key: cord-1046793-3wcc9r5g
authors: Fairman, K. A.; Goodlet, K. J.; Rucker, J. D.
title: Changes in Cause-of-Death Attribution During the Covid-19 Pandemic: Association with Hospital Quality Metrics and Implications for Future Research
date: 2020-07-28
journal: nan
DOI: 10.1101/2020.07.25.20162198
sha: 020b03ed30ea568da94b34216c7b6bbbd2b750ff
doc_id: 1046793
cord_uid: 3wcc9r5g

Background: Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is often comorbid with conditions subject to quality metrics (QM) used for hospital performance assessment and rate-setting. Although diagnostic coding change in response to financial incentives is well documented, no study has examined the association of QM with SARS-CoV-2 cause-of-death attribution (CODA). Calculations of excess all-cause deaths overlook the importance of accurate CODA and of distinguishing policy-related from virus-related mortality. Objective: Examine CODA, overall and for QM and non-QM diagnoses, in 3 pandemic periods: awareness (January 19-March 14), height (March 15-May 16), and late (May 17-June 20). Methods: Retrospective analysis of publicly available national weekly COD data, adjusted for population growth and reporting lags, October 2014-June 20, 2020. CODA in 5 pre-pandemic influenza seasons was compared with 2019-20. Suitability of the data to distinguish policy-related from virus-related effects was assessed. Results: Following federal guidance permitting SARS-CoV-2 CODA without laboratory testing, mortality from the QM diagnoses cancer and chronic lower respiratory disease declined steadily relative to prior-season means, reaching 4.4% less and 12.1% less, respectively, in late pandemic. Deaths for non-QM diagnoses increased, by 21.0% for Alzheimers disease and 29.0% for diabetes during pandemic height. Increases in competing CODs over historical experience, suggesting SARS-CoV-2 underreporting, more than offset declines during pandemic height. However, in the late-pandemic period, declines slightly numerically exceeded increases, suggesting SARS-CoV-2 overreporting. In pandemic-height and late-pandemic periods, respectively, only 83.5% and 69.7% of increases in all-cause deaths were explained by changes in the reported CODs, including SARS-CoV-2, preventing assessment of policy-related mortality or of factors contributing to increased all-cause deaths. Conclusions: Substitution of SARS-CoV-2 for competing CODs may have occurred, particularly for QM diagnoses and late in the pandemic. Continued monitoring of these trends, qualitative research on pandemic CODA, and the addition of place-of-death data and psychiatric CODs to the file would facilitate assessment of policy-related and virus-related effects on mortality.

Ascertainment of the number of deaths from severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is foundational to understanding the severity, scope, and spread of the infection. Despite its importance, estimation of SARS-CoV-2 deaths is challenging because advanced age, genetic polymorphisms, and obesity-related comorbidities that predispose to inflammatory states increase the likelihood of dysregulated immunological function, severe respiratory distress, and mortality from infectious respiratory illness. 1, 2 These host factors represent competing potential causes of death (COD). For example, 98.8% of Italy's SARS-CoV-2 deaths occurred in persons with >1 comorbidity, 48.6% with >3 comorbidities, and median decedent age was 80 years. 3 Similarly, of U.S. SARS-CoV-2 deaths reported as of May 28, 2020, 93% involved other CODs (mean 2.5 additional causes), and 60% occurred in persons aged >75 years. 4 This pattern of multiple contributing CODs is common in respiratory infection-related mortality. 5 In death certificate issuance during the pandemic, methods to account for this pattern varied, as no single standard for SARS-CoV-2-attributable death exists. In Italy, all deaths in patients testing positive for SARS-CoV-2 were attributed to the infection despite high prevalence rates for comorbid conditions, measured in early deaths: ischemic heart disease (30%), diabetes (36%), cancer (20%), and atrial fibrillation (25%). 6 The U.S. National Center for Health Statistics (NCHS) issued death-certification guidance on March 4, 2020, indicating that SARS-CoV-2 should be reported if "the disease caused or is assumed to have caused or contributed to death." 7 Follow-up guidance issued on April 3 indicated that it was "acceptable to report COVID-19 on a death certificate without [laboratory test] confirmation" if circumstances indicating likely infection were "compelling within a reasonable degree of certainty." 8 This nonspecific guidance should be interpreted in light of previous research findings that COD attribution (CODA) errors are common on death certificates, particularly in infectious disease and septic shock. 9, 10 In one survey of New York City (NYC) resident physicians in 2010, 49% indicated they had knowingly reported an inaccurate COD on one or more certificates, often (54%) at the behest of hospital staff, and 70% reported they had at least once been unable to report septic shock "as an accepted cause of death" and had been "forced to list an alternate cause." 9 In an audit of NYC data from 2010-2014, 67% of pneumonia death certificates contained >1 error, compared with 46% for cancer and 32% for diabetes. 10 Such CODA ambiguities are often addressed by calculating "excess deaths," defined as all-cause deaths exceeding those projected from historical experience. 11 This method, recently used to estimate that official tallies of SARS-CoV-2 deaths represented only about 66%-78% of the disease's true mortality impact, 12, 13 is potentially advantageous in estimating SARS-CoV-2 impact by accounting for deaths that may not have been explicitly coded as infection-related. 14 Examples include deaths from cardiac events to which undetected SARS-CoV-2 may have contributed 15 or out-of-hospital deaths occurring without medical care because of health-system overcrowding. 16 Despite these advantages, the method is compromised by 3 considerations when applied to SARS-CoV-2 CODA, which should be quantified to inform future policy.

