Changes in reporting for unintentional injury deaths, United States of America

Abstract Objective To quantify how changes in reporting of specific causes of death and of selecting underlying cause from among multiple causes of death contribute to trends in mortality from unintentional injury in Americans aged 65 years or older. Methods We extracted age-standardized unintentional injury mortality data in the United States Centers for Disease Control and Prevention online databases from 1999 to 2016. We used an attribution method to calculate two indicators: the proportion of mortality with specific codes out of all mortality; and the proportion of mortality with underlying cause of death selected from multiple causes of death. We conducted a linear regression to examine the changes over time in these proportions and in reported and age-adjusted mortality. Findings From 1999 through 2016, the proportion of cause-specific unintentional injury mortality in this age group increased from 74% in 1999 (136.9 out of 185.0 per 100 000 population) to 85% in 2016 (143.0 out of 169.1 per 100 000 population) based on multiple causes of death codes. The proportions of mortality with underlying cause of death selected out of multiple causes of death rose in all specific causes of unintentional injury except motor vehicle crash. Age-standardized mortality attributed to reporting changes increased steadily between 1999 and 2016. The increases for overall unintentional injury, fall, motor vehicle crash, suffocation, poisoning and fire or hot object were 24.2, 13.5, 2.1, 2.3, 1.6 and 0.4 deaths per 100 000 persons, respectively. Conclusion Changes in data reporting affect trends in overall and specific unintentional injury mortality over time for older Americans.


Introduction
Proper interpretation of reported injury mortality rates and their changes over time are important for assessing the effects of sociodemographic factors on injury risk, developing injury control and prevention efforts, and prioritizing policy interventions. The underlying cause of death is defined as the disease or injury that initiated the train of events leading directly to death, or the circumstances which produced a fatal injury. 1 Underlying cause-of-death statistics, from which injury mortality rates are derived, are therefore often used to inform researchers, policy-makers and the public. Most users assume the statistics are valid.
The validity of underlying cause-of-death statistics, however, depends highly on the quality of multiple causes of death listings on death certificates. Two factors important for quality reporting are the correctness and specificity of the data on multiple causes of death and the correct selection of the underlying cause of death from the multiple causes of death. 1,2 The correctness and specificity of multiple causes of death is primarily determined by clinical diagnosis and the skill of staff members who are responsible for completing death certificates. This includes both their skill in proper recording of each cause of death and in proper ranking of the sequence of multiple causes of death. 1,2 Selection of the underlying cause of death from multiple causes of death is typically made according to the sequence of multiple causes of death on the death certificate, with the first cause in the sequence usually chosen. 2 Over the last two decades, several efforts have been made to improve mortality reporting in the United States of America.
In 1996, an international collaborative effort on automating mortality statistics was implemented in Washington, DC. The aim was to develop automated coding systems that assist in performing causal death coding, editing, selection of underlying cause of death and classification of multiple causes of death based on standardized decision tables. 3 Since that time, automated computer systems such as the Automated Classification of Medical Entities system have been established, refined and expanded. 4 The systems were demonstrated to be effective adjuncts to the traditional process of coding and classifying the underlying cause of death. 2 Starting in 2001, the National Center for Health Statistics (NCHS) has conducted regular training on automated computer systems for staff members who conduct mortality coding and selection, plus substantial curriculum revision for that training. 4 These efforts appear to have improved the accuracy of data reporting of specific causes of death and selecting underlying cause of death from multiple causes of death in the country. [3][4][5][6][7][8] One specific challenge in coding of injury-related deaths is that data-reporting practices on death certificates vary across age cohorts. 9 Elderly individuals have complex patterns of comorbidity, and their death may by delayed following a minor injury. For older people, therefore, coders less commonly attribute the underlying cause of death to the specific cause of the unintentional injury. 10 As an example, three recent publications suggest that changes in data reporting in the United States appear to have had a potentially misleading impact on the increase in deaths from unintentional falls among older Americans. [11][12][13] Reporting for unintentional injury deaths, United States of America Peishan Ning et al.
Despite the evidence about falls, the impact of recent changes in mortality data reporting has not been systematically examined in the United States. We do not know the extent to which these changes might affect trends in mortality, especially for older Americans. We therefore conducted this study to assess the change in two common data reporting practices: (i) cause specificity, and (ii) selection of underlying cause of death from multiple causes of death. We also aimed to quantify the contributions of these practices to reported unintentional injury mortality for Americans older than 64 years from 1999 to 2016.

