Table 4 Considerations when undertaking systematic reviews of prognostic factors
Considerations | Description |
|---|---|
Primary study identification | Studies are not necessarily ‘badged’ as prognostic/predictive and a variety of terms are inconsistently used (e.g. risk, association, relationship etc.) |
Using prognostic filters substantially reduces the volume of search hits, but it is likely that relevant studies will be missed | |
Study selection | Selection criteria are not consistently reported. This may be particularly important in terms of specifying study design (retrospective/prospective) |
Hierarchy of studies | Where large numbers of (poor quality) primary studies are identified, a step-wise approach to inclusion may be feasible: (i) inclusion only of studies reporting a prognostic model/ results adjusted for other prognostic factors, (ii) inclusion of prospective studies reporting on a single prognostic factor and (iii) inclusion of all studies reporting on a single prognostic factor |
Definition of prognostic factor | If identifying a potential prognostic factor is dependent on a diagnostic test, then diagnostic accuracy aspects of one or more tests may need to be assessed in a separate exercise (the QUADAS tool [19] may be appropriate for this) |
Consider whether it is clinically appropriate to dichotomise prognostic factor or whether it should be used as a continuous variable (particularly in a model) | |
Quality assessment | The QUIPS tool [22] should be used to inform quality assessment rather than tools relating to specific study design; further tailoring may be necessary depending on topic specific issues |
Analysis | Meta-analysis should only be undertaken after extensive consideration of clinical and methodological heterogeneity |
Data for meta-analysis can potentially be maximised by converting outcome statistics, which may also allow exploration of sensitivity of results to use of statistics | |
Meta-analysis results should be made specific to particular threshold values or ideally for the factor left on its continuous scale | |
Adjusted results should be presented where possible | |
Time-to-event analyses should be considered when accounting for different lengths of follow-up | |
Small study effects (potential publication bias) should be examined in those meta-analyses containing ten or more studies | |
Models based on individual patient data should be considered |