First, the method should distinguish natural from societal causes to account for possible consequences of policy decisions and fears that, although prompted by anticipated effects of SARS-CoV-2, were not direct or inevitable viral sequelae. Examples include suicides from stay-at-home order-related labor market contraction 17 and social isolation, 18 increases in domestic violence, 19 overdoses due to interruptions in substance use disorder treatment, 20 and delays in emergency care for life-threatening conditions [21] [22] [23] in geographic areas where health-system overcrowding was expected but not realized. 24, 25 To promote evidence-based public health policy, population-level disease-mitigation strategies that go beyond traditional practices of isolating the sick and quarantining those exposed to disease merit empirical investigation. 26, 27 Second, the method should reflect the effects that financial incentives around hospital quality metrics (QM), which are commonly associated with provider coding practices, may have on CODA. [28] [29] [30] For example, in United Kingdom hospitals, increases in coding for palliative-care admissions produced a severity-adjusted mortality-rate decline of 50% over the 5-year period ending in 2009, while the crude death rate remained unchanged. 31 Although we are not aware of studies linking QMs to CODA, it is known that CODA errors are more likely to occur in hospitals than elsewhere, 32 with an 85% error rate reported in comparisons of death certificates with autopsy findings at one regional academic institution. 33 The potential effect of financial incentives on CODA is particularly important for SARS-CoV-2 because several competing CODs, including chronic lower respiratory disease (CLRD), acute myocardial infarction, heart failure, pneumonia, and stroke, are included in Medicare 30-day mortality measures used to calculate prospective payment rates. 34 All but one of these (CLRD) is included in Agency for Healthcare Research and Quality inpatient quality indicators. 35 Sepsis and cancer, other competing causes of death, are also the target of QM. [36] [37] [38] Although not affecting all-cause death counts, the incentive to substitute SARS-CoV-2 for another COD could affect the accuracy of the SARS-CoV-2-attributed count.