Data source
We used data from the Wide-ranging OnLine Data for Epidemiologic Research databases 2 of the United States Centers for Disease Control and Prevention (CDC) web portal. These online databases provide public access to ad hoc queries, summar y statistics, maps, charts and data extracts for many health-related data. The databases provide data on the number of deaths, crude death rates, age-standardized death rates and 95% confidence intervals (CI) for death rates based on death certificates for United States residents in the 50 States and the District of Columbia. Each entry for a death contains a single underlying cause of death; up to 20 additional multiple causes of death; and demographic data of the person, including place of residence, age, race, Hispanic or non-Hispanic ethnicity, gender, place of death, time of death and whether an autopsy was performed. Mortality data from the death certificates are either entered by the physician certifying a death, coded by the States and provided to NCHS through the Vital Statistics Cooperative Program; or they are coded by NCHS from copies of the original death certificate provided to NCHS by the State registration offices. 2,14 Causes of death are coded according to the International Statistical Classification of Diseases and Related Health Problems, 10th revision (ICD-10). 2 The age-standardized mortality rates in the databases are calculated by multiplying the age-specific death rate for each age group (classified by 10-year intervals, except for people aged under 1 year and 85 years and over) by the corresponding weight from the specified standard population (the year 2000 projected population). These are summed across all age groups and the results are multiplied by 100 000. 2 We extracted age-standardized mortality rates, and their standard errors, for deaths coded as unintentional injury for Americans aged 65 years and older from 1999 (the earliest data) to 2016. We extracted data for both underlying cause of death and multiple causes of death. Data extraction was completed from January to April 2018. Based on the injury mortality matrix for ICD-10 recommended by the CDC, 15 we classified the external causes of unintentional injury as: all unintentional injury; unspecified unintentional injury; and five major causes including fall, motor vehicle crash, suffocation, poisoning and fire or hot object (Box 1).

Attribution model
During the reporting process of an injury-related death, the determination of the underlying cause of death from multiple causes of death can be divided into two consecutive and independent steps.
The first step is the reporting of multiple causes of death. To assess diagnoses, record, classify and sequence multiple causes of death, coders usually rely on external criteria, which are based on autopsy reports, physician review panels and querying of complete medical records (i.e. contacting certifiers for clarification). 7,8 When detailed medical documents are lacking, deaths are typically recorded by certifiers as unspecified. 2 Following previous studies, [5][6][7][8] we obtained the unintentional injury mortality with specified causes recoded among multiple causes of death (MCD) codes and calculated this as the proportion of all-cause unintentional injury mortality (A): where y represents mortality, u represents unspecified-unintentionalinjury and t all-unintentional-injury. This indictor approximately measures the specificity of injury causes; an increase or decrease in the proportion suggests improvement or deterioration in cause-specific reporting, respectively. 5,7 The second step is selecting the underlying cause of death from multiple causes of death. How coders do this is determined by the sequence of conditions on the certificate, the provisions of the ICD-10 and the associated selection rules and modifications. 2,16 According to the ICD rules, the sequencing of multiple causes of death determines the choice of the underlying cause of death in most cases. 14 Therefore, the process reflects the quality of the sequencing of multiple causes, which depends on the knowledge, skill and diligence of the coders. We obtained the unintentional injury mortality recorded in underlying cause-of-death (UCD) codes and calculated this as a proportion of unintentional injury mortality recorded in multiple causes of death (MCD) codes (B): This indictor roughly reflects the quality of sequencing of multiple causes of death; a higher proportion indicates better quality coding.
We then performed two steps to estimate the contribution of changes in data reporting (of cause specificity and selection of underlying cause of Box 1. ICD-10 codes used in the study of unintentional injury mortality attributed to changes in data reporting for Americans aged 65 years and older, 1999-2016

