4.1-1 CRAN
4.1 svyquantile() has been COMPLETELY REWRITTEN. The old version is available
as oldsvyquantile() (for David Eduardo Jorquera Petersen)
svycontrast()'s improvements for statistics with replicates are now also there with
svyby(), for domain comparisons (Robert Baskin)
svyttest() now gives an error message if the binary group variable isn't binary
(for StackOverflow 60930323)
confint.svyglm Wald-type intervals now correctly label the columns (eg 2.5%, 97.5%)
(for Molly Petersen)
svyolr() using linearisation had the wrong standard errors for intercepts
other than the first, if extracted using vcov (it was correct in summary() output)
svyglm() gave deffs that were too large by a factor of nrow(design). (Adrianne Bradford)
svycoxph() now warns if you try to use frailty or other penalised terms, because they
just come from calling coxph and I have no reason to believe they work correctly
in complex samples (for Claudia Rivera)
coef.svyglm() now has a complete= argument to match coef.default(). (for Thomas Leeper)
summary.svyglm() now gives NA p-values and a warning, rather than Inf standard errors,
when the residual df are zero or negative (for Dan Simpson and Lauren Kennedy)
In the multigroup case, svyranktest() now documents which elements of the 'htest'
object have which parts of the result, because it's a bit weird (for Justin Allen)
svycontrast() gets a new argument add=TRUE to keep the old coefficients as well
twophase() can now take strata= arguments that are character, not just factor
or numeric. (for Pam Shaw)
add reference to Chen & Lumley on tail probabilities for quadratic forms.
add reference to Breslow et al for calibrate()
add svyqqplot and svyqqmath for quantile-quantile plots
SE.svyby would grab confidence interval limits instead of SEs if vartype=c("ci","se").
svylogrank(method="small") was wrong (though method="score" and method="large" are ok),
because of problems in obtaining the at-risk matrix from coxph.detail. (for Zhiwen Yao)
added as.svrepdesign.svyimputationList and withReplicates.svyimputationList
(for Ángel Rodríguez Laso)
logLik.svyglm used to return the deviance and now divides it by -2
svybys() to make multiple tables by separate variables rather than a joint table
(for Hannah Evans)
added predictat= option to svypredmeans for Steven Johnston.
Fixed bug in postStratify.svyrep.design, was reweighting all reps the same (Steven Johnston)
Fix date for Thomas & Rao (1987) (Neil Diamond)
Add svygofchisq() for one-sample chisquared goodness of fit (for Natalie Gallagher)
confint.svyglm(method="Wald") now uses t distribution with design df by default.
(for Ehsan Karim)
confint.svyglm() checks for zero/negative degrees of freedom
confint.svyglm() checks for zero/negative degrees of freedom
mrb bootstrap now doesn't throw an error when there's a single PSU in a stratum
(Steve White)
oldsvyquantile() bug with producing replicate-weight confidence intervals for
multiple quantiles (Ben Schneider)
regTermTest(,method="LRT") didn't work if the survey design object and model were
defined in a function (for Keiran Shao)
svyglm() has clearer error message when the subset= argument contains NAs (for Pam Shaw)
and when the weights contain NAs (for Paige Johnson)
regTermTest was dropping the first term for coxph() models (Adam Elder)
svydesign() is much faster for very large datasets with character ids or strata.
svyglm() now works with na.action=na.exclude (for Terry Therneau)
extractAIC.svylm does the design-based AIC for the two-parameter Gaussian model, so
estimating the variance parameter as well as the regression parameters.
(for Benmei Liu and Barry Graubard)
svydesign(, pps=poisson_sampling()) for Poisson sampling, and ppscov() for
specifying PPS design with weighted or unweighted covariance of sampling indicators
(for Claudia Rivera Rodriguez)
4.0 Some (and eventually nearly all) functions now return influence functions when
called with a survey.design2 object and the influence=TRUE option. These allow
svyby() to estimate covariances between domains, which could previously only be
done for replicate-weight designs, and so allow svycontrast() to do domain contrasts
- svymean, svytotal, svyratio, svymle, svyglm, svykappa
Nonlinear least squares with svynls() now available
Document that predict.svyglm() doesn't use a rescaled residual mean square
to estimate standard errors, and so disagrees with some textbooks. (for Trent Buskirk)
3.38 When given a statistic including replicates, svycontrast() now transforms the replicates
and calculates the variance, rather than calculating the variance then using the
delta method. Allows geometric means to exactly match SAS/SUDAAN (for Robert Baskin)
vcov.svyrep.design to simplify computing variances from replicates (for William Pelham)
svykm() no longer throws an error with single-observation domains (for Guy Cafri)
Documentation for svyglm() specifies that it has always returned
model-robust standard errors. (for various people wanting to fit relative risk
regression models).
3.37 RODBC database connections are no longer supported.
Use the DBI-compatible 'odbc' package
set scale<-1 if it is still NULL after processing, inside svrepdesign()
[https://stats.stackexchange.com/questions/409463]
Added withPV for replicate-weight designs [for Tomasz Żółtak]
svyquantile for replicate-weight designs now uses a supplied alpha to get
confidence intervals and estimates SE by dividing confidence interval length
by twice abs(qnorm(alpha/2)). [For Klaus Ignacio Lehmann Melendez]
All the svyquantile methods now take account of design degrees of freedom and
use t distributions for confidence intervals. Specify df=Inf to get a Normal.
[For Klaus Ignacio Lehmann Melendez]
svyivreg() for 2-stage least-squares (requires the AER package)
warn when rho= is used with type="BRR" in svrepdesign [for Tomasz Żółtak]
Add "ACS" and "successive-difference" to type= in svrepdesign(),
for the American Community Survey weights
Add "JK2" to type= in svrepdesign
Warn when scale, rscales are supplied unnecessarily to svyrepdesign
More explanation of 'symbolically nested' in anova.svyglm
Link to blog post about design df with replicate weights.
Chase 'Encyclopedia of Design Theory' link again.
The canonical form [1] of an R package Makefile includes the
following:
- The first stanza includes R_PKGNAME, R_PKGVER, PKGREVISION (as
needed), and CATEGORIES.
- HOMEPAGE is not present but defined in math/R/Makefile.extension to
refer to the CRAN web page describing the package. Other relevant
web pages are often linked from there via the URL field.
This updates all current R packages to this form, which will make
regular updates _much_ easier, especially using pkgtools/R2pkg.
[1] http://mail-index.netbsd.org/tech-pkg/2019/08/02/msg021711.html
Summary statistics, two-sample tests, rank tests, generalised linear
models, cumulative link models, Cox models, loglinear models, and
general maximum pseudolikelihood estimation for multistage stratified,
cluster-sampled, unequally weighted survey samples. Variances by
Taylor series linearisation or replicate weights. Post-stratification,
calibration, and raking. Two-phase subsampling designs. Graphics. PPS
sampling without replacement. Principal components, factor analysis.
