Analysis of Complex Survey Samples

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.


3.35-1 svytable() could give an integer overflow with a replicate-weight design having integer weights, such as CHIS. (Elizabeth Purdom)

   empty factor levels in the strata= argument to svydesign() no longer create strata

   The twophase() function gave errors when the first phase of sampling had multiple
   stages (reported by Pedro Luis Baldoni)

3.35 The use of RODBC is DEPRECATED. If possible, I want to move to just supporting the R-DBI interface; you can use ODBC connections with the DBI-compatible 'odbc' package

   Some instances of deparse() needed to be paste(deparse(), collapse=""), notably in svyciprop() (Boris Fazio)

   Fix to svycontrast had broken the no-names case (Brian Miner)

   More helpful error message with missing values in replicate weights (Antony Damico)

   svyglm() now uses rather than for storing the call (for Michael Laviolette)

   svystandardize() now takes over=~1 for the whole population (for Michael Laviolette)

   svycralpha() does Cronbach's alpha (for Franziska Kößler)

   many tests based on the printed output are moved to tests/testoutput because they
   differ trivially between platforms and so aren't CRAN-compliant.

3.34 removed duplicate definition of svycontrast.svystat

   fix match.names() for the case of all the same names in different order (Sebastien Lucie)

   make rescaling of weights optional in svyglm and svycoxph (Greg Ridgeway)

   USER VISIBLE CHANGE: default rescaling of weights in svyglm and svycoxph with replicate
   weights is now mean=1, as it always was with svydesign objects, rather than sum=1.
   This does not affect estimates or tests or comparisons, but the AIC and its effective
   degrees of freedom will look more plausible, as will the deviance.

   Fix handling of missing data for calibrated designs in svyranktest (for Brad Biggerstaff)

   handle empty subsets in onestage, because svyby().  (for Greg Freedman)

   allow offset() in svymle()   (Patrick Brown)

   svyhist() returns the same as hist(), plus when freq=TRUE a component
   count_scale giving the scale factor between counts and density (for Ward Kingkade)

   example of geometric means added to ?svycontrast (for Irene van Woerden)

   move check for missing repweights later in the function to pick up more cases (Anthony Damico)

   fix printing of predicted values from predict.svrepglm  (for Anthony Damico)

   added minqa::newuoa and minqa::bobyqa as options for svymle(), and made 'newuoa'
   the default when a gradient is available

3.33-2 confint method for svyttest (for Brian Guay)

   partial fix in start= argument for svyglm() [still doesn't work inside function] 

3.33-1 two minor bug fixes for ABS contributions to calibrate()

3.33 Add explicit family= argument to svyglm() methods because of strange scoping problems (for Thomas Leeper)

   calibrate() now displays differences between sample and population names (for Stas Kolenikov)

   cal_names() displays what the auxiliary variable names will be for a formula

   regTermTest() now does Rao-Scott-type Wald tests, which I think are the same as SUDAAN's
   "Satterthwaite Adjusted Wald Tests" (but I don't have an example to verify).

3.32-2 Actually add the hyperbolic sine distance function

   Add AIC method for svycoxph

3.32-1 pseudo-rsquared paper is out: update reference

   Add cal.sinh for the hyperbolic sine distance function (used in CALMAR2) with code from Maciej Beręsewicz

   Coerce tbl_dfs to data frames in svydesign, svrepdesign, twophase, because they aren't actually a drop-in replacement.

3.32 Add diffs for calibration from Daniel Oehm at ABS

    - Sparse matrix support
- Heteroscedasticity parameters
    - more flexible boundary constraints

    Update references to AIC/BIC paper, Statistical Science review paper

    More information on accuracy of pchisqsum methods

    svyglm() didn't work with missing values in database-backed designs (Anthony Damico)

3.31-8 Add svysurvreg() (for Pam Shaw and Eric Oh)

    Remove the tests that don't produce character-identical output on all platforms (for CRAN)

3.31-7 Add deff="replace" as an option for svyratio (for Chris Webb)

    Add psrsq() for pseudo-R^2 statistics (for Ward Kinkade)

3.31-6 Make database-backed svydesign work with no design variables (for Anthony Damico and Guilherme Jacob)

3.31-5 predict.svyglm() uses object$xlevels and object$contrasts and so should be able to guess the right factor levels when they aren't supplied in newdata= (for @thosjleeper)

fix return() without parentheses in svykm.R

3.31-4 svyciprop(,method="like") and confint.svyglm(method="like") work even when the design effect is large.

3.31-3 svyciprop has method="xlogit" that reproduces what SUDAAN and SPSS give. (for Rex Dwyer)

Added reference for svylogrank

Added example from YRBS for software comparison

Copied some names from NEWS into THANKS

3.31-2 explicitly dropping dimensions on a 1x1 matrix

3.31-1 Allow for incompatible change in output of CompQuadForm::farebrother()

3.31 update isIdValid() to dbIsValid() for DBI changes.

explicitly :: or import ALL THE THINGS.

mse option for svrepdesign.character and svyrepdesign.imputationList was ignored (Antony Damico)

confint works on output of svycontrast (for Michael Laviolette)

denominator df fixed in confint.svyglm (Joey Morris)

svyboxplot rule for which lower-tail points are outliers was wrong (David Collins)

calibrate() with variable-specific epsilons and zero sample totals didn't work (Alex Kowarik)

document that regTermTest(method="LRT") can't handle models with a start= argument
and document how to use anova.svyglm instead. (Brad Biggerstaff)

update tests output for new formatting in current R.