Research
Reporting for unintentional injury deaths, United States of America Peishan Ning et al.
death from multiple causes of death) to mortality data coded as unintentional injury. First, we adopted the widely used strategy of age-standardization to calculate adjusted unintentional injury mortality by assuming that data reporting remained unchanged from 1999 to 2016. The adjusted mortality (aM) from the underlying cause of death (UCD) for a specific cause i in a certain year m (mortality UCD i,m ) can be calculated as: 1999 1999 (3) In this equation i means specific injury cause, with i = 1, …, 5, representing five major causes of unintentional injury (fall, motor vehicle crash, suffocation, poisoning and fire or hot object, respectively). m denotes the year, with m = 1999, …, 2016. A 1999 denotes the proportion of unintentional injury mortality from multiple causes of death (mortality MCD t ) with specific ICD-10 codes out of all-cause unintentional injury mortality from multiple causes of death in 1999. C i,m represents the proportion of injury mortality recorded as the cause i in injury mortality with specific ICD-10 codes in the year m. This proportion reflects the combined impact of risk factors, injury control efforts and quality of data reporting (typically, misclassification across causes). 17,18 Due to a lack of relevant data, we could not separate the effects of data reporting from that of risk factors and injury control efforts, and thus we used the proportion in year m here. B i,1999 means the proportion of underlying cause of death selected from multiple causes of death for a given specific cause i in 1999.
Second, we calculated the difference between adjusted mortality UCD i,m and reported mortality UCD i,m . The difference reflects the contribution of changes in data reporting to unintentional injury mortality for a specific cause i in year m compared with year 1999.

Statistical analysis
We conducted the analyses in three steps. First, we plotted line graphs to demonstrate changes in the proportion of unintentional injury mortality with (i) specific external causes (i.e. cause specificity) and (ii) proportion of underlying cause of death selected out of multiple causes of death over years. The 95% CI of the two proportions were estimated using normal approximation due to large sample size. Second, we plotted stacked bar charts to show unintentional injury mortality attributed to data reporting changes from 1999 to 2016. Finally, we made a linear regression to examine changes in the proportion of unintentional injury mortality with specific external causes; selection proportion of underlying cause of death from multiple causes of death; reported and adjusted unintentional injury mortality; and unintentional injury mortality by data reporting changes from 1999 to 2016. We calculated the robust standard errors to provide valid inferences even under model misspecification. 19 Because differences between robust and classical standard errors of regression coefficients were extremely small (from 0.00001 to 0.01944), we decided to use classical methods to estimate the standard errors of regression coefficients (β). We calculated percentage changes in injury mortality between 1999 and 2016 as: We performed all data analyses using Stata, version 12.1 software (Stata Corp., College Station, United States). We considered differences to be statistically significant in two-tailed tests if P-values were less than 0.05.

Ethical statement
This study used anonymous open-access data and did not involve personal information from individuals. Our use of data complied strictly with the terms of the data use restrictions of the CDC online databases. The research proposal was approved by the ethics committee of Xiangya School of Public Health, Changsha, China (no. XYGW-2016-28).
Age-standardized mortality attributed to data reporting changes (namely, improvement in cause specificity and selection of underlying cause of death based on multiple causes of death) rose for overall and specific injury deaths from 1999 to 2016 (P < 0.05). Compared with 1999, data reporting changes contributed 24.2 per 100 000 persons to reported mortality rates based on underlying cause of death in 2016 for all unintentional injuries (Fig. 3). Corresponding figures for the five major causes of injury were 13.5 per 100 000 persons for fall, 2.1 for motor vehicle crash, 2.3 for suffocation, 1.6 for poisoning and 0.4, for fire or hot object due to data reporting changes.
Trend analysis showed data reporting changes significantly distorted and even reversed trends in overall and cause-specific mortality from underlying cause of death. The percentage change in mortality between 1999 and 2016 greatly differed between reported versus adjusted unintentional injury mortality: 18% versus −10% for overall unintentional injury (Table 2 and Fig. 3). Changes for the five major causes of injury were: 113% versus 63% for fall; −35% versus −44% for motor vehicle crash (Fig. 4); −18% versus −48% for suffocation; 134% versus 64% for poisoning; and −32% versus −42% for fire or hot object (Table 2 and Fig. 5).

