Reclassification and Weighting of Multiple Causes of Death: US Death Certificates 2003-2023

This study analyzes 21 years of US death certificates to demonstrate that reclassifying causes of death and applying weighting schemes to distribute mortality burden across all listed conditions significantly alter disease burden estimates, particularly for COVID-19, falls, and external causes, with profound implications for public health resource allocation.

The Gist

The Big Picture: A Massive Re-Counting of the Dead

Imagine the United States has a giant, 21-year-long ledger of every single person who passed away (about 57 million people). For decades, we've been reading this ledger the "official" way: looking at the very first cause of death a doctor wrote down and counting that as the only reason the person died.

This study, conducted by researchers from Stanford and other institutions, decided to flip the script. They asked two big questions:

1. Did the official "cleaned" version of the ledger match what the doctors actually wrote? (They found out the answer is often no).
2. What happens if we stop counting just one cause of death and start giving credit to all the reasons a person died? (They found out this changes the story completely).

Here is how they did it, broken down into three simple concepts.

---

1. The "Editor" vs. The "Author" (Re-classification)

The Analogy: Imagine a student (the doctor) writes a final exam answer sheet (the death certificate). Then, a strict editor (the government coding system) takes that sheet, reads it, and rewrites the answer to fit a specific textbook format before putting it in the official grade book.

What the study found:
The "Editor" (the Record Axis) often changes the "Author's" (the Entity Axis) answer.

• The COVID-19 Surge: When the pandemic hit, the Editor was very aggressive. If a doctor wrote "Pneumonia" or "Respiratory Failure" but the patient had COVID, the Editor changed it to "COVID-19." This made the official count of COVID deaths double compared to what doctors originally wrote.
• The "Other" Shuffle: The Editor also moved many deaths from a vague bucket called "Other External Causes" into specific buckets like "Car Accidents," "Suicide," or "Falls."
• The Takeaway: The official numbers we see in the news are often the result of a computer program re-writing the doctor's original notes. About 1 in 7 deaths gets a different "official" label than what was originally written.

2. The "Cake" vs. The "Slice" (Weighting)

The Analogy: Imagine a person dies from a complex mix of problems: they had a heart attack, but they also had severe diabetes, and they were in a car crash.

• The Old Way (Unweighted): We only count the "biggest slice" of the cake. If the heart attack is listed as the main cause, we give that disease 100% of the credit for the death. The diabetes and the car crash get zero credit.
• The New Way (Weighted): We realize the whole cake caused the death. So, we cut the cake into equal pieces. If there are 3 causes, each gets 33% of the credit. If there are 2 causes, each gets 50%.

What the study found:
When they started "sharing the cake" (weighting the causes), the story of American death changed dramatically:

• COVID-19 Shrinks: Because so many people who died of COVID also had other serious illnesses (like heart disease or diabetes), giving credit to all of them meant the "COVID slice" got smaller. The study suggests the official COVID death count might be 40% to 60% too high if we only count the virus as the sole killer.
• "Other" Causes Explode: Categories like "Other External Causes" (which includes things like choking, falls, or injuries that don't fit neatly elsewhere) tripled in size. This is because these conditions often appear as contributing factors (the second or third cause) rather than the main one.
• Falls and Transport: Deaths from falls and car accidents also dropped significantly when shared with other causes, because they were often listed as secondary issues.

3. The "Seasonal Rhythm" (Why it Matters)

The Analogy: Think of the seasons. In winter, we expect more respiratory illnesses (colds, flu). In summer, we expect more car accidents.

• The Problem: During the pandemic, the official "unweighted" numbers for respiratory deaths looked flat and weird. It was like the seasons stopped working.
• The Fix: When the researchers used the "Weighted" method (sharing the cake), the seasonal rhythm came back. The data showed the usual winter spikes in respiratory deaths, proving that the virus wasn't the only thing killing people in the winter; the underlying seasonal illnesses were still there, just hidden by the "single cause" counting method.

