Assessing Inmate Cause of Death - Deaths in Custody Reporting Program and National Death Index , DOJ BJS, 2016

• May 26, 2016 • Locations: United States of America • Topics: Statistics/Trends, Wrongful Death

Download original document:
Brief thumbnail

Document text

This text is machine-read, and may contain errors. Check the original document to verify accuracy.

U.S. Department of Justice
Office of Justice Programs
Bureau of Justice Statistics

TECHNICAL REPORT

April 2016, NCJ 249568

Assessing Inmate Cause of Death: Deaths in Custody
Reporting Program and National Death Index
Zhen Zeng, Ph.D., Margaret Noonan, and E. Ann Carson, Ph.D., BJS Statisticians
Ingrid Binswanger, M.D., BJS Visiting Fellow
Patrick Blatchford, Ph.D., Colorado School of Public Health
Hope Smiley-McDonald, Ph.D., and Chris Ellis, RTI International

Introduction
The U.S. Department of Justice’s Bureau of Justice Statistics
(BJS) has collected data annually on inmates who died in
state prison and local jail and the circumstances surrounding
these deaths since the Death in Custody Reporting
Act (P. L. 106–297) was passed in 2000. There were no
standardized requirements for prison and jail administrators
to report inmate deaths prior to the passage of the act.
Under BJS’s Deaths in Custody Reporting Program
(DCRP), respondents from jails and state departments of
corrections are asked to report causes of deaths (COD),
identifiers, and characteristics of each inmate who died.
This includes personal data (i.e., name, date of birth, date
of death, and sex), correctional information (i.e., length
of stay, legal status, and criminal offense of the deceased),
and the circumstances of the death (i.e., medical treatment
before death, pre-existing conditions, and time and place of
death incident).
The DCRP had a 100% response rate among 50 state
departments of corrections and 99% among the
approximately 2,800 jail jurisdictions. Item response rates
were also generally close to 100%. However, BJS had some
concerns about the collection of COD data in the DCRP.
Due to changes in data collection agents in 2008, data
quality follow-up was cut short, resulting in 21% of missing
COD information for jail deaths that year, compared to
2% in 2007 and 2009. This data anomaly significantly
compromised BJS’s ability to report death trends.
Although the DCRP has collected inmate COD data since
2000, the extent to which respondents accurately reported
this information, especially in multiple-cause deaths, is
unknown. Various correctional professionals may complete
the DCRP records, including jail administrators and
correctional medical and line staff, while only the treating
physician, a forensic pathologist, coroner, or funeral director
may complete and file death certificates in the United States.

The DCRP did not collect information on the manner of
death for suicides (e.g., hanging) until 2008 and for alcohol
and drug intoxication (e.g., ethanol toxicity and heroin
toxicity) until 2009.
DCRP respondents are instructed to report the final COD
(known as the immediate COD on the death certificate) as
recorded by a medical examiner or through another official
medical investigation. (See Collecting cause of death data
in the DCRP and NDI textbox.) In comparison, U.S. death
certificates capture “the chain of events—diseases, injuries,
or complications—that directly caused the death” ranked
in order from the immediate to underlying. Contributing
causes, which did not result in the underlying cause, are
also collected in the National Center for Health Statistics’
(NCHS) National Death Index (NDI).
COD is one of the items that require the highest level of
data quality follow-up in the DCRP collection due to item
nonresponse or invalid data. If an autopsy is requested—
as it is for nearly all jail and most prison deaths—DCRP
respondents must wait for the autopsy report and all
associated toxicology and lab tests to be made available. In
some cases, respondents cannot obtain COD information
from the medical examiner and will submit a record that is
missing the COD. Missing data require follow-up and can
delay the final death record for weeks or months.
To address these issues, BJS examined the NDI as an
alternative means of capturing COD data. The NDI is
a centralized database of death certificate information
abstracted from state vital statistics offices. Through NDI
searches, health and medical researchers can determine
whether persons in their studies have died or establish
COD for persons known to be deceased. The NDI is
considered the most comprehensive and accurate collection
of death identification and COD information. Both the
source (e.g., death certificates) and coverage are better
than other mortality death files, such as the Social Security
Administration’s Death Master File.1
1Fillenbaum,

G. G., Burchett, B. M., & Blazer, D. G. (2009). Identifying
a National Death Index match. American Journal of Epidemiology, 170(4),
515-518.