Primary findings
Using the latest mortality data, we assessed two changes in data reporting -cause specificity and selection of underlying cause of death from multiple causes of death -and quantified their impact on unintentional injury mortality reports for Americans aged 65 years and older. The results reveal that a gradual increase in recording specificity of cause of death and an improvement in selecting the underlying Reporting for unintentional injury deaths, United States of America Peishan Ning et al.
cause of death from 1999 to 2016 caused a growing contribution to overall and specific unintentional injury mortality over time. These reporting changes altered published trends in unintentional injury mortality for older Americans, overestimating increases in mortality from overall unintentional injuries, fall and poisoning, and underestimating decreases in mortality from motor vehicle crash, suffocation and fire or hot object to some extent.

Validity of attribution method
The attribution method we propose is based on the built knowledge on roughly measuring cause specificity and selection of underlying cause of death from multiple causes of death through two surrogate indicators (cause specificity and selection of underlying cause of death from multiple causes of death). The two indicators are especially valuable when external and complete information for quality assessment is unavailable or inaccessible.
The proportion of mortality with specific ICD-10 codes out of all mortality approximately reflects the specificity of multiple causes of death, so it measures the quality of codes and sufficiency of information on the death certificates. [5][6][7][8] The proportion of underlying cause of death selected from multiple causes of death depends on the sequencing of multiple causes of death in most cases and follows the selection rules  16 All unintentional injury mortality is the total mortality from all causes recorded in multiple causes of death codes. The cause specificity indicator is the proportion of unintentional injury mortality with specific causes recorded out of all unintentional injury mortality recorded in multiple causes of death codes. The cause selection indicator is the proportion of unintentional injury mortality recorded in underlying cause of death codes selected from unintentional injury mortality recorded in multiple causes of death codes.

Interpretation of findings
This study replicates previous findings suggesting that reporting of specific causes of death for unintentional injury among older adults has improved. This change in reporting has had an impact on reported cause-specific unintentional injury mortality in the United States. [6][7][8] Our study extends those findings in several ways, by visualizing the improvements in selection of underlying cause of death from multiple causes of death. We also describe a strategy to quantify the combined contribution of changes in cause specificity and selection of underlying cause of death from multiple causes of death to the reported mortality from multiple causes of death from 1999 to 2016. The data reporting improvements we observed are likely to be a result of continuous and enhanced efforts by federal, state and local governments in the United States over the past three decades. These efforts include improved quality of training and practice for coding and selection [23][24][25][26] and the development of automated coding and classification systems using computer procedures that implement standardized selection rules consistent with ICD-10 modifications. 3,4,14,16,27 For example, New York city created an eLearning instructional course on cause-of-death reporting in 2008 and mandated this course for users of the electronic death registration system in 2010. 3,4 Empirical research demonstrated these changes were successful in teaching medical professionals to report cause of death accurately. 23,24

