Skip to main content

Advertisement

Table 1 Estimates of associations between predictors and health outcome at 70% response-rate

From: Preventing bias from selective non-response in population-based survey studies: findings from a Monte Carlo simulation study

bnon   10% most extreme values on the health outcome totally missing Extreme values on the health outcome not totally missing
CC FIML MI CC FIML MI
×1,×2,× 3 Y Pred. Bpred (SE) 95% coverage Bpred (SE) 95% coverage Bpred (SE) 95% coverage Bpred (SE) 95% coverage Bpred (SE) 95% coverage Bpred (SE) 95% coverage
0.3 0.0 ×1 .17 (.04) 88 .18 (.04) 92 .17 (.04) 86 .20 (.04) 92 .20 (.04) 92 .20 (.04) 95
   ×2 .25 (.04) 86 .26 (.04) 92 .25 (.04) 76 .29 (.04) 94 .30 (.04) 93 .29 (.04) 95
0.1 0.1 ×1 .17 (.04) 88 .17 (.04) 91 .17 (.04) 88 .20 (.04) 94 .20 (.04) 94 .20 (.04) 95
   ×2 .25 (.04) 81 .26 (.04) 91 .25 (.04) 79 .29 (.04) 95 .29 (.04) 93 .29 (.04) 96
0.3 0.1 ×1 .16 (.04) 84 .17 (.04) 88 .16 (.04) 83 .18 (.04) 93 .19 (.04) 93 .18 (.04) 93
   ×2 .24 (.04) 80 .25 (.04) 88 .24 (.04) 69 .27 (.04) 93 .28 (.04) 95 .28 (.04) 91
0.0 0.3 × 1 .18 (.04) 88 .18 (.04) 91 .17 (.04) 88 .20 (.04) 95 .20 (.04) 93 .20 (.04) 96
   ×2 .26 (.04) 84 .26 (.04) 93 .26 (.04) 82 .29 (.04) 97 .29 (.04) 94 .29 (.04) 98
0.1 0.3 ×1 .17 (.04) 88 .17 (.04) 88 .17 (.04) 88 .18 (.04) 93 .19 (.04) 95 .18 (.04) 94
   ×2 .25 (.04) 77 .25 (.04) 88 .25 (.04) 76 .28 (.04) 93 .28 (.04) 93 .28 (.04) 93
0.3 0.3 ×1 .14 (.04) 70 .14 (.04) 72 .13 (.04) 63 .15 (.04) 75 .15 (.04) 78 .14 (.04) 71
   ×2 .22 (.04) 56 .23 (.04) 71 .21 (.04) 48 .23 (.04) 74 .24 (.04) 83 .23 (.04) 59
  1. The true population values are bpred ×1 = 0.20, bpred ×2 = 0.30. Different degrees of dependency between study variables and non-response are modeled.
  2. Y = health outcome. bnon = regression coefficients of normally distributed liability of non-response (L) on predictors (× 1, × 2, ×3) and on health outcome.
  3. bpred = coefficients for the regression of health outcome on × 1 and × 2, when the outcome is treated as a categorical variable, and the probit-link is used.
  4. SE standard error, 95% coverage percentage of the randomly drawn samples providing a 95% confidence interval containing the true population value. FIML full information maximum likelihood, MI multiple imputation (predictive mean matching). N in the original sample before non-response = 1000. × 3 was included as auxiliary variable in FIML and as predictor in MI. 50 data sets were imputed.