Good coding cannot rescue a bad death certificate

#Question

Why is coding accuracy only one part of mortality-data quality?

#Answer in brief

Because mortality data are created in stages. A clinician certifies conditions and their sequence; civil-registration and health-information systems capture the death; ICD coding standardizes the recorded information; mortality rules select an underlying cause; analysts aggregate and compare results. WHO's current cause-of-death framework explicitly treats certification, coding, selection and quality analysis as parts of one pipeline.^[1]

If the certificate is vague, a perfect coder can only code vague information perfectly.

#Coding quality and information quality are different

Imagine two certificates.

Certificate A

Acute myocardial infarction due to atherosclerotic coronary artery disease.

Certificate B

Cardiac arrest.

Suppose both are coded exactly according to the applicable rules.

Coding accuracy could be excellent for both.

Yet Certificate B carries far less etiological information. The limitation is upstream.

This distinction prevents a common error: treating agreement between coders as proof that the mortality data are epidemiologically informative.

#Six dimensions of mortality-data quality

#1. Registration completeness

Were all deaths in the target population registered?

WHO notes that cause-of-death statistics depend on civil-registration systems and that incomplete coverage can limit interpretation.^[2]

Coding quality cannot compensate for deaths that never entered the system.

#2. Certification quality

Did the certifier record a medically meaningful causal sequence?

Ill-defined mechanisms, abbreviations, impossible sequences or omission of important antecedent conditions reduce what can be derived downstream.

WHO's cause-of-death resources place medical certification at the start of the mortality information chain for this reason.^[1]

#3. Coding quality

Were certified terms mapped correctly to ICD categories under the applicable version and coding rules?

This is the dimension most often associated with "coding accuracy," but it is only one stage.

#4. Selection quality

Was the underlying cause selected consistently according to mortality rules?

Automated selection can improve reproducibility, but reproducible application of rules still depends on the certified input.

#5. Comparability

Can rates be compared across time, place or populations?

Changes in ICD revision, automatic coding, certification practice or registration coverage can alter statistics independently of disease occurrence.

Eurostat metadata explicitly document such comparability issues and technical changes in national cause-of-death series.^[3]

#6. Fitness for purpose

Is the dataset adequate for the question?

A dataset suitable for national long-term trend analysis may be inadequate for small-area comparisons, rare causes, multiple-cause research or recent rapid monitoring.

Quality is therefore relative to intended use.

#A technically valid code can preserve an uninformative statement

This point deserves emphasis.

If "senility," "cardiac arrest" or another nonspecific formulation is accepted as source text, mapping it correctly to an ICD category does not make the underlying medical statement more informative.

The coding system is not a diagnostic inference engine that should invent missing causal history.

That boundary protects integrity: the coder translates and applies rules; the certifier supplies medical information.

#Missingness can be semantic

Mortality datasets can have no blank fields and still be information-poor.

For example:

• every record has a cause code;
• every date is valid;
• every geographic code is present.

Yet the causes may contain an unusually high proportion of ill-defined categories. A systematic review of death-certificate studies found recurring errors including incorrect underlying-cause statements, improper sequences, mechanisms recorded without an informative underlying cause, and nonspecific or ill-defined conditions.^[4]

Traditional completeness checks would pass. Semantic quality would remain weak.

A quality dashboard should therefore include both structural and epidemiological indicators.

#Comparability can fail without any error

Suppose a country introduces automated cause selection.

Before implementation, manual coders resolve ambiguous certificates using established practice. After implementation, rule application becomes more standardized.

Neither period is necessarily "wrong."

But the change can shift assignment among categories.

A discontinuity around implementation may reflect improved standardization rather than altered mortality.

Eurostat's cause-of-death metadata show why methodological and coding transitions need to accompany statistical series.^[3]

#A synthetic audit

A mortality unit reports:

• 99.8% valid ICD codes;
• 99.5% complete age and sex;
• low duplicate rate.

The dataset appears excellent.

A deeper audit finds:

• one hospital has a very high proportion of poorly specified causes;
• recent-year registration is incomplete;
• a coding software change coincides with a category shift;
• multiple-cause fields are inconsistently populated.

The correct conclusion is not "quality is poor."

It is:

structural coding validity is high, while certification specificity, recent completeness, temporal comparability and multiple-cause availability have material limitations.

That statement tells users what the data can and cannot support.

#Quality improvement should target the stage that failed

If the problem is certification:

• improve certifier training;
• redesign prompts;
• provide feedback on common errors.

If the problem is coding:

• audit code assignment;
• maintain coding standards;
• review software and rule versions.

If the problem is registration:

• improve capture and reconciliation across civil and health systems.

If the problem is comparability:

• document breaks;
• run bridge coding or sensitivity analyses where feasible;
• avoid false trend continuity.

If the problem is fitness for purpose:

• change the analysis or collect additional information.

A generic "data quality project" risks treating all defects as interchangeable.

#Why multiple-cause information matters

Underlying-cause statistics deliberately select one primary cause for tabulation.

That is useful but lossy.

For multimorbidity, ageing and complex causal pathways, contributing conditions can provide important context. A dataset optimized only for underlying-cause production may be insufficient for analyses requiring multiple-cause information.

The preceding article A cause of death is a constructed statistical object explains why underlying cause and multiple causes answer different questions.

#Quality should be visible to users

Mortality outputs should ideally document:

• registration coverage;
• ICD version;
• coding method;
• underlying-cause selection method;
• proportion of ill-defined causes;
• known breaks in series;
• completeness of recent periods;
• whether multiple-cause data are available;
• revisions.

WHO's Mortality Database itself warns users that reported data are not automatically adjusted for incompleteness.^[2] That kind of transparency is part of quality, not an admission of failure.

#Quality indicators can conflict

Improving one mortality-quality metric can make another look worse.

For example, better investigation can reduce ill-defined causes while increasing the proportion assigned to a more specific chronic category. A time series may become less comparable exactly because source quality improved.

Quality assessment should therefore record interventions that change certification or coding practice and interpret resulting trend breaks explicitly.

#Training should be evaluated by downstream information quality

Certifier training is often proposed when certificates are nonspecific.

Attendance at training is an output.

A stronger evaluation asks whether the intervention changes certificate structure, reduces unusable mechanisms, improves causal sequencing and does so without creating new systematic miscoding.

The programme-evaluation distinction between activity and outcome applies to mortality data quality too.

#Automated coding changes the error landscape

Automation can improve consistency and reproducibility in rule application. It cannot make a poorly certified causal sequence informative by itself: experimental recoding work shows that some certification errors materially alter the underlying cause selected, while others are partly mitigated by correct coding rules.^[5]

Automation can also scale systematic errors when input parsing, dictionary mappings or rule implementations are wrong.

A quality programme should therefore distinguish:

• reproducibility of the coding engine;
• validity of the input certificate;
• validity of mappings and rule implementation;
• monitoring of unexpected category shifts after software changes.

Automation moves some quality control from individual coders to software validation and change governance. It does not remove the need for review.

#Limitations and uncertainty

The examples here are synthetic and do not assess any national or regional mortality system. Indicators of poor certification, registration completeness and ill-defined causes require jurisdiction-specific definitions and thresholds. Coding and certification practice also change as ICD implementations evolve. The article addresses epidemiological interpretation, not professional instruction for certifying or coding an individual death.^[1]