Third, the method should account for baseline life expectancies among those whose deaths were reported as caused by SARS-CoV-2. For example, at age 80 years, the 1year probability of death is 5.8% for males and 4.3% for females, higher in those with cardiovascular comorbidities. 39, 40 In that age group, the population-level risk of a SARS-CoV-2 death in New York City, a pandemic epicenter, was 1.5% in about 3 pandemic months through June 17, 2020. 41 Thus, deaths from competing CODs would be expected to decline late in the pandemic and in subsequent months. From a policy perspective, quantifying this effect is consistent with the quality-adjusted life year approach in evidence-based medicine, which considers future life expectancy in assessing the effects of disease and disease-mitigation interventions. 42 To permit assessments of SARS-CoV-2-related mortality, publicly available NCHS data include weekly aggregated totals for all-cause deaths, natural-cause deaths, and selected categories of CODs, reported as final data for 2014-2018 and provisional data for 2019-2020. 4 These data, which are updated weekly, have important limitations. First, International Classification of Diseases (ICD)-10 diagnosis codes are grouped into broad categories, rather than the individual ICD-10 codes available in full COD files (Appendix 1). Second, only 11 selected diagnostic categories are reported. Third, although 63% of deaths are reported within 10 days, reporting lags vary by state. 4 Reporting delays for injurious deaths are greater because they require investigation (e.g., forensic toxicology). 43 Pending investigation, these deaths are often assigned ICD-10 code R99, "ill-defined and unknown cause of mortality." 43 In this exploratory study, we used these files to provide preliminary evidence on the following: (1) change in CODA compared with historical experience; (2) association of CODA with QM; and (3) suitability of the files to distinguish policy-related from virusrelated effects. All analyses were adjusted for population and reporting lags and based on comparisons of 2020 with equivalent weeks in the 5 most recent years. We hypothesized that if substitution of SARS-CoV-2 for alternative CODs occurred, death counts for competing diagnoses would decline relative to historical experience during the pandemic, especially after issuance of the NCHS death-certification guidance; these declines would be greater for QM than for other conditions; and they would accelerate late in the pandemic as earlier SARS-CoV-2 deaths offset later deaths from competing causes.

Publicly available COD data for October 2014 through June 20, 2020, were downloaded from the NCHS website on July 10, 2020. The outcome was reported underlying COD (UCOD), defined as "the disease or injury which initiated the train of morbid events leading directly to death." 44 Three states were excluded because of incomplete reporting in 2020 (Connecticut and North Carolina) or in 2015 (Wisconsin) by subtracting their death counts and population counts from U.S. figures. Raw death counts were adjusted to July 2019 population using annual census data (Appendix 2). 45 Reporting lag adjustment was made by calculating percentage changes in death counts, overall and by ICD category, for each of 7 weekly updates from May 27 to July 8, 2020, as a function of the number of weeks from the reported week to the update week (e.g., deaths in week ending May 9, updated on May 20 = 1.6-weeks). For each of 16 weeks prior to the July 8 update, we used mean percentage change values as inflationary factors to calculate compounded adjustment multipliers (Appendix 3).

Time series analyses were performed using standard methods, beginning with descriptive analysis. 46, 47 We first plotted mortality curves over time, then compared pandemic periods with mean values for the same weeks in prior seasons. Absolute and percentage changes in counts of deaths, comparing 2020 with 5-year means, were calculated for 3 pandemic periods: pandemic awareness beginning January 19, based on media-report analysis of news coverage volume; 48 

Of approximately 15 million population-adjusted all-cause deaths reported in the 47 study states during the 5.8-year study period, 10.7 million (71%) were reported in one of the ICD categories included in the COD file, of which 400,000 were reported as a death NEC, including the ICD code R99. Counts formed a seasonal annual pattern, with defined peaks and troughs at similar times each year (Appendix 4). Exceptions in 2020 included a higher and later peak in all-cause deaths in April; a later peak for heart disease deaths, also in April; a marked increase in NEC deaths; and declines in deaths from cancer and CLRD. The 1-week maximum peak in all-cause deaths increased Table 1 ). Influenza/pneumonia deaths peaked later in the season in 2020 than in prior years, despite similar magnitudes. NEC deaths including R99 increased rapidly beginning in January, exceeding 5-year means by 53% during the pandemic awareness period and by >100% during and after pandemic height.

Following issuance of NCHS reporting guidance in the first week of April 2020, deaths for 2 of 3 QMs, cancer and CLRD, but not for sepsis, steadily declined relative to prior year means ( Figure 1 , Panel 2; Table 1 ). Although these declines were initially small (1%-2%) and within the fluctuation observed during pandemic awareness, latepandemic cancer deaths were 4.4% less and CLRD deaths 12.1% less than in prior years (3.2% and 11.4%, respectively, in the sensitivity analysis excluding the final study week). Measured from guidance issuance to study end, these changes represented declines of 5,937 deaths (4,566 in sensitivity analysis) for these 2 QMs, compared with prior years.

