What is Meta-Analysis?
Meta-analysis is a statistical technique that combines results from multiple studies to produce a single summary estimate. It increases statistical power and precision beyond individual studies and can resolve inconsistencies across studies.
When to Use Meta-Analysis
- You have 2+ studies measuring the same outcome
- Studies are sufficiently similar (population, intervention, design)
- Outcome data is reported in combinable format
- Heterogeneity is not prohibitively high
- Pooling makes clinical/scientific sense
When NOT to Meta-Analyze
- Only one study available
- Studies too heterogeneous (different populations, interventions, outcomes)
- Poor quality studies that would bias results
- Insufficient data reported for effect size calculation
- Pooling wouldn't answer a meaningful question
Remember: Just because you can meta-analyze doesn't mean you should. Narrative synthesis may be more appropriate for highly heterogeneous studies.
Types of Meta-Analysis
Fixed-Effect Model
Assumes all studies estimate the same underlying effect:
- Use when: Studies are very similar, heterogeneity is low
- Gives more weight to larger studies
- Narrower confidence intervals
- Less commonly used (most reviews have some heterogeneity)
Random-Effects Model
Assumes studies estimate different but related effects:
- Use when: Some heterogeneity expected (most cases)
- Accounts for between-study variance
- Wider confidence intervals (more conservative)
- Generally preferred for systematic reviews
Effect Size Metrics
For Continuous Outcomes
Standardized Mean Difference (SMD):
- Use when: Different scales measuring same construct
- Example: Various depression scales (BDI, HAM-D, PHQ-9)
- Expressed in standard deviation units
- Cohen's d interpretation: 0.2=small, 0.5=medium, 0.8=large
Mean Difference (MD):
- Use when: All studies use same scale
- Example: All studies report HbA1c in %
- Results in original units (clinically interpretable)
- Preferred when possible
For Binary Outcomes
Risk Ratio (RR) / Relative Risk:
- Ratio of event probability in intervention vs control
- Easy to interpret
- RR = 1 means the event is equally likely in both groups
- RR < 1 means the event is less likely in the intervention group; RR > 1 means it is more likely
- Whether either direction is favorable depends on the outcome and how the groups are ordered
Odds Ratio (OR):
- Ratio of odds of event in intervention vs control
- Useful for case-control studies
- Less intuitive than RR
- Similar to RR when events are rare (<10%)
Risk Difference (RD):
- Absolute difference in event rates
- Most clinically interpretable
- Example: RD = -0.10 means 10% absolute risk reduction
Meta-Analysis in Scholara
- 1Ensure outcome data has been extracted via the Analysis Chat (or ask the AI to extract it)
- 2In the Analysis Chat, ask the AI to run a meta-analysis (e.g., 'Run a random-effects meta-analysis for depression outcomes')
- 3The AI extracts the needed data, selects the appropriate effect size metric, and runs the analysis
- 4A forest plot and summary statistics appear in the chat as interactive assets
- 5Refine by asking follow-up questions (e.g., 'Use fixed-effect model instead', 'Add subgroup analysis by study design')
- 6Export the forest plot and results table in your preferred format
Interpreting Results
Summary Effect
- Point estimate: Average effect across studies
- 95% CI: Range of plausible values
- If CI excludes null (0 for MD/SMD, 1 for RR/OR), effect is statistically significant
- p-value: Assuming the null hypothesis and statistical model are correct, the probability of results at least as extreme as those observed; it does not measure clinical importance
Heterogeneity Statistics
I² (I-squared):
- % of variability due to heterogeneity vs chance
- 0-40%: Low heterogeneity
- 30-60%: Moderate heterogeneity
- 50-90%: Substantial heterogeneity
- 75-100%: Considerable heterogeneity
Tau² (Tau-squared):
- Variance of true effects across studies
- Used in random-effects model
- Larger values = more heterogeneity
Cochran's Q test:
- Tests null hypothesis of homogeneity
- p < 0.10 suggests significant heterogeneity
- Note: Low power with few studies, high power with many
Dealing with Heterogeneity
If heterogeneity is high (I² > 50%), consider:
- 1Check for data entry errors
- 2Investigate sources (subgroup analysis, meta-regression)
- 3Sensitivity analysis excluding outliers
- 4Use random-effects model (already accounts for heterogeneity)
- 5Consider narrative synthesis instead of pooling
- 6Report heterogeneity and discuss limitations
Publication Bias
Small, non-significant studies are less likely to be published, potentially biasing results. Scholara provides tools to assess:
- Funnel plot: Visual inspection for asymmetry
- Egger's test: Statistical test for funnel plot asymmetry
- Trim-and-fill: Estimates effect after adjusting for missing studies
- Fail-safe N: How many null studies needed to change conclusion
Quality of Evidence
Use GRADE approach to rate certainty of evidence:
- High: Very confident in estimate
- Moderate: Moderately confident; true effect likely close
- Low: Limited confidence; true effect may differ substantially
- Very low: Very little confidence; true effect likely substantially different
- Downgrade for: risk of bias, inconsistency, indirectness, imprecision, publication bias
Meta-analysis provides a precise summary, but precision ≠ accuracy. Always interpret results in context of study quality, heterogeneity, and potential biases. A well-done narrative synthesis is better than a poorly done meta-analysis.