Re-classification and Weighting of Multiple Causes of Death: US Death Certificates 2003–2023

Abstract

Death certificates record causes of death as reported by certifiers (Entity Axis) and as standardized by mortality coding rules (Record Axis). Conventional mortality statistics then reduce these to a single underlying cause, ignoring other contributing conditions; weighting schemes can instead distribute the mortality burden across all listed causes. We evaluated the impact of re-classification and weighting across all 56,986,831 US death certificates from 2003–2023, mapping ICD-10 codes to 14 broad disease categories and testing three weighting schemes: W1 (50% weight to the underlying cause, 50% shared equally among contributing causes), W2 (equal weighting across all causes), and W2A (equal weighting at the ICD-10 code level). Entity and Record Axes agreed on underlying cause in 48,313,403 deaths (84.8%) by broad disease category and in 39,260,709 deaths (68.9%) by ICD-10 code (Supplementary Tables S5A, S5B); concordance reached 70.4% when using 3-character rather than 4-character ICD-10 categories. Re-classification markedly increased COVID-19 deaths (+92%) and Transport deaths (+43%) while markedly decreasing Other External Causes (−54%). Weighting substantially altered death burden attribution: COVID-19 (−44–63%) and Falls (−46–66%) decreased, Other External causes more than tripled (+204–254%), and deviations from unweighted counts were more pronounced with W2 and W2A than with W1. Weighting also brought disease-category counts closer to Entity Axis values and restored Respiratory seasonality patterns suppressed during the pandemic years. Systematic differences between reported and re-classified causes of death, and the choice of weighting scheme, profoundly alter disease burden estimates for several causes, with major implications for resource allocation and public health priorities.

INTRODUCTION

Death certificates are widely known to have high rates of inaccuracies and inconsistencies (1–5), but they serve as the foundational data source for mortality statistics and for public health policies that depend on these statistics. Problems often arise at the level of filling out the certificate forms by physicians and other eligible health care practitioners who certify the causes of death. The certifiers may lack experience, make inadvertent errors, or have less than optimal knowledge of the complete death situation, circumstances, and past medical history of the deceased. Autopsies are uncommonly performed. To enhance standardization, for several decades now systems have been in place that process the crude reported death certificates and re-classify the causes of death with specific mortality coding rules (6,7). The certifier-entered information is termed Entity Axis, and the re-classified standardized information is termed Record Axis. The data of the Record Axis are the ones endorsed officially to represent the cleaned information on underlying causes of death and on other contributing causes. The mortality coding rules for the cleaning and re-classification process from the Entity to the Record Axis are laid out in detail (6,7). However, they are sometimes revised, when a new version of the International Classification of Diseases (ICD) is released and when new ICD codes are added or existing ones are modified. The introduction of COVID-19 coding in the ICD system during the pandemic was such an example of a prominent modification (8).

Even with the Record Axis standardization, major inaccuracies may persist if the original input at the Entity Axis is flawed in ways that cannot be amended by the re-classification rules. Studies of in-depth medical record audits demonstrate large discrepancies between death certificates and more appropriate determination of causes of death based on the in-depth assessments; this has been particularly prominent for deaths where multi-morbidity is involved such as COVID-19 (9–11). A pervasive issue relates to the fundamental traditional rule allowing only one underlying cause. In reality, many deaths have substantial contributions from multiple causes (12–22). The preference for listing a specific cause as underlying rather than contributing may change over time, depending on physician awareness and on how lethal that cause is perceived to be relative to competing causes (16). Evolving perceptions may create artefacts on the increase or decrease of specific diseases as underlying causes of death. Misclassifications of underlying and contributing causes become a greater challenge with increasing multimorbidity (17). For people with co-existing morbidities that have eroded their health, it may be best to give weight not just to one cause but many. Absent in-depth medical record audits, the listing of additional causes in the death certificates may provide some indication of the broader background for each death. Several authors have proposed weighting schemes where multiple contributing causes get a share of the death attribution (19–22).

Here we aimed to analyze all death certificates in the USA between 2003 and 2023 to empirically assess how common are changes between Entity and Record Axes and whether specific major patterns exist in these re-classifications. Moreover, we aimed to assess whether different weighting schemes for multiple causes of death may substantially affect the estimated burden of death from specific causes. We examined evolving patterns over time during this extended period, and separately in elderly and non-elderly populations paying attention to the pandemic period 2020–2023 and to COVID-19 deaths. COVID-19 represented a major change in cause of death classifications with extensive debate about the over- or under-counting of COVID-19 deaths (23).

METHODS

Death certificate data

Raw US death certificate data covering the 21-year period from 2003 to 2023 were obtained from the publicly accessible Centers for Disease Control and Prevention (CDC) National Center for Health Statistics (NCHS) FTP server (24). The dataset consists of annual mortality files (mort2003us.zip through mort2023us.zip). These archives were decompressed to yield fixed-width text files (VS23MORT.DUSMCPUB) containing death certificate data in specified columns. For each death, we extracted the complete information on ICD codes from both the Entity Axis and the Record Axis. We then classified the ICD codes into 14 broad disease categories, 7 natural ones (Circulatory, Cancer, Respiratory, Endocrine/Nutritional/Metabolic, Digestive, COVID-19, and Other Natural) and 7 external ones (Drug Poisoning, Transport, Suicide, Alcohol-induced, Falls, Homicide, Other External), following the classification scheme of Wrigley-Field et al (25) that simplifies a previously proposed grouping of the National Academies of Science, Engineering and Medicine (26). Our grouping is the same as in (25) except that we also have Falls as a separate broad category rather than merged with Other External, since reported Falls deaths have rapidly increased in the 21^st century and thus would be worth closer examination. For COVID-19 we used three ICD-10 codes (U07.1, U07.2, and U09.9) as done before (25). The few (n=518) occurrences of uncertain values (all were string “1”) under ICD-10 codes in the Record Axis were ignored.