3.30-4 svypredmeans() does the same things (together with svycontrast()) as as PREDMARG in SUDAAN (for Thomas Yokota and Anthony Damico)

3.30-4 confint.svystat was handling denominator degrees of freedom wrongly for the Wald method (Jared Smith)

3.30-3 vcov.svrepstat does more sensible things when covariances aren't estimated (eg for quantiles). This fixes issues with svyby

    dropped support for old version of hexbin

3.30-1 Fix example(svyplot) now that "hexbin" package no longer loads grid package

3.30 svyranktest() now allows k-sample tests (eg Kruskal-Wallis)

    svylogrank() does the generalised G-rho-gamma logrank tests
    [methods from Rader and Lipsitz (and probably al)]

various CRAN fixes

3.29-9 AIC.svyglm, based on Rao-Scott approximation BIC.svyglm, based on multivariate Gaussian likelihood for coefficients checks values are finite before replicating

    calibrate() using a list of margins now allows named vectors for 1-d margins

3.29-8 svyhist(freq=TRUE) works with replicate-weight designs (for Ward Kinkade)

3.29-7 svyranktest() works with replicate-weight designs (for Matthew Soldner)

    reference to the lavaan.survey package in ?svyfactanal

3.29-6 svyby() now always includes within-domain covariances

3.29-5 Change from multicore to parallel.
Parallel processing is now only available with R >=2.14

ddf argument really works now in confint.svyglm (Anthony Damico)

colour specification in plot.svykmlist now works (Mark Rosenstein)

    svyplot() documentation explains how to annotate a hexbin plot

3.29-4 add symmetric=TRUE to eigenvalue calculation in anova.svyloglin, to improve numerical stability

subset.svyimputationList now allows the subsets to end up with 
different observations in them (for Anthony Damico)

subset.svyDBimputationList now gives an error if the subsets 
have different observations, not just a warning that people

svydesign gives an error if there is only one PSU, to catch
omission of tilde in svydesign(id=~1,...) (Milan Bouchet-Valat)

3.29-2 confint.svyglm(method="Wald") wasn't using its ddf= argument, because confint.default() doesn't (Anthony Damico)

3.29 svystandardize() for direct standardization over domains

    withReplicates() has a method for svrepstat objects

added predict.svrepglm(), which can return replicates

    saddlepoint approximation to sum of chisquares works further out into the tails

    fixed bug in rescaling in calibrate() when initial weights are very wrong (Takahiro Tsuchiya)

    documented df= argument in svyciprop(method="mean") (Anthony Damico)

    added df= argument to other svyciprop methods for Stata compatibility (Anthony Damico)

3.28-3 svykappa didn't work for larger than 2x2 tables. (Jeffery Hughes)

    svyby didn't allow deff="replace"  (Francisco Fabuel)

svrepdesign(,type="other") now warns if scale, rscales arguments are not given

svystat, svrepstat objects now have a plot method (barplot, currently.)

    svyplot(,type="bubble") now uses the basecol= argument for colors.

    postStratify() now works when some input weights are zero

3.28-2 calibrate() prints out sample and population totals when the lengths disagree

calibrate() is more stable when the initial weights are wrong by orders of
magnitude (for Kirill Mueller)

calibrate() can now take a list of margins as input, similar to rake()
(for various people including Kirill Mueller)

3.28-1 SE now works with output of predict.svyglm (Kieran Healy)

   make.panel.svysmooth() sometimes had invalid bandwidth choices.

   as.svrepdesign() now allows for fpc information not present in 
   the design object (Alistair Gray)

   regTermTest(,method="LRT") works for svyolr(), and
   method="Wald" now doesn't need user-specified df (for Zachary Grinspan)

   svrepdesign() checks the length of the rscales= argument  (Ward Kinkade).

   Document the problem with in-line data-dependent variable construction
   in svyby()  (Anthony Damico)

   Check for completely-missing groups in svyby 

3.28 svyvar() for replicate-weight designs now returns whole matrix

   withReplicates() has method for svyvar() output, to simplify multivariate analyses.

   design effect estimate for svytotal with replicate weights was wrong (Daniel Fernandes)

   transform() is now a synonym for update().

   lots of partial argument matching removed to keep CRAN happy.

3.27 added anova.svyglm() for Wald tests and Rao-Scott LRT. anova(model1, model2) works for nested models even if not symbolically nested.

   formula component of svyglm$call is now always named, so update() will work.

   svyboxplot(,all.outliers=TRUE) didn't work for single boxes (Takahiro Tsuchiya)

3.26-1 Better missing-value handling with replicated weights in svyquantile

   svyboxplot() has all.outliers= argument to plot all outliers

3.26 Added Preston's multistage rescaled bootstrap (for Alois Haslinger)

   The multistage bootstrap can use the multicore package if available.

   calibrate() can take a vector of tolerances (for Alois Haslinger)
   [this actually used to work by accident, but now it's documented]

   Clearer error messages when post-strata contain NAs.

