/passive rows and columns for rc and
- For association model plots, make area of symbols proportional to the
weights rather than to their square.
- Plot average association by default for RC-L models, and support
rotating the axes to those with highest variance.
- New 'add' argument to plot.unidiff() to allow representing several
models in one figure; also change the default plot type to points.
- Accept start=NA to use good default values.
- Use eliminate more aggressively to greatly improve performance.
- Make fitting UNIDIFF models more robust by not constraining coefficient
which do not stricly need to be constrained (reported by Gina Potarca).
- Fix using multiple CPUs for jackknife/bootstrap.
- Drop snow support (the parallel package must be used instead).
- Use load balancing for large models with jackknife/bootstrap.
- Do not fail with tables without dimension names.
- Allow passing checkEstimability=FALSE to unidiff() for performance.
- Accept nd.symm=0 in hmskew() and hmskewL().
- Print layer coefficients from print.unidiff().
- Documentation fixes.
2012-09-29 - Version 0.5