When several ICD codes were mapped to the same broad disease category, they were usually counted only once in all analyses that used the broad disease categories (W2A is the exception). We used the broad-level classification to focus on major shifts across major groups of causes.

Entity Axis versus Record Axis comparisons

We noted for each broad disease category how often it is listed as what might be assumed to be the underlying cause of death in the Entity Axis (i.e. appearing in the line in Part I of the death certificate filled out by certifying physician or other eligible health care practitioner) versus as underlying cause in the Record Axis. We also examined in each Axis for each broad disease category how often it was listed but not as underlying cause; how often it was listed at all (either as underlying cause or other); and how often it was listed as the only broad disease category in the death certificate. We then evaluated the underlying causes of death according to the two Axes and noted how often they agreed at the broad disease category level and at the specific ICD level. Focusing on the broad disease categories, we also recorded how often each cause was promoted to underlying cause of death from the re-classification of Entity Axis into Record Axis and how often it was demoted from underlying cause of death during that re-classification, the ratio of promotions to demotions, and the specific pairs of broad disease categories on which major shifts occurred. Specifically for COVID-19 we also examined the specific ICD categories involved in such promotion and demotion switches.

When there was more than one ICD code listed in the last line of Part I in the Entity Axis, we used the one listed last for the main analyses. As a sensitivity analysis, we examined the one listed first in the last line of Part I.

Of note, Entity Axis is not normally used for underlying cause reporting, therefore the comparison between Entity and Record Axis is done primarily to examine the effect of the re-classification process that generates the official cause data (Record Axis). Moreover, examination of the Entity Axis data allows to see whether weighted analyses considering multiple causes of death bring the death counts closer to the original Entity Axis entries.

Weighting of multiple causes of death

We generated data for death counts comparing different weighting schemes using information from the Record Axis: the baseline W0 (counting only the listed underlying cause of death in the Record Axis); a hierarchical weighting scheme W1, where the underlying cause of death is given 50% of the weight, and the remaining contributing causes of death share equally among them the other 50% (e.g. if there are 2 contributing causes, then each of them gets 25% of the weight and the underlying cause gets 50% of the weight); an equal weighting scheme W2 where all causes share equally the weight (e.g. with two contributing causes, each of them gets 33.33% of the weight and the underlying cause also gets 33.33%) and a variant of W2 called W2A where all the causes (including repeated occurrences in the same broad category that are normally excluded) share the weight equally. Similar weighting schemes have been previously proposed in analyses of multiple cause of death data (19–22) and have used different levels of grouping of causes. Hence, we explored both the high-level broad categories grouping (W1, and W2) and the more granular ICD-level analyses (W2A). We estimated the absolute and relative magnitude of change between W1, W2, and W2A versus the baseline W0 (which represents the traditionally presented mortality statistics) for the number of deaths attributed to each of the broad death categories.

We mapped the baseline versus the W1, W2 and W2A estimates for each of the broad death categories per month in the 2003–2023 period. We also performed similar subgroup analyses for age groups of elderly (greater than 65 years old, GE65) and non-elderly (less than 65 years old, LT65). The plotted estimates were population-adjusted using annual mid-year US population estimates from CDC Wonder available through the Human Mortality Database (https://www.mortality.org/File/GetDocument/hmd.v6/USA/STATS/Population5.txt), with 2023 as the reference year. Scaling was performed separately for the overall population and for each age subgroup using overall and subgroup-specific population denominators, such that observed deaths were multiplied by the ratio of 2023 population to that year’s population within each stratum.

For the weighted analyses, we also performed sensitivity analyses using the underlying cause information from the Record Axis and as additional causes the Part II causes listed from the Entity Axis, so as to avoid the potential immediate causes of deaths (for example, sepsis) that might have otherwise contaminated the listing of other causes in the Record Axis.

Software

All analyses were performed using Python scripts developed iteratively with LLM assistance (Claude, Anthropic). The LLM served as a coding assistant – the investigators specified the analytical requirements, reviewed all generated code for correctness, and validated results through multiple cross-checks including verifying that weighted totals sum to total deaths, comparing results across independent implementations, spot-checking calculations manually on subsets of data, and confirming concordance with published CDC statistics where available. The final Python scripts are deterministic and reproducible; they are provided along with the key data in https://github.com/scientific-computing-user/Reclassification-and-Weighting-of-Multiple-Causes-of-Death-US-Death-Certificates-2003-2023/.

RESULTS

Entity versus Record Axis

56,986,831 U.S. death certificates were issued in 2003–2023. Comparison of Entity and Record Axis data (Table 1) shows substantial changes for some broad categories of death such as COVID-19 and Other External causes. In relative terms, re-classification from Entity Axis to Record Axis almost doubles the number of COVID-19 deaths (+92%) and creates also major increases in the number of Falls deaths (+69%), Transport deaths (+43%), and modest increases in the number of Suicide (+25%), Homicide (+30%), Endocrine (+16%) and Cancer deaths (+12%). Re-classification leaves Circulatory, Digestive, Alcohol-related and Drug Poisoning deaths largely unchanged. Conversely, it causes major decreases in Other External causes (−54%), and modest decreases in Other Natural causes (−14%), and Respiratory deaths (−11%).

Table 1.

Causes of death classified in broad disease categories⁺⁺ in Record Axis and Entity Axis, USA 2003–2023 (when there was more than one ICD code listed in the last line of Part I in the Entity Axis, we used the one listed last).