3.25-1 The ... argument to svytable() is now passed to xtabs()

   Clearer documentation about graphing discrete variables.

3.25 svyhist() didn't work for two-phase designs.

   added svylogrank() for logrank test for survival data.

   added svyranktest() for two-sample rank tests.

   svrepdesign() and as.svrepdesign() now have mse= argument to request 
   replicate-weight variances centered around the point estimate rather
   than the mean of the replicates.  The default is controlled by 
   options(survey.replicates.mse), which defaults to FALSE, consistent with
   previous versions. (For Anthony Damico, among others)

3.24-1 CHANGE: svychisq() statistic="lincom" and "saddlepoint" now use the linear combination of F statistics from pFsum().

3.24 Rao-Scott test based on linear combinations of Fs is now also available in regTermTest

   Algorithms from CompQuadForm (AS155 and AS204) now used for method="integration" 
   in pFsum and pchisqsum.  These are more accurate and faster than the previous
   implementations. If you use CRAN binary packages you will need at least R 2.12. 

   pFsum() saddlepoint and Satterthwaite methods are also much faster. The 
   saddlepoint approximation now works for the whole range, not just the right tail.

3.23-3 Some vignettes didn't load the package (Brian Ripley)

   Added pFsum() for linear combination of F distributions with same denominator.

   better example (quantile regression) in withReplicates().

3.23-2 svyhist() didn't handle include.lowest= correctly. (Chris Wild)

   svyby(, return.replicates=TRUE) now returns the replicates in the same
   order as the printed output, and labelled. (for Bob Fay)

3.23-1 svycdf() wasn't handling replicate weights correctly.

   Change in svyquantile() for replicate weights when using type="quantile". 
   Point estimate used to be mean of replicates, now is ordinary weighted quantile.
   (for Bob Fay)

   Small changes in handling of zero weights in svyquantile().

3.23 two-sample svyttest() didn't work with replicate weights. (Richard Valliant)

3.22-4 postStratify now allows 1-d matrix as well as vector in data frame of population counts. (for Jean Opsomer)

   print.summary.pps wasn't being exported (Gonzalo Perez)

   svyhist() ignored right= argument

   predict.svycoxph() was slightly overestimating standard errors for survival curves.

   [.pps and [.twophase2 crashed when no observations were removed (Gonzalo Perez)

3.22-3 bug in trimWeights (Richard Valliant), also add warning for attempts to trim past the mean weight.

3.22-2 bug in the argument to svyby() (Trevor Thompson)

  regTermTest() now does F tests by default (for Chris Wild)

3.22-1 added df= argument to confint() methods for svystat, svyrepstat, svyby, svyratio (for Richard Valliant)

   added argument to svyby(), to drop groups defined by
   missing values of by= variables.

   confint.svyby() uses SE(), not vcov(), so undefined values in replicates
   are handled on a per-group basis.

   svysmooth(,method="locpoly") now has automatic bandwidth choice, and
   make.panel.svysmooth() will use this choice by default.

3.22 added stratsample() to take stratified samples.

   fixed bug in design effects for subsets of calibrated or 
   database-based surveys

   changed scaling in biplot.svyprcomp so area is proportional to 
   weight, rather than height proportional to weight.

3.21-3 svyratio() can now estimate design effects (for Scott Kostyshak)

3.21-2 Rao & Wu bootstrap wasn't sampling n-1 PSUs (Richard Valliant)

3.21-1 bug in printing variances for three or more variables (Corey Sparks)

   svyquantile() reliably returns NA for NAs in data when na.rm=FALSE.

   svymle() was not using analytical gradients with nlm() (Christian Raschke)

3.21 added trimWeights() to trim weights, and trim= option to calibrate (for Richard Valliant)

   clearer documentation that svyquantile() needs ci=TRUE or keep.var=FALSE
   to work with svyby()

   added a simple random sample to data(api) as promised in book (Djalma Pessoa)

3.20 in svycoxph() modify the rescaling of weights to avoid very small weights because of convergence problem in coxph() with counting-process data (for Tapan Mehta)

   added some multivariate statistics: 
       svyprcomp(): principal components, svyfactanal(): factor analysis.

   added heuristic check that combined.weights= has been specified correctly.

   confint.default wouldn't give CIs for multiple parameters with replicate weights, because
   the vcov matrix didn't have variable names. (Art Burke)

   More of the svyciprop() methods now work for replicate-weight designs.