BJS had several objectives for this project: obtain the COD
codes reported on death certificates, evaluate the quality of
the matches, and assess the agreement in COD between the
matched DCRP and NDI records.
This report summarizes the results of these analyses and
provides recommendations for improving COD data by
linking DCRP and NDI data.

DCRP and NDI cause of death coding
Both the DCRP and NDI use codes for COD from the
10th revision of the International Statistical Classification of
Diseases and Related Health Problems (ICD-10). However, the
two collections differ in how they assign ICD-10 codes and
identify the most important COD when multiple conditions
are present. These differences play an important role in COD
mismatch between the DCRP and NDI (table 1).
A nosologist (i.e., a clinical coder trained by the NCHS)
processes DCRP data based on the codes listed on DCRP
death forms. The nosologist assigns a maximum of five
codes to each DCRP death case. The NDI provides up to
20 ICD-10 codes for the various conditions a decedent had,
with a maximum of 14 for this dataset. While the NDI assigns
ICD-10 codes for all deaths, the DCRP assigns ICD-10 codes
only to natural deaths, such as heart disease, cancer, and liver
disease. DCRP respondents are instructed to check a box for
AIDS-related deaths, but no additional COD information is
captured for these deaths. Unnatural deaths, such as drug or
alcohol intoxication, homicides, suicides, and accidents, are
not assigned ICD codes in the DCRP.

The NDI uses the Automated Classification of Medical Entities
(ACME) program to identify the underlying COD, which
is the international standard for automated selection of the
underlying COD. Studies have shown little discrepancy (up
to 7%) between underlying causes assigned by ACME and a
professional nosologist, but ACME offers considerable savings
both in cost and efficiency.2
The inputs to ACME are the multiple ICD codes assigned to
each entity (e.g., disease condition, accident, or injury) listed
on death certificates, in a particular order as reported. ACME
then applies the World Health Organization rules to the ICD
codes and selects an underlying COD. Underlying COD is
defined as “(a) the disease or injury which initiated the train of
events leading directly to death, or (b) the circumstances of the
accident or violence which produced the fatal injury.”3
The DCRP uses a different system for identifying the
underlying COD. Specifically, respondents are instructed to
report a final COD as determined by a medical examiner.
Cases with a single COD are ranked by frequency separately
for male and female decedents. When a death is caused by
multiple illnesses, the most common sex-specific COD among
those with only one COD is designated the underlying COD.
For example, a male decedent with two conditions—liver
disease and cancer—is counted as a cancer death because
cancer is a more common COD than liver disease among men
who die from a single COD.
2 Doody,

M. M., Hayes, H. M., & Bilgrad, R. (2001). Comparability of
National Death Index Plus and standard procedures for determining
causes of death in epidemiologic studies. Annals of Epidemiology, 11(1),
46-50.
3 World

Health Organization. (1975). Manual of the international
statistical classification of diseases, injuries, and causes of death: World
Health Organization.

TABLE 1
Comparison of DCRP and NDI cause of death data
Data source
Data provider
Availability of ICD-10 codes
Maximum number of ICD-10 codes
Selection of the underlying COD
Categorization of the underlying COD
Time lag

DCRP
DCRP death forms
Prison and jail administrator or correctional medical staff
Illness deaths only
5
Most common cause for single-cause deathsb
11 major categories of death
9 months from end of calendar year

NDI
Death certificates
Treating physician, forensic pathologist, coroner, or funeral director
All deaths
14a
Automated Classification of Medical Entities (ACME) program
ICD-10 code, 113 categories, and 11 major categories of death
12 months from end of calendar year

Note: COD = cause of death.
aThe NDI records a maximum of 20 International Classification of Diseases (ICD-10) codes. The matched DCRP 2007–10 death records captured a maximum of 14 codes.
bCases with a single COD are ranked by frequency separately for male and female decedents. When a death is caused by multiple illnesses, the most common sex-specific COD
among those with only one COD is designated the underlying COD.
Sources: Bureau of Justice Statistics, Deaths in Custody Reporting Program (DCRP), 2007–2010; and National Center for Health Statistics, National Death Index (NDI),
2007–2010.