Implications
Our findings have several implications. First, we documented that reporting of unintentional injury mortality among older Americans has gradually improved, yielding significant effects on reported mortality from the underlying cause of death. For this reason, data reporting changes must be considered when using mortality from underlying cause of death to examine the extent of injury mortality and the effectiveness of interventions over time, as well as to develop and evaluate national and local health initiatives such as Healthy People 2020 and 2030. 28 In practice, two strategies can be used to address data reporting changes. The first strategy is to completely and routinely assess data reporting quality (including cause specificity, selection of the underlying cause of death from multiple causes of death, and completeness and misclassification). Its impact on reported injury mortality from By data reporting change Adjusted underlying cause of death mortality Notes: We defined adjusted underlying cause of death mortality as mortality from underlying cause of death calculated assuming that data reporting (cause specificity and selection of underlying cause of death from multiple causes of death) remained unchanged from 1999 to 2016. We defined data reporting change as the difference between adjusted underlying cause of death mortality and reported underlying cause of death mortality.
Reporting for unintentional injury deaths, United States of America Peishan Ning et al.
the underlying cause of death can then be adjusted at federal, state and local levels before releasing data. An alternative strategy is to provide users with warnings that the mortality rates of specific subgroups or certain causes might be heavily affected by changes in data reporting. The first option requires considerable resources, but may be more effective; the second option may be more feasible in many lower-resource environments.
Second, our findings reinforce the vital role that data quality plays in evidence-based research and decisionmaking. Although our findings were based on American data, the results are relevant globally, as many other countries undergo regular data reporting changes. For example, Australia, France, Italy, the Netherlands and Sweden, and have successively introduced and converted to automated computer systems over the past two decades. 4 These transitions have been shown to yield data changes (e.g. in pneumonia deaths in Australia after the introduction of automated computer systems). 29 Similarly, a recent analysis of 195 countries reported that 26 countries had improved data consistency between health and nonhealth data for road traffic mortality from 1985 to 2013, and nine countries had decreased data consistency. 30 Last, although the changes in data reporting we report here yield a growing impact on overall and specific unintentional injury mortality, the actual mortality caused by unintentional injury remains high among older Americans. Continued efforts to develop and disseminate effective prevention strategies are needed in addition to efforts to improve data quality.

Study limitations
These findings should be viewed in the light of some limitations of the research. First, although we quantitatively attributed changes in specificity and selec-  CI: confidence interval. a Adjusted mortality from the underlying cause of death was calculated by assuming that data reporting (cause specificity and selection of underlying cause of death from multiple causes of death) remained unchanged from 1999 to 2016. Notes: We extracted age-standardized mortality data from of the United States Centers for Disease Control and Prevention online databases; 2 causes of death are recorded by International Statistical Classification of Diseases and Related Health Problems (10th revision) codes. 16

Research
Reporting for unintentional injury deaths, United States of America Peishan Ning et al.
tion of reporting the underlying cause of unintentional injury deaths from multiple causes of death, such attribution assumes all other factors remain unchanged over the study period. In reality, data reporting changes are likely interacting with other factors that we did not consider. Second, due to lack of detailed data, we could not quantify the contribution of other attributes of data reporting (e.g. completeness of reporting and misclassification across diseases and injuries). 31 Further studies are needed to systematically assess recent changes in data reporting and their impact on injury mortality statistics.

Conclusions
Injury mortality data reporting in the United States has changed significantly over 1999-2016 in both the reporting of the specific causes of death and in the selection of underlying cause of death from multiple causes of death. These changes created an apparent contribution to increasing unintentional injury mortality for Americans aged 65 years and older and significantly altered trends in overall and specific unintentional injury mortality. The potential impact of changes in data reporting should be considered by researchers, policy-makers and other stakeholders who interpret and use the data.

Research
Reporting for unintentional injury deaths, United States of America Peishan Ning et al.

Cambios en la manera de reportar muertes por lesiones no intencionales, los Estados Unidos de América
Objetivo Cuantificar cómo los cambios en el reporte de las causas específicas de muerte y de la selección de la causa subyacente de entre las múltiples causas de muerte contribuyen a las tendencias en la mortalidad por lesiones no intencionales en los estadounidenses de 65 años de edad o más. Métodos Se extrajeron datos de mortalidad por lesiones no intencionales estandarizados por edad de las bases de datos en línea de los Centros para el Control y la Prevención de Enfermedades de los Estados Unidos desde 1999 hasta 2016. Se utilizó un método de atribución para calcular dos indicadores: la proporción de mortalidad con códigos específicos de entre toda la mortalidad; y la proporción de mortalidad con causa subyacente de muerte seleccionada entre múltiples causas de muerte. Se realizó una regresión lineal para examinar los cambios a lo largo del tiempo en estas proporciones y en la mortalidad informada y ajustada por edad.