Should I delete a row with missing data?

Find the variable’s valid range and missing-value codes in the codebook. A special value recognised as missing by one program may be read as an ordinary number after import into another. In the UK Data Service example, the missing-value codes −9 and −8 need to be explicitly recoded before analysis. These are not universal codes. UK Data Service, pp. 1–3

The six records below are entirely fictional. Each score is the number of correct answers, ranging from 0 to 10. A dash denotes an unrecorded score in this display; the missing-value representation in the actual analysis file must be defined separately.

Deleting record 03’s zero would discard an observation. Replacing the missing scores for records 02 and 05 with zero would record an unmeasured outcome as “no correct answers”. In this example, the four observed final scores sum to 18 and have a mean of 4.5. Counting the two missing scores as zero produces a mean of 3: you have changed the result without obtaining the missing measurements.

Keep the raw file. Document recoding and the reasons for missingness separately. Do not fill an unknown reason with a guess.

These categories are not diagnoses that can be established simply by inspecting a file. In particular, MAR cannot be verified from the observed data alone. National Research Council, Chapter 4, pp. 50–54

Do not stop at “file not found”: which class, device or date was involved? If that information is unavailable, record the uncertainty.

1. Variable summary: Report the number of valid and missing records for each variable, with the denominator used. A single percentage across all cells can conceal missing final scores.
2. Pattern table: Identify the groups, measurement occasions or data-collection sessions in which values are missing. Distinguish participants missing entirely from those with one unanswered item.
3. Process note: Record known reasons, the source of that information and explanations that remain unverified. Keep participant statements separate from researcher interpretations.
4. Analysis decision: State the target quantity, chosen method, underlying assumption and alternative assessment. For example, a “mean final score” and “change from baseline to the final assessment” do not require identical analysis inputs.

This decision record is AcademicFix’s suggested workflow.

Complete-case analysis uses records with no missing values in the relevant analysis variables. Its validity depends on the target, model and missingness process; it is not automatically right or wrong. National Research Council, Chapter 4, pp. 55–56

Multiple imputation creates several plausible completed datasets, analyses them separately and pools the results. For example, mice provides different imputation models by variable type, predictor selection and diagnostic plots. Having these options does not mean that their assumptions are automatically satisfied. Official mice documentation

If you use the method, record the variables included in the model, imputation methods, numbers of imputations and iterations, diagnostic checks and pooling step. Do not assume the software’s default number of imputations is adequate for every study. Do not report a single completed table as the “true data”.

The NRC source concerns clinical trials; this guide draws on its general statistical principles. The student example is not clinical regulatory advice.

Return to the fictional six-person table. If x is the mean of the two missing final scores, the mean final score for all six people is (18 + 2x) / 6. Since x must lie between 0 and 10 on this scale, these facts alone place the full-group mean between 3 and 38/6 (approximately 6.33).

This range is not a confidence interval, a multiple-imputation result or a recommended estimate. It is only a mathematical bound illustrating why information about the two unobserved values matters. In a real analysis, plausible alternative assumptions need to reflect the data-collection context. Sensitivity analysis examines how much the inference changes when those assumptions change. National Research Council, Chapter 5

Fictional opening sentence: “The final measurement was missing for 20 of the 100 students with a baseline record (20%). Missing counts were reported by group and measurement occasion. Known reasons were recorded separately.”

Follow this with the method actually used, the rationale for its assumptions, the number of records analysed and the sensitivity assessment. Do not describe analyses that have not been performed as completed work.