Table 3

Kappa statistics for different imputation strategies when missingness is covariate dependent

Imputation Level

Imputation strategies

Percentage of missingness


5%

10%

15%

20%

30%

50%


Within cluster

Logistic regression

0.949

0.902


Propensity score

0.947

0.899

0.850

0.801

0.706

0.524


MCMC1

0.948

0.901

0.854

0.806

0.714

0.535


Across cluster

Propensity score

0.949

0.903

0.853

0.805

0.713

0.529


Random-effects logistic regression

0.951

0.908

0.859

0.808

0.717

0.538


Fixed-effects logistic regression

0.949

0.899

0.850

0.801

0.707

0.528


Ignore cluster

Logistic regression

0.947

0.895

0.844

0.793

0.695

0.508


Propensity score

0.945

0.891

0.839

0.787

0.688

0.495


MCMC1

0.946

0.893

0.841

0.790

0.691

0.501


Note:

1 MCMC = Markov chain Monte Carlo. For the MCMC methods, we round the imputed values to 1 if it is equal or greater than 0.5 and to 0 otherwise.

Ma et al. BMC Medical Research Methodology 2011 11:18   doi:10.1186/1471-2288-11-18

Open Data