For non-QM diagnoses, these declines did not occur; instead, deaths increased beginning at pandemic start, including increases of 21.0% for Alzheimer's disease Table 1 ). Deaths from other competing CODs, for which some but not all diagnoses are QMs, increased by smaller amounts than non-QM diagnoses (CBVD by 8.7%, heart disease by 9.2%; Figure 1 , Panel 4; Table 1 ).

During pandemic height, increases in the non-QM or partial-QM diagnoses more than offset declines in the QM diagnoses (net effect of +21,714 deaths excluding respiratory illness; not shown in Table 1 ). However, of the total increase of 123,247 all-cause deaths compared with prior-year mean, only 102,956 deaths (83.5%) were explained by changes in the specific CODs included in the NCHS files, leaving 16.5% of the increase in all-cause deaths unexplained.

By the late pandemic period, 2 additional competing CODs had declined compared with historical experience: influenza/pneumonia (-11.6%) and other acute respiratory illnesses (-3.8%), whereas other competing CODs were relatively unchanged (+/-1%, heart disease, septicemia, and kidney disease) or remained elevated (Alzheimer's disease, 10.6%; diabetes 15.1%; CBVD 5.4%; not all percentages shown in Table 1 ). In this time period, the net effect of competing CODs was a small decline (-1,232 deaths), and 30.3% of the increase in all-cause deaths was unexplained by specific CODs.

Time series models using AR1 terms fit the data well during the calibration period (October 2014-September 2019; Appendix 5). However, when equations were applied to the 2019-20 influenza season, results were highly sensitive to modeling method (Appendix 6). When modeling was performed using AR1 terms to account for first-order autocorrelation, results suggested only 21,833 excess all-cause deaths during pandemic height. When the nonautocorrelation assumption was intentionally violated by excluding AR1 terms, that estimate increased to 87,943 during pandemic height and 42,964 in late pandemic, for a total of 130,907. Despite this major discrepancy, both models showed greater magnitudes in CODA for non-QM than for QM conditions, as well as less-than-expected death counts for QMs in the late pandemic period.

In this exploratory analysis of NCHS COD files, we found evidence suggesting SARS-CoV-2 may have been substituted for 2 of 3 competing CODs associated with QMs after the NCHS published guidance specifying no laboratory testing requirement for presumptive SARS-CoV-2 UCOD. We also found decreases in several competing CODs relative to prior year means, offset by marked increases in reported deaths from other conditions during pandemic height. Although validating assertions that for patients with certain comorbidities, SARS-CoV-2 may have been underreported as a UCOD, 51 our findings suggest these assertions are oversimplified because of overreporting of SARS-CoV-2 for those with other comorbidities including the second and third top causes of natural death in the United States, cancer and CLRD. 52 Despite concerns about underreporting of SARS-CoV-2 on death certificates, 53 overreporting began to slightly numerically exceed underreporting by the late pandemic period.

Our methods were generally similar to those of the previously reported studies of excess deaths, 12,13 although our method of adjustment for reporting lags was simpler and our study period was longer. The primary strength of the present study was our examination of specific CODs, rather than all-cause deaths alone, because from a policy perspective, the quality of mortality data is important. 53 Moreover, we assessed the suitability of the files to distinguish policy-related from virus-related mortality.

Information of this type should be viewed as essential in considering responses to future pandemics.

Like previous investigators of excess deaths, 12 we found large (122,300 in previous analysis, 130,907 in our analysis) increases in all-cause deaths compared with prior years in Poisson regression models adjusted for seasonality using Fourier terms, but only when our model, like the previous model, included no term to control for autocorrelation. In the model including an AR1 term, we found only a small (21, 833) increase in all-cause deaths. It is possible the AR1 model better accounted for the gradually increasing peak in all-cause mortality that occurred biannually or triannually from 2014-15 to 2019-20. However, it is also possible that the model is misleading because of excessively data-driven specification. 46 Therefore, we view the time series models as inconclusive, basing our interpretation primarily on descriptive analyses of the population-adjusted data.

