The decision-tree algorithm divides the data to reduce the deviance, and classifies them into the pre-defined categories as many tree branches.
(iii) Variables with dichotomized P values>0.3 were added to the model, one by one, and any variable that reduced the deviance significantly was retained in the model.
Our Risk Difference Model allows for the initial risk difference between the intervention groups and control groups and reduced the deviance from 23.54 in the Simple Model to 18.5, and thereby achieved the smallest deviance measure among all six models (ΔDev = 5.04, Δdf = 1, p = 0.02).
The addition of the prognostic score to the model with clinical covariates reduced the residual deviance with a X1 = 2.96, p = 0.09.
The tree construction process attempts to reduce the node deviance, defined as ∑ y − y ¯ 2, where y is the relative bias and y ¯ is the mean relative bias within a node.
The inclusion of the additional interaction term reduced the residual deviance by 0.38, 2.80 and 0.08 respectively on 4, 1 and 1 degrees of freedom; these reductions are all small in magnitude and none are statistically significant.
In particular, adding SSR type as an explanatory factor to the null model reduces the residual deviance by over 1500, which is clearly statistically significant (p < 0.0001 when compared to a chi-squared distribution with 18 degrees of freedom).
For COPD the deviance reduces from 1,038.1 on 128 degrees of freedom to 421.8 on 112 degrees of freedom on fitting the basic model, substantially reducing, but not eliminating, the heterogeneity.
For CB the deviance reduces from 657.1 on 113 degrees of freedom to 433.3 on 103 degrees of freedom on fitting the basic model, again substantially reducing, but not eliminating, the heterogeneity.
