Small studies that find nothing tend not to get published. So the literature looks like a funnel with one side missing — and the pooled estimate is pulled toward a lie. Slide (or tap) to restore the suppressed studies and watch the funnel go symmetric and the estimate fall back to the truth.
A funnel plot: effect size on x, precision (1/SE) on y, so precise studies sit at the top and scatter narrows upward. The truth here is a small real effect (μ₀=0.20). Publication bias suppresses the small, imprecise, non-significant studies — the bottom-left — leaving an asymmetric funnel whose published-only inverse-variance pool is INFLATED. Restore them and the funnel is symmetric and the pool returns to μ₀. A fail-loud self-check throws unless the published-only pool exceeds the full-data pool.
The study cloud is a fixed synthetic sample around μ₀=0.20 with a simple 'significant-or-large-only gets published' suppression rule; real bias is messier and detected statistically (Egger's test, trim-and-fill). The pooling arithmetic and the direction of the inflation are exact.