RS517 Chap.10 Evidence Synthesis and Research Inference
Evidence Synthesis and Research Inference
Review before pooling
A systematic review uses an explicit question, eligibility criteria, search, selection, appraisal and synthesis. Meta-analysis is an optional combination inside that process. A pooled estimate cannot repair an incomplete or biased evidence set.
Heterogeneity carries meaning
Studies differ in participants, interventions, comparators, outcomes, follow-up and bias.
Statistical heterogeneity is one signal of those differences. A random-effects model permits varying effects but does not explain them or make incompatible outcomes comparable.
Different designs answer different questions
Survival analysis preserves event timing and censoring. Qualitative research can illuminate experience, mechanisms and implementation without estimating prevalence.
Evidence hierarchies must be applied to the question, because design labels do not erase execution or relevance.
Build the review before pooling estimates
A systematic review requires a focused question, eligibility criteria, reproducible search, transparent selection, critical appraisal and structured synthesis. Meta-analysis is optional.
Studies must be comparable enough in population, intervention, comparator, outcome, timing and effect scale for a pooled average to answer something coherent. Statistical precision cannot repair an incomplete search, selective reporting or shared design bias.
Keep study-level evidence visible beside any diamond.
Explain heterogeneity instead of hiding it
A random-effects model permits variation in true effects but does not explain that variation or make incompatible measurements equivalent. Examine clinical and methodological differences first, then statistical heterogeneity.
Subgroup analysis and meta-regression need prespecified hypotheses and enough studies; otherwise chance patterns are easy to overread.
A prediction interval can describe the range of effects expected in a comparable new setting and is often more relevant to transport than the confidence interval around the pooled mean.
Match synthesis to the evidence type
Qualitative synthesis can explain experience, mechanism, implementation and context without estimating prevalence.
Its credibility depends on sampling for relevant variation, transparent analysis and support for interpretations. Time-to-event evidence preserves follow-up duration and censoring: Kaplan-Meier curves estimate survival, while hazard models compare instantaneous rates under assumptions. Competing events may prevent the outcome rather than act as ordinary censoring.
Each method answers a different question, so design hierarchies must be applied to the question being asked.
Revision drill
Read a forest plot in this order: study populations, effect scales, intervals, risk of bias, heterogeneity, pooled result and applicability. Write a conclusion separating statistical evidence from clinical importance.
Then describe a qualitative finding without turning it into a percentage, and interpret a hazard ratio without calling it a final risk ratio. The closing sentence should be no stronger than the weakest essential link in the evidence chain.
What this chapter covers
- 01
Systematic review question and search
- 02
Eligibility and critical appraisal
- 03
Forest plots and study weights
- 04
Heterogeneity and random effects
- 05
Survival analysis and censoring
- 06
Qualitative inference and evidence levels
Worked example · free
Read a pooled estimate against clinical importance
- 4Separate statistical evidence from clinical threshold.
- 4Include heterogeneity and applicability.
Key terms
- Systematic Review
- A reproducible process for finding, appraising and synthesising evidence for a defined question.
- Meta-analysis
- Statistical combination of compatible study estimates under an explicit model.
- Heterogeneity
- Variation in study effects or characteristics beyond a single common description.
- Censoring
- Knowledge that an event was not observed up to a participant's final follow-up time.
- Forest Plot
- A display of study estimates and intervals together with any pooled estimate.
Evidence Synthesis and Research Inference FAQ
Is meta-analysis always better than narrative synthesis?
No. Pooling is useful when study questions, outcomes and effect measures are sufficiently compatible. A statistical model cannot make incomparable studies equivalent. When heterogeneity reflects important clinical differences, structured narrative synthesis or separate pools may be more informative.
What does a random-effects model solve?
It allows the modelled true effect to vary across studies and incorporates between-study variation into uncertainty. It does not explain the source of variation, remove bias or guarantee that an average applies to a particular setting. Clinical comparison remains essential.
Can qualitative research support a clinical proposal?
Yes, when the question concerns experience, implementation, acceptability or mechanism. Judge purposeful sampling, analytic transparency, reflexivity and support for interpretations. Do not convert themes into prevalence claims, but use them to explain why an intervention may work differently across contexts.
Exam move
Read forest plots from study rows to pooled diamond, noting population, effect direction, interval and risk of bias. Practise one survival question by stating the event, time origin and censoring rule. Finish each evidence answer with the population and decision to which the conclusion applies.
Working through Evidence Synthesis and Research Inference in RS517? Sia is AskSia’s AI Statistics tutor — ask any RS517 Evidence Synthesis and Research Inference question and get a clear, step-by-step explanation grounded in how RS517 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.