Survey package news

New features

• svyquantile now takes account of design degrees of freedom in computing confidence intervals and in turning confidence intervals into standard error estimates. This means results will change (slightly, and for the better).
• svyivreg for two-stage least squares with instrumental variables. (described here)
• withPV for ‘plausible value’ analyses now supports replicate-weight designs. ‘Plausible values’ are how education people describe multiple imputation. The difference between this and the functions based on the mitools package is that withPV works with multiple columns in the same data frame (having, obviously, different names for a given variable in each imputed set) but with.svyimputationList works with multiple complete data frames (having the same name for a given variable in each imputed set). The basic withPV ideas are described here
• svrepdesign has specific options type="JK2 for JK2 replicate weights (used by the California Health Interview Survey) and type="ACS" for the replicate weights in the American Community Survey¹
• svrepdesign is now much more likely to give warnings if you unnecessarily specify the scale= and rscales= arguments when it could have worked them out itself – because you might have got it wrong.

New non-features

You can’t use RODBC database connections as data tables any more. Instead, you can use the odbc package, which follows the R-DBI interface.

Non-new non-features

Things that are still on the to-do list

• Make all the functions produce influence functions so that svyby can compare arbitrary statistics between groups without needed replicate weights. This isn’t especially hard, but it’s a lot of changes that need to be got right.
• Linear mixed models: I have two students working on improvements to the svylme package. It might end up as part of survey after that. Or not.
• Nonlinear least squares. If I get motivated, or if the people who complained about its non-existence on Stack Overflow actually ask for it.