Missing data: deletion, imputation and reporting

Problem

Missing data are common in real research datasets, but students often handle them casually. They may ignore missingness, delete incomplete rows automatically or replace missing values with simple averages without considering why the data are missing. These choices can change the sample, reduce power, introduce bias and affect conclusions. Missing data handling is therefore not just a technical cleaning step; it is part of the validity of the analysis.

• Missing values are sometimes hidden as codes such as 99, 999 or -1.
• Students often delete incomplete rows without reporting how many were removed.
• Mean imputation is used without understanding its consequences.
• Different analyses may use different sample sizes without explanation.
• Missingness may be related to the outcome or exposure.
• Complete-case analysis can reduce power and bias estimates.
• The missing data method is often absent from dissertation reports.

Intuition

The key question is not only how much data are missing, but why they are missing. If data are missing completely at random, deletion may be less problematic. If missingness is related to observed variables, methods such as multiple imputation may be more appropriate. If missingness depends on unobserved values, no simple method fully solves the problem and sensitivity analysis becomes important.

• Missing completely at random means missingness is unrelated to observed or unobserved data.
• Missing at random means missingness can be explained by observed variables.
• Missing not at random means missingness depends on unobserved values or the missing value itself.
• Small amounts of missing data can still matter if missingness is systematic.
• Large amounts of missing data can reduce precision and credibility.
• The missingness pattern should be described before choosing a method.

Method

A sensible missing data workflow begins by identifying missing values correctly. Then the analyst should summarise the amount and pattern of missingness, compare complete and incomplete cases where useful, decide on a handling method and report the decision clearly. The method should be justified based on the research question, amount of missingness, likely missingness mechanism and complexity of the analysis.

• Step 1: Identify all missing value codes in the raw dataset.
• Step 2: Convert hidden missing codes into proper missing values.
• Step 3: Count missing values for each variable.
• Step 4: Examine missingness patterns across key variables.
• Step 5: Check whether missingness is related to observed characteristics.
• Step 6: Decide whether complete-case analysis is defensible.
• Step 7: Consider multiple imputation when missingness is substantial and plausibly related to observed variables.
• Step 8: Avoid simple mean imputation for formal inference.
• Step 9: Compare results across methods where possible.
• Step 10: Report missing data handling transparently.

Working

Suppose a health study examines whether BMI is associated with blood pressure, but BMI is missing for 18% of participants. If those with missing BMI are similar to those with observed BMI, complete-case analysis may be reasonable, although power is reduced. If BMI is more likely to be missing among older participants or those with severe disease, deletion may bias results. Multiple imputation may be considered if the missingness can be explained using observed variables.

• First, report how many participants have missing BMI.
• Second, compare key characteristics between complete and incomplete cases.
• Third, decide whether deletion changes the analysis sample substantially.
• Fourth, consider whether missingness is plausibly related to observed variables.
• Fifth, use complete-case analysis only if the assumptions are acceptable.
• Sixth, consider multiple imputation for more defensible analysis when appropriate.
• Seventh, report whether conclusions changed under different approaches.

Limitations

No missing data method is magic. Complete-case analysis can be biased. Simple imputation can underestimate variability. Multiple imputation relies on assumptions and must be implemented carefully. If data are missing not at random, standard methods may still be biased. The most important principle is transparency: readers should know what was missing, how it was handled and how it may affect conclusions.

• Complete-case analysis can waste data and introduce bias.
• Mean imputation distorts variability and relationships between variables.
• Multiple imputation requires careful model specification.
• Imputation cannot recover information that was never collected.
• Missing not at random mechanisms are difficult to handle.
• High missingness can weaken credibility even after imputation.
• Sensitivity analysis may be needed when assumptions are uncertain.

Discussion

A strong dissertation or research report should include a short missing data statement. This should describe the extent of missingness, the method used, the rationale for that method and the possible impact on interpretation. Missing data should not be hidden because unexplained changes in sample size can undermine confidence in the analysis.

• Report the number and percentage missing for important variables.
• Explain whether the analysis used complete cases or imputation.
• Justify the chosen method briefly.
• Mention whether missingness may have biased the findings.
• Avoid pretending missing data do not matter.
• Use tables or flow diagrams where sample exclusions are important.
• Discuss missing data as a limitation when relevant.

How this connects to learning

Use the guide as a bridge between theory and application.

A resource guide should not replace a full course or live teaching session. Instead, it helps you organise your thinking. Use it to identify what you understand, what feels unclear, and what questions you should ask before applying a method to real data.