	Record Axis				Entity Axis
Disease	Underlying (% change)	Contributing	Only+	Any#	Underlying	Contributing	Only	Any
Circulatory	17,796,602 (0%)	13,510,161	7,678,693	31,306,763	17,793,268	13,582,667	7,521,247	31,375,935
Cancer	12,576,466 (12%)	1,435,404	5,482,941	14,011,870	11,214,012	2,817,512	5,476,807	14,031,524
Other Natural	10,708,481 (−14%)	18,124,458	4,662,437	28,832,939	12,386,337	17,171,701	4,603,744	29,558,038
Respiratory	5,303,862 (−11%)	9,896,044	1,295,750	15,199,906	5,978,426	9,277,447	1,295,638	15,255,873
Endocrine *	2,560,585 (16%)	6,780,460	156,081	9,341,045	2,209,113	7,136,922	124,334	9,346,035
Digestive	1,731,012 (3%)	2,394,730	346,257	4,125,742	1,675,880	2,692,069	346,111	4,367,949
COVID-19	1,004,208 (92%)	165,605	58,208	1,169,813	522,512	647,301	58,211	1,169,813
Drug Poisoning	1,229,766 (9%)	241,136	701,531	1,470,902	1,130,962	342,169	701,529	1,473,131
Transport	915,055 (43%)	19,491	853	934,546	637,778	296,760	852	934,538
Suicide	762,109 (25%)	3,190	662	765,299	609,437	155,860	662	765,297
Alcohol-related	676,440 (−2%)	679,715	155,262	1,356,155	686,868	669,287	102,929	1,356,155
Other External	671,615 (−54%)	4,197,101	269,509	4,868,716	1,451,348	3,420,721	269,514	4,872,069
Falls	654,100 (69%)	159,415	161	813,515	386,291	427,221	161	813,512
Homicide	396,530 (30%)	4,118	275	400,648	304,599	96,049	275	400,648
Uncertain	0	518	0	518	0	0	0	0
Total	56,986,831	57,611,546	20,808,620	114,598,377	56,986,831	58,733,686	20,502,014	115,720,517

⁺⁺The mapping of the ICD-10 codes onto broad disease categories is given in Table S9.

⁺Only means that there only one broad category cause of death for the particular axis. There are 309,574 death certificates with more than a single unique broad cause of disease for the record or the entity axis when the other axis has only a single cause. For the vast majority of certificates (308,090), it is the Entity Axis that has multiple causes.

^#Any means that we code all broad disease categories.

^*endocrine, nutritional and metabolic

The availability of additional contributing causes of death more than doubles the total number of available causes (n=57,611,546 more) even with the clustering in broad categories. For both Axes, slightly over a third of the deaths (n=20,502,014 or 20,808,620) have only one of the broad disease categories represented in the death certificate. The relative ranking changes substantially for some categories such as COVID-19 and Cancer when any cause is considered instead of only the underlying cause. There is a large diversity across the broad categories on the ratio of their appearance as contributing causes rather than as underlying causes (ranging from near zero to 6.25, see Supplementary Table S1).

In 7,749,865 (13.6%) of the 56,986,831 deaths, the last line of Part I in the death certificate had multiple ICD codes listed. When assuming as underlying cause of death the first ICD listed in the last line of Part I instead of the last ICD listed in that line, the respective results and changes versus the main analysis are shown in Supplementary Tables S2 and S3.

Table 2 shows the specific transitions across the broad disease categories between the Entity Axis and the Record Axis for the underlying cause of death. Entity and Record Axes agreed on underlying cause in 48,313,403 (84.8%) of the total deaths based on broad disease categories and in 39,260,709 (68.9%) based on ICD-10 codes (Supplementary Table S5A) and in 30,806,674 (70.4%) when we used 3-character ICD-10 categories instead of 4-characters (e.g. J44 instead of J449). Concordance was 85.7% and 70.1%, respectively, in pre-pandemic years (2003–2019) and 81.6% and 65.0%, respectively, during pandemic years (2020–2023). Discordance rates slowly increased between 2003 and 2019 (by about 15%) and then had a marked increase in 2020 and 2021 and partial recovery in 2022–2023.

Table 2.

Transition matrix from Entity Axis (columns) to Record Axis (rows). When there was more than one ICD code listed in the last line of Part I in the Entity Axis, we used the one listed last.

ENTITY	Circulatory	Cancer	Other Natural	Respiratory	Endocrine*	Digestive	COVID-19	Drug Poisoning	Transport	Suicide	Alcohol	Other External	Falls	Homicide	Total Record Axis
RECORD
Circulatory	15,936,312	61,223	1,092,514	408,494	143,301	49,031	7,330	16,246	1,313	104	12,981	63,365	4,047	341	17,796,602
Cancer	344,036	11,020,772	598,117	360,314	109,320	104,079	3,591	1,323	135	26	14,196	19,895	640	22	12,576,466
Other Natural	628,963	53,307	9,350,816	406,374	130,961	60,841	4,473	7,028	714	68	16,553	45,311	2,840	232	10,708,481
Respiratory	297,054	29,430	506,545	4,346,838	57,638	17,118	2,539	2,832	239	14	4,870	37,738	966	41	5,303,862
Endocrine *	379,340	13,610	364,998	60,867	1,711,842	11,766	1,924	2,013	111	8	4,529	9,227	329	21	2,560,585
Digestive	68,911	14,182	187,179	54,731	15,488	1,372,198	338	1,017	159	8	2,568	13,926	257	50	1,731,012
COVID-19	54,475	10,168	119,181	288,936	21,328	5,017	501,747	204	48	5	567	2,251	271	10	1,004,208
Drug Poisoning	16,732	1,578	13,339	7,380	2,855	1,906	125	1,088,054	96	482	87,860	9,181	83	95	1,229,766
Transport	4,014	169	3,570	1,686	455	242	26	1,086	634,750	1	2,378	266,638	31	9	915,055
Suicide	531	443	9,809	359	121	78	14	1,943	5	608,669	966	139,167	3	1	762,109
Alcohol-related	17,325	2,231	48,391	11,891	8,062	48,038	87	3,809	50	31	530,942	5,339	233	11	676,440
Other External	29,504	4,740	63,407	22,983	5,584	4,561	89	4,721	107	14	4,980	530,516	391	18	671,615
Falls	15,552	2,144	27,326	7,253	1,974	929	229	426	39	2	3,402	218,625	376,199	0	654,100
Homicide	519	15	1,145	320	184	76	0	260	12	5	76	90,169	1	303,748	396,530
Total Entity Axis	17,793,268	11,214,012	12,386,337	5,978,426	2,209,113	1,675,880	522,512	1,130,962	637,778	609,437	686,868	1,451,348	386,291	304,599	56,986,831

