Track Complete-Case Denominators for Each Planned Model

Define completeness for a specific analysis

A record is complete only relative to a specified set of required fields. A participant with an observed outcome and exposure may be usable for one model but lack a covariate required by another. Avoid a universal complete-case flag created from every column in the database, including fields irrelevant to the analysis.

Write the planned outcome, exposure, adjustment variables, population restrictions, and any transformation requirements. A nonmissing raw value can still become unusable after a transformation, such as taking the logarithm of a nonpositive value. Keep those reasons distinct from ordinary missingness.

Build an explicit membership audit

For each model, count eligible records, records with each required field available, and records satisfying the full conjunction of requirements. Preserve stable identifiers in the approved analysis environment so changes in membership can be inspected without publishing identifying details.

The official pandas dropna documentation shows that removal can depend on a specified subset of columns. That software operation makes the selection reproducible, but does not establish that complete-case analysis is scientifically appropriate or unbiased.

Reconcile fictional models

Consider six fictional participants. Outcome Y and exposure X are both observed for five. A planned covariate Z is missing for two of those five, so a model requiring Y, X, and Z has three complete records. A separate unused laboratory field is missing for everyone; it should not reduce either model's denominator merely because it is present in the export.

The missing-pattern tool can illustrate combinations of observed and missing fields with synthetic inputs. For the field order Y, X, Z, enter 000 with count three, 001 with count two, and 100 with count one: here 1 means missing and 0 means observed. The model-specific denominator comes from the joint pattern, not from subtracting each variable's missing count independently.

Interpret differences between analyses carefully

If a coefficient changes after adding Z, two things may have changed: the adjustment set and the participants contributing to the fit. An additional comparison on a common eligible subset can help separate those influences descriptively, but it does not automatically remove selection bias or justify the final model.

Use the missing-pattern worksheet to document model labels, required variables, exclusions, and overlap between samples. Reconcile these counts with the software's actual fitted-sample count, including any rows removed for invalid values or estimation-specific requirements.

Cochrane's missing-outcome bias discussion explains why missingness and its relationship to outcomes require substantive assessment. A small percentage or a tidy membership table cannot prove the assumptions behind a chosen handling method.

Finish with a model-by-model denominator table and reasons for changes. Discuss the analysis question, missing-data assumptions, and appropriate sensitivity analysis with the study team before treating complete-case filtering as the default answer. Reproducible exclusion is necessary for auditing; it is not sufficient for valid inference.

Sources checked on 3 November 2026; fictional model membership independently enumerated.