Our finding of potential underreporting of QM-associated deaths compared with prior years is consistent with our a priori hypothesis, which was based on previous research documenting institutional sensitivity of coding practices to policy and financial incentives. [28] [29] [30] [31] The present study adds to the body of literature on this topic by associating these effects with death certification. This finding suggests a need for qualitative assessment of how CODAs were made during the pandemic, perhaps conducted by the Office of Inspector General, as QM data are used to set Medicare prospective payment rates. 34 A potential pragmatic contribution of that research is a specific CODA algorithm to be used in future pandemics, facilitating valid and reliable reporting across sites (e.g., hospitals versus nursing homes) and geographic regions. Important limitations of this exploratory work should be noted. Foremost, findings of this study should be reassessed using data from later weeks to verify or refute the preliminary trends described here. Additional limitations suggest improvements that might be made to the COD files pending availability of the full, final COD data. These full COD data, which are currently available only through 2018, likely will not be posted until 2022.

First, a large number of deaths had no COD or a nonspecific COD recorded in the file.

Compared with prior years, NEC deaths more than doubled in a progression that began early in 2020, as pandemic awareness was growing but before SARS-CoV-2 would Second, we imputed possible effects of financial incentives on CODA but did not measure incentives directly. To support or refute our findings, the addition of place of death to the COD files would allow for assessment of whether the QM associations we observed were greater in hospital than in nonhospital settings. Additionally, this change would facilitate assessment of changes in out-of-hospital deaths from heart disease or cerebrovascular disease, [21] [22] [23] informing the question of whether stay-at-home orders exacerbated deaths from these causes in geographic areas where hospital bed supply was not limited during the pandemic. [24] [25] Third, neither the study outcome measure nor recommended changes to the COD files would account for long-term physiological effects of SARS-CoV-2, which are the subject of an ongoing cohort study; 56 of suboptimal policy decisions that could have increased SARS-CoV-2 deaths, (e.g., mandatory posthospitalization release of infected patients to nursing facilities); 57 or of long-term effects of stay-at-home order-related declines in preventive care (e.g., pediatric immunizations). 58 Finally, it should be noted that although recent comparisons of infection-fatality and case-fatality rates for influenza and SARS-CoV-2 have adopted an implicit assumption that the respective infections' numerators are equivalent, 59 the 2 estimates are based on markedly different methods that make direct comparison problematic. National estimates of influenza hospitalizations and deaths are based on a statistical model incorporating ambulatory and inpatient data, in addition to the raw death certificate counts reported here. 60 In contrast, SARS-CoV-2 death counts rely solely on the judgment of the reporter and are, at present, not clinically validated. Pending assessment of the reliability and validity of SARS-CoV-2 death reporting, policy making death-certificate counts should be circumspect, recognizing both the possible vulnerability of the SARS-CoV-2 data to financial incentives and detection bias and the absence of important COD information from the currently available NCHS files.

Alzheimer's disease and diabetes, and overreported with others including cancer and CLRD, the second and third leading causes of natural death in the United States. 48 Season week 25 (March 15) is the approximate start of the pandemic height period. Season week 28 (April 5) corresponds to the issuance of NCHS guidance stating laboratory testing was recommended but not required to name SARS-CoV-2 as the underlying cause of death. Season week 34 (May 17, 2020) is the start of the late pandemic period. Acute respiratory (Panel 1) is the sum of SARS-CoV-2, influenza/pneumonia, and other respiratory illnesses, nearly all of which are acute (e.g., pharyngitis). Not elsewhere classified (Panel 1) includes R99, the code used for non natural-cause deaths pending forensic investigation. QM diagnoses include cancer, chronic lower respiratory disease, and septicemia. Non-QM diagnoses include Alzheimer's disease, diabetes, kidney disease, and other respiratory illnesses. QM=quality metric; SARS-CoV-2=severe acute respiratory syndrome coronavirus 2. 