   The book of the package is now available! (see

3.19 svrepdesign() can specify replicate-weight columns with a regular expression

   svrepdesign() can produce database-backed designs

   svyquantile() has a df argument to use a t distribution in
    Woodruff's method (for Wade Davis)

   calibrate() doesn't require an intercept in the calibration model (for Richard Valliant)

   regTermTest() and model.frame() work with svyolr() (for Michael Donohue)

   better printing of svyvar() output (for Brad Fulton)

   twophase() documents more clearly that method="simple" is preferred for standard epi
   designs where it works.

   better error messages when a database-backed design has a closed connection

3.18-1 documented the need to use quasibinomial/quasipoisson in svyglm

   improved the description of confidence intervals and standard errors for

3.18 Changed the default to combined.weights=TRUE in svrepdesign()

   Fixed bug in multiple imputation analysis with multicore package.

   The check for PSUs properly nested in strata had some false negatives.

3.17 Under Linux, Mac OS, and most Unix systems, multiple processors can be used for the subgroups in svyby(), the imputed data sets in with.svyimputationList and with.DBsvyimputationList, and the replicate weights in and This requires the 'multicore' package and the argument multicore=TRUE to the functions (in the absence of the multicore package, the multicore=TRUE option is just ignored). handled NA values incorrectly (Arthur Burke)

   print.summary.twophase2 wasn't exported, so summary(twophase.object) 
   gave Too Much Information  (Norman Breslow) labelled the statistic it computed as 'mean',
   although it really was the correct total. (Arthur Burke)

   detection of PSUs not nested in strata was incorrect in some cases.

   added xbins= option to svyplot for hexbin styles (for Bryan Shepherd)

   print() method now has strata in a more predictable order (for Norman Breslow)

   regTermTest(,method="LRT") now does Rao-Scott-type tests based on the estimated 
   loglikelihood ratio, for generalized linear models and the Cox model. Similarly,
   confint.svyglm(,method="likelihood") does confidence intervals based on the 
   Rao-Scott-type likelihood ratio test.

   Updated marginpred() to work with survival 2.35-7

   Documentation fixes revealed by the new R pre-2.10 help parser

   Added unwtd.count() to count the raw number of non-missing observations.

   The new PPS designs now work with subset().

3.16 PPS designs without replacement, based on the weighted covariance of sample indicators: Horvitz-Thompson and Yates-Grundy estimators, Overton's approximation, Hartley-Rao approximation, a modified Hartley-Rao approximation that depends only on sample data.

3.15-1 The new two-phase designs added in 3.15 are now exported properly.

3.15 Full multistage sampling now possible at both phases of a two-phase design, and the standard errors now exactly match Sarndal et al. The underlying algorithms use sparse matrices to store the weighted covariance of sampling indicators, and so require the Matrix package. Use method="approx" in twophase() to get the old methods, which use less memory.

   added marginpred() for predictive margins, ie, predictions after
   calibration/direct standardization on confounder distribution.

   standard errors for predict.svyglm(,type="response") were 
   printing incorrectly. now works when the result has 

   The separate package odfWeave.survey provides methods for odfWeave::odfTable
   for some survey objects.

   formula() now works correctly on svykmlist objects with standard errors.

3.14 predict.svycoxph() now does fitted survival curves with standard errors for the Cox model. (for Norman Breslow)

   standard errors for svykm use a bit less memory.

   quantile.svykm can do confidence intervals

   added some references on svykm standard errors.

   tidied up some help pages.

3.13 Add standard errors to svykm() (for Norman Breslow)

   fix typo in svyquantile(interval.type="betaWald") and add
   'degrees of freedom' correction to the effective sample size.

   add 'degrees of freedom' correction to effective sample size
   in svyciprop, type="beta".

   SE, coef for svyratio objects now optionally convert to a vector
   and confint() now works on ratios.

3.12 Add svyttest() for t-tests, as a wrapper for svyglm

   Add svyciprop() for confidence intervals for proportions,
   especially near 0 or 1

   confint() works with svycontrast(), svyquantile(), 
   svyciprop() output.

   bug fix for updates to ODBCsvydesign objects.

   Add example of PPS sampling to example(svydesign), and link to 
   help for variance estimation.  Add Berger(2004) reference.

   svyby() now has vartype="ci" to report confidence intervals 
   (for Ron Burns)

   update survival examples to work with new version of survival

3.11-2 Document that calibrate() to PSU totals requires at least as many observations as calibration variables

   pchisqsum(,type="saddlepoint") now works down to mean x 1.05 
   rather than mean x 1.2

   The breaks= argument to svyhist() now works (Stas Kolenikov)

   svyhist() works on database-backed designs.

3.11-1 svyglm() [and svyratio()] gave an error for post-stratified designs with missing data (Shelby Chartkoff)

   svycoxph() gives a clearer error message for negative weights.

   svyquantile() now has a 'betaWald' option, as proposed  
   by Korn & Graubard (1998), and has an option for handling 
   ties that appears similar to (some versions of) SUDAAN
   (for Melanie Edwards)

   plot.svycdf() has an xlab argument to override the default labels

3.11 as.svrepdesign now has type="subbootweights" for Rao and Wu n-1 bootstrap

   An approximation for PPS without replacement due to Brewer 
   is available in svydesign()

   svydesign() no longer warns if some fpc are exactly zero, but
   still warns if they are suspiciously large or small

3.10-1 svycoplot can now pass ... arguments to xyplot(), not just to panel.

   svycontrast() has a 'default' method that assumes only a coef() 
   and vcov() method are available.