Assessing Inmate Cause of Death: Deaths in Custody Reporting Program and National Death Index | April 2016

In the NDI, the underlying COD variable is assigned an
ICD-10 code. This variable is collapsed into broader COD
categories: first into a 113-category Level-2 COD variable and
then into an 11-category Level-1 COD variable. For instance,
ICD-10 codes E10 through E14 (i.e., insulin-dependent
diabetes mellitus, non-insulin-dependent diabetes mellitus,
malnutrition-related diabetes mellitus, other specified diabetes
mellitus, and unspecified diabetes mellitus) are grouped
into diabetes mellitus, a Level-2 category. The Level-1 CODs
include 11 major categories of death:

homicide

heart disease

In the DCRP, the underlying COD is a categorical variable with
the same 11 groups as the Level-1 underlying COD in the NDI;
the DCRP does not offer ICD-10 codes or detailed categories
for the underlying COD. Because BJS publishes inmate death
rates by major categories of COD annually, this study focuses
on the 11-category underlying COD variable when comparing
DCRP and NDI data. Comparisons of multiple causes using all
available ICD-10 codes were also conducted.

AIDS-related
cancer
liver disease
respiratory disease
other illness

drug and alcohol intoxication
accident
other.

Both Level-1 and Level-2 categories are standard classifications
of CODs that are used by epidemiologists to categorize CODs
and by the Centers for Disease Control and Prevention (CDC)
and the NCHS.

suicide

DCRP form

Collecting cause of death data in the DCRP and NDI

15. What was the cause of death? *** Please SPECIFY cause of death—it is critical information ***

Illness—Exclude AIDS-related deaths [Specify] ______________________________________________________________
Acquired Immune Deficiency Syndrome (AIDS)
Accidental alcohol/drug intoxication [Describe] _____________________________________________________________
Accidental injury to self [Describe] ________________________________________________________________________
Accidental injury by other (e.g., vehicular accidents during transport) [Describe] ___________________________________
Suicide (e.g., hanging, knife/cutting instrument, intentional drug overdose) [Describe] ______________________________
Homicide [Describe] ___________________________________________________________________________________
Other cause(s) [Specify] _________________________________________________________________________________

U.S. standard certificate of death

CAUSE OF DEATH (See instructions and examples)
32. PART I. Enter the chain of events—diseases, injuries, or complications—that directly caused the death. DO NOT enter terminal
events such as cardiac arrest, respiratory arrest, or ventricular fibrillation without showing the etiology. DO NOT ABBREVIATE. Enter
only one cause on a line. Add additional lines if necessary.
IMMEDIATE CAUSE (Final disease or condition resulting in death ---------> a. _______________________________________________
Sequentially list conditions, if any, leading to the cause listed on line a. Enter the UNDERLYING CAUSE (disease or injury that initiated
the events resulting in death) LAST.
b. ___________________________________________________________________________________________________________
c. ___________________________________________________________________________________________________________
d. ___________________________________________________________________________________________________________
PART II. Enter other significant conditions contributing to death but not resulting in the underlying cause given in PART I.

Assessing Inmate Cause of Death: Deaths in Custody Reporting Program and National Death Index | April 2016

Matching procedure
DCRP death records from 2007 through 2010 were sent to
the NDI Plus service, which provides COD listings for each
potential match. Subjects in the DCRP were considered known
decedents by the NDI, meaning that the DCRP had established
the fact of death. Known decedent searches result in a higher
likelihood of true matches than searches for subjects with an
unknown vital status.
The NDI prefers using Social Security numbers (SSNs) as
identifiers. Neither SSNs nor maiden names of female inmates
are collected by DCRP and could be linked with the NDI. The
match between DCRP and NDI records used the following
variables: decedents’ birth and death dates, first and last names,
sex, race or Hispanic origin, and state of residence.
The DCRP 2007–10 file contained 17,420 records. Analysis
was limited to those records with complete information on
birth and death dates and unique birth date and death date
combinations. Some pairs of subjects with the same birth and
death dates also matched on last name, sex, race or Hispanic
origin, and state of residence, differing only in first name. It is
likely that these were duplicate records, and the difference in
the first names was due to the inmate’s use of an alias. The final
study cohort contained 17,358 decedents (figure 1).
The search for 17,358 DCRP subjects in the NDI database
resulted in 37,868 possible matches. Each possible match was
assigned a probabilistic match score by the NDI system. The
probabilistic score is calculated by summing the individual
weights applied to each data element (e.g., decedents’ sex, race
or Hispanic origin, date of birth, state of residence) used in

