Meta-Analysis

Use fixed-effects if studies are presumed to share a single true effect and differences are due to sampling error only. Use random-effects if effects are expected to vary by context (population, setting, duration), which is common in practice. Random-effects yields wider confidence intervals, accounting for unexplained heterogeneity. Most guidelines recommend random-effects as default; examine Q-test and I² to assess heterogeneity magnitude. Sensitivity analysis comparing both models is prudent.

I² (Inconsistency Index) quantifies the proportion of variation across studies due to heterogeneity rather than chance: I²=0% all variation is chance; I²=100% all variation is real heterogeneity. Interpretation: <25% low, 25–50% moderate, 50–75% substantial, >75% very high. I² can be calculated from Q (Cochran statistic) and degrees of freedom. Important: I² depends on number and precision of studies; high I² with few studies may inflate uncertainty, while low I² with many studies may hide important variation. Always interpret alongside forest plot visual inspection and subgroup analysis.

Studies with zero events in both arms contribute no information to odds ratio meta-analysis; exclude them. If one arm has zero events, add a continuity correction (e.g., 0.5 to all cells) to compute odds ratio; most software does this automatically. Document how zero-event studies were handled. Consider reporting incidence rates instead of odds ratios if many studies have rare events. Mantel-Haenszel or Peto methods are alternatives for rare events.

Asymmetry in the funnel plot suggests publication bias (positive studies preferentially published) or heterogeneity. Conduct statistical test (Egger's regression). If bias is suspected, discuss in the interpretation—the pooled estimate may overestimate true effect. Perform trim-and-fill analysis (estimates what effect would be after 'filling in' missing negative studies, though this is controversial). Acknowledge that absence of small negative studies weakens evidence certainty. Always note: funnel asymmetry can also reflect real heterogeneity or small-study effects (e.g., lower quality in small studies).

Meta-analysis with ≥2 studies is technically possible, but limited power and generalizability. With two studies, subgroup analysis and publication bias assessment are not meaningful. Most guidelines recommend ≥10 studies for robust conclusions. If only 2–3 studies exist, emphasize precaution in interpretation and note that results should be considered preliminary pending additional evidence. Narrative synthesis may be more appropriate than pooling.