^*endocrine, nutritional and metabolic

In sensitivity analysis using the first ICD listed in the last line of Part I instead of the last listed one, concordance rates were slightly lower (82.3% for broad category and 67.7% for ICD level) (see Supplementary Table S5A) and 69.1% with ICD-10 at 3-character categories.

Focusing on the broad death categories in Table 2 that have the most striking relative increases (COVID-19, Transport) or relative decreases (Other External, Alcohol) shows that a few specific category transitions are mostly responsible for the changes. For COVID-19, 288,936 deaths (29% of the 1,004,208 total) in the Record Axis are imported from the Respiratory category in the Entity Axis and another 119,181 deaths (12%) are imported from the Other Natural deaths in the Entity Axis. Transport deaths in the Record Axis receive 266,638 deaths (29% of the 915,055 total in the Record Axis) from the Other External causes in the Entity Axis. Other External causes get often re-classified also as Falls (218,625, 33% of Falls total deaths of 654,100 in Record Axis) and Suicide receives 139,167 deaths (18.2% of its total of 762,109 in the Record Axis) from Other External causes in the Entity Axis. Alcohol deaths in the Entity Axis often get re-classified as Drug Poisoning (87,860 deaths) in the Record Axis.

For COVID-19 deaths, the number of promotions to underlying cause of death (n=502,461) outnumber demotions (n=20,765) by 24-fold in the Entity to Record Axis re-classifications. Promotions largely re-classify the underlying cause of death away from respiratory and infection syndromes (Table S4). Most prominently, the ICD-10 J189 code of pneumonia represents 43.8% of promotions. The demotions are much rarer and shift the underlying cause of death toward very diverse chronic disease causes.

Weighted causes of death

Table 3 panel 1 shows the unweighted (W0) and W1-, W2-, and W2A-weighted counts for deaths in broad disease categories for 2003–2023 and separately for the pandemic years 2020–2023 and Table 4 does the same separately for age strata. The weighting results in major changes for the relative contribution of some disease categories to the overall death burden and usually changes are more prominent with W2 and W2A.

Table 3.

Death counts for broad disease categories for 2003–2023 and for the pandemic years 2020–2023 with different weighting schemes (W0: no weighting, only the underlying cause is counted; W1: 50% weight for the underlying cause and the other 50% shared equally among other causes; W2: weight shared equally among all causes; W2A: weight shared equally among all causes but considering ICD-level causes rather than just broad categories). The upper panel weights the counts ot ICD on the Record Axis. The lower panel weights the counts for the underlying ICD on the Record Axis with the contributing ICD on the Entity Axis.

