professor research

Listwise deletion is a straightforward method that produces unbiased estimates when data are MCAR; however, it severely diminishes statistical power and biases results when the MCAR assumption does not hold. Mean substitution maintains sample size, but it does not add any new information, underestimates standard errors, and can lead to serious biases with non-random missing (Kang, 2013). While easy to implement and understand, Last Observation Carried Forward (LOCF) assumes that outcomes do not change over time and biases treatment effect estimates.

Multiple imputation, unlike the other methods, accounts for uncertainty of missing values and generates proper statistical inferences. It works well for MAR data and is not sensitive to violations of normality assumptions. On the other hand, they are more computationally intensive, and require advanced statistical knowledge and more computational power (Kang, 2013). Maximum likelihood methods, such as Expectation-Maximization algorithms, produce parameter estimates that are more theoretically justified, but their computation may be too expensive, especially for large datasets with a lot of missing data, where convergence would be slow as well.

When one use as a threshold or guideline in terms of might when missing data be estimated vs. deleted

A set of useful, though not universal, heuristics is diligent to the literature. When missing data is between 20-30%, deletion is likely to be deleterious, on the account of power loss and bias. If the missing data pattern is less important to the decision and more focus on the scattered missing that is easier to control in multiple variables than with the key ones, the decision should incorporate its pattern.

More importantly than any arbitrary percentage, researchers should prioritize understanding why data are missing. If missing appears related to study outcomes or participant characteristics, imputation methods become essential regardless of the proportion missing (Osborne & Overbay, 2004). On the other hand, when data are genuinely MCAR and the sample sizes continue to be sufficient for the analysis, deletion is likely to be fine even for mild to moderate levels of missing. In conformity with the research context, analytical goals, and the implications of the potential biases, the decisions concerning the standards.

References