Skip to main content

Advertisement

Table 2 Summary statistics of simulationsa with missing data patterns (MCAR, MAR, MNAR)

From: How to deal with missing longitudinal data in cost of illness analysis in Alzheimer’s disease—suggestions from the GERAS observational study

Imputation method for missing data pattern 10 % missing 20 % missing 30 % Missing 40 % missing
Mean cost (€) Bias (%) SSE SEE CP Mean cost Bias (%) SSE SEE CP Mean cost Bias (%) SSE SEE CP Mean cost Bias (%) SSE SEE CP
Complete sample 2102 59 61 59 61 59 61 59 61
MCAR
 Complete cases 2102 −0.7 (−0.03 %) 63 64 1.00 2121 18 (0.9 %) 68 69 1.00 2156 53 (3 %) 75 76 1.00 2181 79 (4 %) 87 87 0.99
MI MCMC 2150 48 (2 %) 68 58 1.00 2218 115 (5 %) 78 57 0.46 2300 198 (9 %) 93 57 0.03 2410 308 (15 %) 118 57 0.00
MAR
 Complete cases 1871 −231 (−11 %) 56 57 0.00 1723 −379 (−18 %) 54 55 0.00 1624 −478 (−23 %) 59 59 0.00 1499 −603 (−29 %) 59 61 0.00
MI MCMC 2158 55 (3 %) 112 57 0.76 2039 −64 (−3 %) 83 48 0.65 2218 116 (5 %) 202 52 0.38 2683 581 (28 %) 417 69 0.09
MNAR
 Complete cases 1544 −558 (−27 %) 27 28 0.00 1291 −812 (−39 %) 22 22 0.00 1096 −1007 (−48 %) 20 19 0.00 929 −1173 (−56 %) 17 16 0.00
MI MCMC 1602 −501 (−24 %) 28 26 0.00 1383 −720 (−34 %) 22 19 0.00 1212 −890 (−42 %) 20 15 0.00 1043 −1059 (−50 %) 19 11 0.00
  1. % bias was calculated as (estimated−actual)/actual cost × 100), where actual cost was the mean cost for the complete sample
  2. Abbreviations: CP coverage probability, MCMC Markov Chain Monte Carlo, MI multiple imputation, SEE standard error estimate, SSE sampling standard error
  3. a1000 simulations and sample size 1497 for different levels of missing data (10–40 %)