Skip to main content
. Author manuscript; available in PMC: 2018 Jan 1.
Published in final edited form as: Med Care. 2017 Jan;55(1):31–36. doi: 10.1097/MLR.0000000000000599

Table 2.

Fit statistics of hierarchical logistic models predicting vena cava filter use

Model 1: Random effects only Model 1 + level 1 fixed effects Model 1 + level-2 fixed effects Full model: Model 1 + level 1 and level-2 fixed effects a Cancer model b
Intercept (SE) −2.70 (0.09) −5.76 (0.18) −3.38 (0.20) −6.50 (0.26) −7.10 (0.86)
Hospital random effects, τ (SE) 0.45 (0.10) 0.48 (0.11) 0.22 (0.05) 0.25 (0.06) 0.12 (0.04)
ICCc 12.0% 12.7% 6.4% 7.1% 3.5%
C-statistic 0.62 0.81 0.62 0.81 0.81
AICd 50326.46 42570.91 50301.07 42548.15 –
BICd 50330.61 42670.58 50315.61 42658.19 –
a

Level 1 fixed effects are patient level fixed effects including all demographic and clinical characteristics. Level-2 fixed effects are hospital characteristics.

b

Cancer only model included only individuals with cancer and the individual sites of cancer (Table 4). Fit statistics are not included since it was not compared to other models.

c

Intraclass correlation coefficient: The proportion of the model variance explained by the “hospital” parameter; e.g. 12.0% of the Model 1 variance is explained by the hospital where a person is discharged. Calculated by τ/τ+3.29 for a binary logit model. All ICC values between hospitals were significant at p<0.001.

d

Akaike information criterion and Bayesian information criterion fit statistics for comparison between models. Each measures the model fit but penalizes for additional parameters added to each model. Smaller values are preferred; thus, the full model is preferred over Model 1.