New in the survey package

Bell-McCaffrey standard errors for svyglm.

The standard svyglm standard errors are based on sums of squares of PSU-level residuals. As Bell and McCaffrey point out

These sums of squares tend to be too small for two reasons: residuals are generally smaller than true errors due to overfitting, and residuals tend to have lower intra-cluster correlation than the errors.

Also, since residuals will typically not have constant variance, the estimated standard error will have a longer tail than would be estimated simply by counting independent contributions as the usual estimators do. So, we have (optionally) a different standard error estimator and a different degrees-of-freedom estimator for confint.svyglm.

Wilson (‘score’) confidence intervals for proportions

Yet another option for svyciprop, this time extending the “Wilson” intervals that use the score test to define a quadratic whose roots are the interval endpoints. That is, the interval allows for the variance:mean relationship of the score statistic rather than just evaluating the variance at the mean.

In other highlights

• Multiphase designs: The multiphase function defines survey design objects with arbitrarily many phases. I’ve come across two three-phase designs recently, so I felt it was time for this. It’s still experimental, but there’s a vignette describing how the variance estimators are derived
• NA weights to drop observations: some people want to be able to put fictitious zero-weight observations into survey data files, for non-nefarious reasons. In svydesign.default the na_weights argument has options "fail" for the previous behaviour and "warn" and "allow" to drop records with NA weights before defining the survey design (with or without a warning)
• svystat objects lose their variance information when you do arithmetic on them because people might have expected the variances to magically transform
• rake now does not try to construct the full multiway table implied by the raking margins (it might be quite big). As a consequence the stopping criterion for iterative proportional fitting is a bit different. If you have a raked design object where you used a fairly loose convergence tolerance you might get slightly different results (from Ben Schneider)
• The $aic component of svyglm objects, which is undocumented and meaningless, is now set to NA. If this breaks your code, your code was already wrong.