Data set Option Effects on Grand Mean (GM) Effects on approximated Confidence Interval (CI)
Fig. 6 – corMCAR 1) Complete-case analysis Increased in volatility with the percentage of missing data. Increased non-linearly with the percentage of missing data.
2) Unweighted analysis With deviation. Unbiased, except smaller CI for Fisher’s z.
3) SS-weighted analysis With deviation. Unbiased.
Mean value, median value, and random sample imputation Linear regression Predictive mean matching, classification and regression trees, random forest imputation Bayes predictive mean matching Bootstrap expectation maximization Non-parametric random forest Additive regression and bootstrap predictive mean matching Deviation increases approximately linearly with the percentage of missing data, most strongly for missing SDs and for Fisher’s z. Unbiased below ca. 80% of missing data. Unbiased below ca. 40-60% of missing data, then with increasing deviation and volatile with the percentage of missing data Unbiased below ca. 50% of missing SDs in the log response ratio, ca. 80% of missing SDS in Hedges’ d and 80% in Fisher’s z. Unbiased until a threshold of ca. 50% of missing data, above which the algorithm frequently failed to converge. Unbiased below ca. 70-80% of missing SDs in the log response ratio, nearly linear increase in deviation with the percentage of missing SDs in Hedges’ d and missing SSs in Fisher’s z. Unbiased until a threshold of ca. 50-70% of missing SDs in the log response ratio and in Hedges’ d, above which the algorithm frequently failed to converge. Unbiased below ca. 60-70% of missing data in Fisher’s z. Unbiased, except smaller CI for Hedges’ d. Unbiased, except larger above ca. 70% of missing both, SDs and SSs. Unbiased, except larger above ca. 70% of missing both, SDs and SSs. Unbiased, except larger above ca. 70% of missing both, SDs and SSs. Unbiased. Unbiased. Unbiased.