Table 4.
Selection of variables in multivariable analysis
| % (n = 43*) | |
|---|---|
| Selection of variables for inclusion in multivariable analysis | |
| All candidate variables used (no selection) | 26 (11) |
| All candidate variables apart from a few with contra indications** | 5 (2) |
| Without statistical analysis | |
| Previous literature | 5 (2) |
| Previous literature and few variables by investigator choice | 5 (2) |
| By statistical analysis | |
| Screening by univariable analysis - only significant variables | 37 (16) |
| Screening by univariable analysis - significant variables and investigator choice | 11 (5) |
| Unclear/Not reported | 11 (5) |
| Statistical modelling methods used within multivariable analysis | |
| A priori variables fixed, others added | 2 (1) |
| Backward elimination | 14 (6) |
| Forward selection | 5 (2) |
| Other (pairwise multiple testing for categories of variables) | 2 (1) |
| Unclear/Not reported | 77 (33) |
| Methods for inclusion of variables in final model and prognostic index | |
| No selection. All variables kept in model | 14 (6) |
| Retain only significant variables based on P-value | 65 (28) |
| Retain significant variables plus variables based on previous literature | 2 (1) |
| Retain all variables but alter prognostic score after model to include only significant variables and adjust for other variables | 5 (2) |
| Retain only significant variables but alter prognostic score after final model | 5 (2) |
| Retain based on model goodness of fit | 2 (1) |
| Unclear/Not reported | 7 (3) |
* Excluded four studies using recursive partitioning analysis and artificial neural network models
** Contra indications reported as reasons for exclusion of variables were missing data, collinearity and treatment indicator