A cause of death is a constructed statistical object

#Question

What exactly does a cause-of-death statistic represent?

#Answer in brief

It represents a standardized statistical construction built from a chain of evidence and rules. A clinician certifies the sequence of conditions associated with death. Those statements are coded to the International Classification of Diseases. Mortality coding rules are then used to select an underlying cause of death for primary tabulation. WHO defines that underlying cause as the disease or injury that initiated the train of morbid events leading directly to death, or the circumstances of the accident or violence producing the fatal injury.^[1]

That standardized construction is what makes mortality statistics comparable. It also means the final code should not be interpreted as if death had only one biologically meaningful cause.

#Death can involve multiple causal layers

Consider a synthetic death certificate describing:

• chronic diabetes;
• chronic kidney disease;
• bacterial pneumonia;
• acute respiratory failure.

A reader may ask, "Which one caused the death?"

The statistical system asks a more structured question: what sequence is certified, which conditions contributed, and which condition should be selected as the underlying cause according to international rules?

Different analytical questions may legitimately use:

• underlying cause;
• immediate cause;
• contributing conditions;
• multiple-cause information.

Reducing the entire certificate to one code is useful for primary tabulation, but it is not the only information available.

#Certification comes before coding

WHO emphasizes that the medical practitioner completes the cause-of-death certificate by recording conditions that led directly to death and antecedent conditions giving rise to the underlying cause.^[1]

This ordering matters.

A coder cannot reconstruct a causal sequence that was never documented clearly.

If the certificate contains only "cardiac arrest," the coding process has little information about the disease or injury that produced the arrest. The term may be technically encodable, but epidemiologically uninformative.

#Coding is translation

ICD coding translates clinical language into standardized categories.

That operation supports:

• aggregation;
• comparison;
• trend analysis;
• international reporting;
• automated processing.

But translation always depends on the source text.

Coding quality can be high while source certification quality is poor. The companion article Good coding cannot rescue a bad death certificate develops that distinction.

#Selection is a rule-governed operation

For routine mortality statistics, a single underlying cause is selected using standardized mortality rules.

WHO describes an end-to-end process of certification, coding, underlying-cause selection, quality assessment and analysis.^[1]

This rule-governed selection improves consistency. It prevents every analyst from choosing whichever diagnosis seems most salient.

Yet it also means that the underlying cause is a defined statistical variable, not an unmediated observation.

#Why the underlying cause is useful

Public health often needs a stable, mutually exclusive classification for each death in primary tabulations.

Without an underlying-cause convention, one death with five recorded conditions could contribute to five disease-specific death totals, making the totals unsuitable for many standard mortality rates.

The underlying cause provides one primary category per death.

Multiple-cause analysis can then answer a different question: which conditions were present in the certified causal sequence or contributed to death?

#Counts, crude rates and standardized rates answer different questions

A mortality count describes the number of deaths assigned to a category.

A crude mortality rate relates deaths to population size.

An age-standardized rate adjusts for differences in age structure to support particular comparisons.

None is universally "the mortality measure."

For example, a population can have a higher crude death rate largely because it is older, while age-standardized rates show a different pattern.

Interpretation therefore requires both cause definition and rate definition.

#Coverage matters

The WHO Mortality Database includes country-reported cause-of-death data and explicitly warns users to consider incomplete registration and variation in certification systems.^[2]

A perfectly coded subset of deaths is not equivalent to complete mortality surveillance.

When registration coverage is incomplete, the observed cause distribution may not represent all deaths.

#Comparability is maintained, not assumed

Mortality statistics can change because:

• disease patterns change;
• certification practice changes;
• ICD revisions or updates occur;
• automated coding systems are introduced;
• training improves;
• registration coverage changes.

Eurostat metadata illustrate how ICD revisions and coding-system changes can create time-series comparability issues that must be documented.^[3]

A trend break near a coding transition should not automatically be interpreted as a biological event.

#A synthetic time-series problem

Suppose mortality assigned to a chronic respiratory category falls by 15% in one year.

Possible explanations include:

1. true reduction in mortality;
2. a change in certification practice;
3. a coding-rule update;
4. better specification of an underlying condition previously placed in the respiratory category;
5. changes in population structure;
6. incomplete registration in the recent year.

The observed statistical change is real. Its interpretation remains open.

A responsible analysis asks whether metadata support continuity before assigning the change to public-health success or failure.

#Underlying cause and multiple causes can tell different stories

Imagine deaths among older adults in which diabetes appears frequently as a contributing condition but less often as the underlying cause.

Underlying-cause statistics answer:

In how many deaths was diabetes selected as the initiating condition for primary tabulation?

Multiple-cause analysis can ask:

In how many certificates was diabetes present anywhere in the causal sequence or as a contributing condition?

Both are legitimate.

They should not be mixed without explanation.

#Mortality intelligence is a system

WHO's current cause-of-death framework presents mortality as an end-to-end infrastructure, not merely a coding activity: capture, code, select, analyse and improve.^[1]

That framing is methodologically useful.

Weakness at any stage can propagate:

certification
→ coding
→ underlying-cause selection
→ aggregation
→ rate construction
→ interpretation

The final table inherits the limitations of upstream stages.

#Practical questions for a mortality analysis

Before interpreting a trend, ask:

• How complete is death registration?
• Are deaths medically certified?
• Which ICD revision and update applies?
• Was coding manual or automated?
• Did coding practice change?
• Is the analysis using underlying or multiple causes?
• What population denominator is used?
• Is age standardization appropriate?
• Are recent years complete?
• Are there known comparability breaks?

These questions are not bureaucratic detail. They define what the mortality statistic means.

#Death counts can be revised

Mortality data are not always final when first released.

A late registration, amended certificate, recoding decision or completed investigation can change the assigned cause after an earlier extract was produced.

Analysts should therefore treat extraction date and release vintage as part of the dataset identity.

Two files labelled "2025 mortality" can legitimately differ if one was extracted six months later.

For reproducible trend analysis, preserve:

• extraction date;
• revision status;
• applicable coding rules;
• number of late or revised records where available.

#Small numbers create interpretive instability

Cause-specific mortality becomes especially unstable for rare causes and small populations.

A difference of two or three deaths can create a large relative change in a small area.

This does not mean small-area mortality should never be examined. It means absolute counts, uncertainty and multi-year aggregation may be more informative than percentage change alone.

A dramatic relative increase can coexist with very limited evidence of a sustained population change.

#Certification and coding can create differential misclassification

Misclassification is not necessarily random.

Diagnostic access, autopsy practice, age, care setting and clinician familiarity can influence what is written on certificates.

If those factors also differ across populations being compared, cause-specific mortality differences can partly reflect information processes. Certification errors are not interchangeable: experimental recoding work has shown that some certificate errors can materially change the underlying cause selected by mortality coding rules.^[4]

This is another reason to interpret mortality as a surveillance system rather than a passive mirror of disease.

#Limitations and uncertainty

This article explains general mortality-statistics concepts and does not provide jurisdiction-specific coding instruction. WHO ICD mortality rules and national implementation guidance should be used for actual certification and coding. The Eurostat metadata cited here illustrate comparability issues in European statistics but individual countries have different transition histories. The synthetic examples do not represent real deaths or coding decisions.^[3]