Counts and percentage relative to W0 for the counts of ICD causes from the Record Axis
All Years (2003–2023)					Pandemic Years (2020–2023)
Disease	W0	W1	W2	W2A	W0	W1	W2	W2A
Circulatory	17,797	17,960 (+1)	18,060 (+1)	17,045 (−4)	3,715	3,867 (+4)	3,942 (+6)	3,688 (−1)
Cancer	12,576	9,392 (−25)	8,546 (−32)	8,832 (−30)	2,494	1,827 (−27)	1,615 (−35)	1,678 (−33)
Other Natural	10,708	13,538 (+26)	14,424 (+35)	14,186 (+32)	2,468	3,131 (+27)	3,378 (+37)	3,286 (+33)
Respiratory	5,304	6,122 (+15)	6,287 (+19)	6,481 (+22)	1,046	1,388 (+33)	1,497 (+43)	1,511 (+44)
Endocrine *	2,561	2,653 (+4)	2,707 (+6)	3,404 (+33)	679	733 (+8)	766 (+13)	950 (+40)
Digestive	1,731	1,636 (−5)	1,565 (−10)	1,668 (−4)	394	365 (−7)	347 (−12)	375 (−5)
COVID-19	1,004	566 (−44)	370 (−63)	441 (−56)	1,004	566 (−44)	370 (−63)	441 (−56)
Drug Poisoning	1,230	1,140 (−7)	1,087 (−12)	988 (−20)	426	396 (−7)	377 (−11)	342 (−20)
Transport	915	461 (−50)	380 (−58)	437 (−52)	192	97 (−50)	81 (−58)	91 (−52)
Suicide	762	382 (−50)	308 (−60)	363 (−52)	174	87 (−50)	69 (−60)	83 (−52)
Alcohol-related	676	547 (−19)	512 (−24)	574 (−15)	202	162 (−20)	150 (−26)	170 (−16)
Other External	672	2,039 (+204)	2,376 (+254)	2,104 (+213)	145	454 (+212)	533 (+266)	482 (+231)
Falls	654	351 (−46)	225 (−66)	269 (−59)	180	97 (−46)	59 (−67)	72 (−60)
Homicide	397	200 (−50)	140 (−65)	195 (−51)	97	49 (−50)	34 (−66)	48 (−51)
Counts and percentage change relative to W0 of the underlying Record Axis cause with the contributing causes on Part II ot the Entity Axis
	All Years (2003–2023)				Pandemic Years (2020–2023)
Disease	W0	W1	W2	W2A	W0	W1	W2	W2A
Circulatory	17,816+	16,679 (−6)	16,594 (−7)	16,330 (−8)	3,715	3,556 (−4)	3,568 (−4)	3,484 (−6)
Cancer	12,576	10,789 (−14)	10,396 (−17)	10,514 (−16)	2,494	2,085 (−16)	1,971 (−21)	2,004 (−20)
Other Natural	10,692	13,742 (+29)	14,228 (+33)	13,961 (+31)	2,468	3,191 (+29)	3,336 (+35)	3,250 (+32)
Respiratory	5,308	4,794 (−10)	4,694 (−12)	4,880 (−8)	1,046	965 (−8)	952 (−9)	1,000 (−4)
Endocrine *	2,550	3,448 (+35)	3,706 (+45)	3,865 (+52)	679	920 (+35)	1,002 (+47)	1,046 (+54)
Digestive	1,731	1,662 (−4)	1,641 (−5)	1,695 (−2)	394	378 (−4)	372 (−6)	388 (−1)
COVID-19	1,004	784 (−22)	705 (−30)	733 (−27)	1,004	784 (−22)	705 (−30)	733 (−27)
Drug Poisoning	1,230	1,165 (−5)	1,142 (−7)	1,099 (−11)	426	403 (−5)	394 (−7)	380 (−11)
Transport	915	852 (−7)	839 (−8)	845 (−8)	192	178 (−8)	174 (−9)	176 (−9)
Suicide	762	444 (−42)	428 (−44)	431 (−43)	174	95 (−46)	91 (−48)	92 (−47)
Alcohol-related	680	665 (−2)	655 (−4)	677 (−0)	202	192 (−5)	187 (−8)	195 (−4)
Other External	672	1,246 (+85)	1,306 (+94)	1,273 (+90)	145	295 (+103)	310 (+113)	305 (+110)
Falls	654	478 (−27)	420 (−36)	445 (−32)	180	125 (−31)	105 (−42)	114 (−37)
Homicide	397	240 (−40)	235 (−41)	237 (−40)	97	53 (−46)	52 (−47)	52 (−46)

^*endocrine, nutritional metabolic

⁺There are small differences in W0 counts with the different classification scheme used in this panel.

Table 4.

Death counts for broad disease categories for 2003–2023 and for the pandemic years 2020–2023 with different weighting schemes (W0: no weighting, only the underlying cause is counted; W1: 50% weight for the underlying cause and the other 50% shared equally among other causes; W2: weight shared equally among all causes; W2A: weight shared equally among all causes but considering ICD-level causes rather than just broad categories) separately for deaths under 65 years old and for deaths 65 years old or older. The causes weighted are all on the Record Axis.

Death counts for non-elderly, under 65 years old, in thousands (% of W0)
	All Years (2003–2023)				Pandemic Years (2020–2023)
Disease	W0	W1	W2	W2A	W0	W1	W2	W2A
Circulatory	3,312	3,396 (+3)	3,425 (+3)	3,288 (−1)	695	721 (+4)	734 (+6)	703 (+1)
Cancer	3,635	2,744 (−25)	2,524 (−31)	2,576 (−29)	638	469 (−27)	417 (−35)	428 (−33)
Other Natural	2,086	2,890 (+39)	3,136 (+50)	3,068 (+47)	402	603 (+50)	678 (+69)	654 (+63)
Respiratory	808	1,074 (+33)	1,147 (+42)	1,181 (+46)	168	267 (+59)	302 (+79)	302 (+79)
Endocrine *	721	715 (−1)	713 (−1)	858 (+19)	180	189 (+5)	195 (+8)	232 (+29)
Digestive	532	561 (+5)	557 (+5)	579 (+9)	116	117 (+1)	115 (−1)	122 (+5)
COVID-19	248	138 (−44)	89 (−64)	106 (−57)	248	138 (−44)	89 (−64)	106 (−57)
Drug Poisoning	1,159	1,076 (−7)	1,027 (−11)	934 (−19)	396	368 (−7)	351 (−11)	318 (−20)
Transport	745	374 (−50)	311 (−58)	358 (−52)	153	77 (−50)	65 (−57)	73 (−52)
Suicide	622	312 (−50)	254 (−59)	297 (−52)	139	69 (−50)	56 (−60)	66 (−52)
Alcohol-related	526	426 (−19)	402 (−24)	450 (−15)	152	122 (−20)	113 (−25)	128 (−16)
Other External	356	1,278 (+259)	1,472 (+314)	1,303 (+267)	74	278 (+274)	322 (+333)	290 (+290)
Falls	100	52 (−48)	35 (−65)	42 (−58)	23	12 (−48)	8 (−65)	9 (−59)
Homicide	376	189 (−50)	132 (−65)	185 (−51)	92	46 (−50)	32 (−66)	46 (−51)
Death counts for elderly, 65 years old or older, in thousands (% of W0)
	All Years (2003–2023)				Pandemic Years (2020–2023)
Disease	W0	W1	W2	W2A	W0	W1	W2	W2A
Circulatory	14,485	14,564 (+1)	14,635 (+1)	13,757 (−5)	3,021	3,147 (+4)	3,208 (+6)	2,986 (−1)
Cancer	8,942	6,648 (−26)	6,022 (−33)	6,257 (−30)	1,855	1,358 (−27)	1,197 (−35)	1,251 (−33)
Other Natural	8.623	10,648 (+23)	11,287 (+31)	11,119 (+29)	2,066	2,528 (+22)	2,701 (+31)	2,633 (+27)
Respiratory	4,496	5,048 (+12)	5,140 (+14)	5,299 (+18)	878	1,121 (+28)	1,196 (+36)	1,210 (+38)
Endocrine *	1,839	1,938 (+5)	1,994 (+8)	2,546 (+38)	499	544 (+9)	571 (+14)	718 (+44)
Digestive	1,199	1,076 (−10)	1,008 (−16)	1,089 (−9)	278	248 (−11)	232 (−17)	253 (−9)
COVID-19	756	427 (−43)	281 (−63)	334 (−56)	756	427 (−43)	281 (−63)	334 (−56)
Drug Poisoning	70	64 (−9)	60 (−15)	54 (−23)	30	28 (−8)	26 (−13)	24 (−21)
Transport	171	87 (−49)	69 (−60)	79 (−54)	39	20 (−49)	16 (−60)	18 (−54)
Suicide	140	70 (−50)	54 (−62)	66 (−53)	36	18 (−50)	13 (−62)	17 (−53)
Alcohol-related	150	121 (−20)	110 (−27)	124 (−17)	51	40 (−20)	37 (−28)	42 (−17)
Other External	316	760 (+141)	904 (+186)	801 (+153)	71	176 (+148)	210 (+196)	192 (+171)
Falls	554	299 (−46)	190 (−66)	227 (−59)	158	85 (−46)	51 (−67)	62 (−61)
Homicide +	20	11 (−48)	8 (−63)	10 (−51)	5	3 (−49)	2 (−64)	3 (−52)

