Skip to main content

Table 1 Measures of model fitness

From: Use of days alive without life support and similar count outcomes in randomised clinical trials – an overview and comparison of methodological choices and analysis methods

Model

Root mean squared errors

Median absolute errors

Control group difference in predicted vs. observed means

Intervention group difference in predicted vs. observed means

Control group mean

Intervention group mean

Mean difference

Ratio of means

Linear

12.6 (12.6 to 12.6)

13.1 (12.0 to 13.9)

0.02 (-1.11 to 1.15)

0.01 (-1.09 to 1.13)

14.9 (13.8 to 16.0)

16.5 (15.4 to 17.6)

1.59 (-0.01 to 3.18)

1.11 (1.00 to 1.23)

Hurdle-negative binomial

12.6 (12.6 to 12.7)

13.1 (12.0 to 13.9)

-0.01 (-1.20 to 1.23)

0.00 (-1.21 to 1.20)

14.9 (13.7 to 16.1)

16.4 (15.2 to 17.6)

1.58 (-0.16 to 3.28)

1.11 (0.99 to 1.24)

Zero–one-inflated beta

12.6 (12.6 to 12.6)

13.2 (12.1 to 13.9)

-0.03 (-1.15 to 1.11)

-0.12 (-1.23 to 0.96)

14.8 (13.7 to 16.0)

16.3 (15.2 to 17.4)

1.49 (-0.09 to 3.05)

1.10 (0.99 to 1.22)

Cumulative logistic (death converted to 0 days in predictions)

12.6 (12.6 to 12.6)

13.0 (12.0 to 13.9)

0.13 (-0.99 to 1.25)

-0.17 (-1.24 to 0.90)

15.0 (13.9 to 16.1)

16.3 (15.2 to 17.3)

1.28 (-0.26 to 2.84)

1.09 (0.98 to 1.20)

Cumulative logistic (death predicted as -1 days)

13.0 (13.0 to 13.0)

13.3 (12.3 to 14.4)

0.14 (-1.01 to 1.30)

-0.17 (-1.27 to 0.93)

14.7 (13.5 to 15.8)

16.0 (14.9 to 17.1)

1.32 (-0.27 to 2.92)

1.09 (0.98 to 1.21)

  1. Model fitness, predictions, and effect estimates for all four models fit to the COVID STEROID 2 trial data [15]; days alive without life support at day 28 was modelled after assigning 0 days to non-survivors in all models except the cumulative logistic model, where non-survivors were assigned -1 (i.e., a category worse than all other categories). Values are presented posterior medians with 95% percentile-based credible intervals. Effect estimates were calculated by predicting the mean value for one patient in each group (as no covariate adjustments were used in these models).