Skip to main content

Table 4 Reporting of methods for controlled MI

From: A review of the use of controlled multiple imputation in randomised controlled trials with missing outcome data

Controlled MI feature n
(N = 16ª)
%
Type of primary outcome
 Binary 4 25%
 Continuous 9 56%
 Time-to-event 3 19%
Type of controlled MI
 Delta-based MI 9 56%
 Reference based MI 7 44%
Method of delta-based MI (N = 9)
 MICEe 2 22%
 MVN imputation (non-monotone missing patterns) and regression MI model (monotone patterns) 1 11%
 Kaplan-Meier MI (KMMI) 2b 22%
 ANCOVA MI 1 11%
 Cox model MI 1 11%
 Not stated 3 27%
Method of reference-based MI (N = 7)
 MCMC or random draws from a normal distribution with mean equal to subject’s own baseline valuef 1 14%
 Linear MMRM 2 29%
 Kaplan-Meier MI (KMMI) 1 14%
 Not stated 3 43%
Specified variables in imputation model 9 56%
 Imputation model incl. All variables in analysis model only 5 31%
 Imputation model incl. All variables in analysis model + auxiliary variables 3 19%
 Imputation model did not include all variables in analysis modelg 1 6%
Did not specify variables in imputation model 7 44%
 Imputation model incl. All variables in analysis model + auxiliary variables 2 13%
Reported the number of imputations 13 81%
No. of imputations
 5 2 13%
 20 1 6%
 100 6 38%
 1000 4 25%
Not stated 3 19%
Specific procedure/command(s) (software) for implementing MI
 Proc MI and Proc MIANALYZE (SAS) 2 13%
 Proc MIXED and Proc MIANALYZE (SAS) 2 13%
 Not stated 12 75%
Rubin’s rules used for inference
 Yesh 9 56%
 Noc 1 6%
 Not stated 6 38%
Analysis status
 Primary 2 13%
 Sensitivity 14 87%
No. scenarios used in sensitivity analysisd Median Range
 Median (range) 3 (1–48)
Performed diagnostic check of imputations 0 0%
  1. ª Denominator for variables 16 unless otherwise indicated. bOne trial used both KKMI and Cox model MI in two separate sensitivity analyses. cOne trial reported using a modified version of Rubin’s rules, “the overall average estimated event rate difference and average estimated variance” (did not incorporate any between imputation variability in the variance calculation). dN = 13. Not clear for 1/14 trials using controlled MI in sensitivity analysis. eNo further details available on types of models utilised within MICE. fMissing data during the on-treatment period were imputed “using the MI SAS procedure (using Markov Chain Monte Carlo)” and values missing values during the post-treatment period were “multiply imputed using random draws from a normal distribution where the mean was equal to subject’s own baseline value.” gOne trial did not include all variables in analysis model and included auxiliary variables in the imputation model. h Explicitly stated (n = 5) or inferable from specified software or reference (n = 4). Percentages are rounded to 0 decimal places so may not sum exactly to 100%