^*endocrine, nutritional metabolic

⁺The percentage changes using the complete counts not those rounded in thousands.

With weighting, COVID-19 (−44–63%), Transport (−50–58%), Suicide (−50–60%), Falls (−46–66%) and Homicide (−50–65%) death counts decrease drastically and there are also major 19–32% decreases for Cancer and Alcohol-related deaths. Conversely, Other External causes more than triple (+204–254%); Other Natural, Respiratory and Endocrine, Nutrition and Metabolic deaths increase by 4–35% in the 2003–2023 data and slightly more prominently (8–44%) in the pandemic years alone. There are small decreases in Drug Poisoning and Digestive deaths, while Circulatory deaths remain largely unchanged. The rank order of causes changes substantially in the weighted analyses. For all years, all three weighting schemes promote Other External causes from rank 12 to 6. At the same time Digestive, COVID-19, Drug Poisoning, Suicide and Transport causes are demoted by between 1 and 3 ranks.

Table 3, panel 2 shows the weighted analyses using the Record Axis underlying cause with contribution causes from Part II of the Entity Axis instead of the causes from Record Axis as presented above: results were largely similar. Moreover, Figure S1 shows how the percent changes to each broad disease category for the different weighting schemes (W1, W2 and W2A) are highly correlated with correlation coefficients for the entire period of 0.809, and of 0.824 for the pandemic years) for calculations done using the causes on Part II of the Entity Axis instead of causes on the Record Axis.

These patterns follow directly from the disparity in count between underlying and contributing causes of death already seen in Table 1 and Supplementary Table S1. If there are equal counts of the two categories, sensible weighting is expected to have little effect (e.g. Circulatory). If there are fewer counts in contributing causes of death, such weighting decreases the death counts (e.g. COVID-19, Transport, Suicide, Falls) and if there are more counts in contributing causes of death, such weighting increases the death counts (e.g. Respiratory, Endocrine, Other External). In the extreme increases with weighting that are seen for Other External causes, the key reasons are respiratory tract obstruction, fractures, and injuries (Supplementary Table S6), and many of these ICD-10 codes probably reflect the nature rather than the cause of the injury.

Among deaths in ages under 65 years (Table 4), Cancer deaths are the most common cause in the unweighted counts, but with weighting the Circulatory category becomes the most common one. With the weighting schemes W2 and W2A Cancer gets to third rank after Circulatory and Other Natural causes. Conversely, with weighting the relative increases in Respiratory deaths and deaths due to Other Natural causes are more prominent in this age stratum, especially in the pandemic years. With weights W2 and W2A COVID-19 deaths are cut the most (by 64% and 57%, respectively) in this age stratum. In the age stratum of age 65 years and older (also in Table 4), for most of the broad death categories the relative change in death counts with weighting are similar as in the non-elderly stratum. However, there are some exceptions. Boosting of Other External causes is less pronounced while boosting of Respiratory, Endocrine and Other Natural causes is more pronounced than in the non-elderly.

The corrections to the Record Axis underlying death counts made by all three weighting schemes are such as to make the counts become more similar to what they were in the Entity Axis of the death certificate. Specifically, the Pearson correlation coefficients of the fractional changes caused by weighting and the fractional change between Entity and Record axes is between 0.96 and 0.97 (Supplementary Table S7).

Figure 2 shows the evolution of death numbers between 2003 and 2023 for each of the broad disease categories using the unweighted W0 death counts and the W1-weighted death counts. There is observed seasonality for several broad categories. For Respiratory, but also for Circulatory, Endocrine Nutrition and Metabolic, Falls, and Other Natural disease causes there are annual peaks in the colder seasons and troughs in the summer, while for Transport and Suicide deaths the opposite seasonality pattern is seen. In the baseline W0 counts, the seasonality is largely lost for Respiratory deaths during the pandemic years, but it is recovered with the weighting. Supplementary Figures S3 and S4 show separate data for the elderly and non-elderly age groups. Similar patterns are seen with W2 and W2A weights (Supplementary Figures S5 and S6, and age subgroups in Supplementary Figures S7, S8, S9, and S10), although the magnitude of the deviations from W0 tend to be bigger with W2 and W2A than with W1.