   Fixed example code for anova.svyloglin. 

   Added predict(,type="terms"), termplot(), residuals(,type="partial") 
   for svyglm.  As a result, the default for se= in
   predict.svyglm has changed. 

   make.panel.svysmooth() makes a weighted smoother as a slot-in
   replacement for panel.smooth(), eg in termplot().

   print.summary.svyloglin was broken (Norm Breslow).

   confint() method for svyglm has both Wald-type and
   likelihood-type confidence intervals (based on Rao-Scott test)

   documented that svykappa() requires factor variables. 

   svysmooth() doesn't fail when data are missing.

   documented that update.svyloglin is faster than fitting a new

   dotchart() methods for svyby, svystat, svrepstat

   svyloglin() handles missing data better.

   svymle() didn't work if constant parameters were in any 
   position other than last.

   svyby() now has a return.replicates argument (for Phil Smith).

   logit and raking calibration could run into NaN problems with
   impossible bounds.  Step-halving seems to fix the problem.

3.10 update() methods for database-backed designs.

   improvements in graphics for subsets of database-backed designs.

   barplot methods for svystat, svrepstat, svyby objects.

   svytable() for database-backed designs

   quantiles work with svyby(covmat=TRUE) for replicate-weight designs.

   fix printing of p-value in svychisq, type="lincom"

   better error messages for misspecified fpc in svydesign()

   database-backed analysis of multiple imputations.

   formatting changes to coef.svyquantile, SE.svyquantile, svyby

   svrepdesign works with multiple imputations (though not with databases)

   fix for missing factor levels in subsets of database-backed designs

   allow svychisq(statistic='lincom') with replicate weights.

   quantile regression smoothers in svysmooth()

   add svychisq.twophase() (for Norm Breslow)

   changed defaults in predict.svyglm so that plot.lm works
     (for Patricia Berglund)

   svyloglin() for loglinear models, with Wald and Rao-Scott tests.

   pchisqsum() (and svychisq, anova.svylogin) have a saddlepoint approximation.

3.9-1 improvments in svyby, degf, svyglm for subsets of calibrated designs or database-backed designs.

   svyboxplot() and svycdf() now work with database-backed designs.

   ODBC support for database-backed designs.

   modified the degrees of freedom calculation in svyglm.

3.9 Added database-backed design objects. The data= argument to svydesign can be the name of a database table in a relational database with a DBI-compatible interface. Only the meta-data is kept in R, other variables are loaded as necessary.

3.8-2 Added svycoplot()

3.8-1 Added subset.svyimputationList

   coef.svyolr returns intercepts as well (by default).

   svyolr() has a method for replicate-weight designs

   print methods for svykm, svykmlist weren't exported.

3.8 svyolr() for proportional odds and related models.

   license is now GPL 2|3 to accomodate code ripped from MASS package

   svykm() for survival curves (no standard errors yet)

3.7 Added style="transparent" to svyplot().

   svyby() and svytable() work on twophase objects.

   svychisq() has statistic="lincom" for linear combination of chisquare, 
   the exact asymptotic distribution. 

   Added interface to mitools package for analyzing multiple imputations

   svykappa() for Cohen's kappa (for Tobias Verbeke)

3.6-13 Change in tolerances so that calibrate() works better with collinear calibration variables (Richard Valliant)

   calibrate() can be forced to return an answer even when the specified
   accuracy was not achieved.

3.6-12 svyhist() handles missing data better.

  Added svycdf() for cumulative distribution function estimate.

3.6-11 postStratify() for repweights was standardizing the replicates to slightly wrong population totals. (Alistair Gray)

  vcov() for two-phase designs gives the contributions from each phase
  for a wider range of statistics. (Norman Breslow)

  fixes for codetools warnings.

3.6-10 Added error message for missing sampling indicator in two-phase design (Lucia Hindorff)

  Added tests/kalton.R with reweighting examples.

  make.calfun() for creating user-specified calibration distances. 

  NOTE: Calling grake() directly now requires a calfun object rather than 
  a string: see help(make.calfun).

3.6-9 Bootstrap weights used last stratum size rather than harmonic mean for n/(n-1) factor (Djalma Pessoa)

  method= argument to svycoxph() didn't work (Lisa McShane)

  svyquantile did not treat missing values as a domain
  (Nicole Glazer)

  fix for change in pmax/pmin (Brian Ripley)

  Add pchisqsum for distribution of quadratic forms.

3.6-8 A fix in 3.6-6 had broken svycoxph when only a single predictor variable was used (Lisa McShane)

3.6-7 svycoxph() is much faster for replicate weights uses a cached value rather than

3.6-6 svyquantile was not passing method= argument to approxfun() (Jacques Ferrez)

  Documented that svyquantile(interval.type="score") may not be any
  more accurate

  Broken link due to typo in svyratio.Rd (Giuseppe Antonaci)

  postStratify could overestimate standard errors for post-strata cutting
  across existing sampling strata. (Ben French)

  svycoxph() would not run for subsets of calibrated designs.
  (Norman Breslow)

3.6-5 Add return.replicates option to svyratio() (for

  Add amount= option to svyplot

  Design effects for totals were wrong for PPS 
  sampling.  (Takahiro Tsuchiya)

3.6-4 rownames fix for svyratio with a single statistic.