The Bottom Line: Why Should You Care?

This paper argues that our current way of counting deaths is like trying to understand a movie by only reading the title of the film. It misses the plot, the characters, and the subplots.

• Resource Allocation: If we think COVID killed 1 million people solely because of the virus, we might pour all our money into antiviral drugs. But if the study is right and many of those deaths were actually a mix of the virus + heart disease + age, we might need to invest more in general heart health and geriatric care.
• Public Perception: We might be over-estimating the danger of some things (like the virus acting alone) and under-estimating others (like the danger of falls or "other" external causes).
• The Future: The authors suggest we stop relying on a single "underlying cause" of death. Instead, we should embrace a system that acknowledges that most people die from a team effort of multiple diseases.

In short: The official death counts are a bit like a filtered photo. This study took the filter off, applied a new lens (weighting), and showed us a much more complex, and perhaps more accurate, picture of how and why people are dying.

Technical Summary

1. Problem Statement

Death certificates serve as the foundational data source for national mortality statistics and public health policy. However, the data suffers from two primary structural issues:

• Re-classification Discrepancies: The "Entity Axis" (the cause of death as reported by the certifying physician) often differs significantly from the "Record Axis" (the standardized cause of death after applying ICD-10 coding rules by the National Center for Health Statistics). These rules often re-classify the underlying cause, potentially introducing artifacts or inaccuracies.
• Single-Cause Limitation: Traditional mortality statistics isolate a single "underlying cause of death," ignoring the reality of multimorbidity where multiple conditions jointly contribute to mortality. This "all-or-nothing" attribution may misrepresent the true disease burden, particularly for complex deaths like those involving COVID-19 or chronic conditions.

The authors aim to quantify the magnitude of re-classification shifts over a 21-year period and evaluate how different weighting schemes for multiple causes of death alter disease burden estimates.

2. Methodology

The study analyzed 56,986,831 US death certificates from 2003 to 2023.

• Data Sources: Raw data were obtained from the CDC National Center for Health Statistics (NCHS). The analysis utilized both the Entity Axis (original certifier input) and the Record Axis (standardized output).
• Classification: ICD-10 codes were mapped into 14 broad disease categories:
  • Natural Causes: Circulatory, Cancer, Respiratory, Endocrine/Nutritional/Metabolic, Digestive, COVID-19, Other Natural.
  • External Causes: Drug Poisoning, Transport, Suicide, Alcohol-induced, Falls, Homicide, Other External.
• Re-classification Analysis: The authors compared the underlying cause of death between the Entity and Record axes to calculate concordance rates and identify specific transition patterns (promotions vs. demotions of specific causes).
• Weighting Schemes: Three weighting schemes were applied to the Record Axis data to distribute mortality burden across all listed causes (underlying + contributing):
  • W0 (Baseline): Counts only the underlying cause (standard official statistics).
  • W1 (Hierarchical): Underlying cause receives 50% weight; remaining 50% is shared equally among all contributing causes.
  • W2 (Equal): All listed causes share weight equally (e.g., 1 underlying + 2 contributing = 33.3% each).
  • W2A (Granular Equal): Similar to W2 but weights at the specific ICD-10 code level rather than broad categories, accounting for repeated codes within the same category.
• Sensitivity Analysis: The study tested variations, such as using the first ICD code listed in Part I of the Entity Axis versus the last, and using Entity Axis contributing causes instead of Record Axis contributing causes.

3. Key Contributions

• Comprehensive Scale: This is the first study to analyze the entire US death certificate dataset (nearly 57 million records) over two decades, providing a robust empirical basis for mortality re-evaluation.
• Quantification of Coding Artifacts: The study provides concrete evidence of how mortality coding rules systematically alter the reported burden of specific diseases, particularly highlighting the impact of the pandemic-era coding changes.
• Weighting Framework: It demonstrates that moving beyond single-cause attribution via weighting schemes significantly alters the ranking and magnitude of disease burdens, offering a more nuanced view of multimorbidity.
• Recovery of Seasonality: The study shows that weighting schemes can restore seasonal patterns (e.g., respiratory deaths) that are suppressed in standard unweighted data during pandemic years.