There are also robust linear and logarithmic increases over time for some categories and age-groups (Supplementary Table S8; Supplementary Figure S11), such as Drug Poisoning (linear trend 6.7% per year), Alcohol (4.7% per year) and Falls (4.3% per year) deaths and less prominently for Other Natural (1.9% per year) and Suicide (2.0% per year). For the elderly, there are gradual decreases over time for Circulatory (−2.5% per year) and Cancer (−1.9% per year for ages 65 or over) deaths. Linear and logarithmic annual percentage changes are very similar and weighting has little clear effect on these rates of change.

DISCUSSION

Our analysis of almost 57 million death certificates in the USA over 21 years offers insights on the impact of re-classifications during processing and of weighting for multimorbidity with multiple causes rather than a single death cause shaping each fatality. One third of deaths change ICD code for the selected underlying cause upon processing from the crude reported certificate to the version used in official national statistics and about half of them involve a shift even to a different broad disease category. Shifts have become more common over time and substantially escalated in the pandemic years, especially in 2020–2021, with the introduction of COVID-19 coding. Re-classification almost doubled the number of deaths where COVID-19 was selected as the underlying cause. Conversely, consideration of multimorbidity with weighting for multiple causes of death cut about half to two-thirds of the COVID-19 burden. It also generated substantive changes in the absolute death counts and in the relative ranking of several other causes of death. Weighting brought again the death counts for different disease categories closer to the Entity Axis. Seasonality peaks for Respiratory deaths were affected by the use or not of weighting.

The relatively high, and increasing, rates of disagreement between Record and Entity Axis are worrisome. They may reflect a high, and increasing, level of inaccuracy in filling death certificates. The escalating lack of concordance during 2020–2021 may reflect lack of acquaintance of certifiers in the new COVID-19 coding; or a particular difficulty in describing COVID-19 deaths and properly assigning the underlying cause of death in the standard framework where only one underlying cause is allowed. COVID-19 is a classic example where almost all deaths occurred in the presence of major co-morbidities, and where several comorbidities had strong associations with fatal outcomes in SARS-CoV-2 infections (27–29). It is thus possible that the WHO rules which were embedded also in the USA re-classification mortality coding rules (much like in other countries) were markedly overestimating COVID-19 death burden (23). Record Axis COVID-19 deaths have included many deaths that had multiple other causes with substantive contribution to the fatal outcome. Even beyond COVID-19, deaths with high levels of multimorbidity are probably becoming more common over time posing a challenge for certifiers to select a single underlying cause. Additional analyses using weighting for multiple causes of deaths need to be conducted across diverse countries, with priority on those that have the most complete information from death certificates. Some of our findings may be more specific to the USA (e.g. drug use and homicide counts) while others, such as changes in recording and coding of COVID-19 were experienced in many countries around the world.

Specific non-COVID-19 re-classifications are heavily represented, especially those involving the shift from Other External causes to Transport, Suicide, and Falls deaths. Many such shifts may be appropriate to capture the more precise initiating event that led to death. Some re-classifications may be debatable, however. For example, Alcohol-related death counts were overall reduced by the re-classification, but this may be underestimating the contribution of alcohol to many other conditions that eventually kill people. Moreover, it is well documented that alcohol is already grossly under-reported in death certificates: for example, traffic census data show that in motor traffic crash deaths with detected high alcohol levels, only 16% of death certificates in the USA mention alcohol, with wide variation across states (2–81%) (30).

The weighting schemes attempt to correct for the multi-factorial nature of deaths, a scenario that has become extremely common, largely the norm rather than the exception, in the 21^st century. The paradigm of listing only one underlying cause of death was quite appropriate for eras where previously healthy, young people were dying from acute infectious diseases or violence. In the current circumstances, this paradigm has very questionable validity. Most people die from the concerted impact of multiple causes that undermine their health over a variable range of time. Some causes act on a shorter time horizon than others. The exact extent to which causes other than the underlying cause should be considered is debated. We tried three weighting options inspired by previous work (19–22), extending them to consider other causes both at the level of broad categories and at ICD level.

The weighted analyses yielded largely similar results in terms of direction, although W2 and W2A tend to produce larger deviations from the baseline unweighted W0 than W1. Weighting tends to bring the corrected death counts closer to what certifiers had selected in the Entity Axis. When multimorbidity is extreme (like in COVID-19), the more prominently altered estimates may be more appropriate. A much lower than officially recorded number of COVID-19 deaths is commensurate also with the small loss of life years estimated for many of those fatalities (31). It also agrees with the findings of several in-depth audit evaluations of medical records in different countries where large percentages of previously officially proclaimed COVID-19 deaths are found to have other primary, underlying death causes (9–11). It is thus likely that the true pandemic burden of COVID-19 corresponded to only less than half or even just a third of the officially assigned COVID-19 deaths in the USA. Probably COVID-19 during 2020–2023 accounted cumulatively for ~100,000 deaths among people less than 65 years old, and ~300,000 deaths among people 65 years old and older.

Changes in the frequency of specific death causes over time may reflect genuine changes due to changing incidence (e.g. infectious disease waves) or improvements in diagnostic and therapeutic management and outcomes. However, changes may also reflect an alteration of the perceived importance of a death cause among health care practitioners and in the community in general. This may affect whether a cause is listed at all and whether it is chosen to be listed as underlying or contributing. An example has been dissected in the case of Alzheimer’s disease where an increase in its listing as underlying cause was accompanied by a decrease in its listing as contributing cause (16). In our data, perception issues may underlie the observed patterns for COVID-19. Apparently, in the very early phase of the pandemic, some COVID-19 deaths were not recorded, either because their etiology was entirely missed or because certifiers may have not considered COVID-19 as the underlying cause. At this early phase, substantial under-reporting of COVID-19 deaths may have been reflected in increased death rates from other listed causes (32). Later the inverse pattern probably predominated, with severe over-counting of COVID-19 deaths in an environment where the pandemic was a dominant pervasive theme in health, media and in society at large, excessive testing was performed (23).

