<p>This paper examines how sample composition, the universe of stocks used to test anomalies, affects their significance. Prior studies bundle data-quality filters, investability screens, value-weighting, and factor adjustment, making the sample composition effect hard to isolate. A cumulative filter ladder addresses this by imposing stricter screens, from data-quality to investability and listing restrictions, while holding portfolio construction fixed. A large fraction of anomalies significant in the full sample lose significance by the end of the ladder. This attrition is driven by the investability and market-access screens, not routine data-quality filters, reflecting genuine alpha loss rather than weaker precision or shifting factor exposures. The most resilient anomalies are based on options, analyst, and price data, whereas many accounting and institutional-ownership signals are fragile, losing significance at different filter steps. The survivors tend to begin from stronger initial alphas, and those that drop out from weaker ones. The broader implication is that much of the factor zoo relies on microcap amplification, leaving the robust set that survives the full ladder substantially smaller than conventional full-sample evidence suggests.</p>