d.f. for mira models (#494)

DESCRIPTION

@@ -1,8 +1,8 @@
 Package: emmeans
 Type: Package
 Title: Estimated Marginal Means, aka Least-Squares Means
-Version: 1.10.2
-Date: 2024-05-19
+Version: 1.10.2.090001
+Date: 2024-05-27
 Authors@R: c(person("Russell V.", "Lenth", role = c("aut", "cre", "cph"), 
     email = "russell-lenth@uiowa.edu"),
     person("Ben", "Bolker", role = "ctb"),

NEWS.md

@@ -2,6 +2,10 @@
 title: "NEWS for the emmeans package"
 ---
 
+## emmeans 1.10.2-09xxxxx
+ * Updated `mice::mira` support to use Barnard-Rubin adjusted d.f. (#494)
+
+
 ## emmeans 1.10.2
 This update is focused mostly on trying to clear up confusion with some users
 on the distinction between `emmGrid` objects and their summaries, since they

R/multiple-models.R

@@ -167,6 +167,20 @@ emm_basis.mira = function(object, trms, xlev, grid, ...) {
     }
     bas$bhat = apply(allb, 1, mean)  # averaged coef
     notna = which(!is.na(bas$bhat))
-    bas$V = V + (k + 1)/k * cov(t(allb[notna, , drop = FALSE]))   # pooled via Rubin's rules
+    bas$dfargs$m = k
+    bas$dfargs$df1 = bas$dffun  # this is the dffun for the 1st analysis
+    bas$dfargs$B = cov(t(allb[notna, , drop = FALSE]))
+    bas$V = bas$dfargs$T = V + (k + 1)/k * bas$dfargs$B   # pooled via Rubin's rules
+    bas$dffun = function(a, dfargs) { # pretty much copied from mice:::barnard.rubin
+        dfcom = dfargs$df1(a, dfargs)
+        with(dfargs, { 
+             b = sum(a * (B %*% a))
+             t = sum(a * (T %*% a))
+             lambda = (1 + 1/m) * b / t
+             dfold = (m - 1)/lambda*2
+             dfobs = (dfcom + 1)/(dfcom + 3) * dfcom * (1 - lambda)
+             ifelse(is.infinite(dfcom), dfold, 
+                    dfold * dfobs/(dfold + dfobs))
+        }) }
     bas
 }

vignettes/models.Rmd

@@ -234,7 +234,9 @@ a smoothing method like `pb()`.
 These objects are the results of fitting several models with different predictor
 subsets or imputed values. The `bhat` and `V` slots are obtained via averaging
 and, in the case of multiple imputation, adding a multiple of the between-imputation 
-covariance per Rubin's rules. In the case of `mira` models with model classes not
+covariance per Rubin's rules, along with an associated degrees-of-freedom
+adjustment (Barnard & Rubin 1999).
+In the case of `mira` models with model classes not
 supported by **emmeans**, [GitHub issue #446](https://github.com/rvlenth/emmeans/issues/446) 
 includes a function `pool_estimates_for_qdrg()` that may be useful for obtaining
 results via `qdrg()`.