Perception, testing, and recognition issues are most stark with COVID-19 but may affect also in more subtle and chronic ways several major causes of deaths from chronic disease. For example, other investigators have noted that perception issues may propel ischemic heart disease, rather than diabetes or hypertension or heart failure to be listed as the underlying cause of death (22). Weighting of multiple death causes may correct some of this injustice of all-or-none attribution. However, it has also been argued that weighting may be inappropriate for some causes that might be truly the main and only determinant of fatality, as for example for many cancers (22). Moreover, some additional listed external causes, such as injuries and fractures, and respiratory tract obstruction may reflect mostly the nature of an injury and death process rather than causes that should get a share of the weight. Furthermore, one could further question whether the chronic trends (with or without weighting) reflect real changes or perception changes. This dilemma could affect even the most common causes of disease, i.e. circulatory and cancer deaths. Certifiers may be certifying causes according to the increase or decrease of the reputational presence of each disease in the environment where they practice. Perception biases could make death certificate data become self-fulfilling prophecies about the burden of specific diseases and the evolution of this burden over time.

Our work has several limitations. First, it should be acknowledged that the selected grouping of death causes can make a difference. Here we used primarily high-level, broad categories of death, to avoid the frequent dilemmas between different ICD-10 codes within the same system, as in ischemic heart disease and hypertension (19–22). This approach captures only the most prominent shifts in causal categorization. At the level of single ICD-10 codes, shifts both with re-classification and with weighting were documented to be even more extreme.

Second, while rational arguments can be made regarding whether re-classified or specific weighted options have more (or less) validity than crude reported numbers of deaths, the real gold standard validation of different estimates requires autopsy studies and in-depth audit of medical records. Error rates regarding causes of death in death certificates are so high that healthy skepticism is warranted. Illustratively, in Vermont, a medical record audit study found that the underlying cause of death reported in death certificates was wrong in 60% and errors in ICD codes appeared in 92% of certificates (33). There is substantial potential for errors even with detailed medical records. For example, a study of medical examiners (arguably, the best-trained group to certify deaths) in Australia found that when they attributed deaths based on detailed perusal of medical records, their selected cause of death was still wrong 28% of the time when compared to autopsy (34). Third, we provide separate results for two large age strata and these already show that the impact of different weighting may vary for some death causes in these age groups. However, we cannot exclude that the patterns generated from weighting may deviate even further from the baseline in more granular substrata, e.g. defined by residence status (community versus nursing homes) or socioeconomic indicators. Documentation of multiple as opposed to single causes of death depends on a wide range of socioeconomic and other factors (35). These associations may reflect both genuine differences in multi-morbidity, as well as differences in the likelihood that multiple causes would be documented even if they exist. For example, multiple causes may be more likely to be documented when patients get better health care with continuity and when they are hospitalized rather than when they have no care and die at home (35).

Acknowledging these sources of complexity, we argue that re-classification patterns and multiple causes of death should be routinely considered in analyses of death certificate data. Different re-classification patterns may point to structural defects in death certification that may require targeted training and education of certifiers; or that specific mortality coding rules may be promoting misleading changes. Statistics based on single underlying causes of death may be grossly inappropriate, more so for some causes like COVID-19 than for others. The reliance on single underlying causes of death has long served to encourage condition-specific focus in service provision, medical training, research and lobbying. Public health and health statistics worldwide should move away from unilateral dependence on this outdated concept and embrace more widely approaches that consider multiple causes of death. Different patterns emerging from weighting may point out that the burden from specific causes and categories of causes may be much larger or smaller than traditional death statistics suggest. This can have profound implications for resource allocation for both research and healthcare. Incorporation of weighted estimates may help shape more appropriate, evidence-based public health priorities. Additional in-depth validation assessments of death causes in diverse populations, including higher rates of autopsies, are also needed to better understand and more properly calibrate death certificate data.

Figure 1.

Population-normalized monthly deaths across 14 broad categories during 2003–2023 for all ages comparing unweighted W0 with W1 weighting. In this figure as well as in supplementary figures S3 to S10, unweighted is a colored line and weighted is a thick black line. Disease categories are grouped by fixed Y-axis scales to allow visual comparison across panels. The top band (COVID-19, Circulatory, Cancer, Other natural, Respiratory) uses 0–130,000 deaths/month; the middle band (Endocrine, Digestive, Drug poisoning, Other external) uses 0–25,000; the bottom band (Transport, Suicide, Alcohol-related, Falls, Homicide) uses 0–6,000. All values are population-scaled to 2023 levels.

SIGNIFICANCE STATEMENT.

Standardized re-classification processes using mortality coding rules recast the selected causes of death in many death certificates. Moreover, vital statistics typically isolate a single underlying cause, while for many deaths multiple causes jointly lead to demise. Analysis of ~57 million deaths in the USA (2003–2023) shows that a large proportion of deaths are re-classified by rules to different causes than those filled by original certifiers. Analyses that give weight not only to recorded underlying cause but also to other listed causes lead to markedly different estimates of deaths from several diseases. For example, the footprint of COVID-19 fatalities during the pandemic years decreases by 44–63%. Re-classification and weighting schemes may have profound impact on disease burden estimates and policy decisions.

Data and code availability:

https://github.com/scientific-computing-user/Reclassification-and-Weighting-of-Multiple-Causes-of-Death-US-Death-Certificates-2003-2023/