3.6-3 raking by rake() now has slightly more accurate (smaller) standard errors. As a result, it can't be used on pre-2.9 svydesign objects.

  calibrate() does not warn about name mismatches when population 
  argument has no names.

  svyCprod, svyrecvar, grake now exported.

3.6-2 covmat=TRUE option for svyratio.

  svycontrast() fix for svyby() with empty groups (

3.6-1 Allow averaged bootstrap weights (as StatCanada sometimes produces) in svrepdesign()

  Fix derivative to get faster convergence in logit calibration 
  (Diego Zardetto)

  svycontrast() can take named vectors of just the non-zero coefficients.

  Nonlinear combinations of statistics with svycontrast()

3.6 Allow empty factor levels in calibration (for Diego Zardetto).

  Work around for strange S4 class/NAMESPACE issue with hexbin
  plots; actual fix requires more understanding.

  regTermTest handles MIresult objects.

  Add dimnames, colnames, rownames methods (for

  svysmooth for scatterplot smoothers and density estimation 
  (needs KernSmooth)

  Give a warning when fpc varies within strata. 

  svycontrast() for linear combinations of survey statistics

  covmat=TRUE option to svyby() for replicate-weight designs, so
  the output can be used in svycontrast().

3.5 Add estWeights for Robins et al way of using auxiliary information (ie AIPW).

  Remove JSS article and survey-vanderbilt.pdf from inst/
  since they are now seriously out of date.

  paley() now gives matrices of order 2^k(p+1), which are
  usually of minimal or near-minimal size.

  Drop 72x72 and 256x256 Hadamard matrices, which are easy
  to recreate, from precomputed set and replace 36x36 with the
  one from Plackett & Burman, which has full orthogonal balance

  Note that changes to svyby now require R 2.2.0 or later.

  predict.svyglm has option to return just variances (rather
  than entire variance-covariance matrix)

  drop.empty.groups now works when the grouping variables 
  are not factors.

  Add a namespace

  Move precomputed Hadamard matrices from inst/hadamard.rda to

3.4-5 Add svyboxplot (for Luke Peterson)

  Add drop.empty.groups option to svyby

3.4-4 Paley construction of Hadamard matrices now knows primes up to 7919, works for larger sizes if the user supplies a suitable prime.

  calibrate() now reorders elements of 'population' to match
  column names of model matrix if necessary.

  predict() method for svyglm (for Phil Smith, Andrew Robinson)

  svyratio() for two-phase designs.

  Added vignette on domain estimation.

  svyby() can report multiple vartypes.

3.4-3 make svyratio work with svyby (for Phil Smith)

  increase default number of iterations in calibrate()

3.4-2 Options for residual df for summary.svyglm, default based on degf Default denominator df for svyglm, svycoxph in regTermTest.

  survey.lonely.psu now applies to as.svrepdesign.

  keep up with changes in all.equal() for R 2.3.0

3.4-1 Speed optimizations for JKn weights with self-representing strata - jackknife replicates are not created for these strata - svytotal does not use these strata in variance calculation. - svytotal, svymean, svyratio,svyquantile,svyglm recognize designs (eg subsets) where all strata are self-representing.

  [.repweights_compressed does less copying and is a lot faster 
  for large designs

  Added verbose= option to svyby() to monitor slow computations.

  Added vartype="cv","cvpct" options for svyby().

  Two-phase designs gave incorrect variances in some cases [they
  were correct if the first stage was infinite superpopulation
  sampling or if all phase 1 ultimate sampling units were
  represented in phase 2].  These are fixed but twophase() now
  limits the first phase to single-stage cluster or element
  sampling. [detailed bug report from Takahiro Tsuchiya]

  added vignette describing estimator of phase-one variance in
  two-phase designs

  minor speedup in svyrecvar() for self-representing strata

  added make.formula() for convenience with many variables.

3.4 twophase() for specifying two-phase designs.

  two vignettes: a simple example and a description of two-phase epi designs

  svyratio handles missing data.

  cv() gives NaN rather than an error when the statistic is zero (for 
  [email protected])

  oldsvydesign() is officially deprecated

  Jackknife variances for strata with a single population PSU were wrong
  (non-zero) ([email protected])

  svyglm refused to work on subsets of calibrated designs

3.3-2 Add cv, SE, coef, and deff methods for svyby (for Ana Quiterio) methods for svystat, svrepstat

  regTermTest can do F-tests now (Daryl Morris).

  fix documentation of value for as.svrepdesign (Alan Zaslavsky)

3.3-1 Make nest=TRUE in multistage designs work when only some initial sampling stages are stratified

  Multistage recursive variances were only going to two stages.