4. Key Results

A. Entity vs. Record Axis Re-classification

• Concordance: The Entity and Record axes agreed on the underlying cause in 84.8% of deaths at the broad category level and 68.9% at the specific ICD-10 code level. Concordance dropped significantly during the pandemic (2020–2023).
• Major Shifts:
  • COVID-19: Re-classification increased COVID-19 deaths by +92%. The Record Axis promoted 502,461 deaths to COVID-19 as the underlying cause (mostly from Respiratory and Other Natural categories), while only 20,765 were demoted.
  • Transport: Increased by +43%, largely due to re-classification from "Other External" causes.
  • Falls: Increased by +69%.
  • Other External: Decreased by -54%, as many cases were re-classified into Transport, Suicide, or Falls.
  • Respiratory: Decreased by -11% in the Record Axis.

B. Impact of Weighting Schemes

Weighting schemes drastically altered the attribution of death burden compared to the baseline (W0):

• COVID-19: Weighting reduced the attributed burden by 44% to 63% (depending on the scheme), suggesting that many "COVID-19" deaths had other primary underlying causes.
• Other External Causes: This category saw the most dramatic increase, tripling (+204% to +254%) under W2/W2A. This is driven by the inclusion of contributing causes like respiratory obstruction and fractures which are often listed as "Other External" but contribute to death.
• Falls, Transport, Suicide, Homicide: All saw significant reductions (-46% to -66%) because these causes often appear as the underlying cause but have fewer contributing causes listed compared to natural diseases.
• Cancer: Decreased by 25–32%, as cancer deaths often have fewer listed contributing conditions compared to natural diseases like Endocrine or Respiratory.
• Ranking Changes: Weighting shifted the relative ranking of causes. For example, "Other External" moved from rank 12 to rank 6 in the overall burden.

C. Demographic and Temporal Patterns

• Age Stratification: In the non-elderly (<65) group, Circulatory diseases became the leading cause of death under weighting (surpassing Cancer), whereas Cancer was the leader in unweighted data.
• Seasonality: Standard unweighted data (W0) showed a loss of respiratory seasonality during the pandemic. Weighted data (W1, W2) successfully restored the expected winter peaks for respiratory deaths.
• Trends: Linear trends for Drug Poisoning (+6.7%/year), Alcohol (+4.7%/year), and Falls (+4.3%/year) remained robust regardless of weighting, suggesting these are genuine increases rather than coding artifacts.

5. Significance and Implications

• Policy and Resource Allocation: The study argues that reliance on single-cause statistics may lead to misallocation of public health resources. For instance, the true burden of "Other External" causes (injuries, accidents) is likely vastly underestimated, while the burden of specific infectious diseases like COVID-19 may be overestimated when considering multimorbidity.
• Validity of Pandemic Data: The findings support the hypothesis that official COVID-19 death counts during 2020–2023 may have been inflated by coding rules that prioritized the virus over comorbidities. The authors estimate the true pandemic burden might be only 1/3 to 1/2 of the official count.
• Methodological Shift: The authors advocate for a paradigm shift in vital statistics: moving away from the "single underlying cause" model toward weighted multiple-cause approaches. This approach better reflects the reality of modern mortality, where most deaths result from the concerted failure of multiple systems.
• Data Quality: The high rate of discordance between Entity and Record axes (especially during the pandemic) highlights a need for improved training for certifiers and potentially a revision of coding rules to better handle complex, multimorbid deaths.

In conclusion, the paper demonstrates that the choice of data processing method (re-classification rules) and statistical attribution (single vs. weighted causes) profoundly impacts our understanding of the leading causes of death, with major implications for epidemiology and public health strategy.