Quality Measure 

Influenza pathogenesis: The role of host factors on severity of disease

COVID-19 and the role of chronic inflammation in patients with obesity

What other countries can learn from Italy during the COVID-19 pandemic

Centers for Disease Control and Prevention. Weekly updates by select demographic and geographic characteristics

The impact of age and comorbidities on the mortality of patients of different age groups admitted with community-acquired pneumonia

Case-fatality rate and characteristics of patients dying in relation to COVID-19 in Italy

Guidance for certifying COVID-19 deaths

Guidance for certifying deaths due to coronavirus disease 2019 (COVID-19)

Survey of New York City resident physicians on cause-of-death reporting

Examination of causeof-death data quality among New York City deaths due to cancer, pneumonia, or diabetes from 2010 to

Excess deaths associated with COVID-19

Estimation of excess deaths associated with the COVID-19 pandemic in the United States

Excess deaths from COVID-19 and other causes

Use of all-cause mortality to quantify the consequences of covid-19 in Nembro, Lombardy: descriptive study

New York City struggles to get accurate coronavirus fatality count as more people die at home. CNBC

Inside a Brooklyn hospital that is overwhelmed with Covid-19 patients and deaths

Preventing suicide in the context of the COVID-19 pandemic

Loneliness: a signature mental health concern in the era of COVID-19

Domestic violence amid COVID-19

When epidemics collide: coronavirus disease 2019 (COVID-19) and the opioid crisis

Collateral effect of Covid-19 on stroke evaluation in the United States

The Covid-19 pandemic and the incidence of acute myocardial infarction

Hospitals report fewer heart attacks and strokes amid COVID-19. Yale Medicine

field hospitals stand down, most without treating any COVID-19 patients. NPR

Hospital ship Comfort departs NYC, having treated fewer than 200 patients. NavyTimes

Lessons from the history of quarantine, from plague to influenza A

Quarantine and isolation: lessons learned from SARS. Institute for Bioethics, Health Policy and Law

Data quality bias: an underrecognized source of misclassification in pay-for-performance reporting? Qual Manag Health Care

Association between health information technology and case mix index

Excluding observation stays from readmission rateswhat quality measures are missing

Patient coding and the ratings game

Death certification errors and the effect on mortality statistics

Death certification: errors and interventions

Yale New Haven Health Services Corporation -Center for Outcomes Research and Evaluation. 2020 condition-specific mortality measures updates and specifications report

Agency for Healthcare Research and Quality. Inpatient quality indicators technical specifications

Specifications manual for national hospital inpatient quality measures

Methods for reducing sepsis mortality in emergency departments and inpatient units

Problems with public reporting of cancer quality outcomes data

The effects of diabetes, hypertension, asthma, heart disease, and stroke on quality-adjusted life expectancy

Population-level COVID-19 mortality risk for non-elderly individuals overall and for non-elderly individuals without underlying diseases in pandemic epicenters

QALYs: the basics

Timeliness of death certificate data for mortality surveillance and provisional estimates. Vital Statistics Rapid Release

Coding and classification of causes of death in accordance with the Tenth Revision of the International Classification of Diseases. National Center for Health Statistics

State population totals and components of change

Interrupted time series regression for the evaluation of public health interventions: a tutorial

Time series regression studies in environmental epidemiology

News coverage of coronavirus in 2020 is very different than it was for Ebola

Centers for Disease Control & Prevention. Coronavirus disease 2019. Cases in the

Using Multivariate Statistics. Essex: Pearson

Coronavirus pandemic led to surge in Alzheimer's deaths

Deaths: leading cause for 2017.National Vital Statistics System

The importance of proper death certification during the COVID-19 pandemic

America's state of mind report

Census Bureau-assessed prevalence of anxiety and depressive symptoms in 2019 and during the 2020 COVID-19 pandemic

Looking forward: understanding the long-term effects of COVID-19

Blame game? Cuomo takes heat over NY nursing home study

Effects of the COVID-19 pandemic on routine pediatric vaccine ordering and administration-United States

Estimating the infection fatality rate among symptomatic COVID-19 cases in the United States. Health Aff (Millwood)

Frequently asked questions about estimated flu burden