  Add "(with replacement)" to output of print.survey.design2 when
  no fpc is specified.

3.3 Added more generalized raking estimators: raking ratio, bounded raking ratio, logit, (for Ana Quiterio)

  svytable() could sometimes leave the class attribute off the result.
  summary() now gives tests of association for svytable().

  svychisq() works for replicate designs

  degf() gives approximate degrees of freedom for replicate designs.

  Clearer error messages when design information is missing.

3.2-1 Fix ordering bug in ftable.svyby (Stefano Calza)

  The "probability" option added to svyquantile for replicate designs
  in 3.1 computed standard errors for the wrong tail. (Gillian Raab).

3.2 Add option to calibrate() to make weights constant within clusters.

  Add bounded regression calibration to calibrate()

3.1-1 Rescale svyvar output by n/(n-1) to match Kish, which makes a small difference to design effect computations. (for Takahiro Tsuchiya)

  Test for presence of intercept in calibrate() was too fussy.

3.1 Quantiles for replicate-weight designs now by default compute confidence intervals on the probability scale and transform, so they are valid for jackknife designs. (as Gillian Raab suggested long ago)

  Analyses on replicate weights should use eg svymean, which has
  methods for replicate weight designs; the old (eg svrepmean) variants
  are now deprecated.

  calibrate() can now use regression models with variance proportional 
  to linear combination of predictors (and so can duplicate ratio 
  estimators of means and totals)

  Prettier labelling of objects created by postStratify(), calibrate(), 
  update(), subset()

  svytotal on replicate weight designs was computing means, not totals
  (probably since 3.0). 

3.0-1 Allow some strata to have an infinite population (zero sampling fraction) (this doesn't happen in reality but is the recommended analysis for handling certainty PSUs in some large NCHS studies).

  Let svyby() handle vectors that are not in the design object (even 
  though they are discouraged)

  calibrate() was working only under stratified/simple random sampling.

  Allow user-supplied Hadamard matrix for brrweights.

  as.svrepdesign gave a spurious warning when converting post-2.9-1
  objects without finite population corrections to BRR.

  Allow multicolumn response variable in svymle() (for survival data)

  Add nlm() as the default optimization method for svymle().

3.0 Added simple GREG (G-calibration) estimators with calibrate()

  Added deff="replace" option to compute design effects comparing to
  simple random sampling with replacement, eg for designs where the weights
  do not sum to the population size. (for Gillian Raab)

  Added more references for median estimation.

  Added separate ratio estimator of totals for stratified
  samples. (for Renzo Vettori)

  cv.svyratio was inverted.

  rake() on survey design objects was accumulating cruft in the
  postStrata component on each iteration.

  Subsetting of raked designs without replicate weights was 
  broken (Steve Roberts)

  Standard errors were wrong for some domain estimates in 
  post-stratified models without replicate weights.
  More extensive tests comparing domain estimates to equivalent 
  ratio and regression formulations.

  Changed default in svyby to keep.var=TRUE

  Prettier stratum labels.

  New homepage at

  svyplot(type="hex") works with both pre1.0 and post1.0 versions 
  of the hexbin package.

  Fixed svychisq denominator degrees of freedom for stratified designs 
  for bug introduced by multistage revision. (Takahiro Tsuchiya)

2.9-1 Fixed typo in description of fpc in svydesign.Rd

  Added inst/twostage.pdf with examples of two-stage analyses.

  Handling of fpc specified as proportion in the absence of weights 
  was wrong.

2.9 Added full multistage sampling, involving a redesign of the object. The old objects are deprecated; they may be converted with as.svydesign2. Use options(survey.ultimate.cluster=TRUE) to get the same one-stage standard errors as earlier versions and options(survey.want.obsolete=TRUE) to turn off the annoying warnings about old-style survey objects. If you must create old-style survey objects use oldsvydesign().

  As a consequence of the redesign, most of the svyxxx functions
  are now generic, with methods for both svydesign and svrepdesign
  objects. Use svymean instead of svrepmean, for example.

  Added more Hadamard matrices, including the Paley construction. 
  brrweights() now finds designs of nearly optimal size for most surveys.

  Faster svymean, svytotal for replicates, with less memory use.

  Added "bootstrap" option for as.svrepdesign

  svyby and ftable.svyby now handle Deff (expanded from a suggestion
  by Tobias Verbeke)

  svyhist() for probability-weighted histograms

  added svycoxph() for replicate weight designs

  The "lonely.psu" corrections will be applied to strata with a single
  PSU in a subset (domain) if options("survey.adjust.domain.lonely") 
  is TRUE.  The default is FALSE. was not working for post-stratified designs.

  Added a PDF file with examples from UCLA ATS web site, including 
  comparisons with WesVar and SUDAAN. (inst/ucla-examples.pdf)

  Added slides from a talk at Vanderbilt University. 

