With lambda = (1 + 1/m) B / T, the classical df is
nu_old = (m - 1)/lambda^2; the Barnard-Rubin small-sample value is
nu = nu_old * nu_obs / (nu_old + nu_obs) with
nu_obs = ((nu_com + 1)/(nu_com + 3)) nu_com (1 - lambda), always
at most nu_com.
Usage
morie_midegf(b, t, m, nu_com = NULL)
Arguments
- b
Between-imputation variance.
- t
Total variance.
- m
Number of imputations (>= 2).
- nu_com
Optional complete-data degrees of freedom.
Value
A list with elements df, df_old,
nu_obs, lambda_, m, method.
References
Barnard, J. and Rubin, D. B. (1999). Small-sample
degrees of freedom with multiple imputation. Biometrika, 86,
948-955. van Buuren, S. (2018). Flexible Imputation of Missing
Data, 2nd ed., Sec. 2.3.6.
Examples
morie_midegf(b = 5L, t = 5L, m = 5L)
#> $df
#> [1] 4
#>
#> $df_old
#> [1] 4
#>
#> $nu_obs
#> NULL
#>
#> $lambda_
#> [1] 1
#>
#> $m
#> [1] 5
#>
#> $method
#> [1] "Barnard-Rubin (1999) adjusted MI df"
#>