  Fixed Deff to use simple random sampling without replacement. 
  Much faster confidence intervals for quantiles based on inverting a
  Wald test are now default. These are less accurate in small
  samples; the old method is still available.
  (based on suggestion from Gillian Raab)

2.8-4 Added a whole lot more references to the documentation.

  data(hospital) now has two sets of weights, one matching the
  UCLA ATS site and one matching the original reference.
  (from Tobias Verbeke) was reporting 1 replicates for compressed 
  weights (but still computing correctly)

2.8-3 postStratify for svydesign objects was giving too large standard errors

  Add deff() to extract design effects.

2.8-2 rewrite cv() to use coef() and SE()

2.8-1 Make Deff estimates work with ftable. (for Gillian Raab)

  ftable.svyby didn't work with a single by() variable (for Gillian Raab)

  Missing values now allowed in svychisq(). (for Lee Sieswerda)

2.8 fix printing of svyby broken in 2.7

  add ftable.svyby

  postStratify for svydesign surveys.

2.7-1 as.svrepdesign was giving the wrong weights for type="Fay" in 2.7

2.7 Option compress=TRUE in as.svrepdesign to reduce size of replicate weight matrix (and in rake(), postStratify()). Also function compressWeights() to do this to arbitrary replicate designs.

  terms() reorders variables in interactions, which confused regTermTest
  (Daniel Almirall)

  Added extractor function SE() for standard errors (Andrew Robinson)

  hadamard() now finds smaller Hadamard matrices.

  svyCprod warns if a subset has only one PSU in some stratum 
  (Gillian Raab)
  Added tests/lonely.psu.R

  Added another option "average" for lonely.psu (Gillian Raab)

  svydesign can now detect from sampling weights or fpc when a stratum
  with a single PSU is self-representing, and in these cases 
  options("survey.lonely.psu") is not used.

  ftable.svystat and ftable.svrepstat to produce better tables of
  percentages and totals.

  Experimental set of functions to help in computing non-response weights 
  (see ?nonresponse for details)

2.6-2 Better handling of NAs in svyby

  Subsetting didn't work right for single-observation subsets.

  svyglm and svycoxph had scoping problems when run inside a 
  function (Daniel Almirall)

  svyglm and svycoxph now accept weights (to be multiplied by 
  the sampling weights)

  With R 2.0.0 less copying will occur, especially when variables=
  is not specified in a design

2.6-1 Totals for factors give cell totals.

2.6 Design effects were broken for multiple means computed at once.

  Add coefficient of variation for mean, total, ratio,...

  variables= argument of svydesign works with missing data (Tobias Verbeke)

  Fix reference to Binder (1991) (Tobias Verbeke)

  Means for factors now give cell means.

  coef and vcov methods for svystat and svrepstat.

  Another tiny example dataset from the VPLX manual
  svrepvar was incorrect for multiple variables simultaneously

  Better error messages for missing data in svrVar.

2.5 Wald tests for association in contingency tables.

  svyplot() for weighted graphics (some of these require "hexbin")

  Examples for rake(), postStratify()

  svyby() works for svrepdesign analyses as well

  svrepvar() added

  Design effects for means and totals. (Gillian Raab)

2.4 Make regTermTest work with svycoxph()

  Clearer output for print.svycoxph() (Daniella Gollinelli)

  Rao-Scott adjusted tests for contingency tables.

  svyby() for tables of means, medians, etc

2.3-2 Fix for svyquantile confidence intervals.

2.3-1 clearer warnings in svrVar when some replicates give NA . (for Gillian Raab)

2.3 svyquantile has confidence intervals, added svrepquantile.

2.2-1 as.svrepdesign didn't pass options to brrweights (for Fred Rohde)

2.2 published in Journal of Statistical Software
- If population size is specified, but not weights or probabilities, work out the probabilities from the population size - Clearer error message when some design information is NA (for Tobias Verbeke) - better update() methods

2.0 Just a numbering change.

1.9-3 Fix svytotal variance estimate as.svrepdesign wasn't handling unstratified cluster samples right. Check for fpc in multistage samples, which we don't handle. add print method for basic survey statistics add rake() California API data.

1.9-2 Added post-stratification of replicate-weights

1.9-1 Bugfix: jknweights was requiring finite population correction.

1.9 - "certainty" option for single-PSU strata - Replication weight analyses (alpha version)

1.4 - I think all the possible permutations of arguments in svydesign now work. - The examples in svyglm incorrectly had a data= argument.

1.3 svydesign wasn't allowing weights to be a vector.

1.2 - svydesign(nest=TRUE) now uses less memory - added regTermTest for testing regression terms.

1.1 Added subset, update methods. Variance estimation is now correct for subpopulations produced with select or subscripting.

1.0 No changes

0.9-5 - finite population correction should be done with PSUs not individuals - added Cox models

0.9-4 - svyCprod was computing n/(n-1) using number of observations, not number of PSUs, and was averaging observations rather than PSU means to compute stratum means. - Bug in handling multiple levels of cluster id in svydesign

0.9-3: Finite population correction. Adjustments for stratum with single PSU (Fred Rohde) Fixed svydesign(nest=TRUE) to work with strata

0.9-1: First release.

Reference manual

It appears you don't have a PDF plugin for this browser. You can click here to download the reference manual.