the NDI match. The weights reflect the degree of agreement
between the information on the submission record and the
NDI record.4 For subjects with multiple possible matches, the
record with the highest probabilistic score was selected. This
procedure reduced the total number of matches to 16,265
deaths, or one match per subject.
In addition to the probabilistic match score, NDI provides
two variables to assess the quality of potential matches: status
code and exact match indicator. Derived from the probabilistic
match score, the status code indicates whether an individual
is either assumed to be alive (status code=0) or dead (status
code=1). In this study, given that all DCRP subjects are known
to be dead, a status code of 0 means that the potential match
has a relatively low probability of being a true match. The
NDI assigns status codes conservatively to avoid incorrectly
assuming that a living person is dead. As a result, while most
of the potential matches assigned a status code of 1 were true
matches, a number of true matches were assigned a status code
of 0.
The exact match variable indicates that all match fields on the
user record agree exactly with the information in the NDI.
Exact match is a more conservative criterion than status code
and may misclassify more true matches (i.e., false negatives)
due to misspellings or nicknames (e.g., Robert compared to
Bob). Given that all subjects were known to be dead, a broad
definition of match was employed to avoid the risk of dropping
eligible deaths, and all 16,265 cases with NDI matches were
included in the analysis.
4 National

Center for Health Statistics. (2013). National Death Index user’s
guide. Hyattsville, MD: U.S. Department of Health and Human Services,
Centers for Disease Control and Prevention.

Figure 1
DCRP–NDI data linkage flowchart
17,420 DCRP deaths
Dropped

Sent to NDI for matching

62 cases with incomplete DOB, DOD, or
non-unique DOB and DOD combinations
Unmatched records
1,093 cases had no matches in the NDI

17,358 cases with complete and
unique DOB and DOD combinations
Matched records
37,868 possible matches found
for 16,265 cases

Best match retained for each record,
for a total of 16,265 matches
Note: DOB = date of birth; DOD = date of death.
Sources: Bureau of Justice Statistics, Deaths in Custody Reporting Program (DCRP), 2007–2010; and National Center for Health Statistics, National Death Index (NDI),
2007–2010.

Assessing Inmate Cause of Death: Deaths in Custody Reporting Program and National Death Index | April 2016

Quality of matching
Overall, the NDI had at least one record available for 94% of
the DCRP subjects. No matches were found for 1,093 DCRP
records (table 2). Eighty-two percent of the records had a
status code of 1 and 67% had an exact match.
The overall match rate (94%) was slightly lower than in other
samples, such as Vietnam era veterans (97%) and refinery and
petrochemical workers (97%).5 However, in the absence of
several important identifiers (e.g., SSN and maiden name for
females), this was a decent match rate.
5 Boyle,

C. A. & Decoufle, P. (1990). National sources of vital status
information: Extent of coverage and possible selectivity in reporting.
American Journal of Epidemiology, 131(1), 160-168; and Acquavella,
J. F., Donaleski, D., & Hanis, N. M. (1986). An analysis of mortality
follow-up through the National Death Index for a cohort of refinery
and petrochemical workers. American Journal of Indian Medicine, 9(2),
181-187.

TABLE 2
Matching results between the DCRP and NDI, 2007–2010
Total
Facility type
State prison
Local jail
Year
2007
2008
2009
2010
Sex
Male
Female
Race/Hispanic origin
Whitec
Black/African Americanc
Hispanic/Latino
Otherc,d
Age
17 or younger
18–24
25–34
35–44
45–54
55 or older

N
17,358

NDI
matched
93.7%

Status
code = 1a
81.9%

Exact
matchb
66.6%

13,470
3,888

94.2%
91.9

82.2%
81.0

67.8%
62.4

4,483
4,397
4,359
4,119

92.7%
93.2
94.7
94.2

79.5%
81.5
83.2
83.5

69.1%
64.9
63.8
68.5

16,321
1,037

93.9%
90.3

81.9%
82.2

66.9%
61.7

9,089
5,930
1,940
354

94.3%
93.6
91.6
92.1

83.1%
81.3
78.4
81.6

71.2%
63.7
60.6
29.1

20
625
1,599
2,803
5,179
7,132

90.0%
91.8
92.5
92.9
93.6
94.5

80.0%
85.9
87.4
82.5
80.9
80.9

55.0%
64.0
65.1
66.3
66.2
67.5

Note: Detail may not sum to total due to missing cases.
aThe NDI assigns a status code for each matched record. A code of 1 indicates a high
probabilistic match score, and a code of 0 indicates otherwise.
bExact match means that all items provided on the user record agree exactly with
the items in the NDI.
cExcludes persons of Hispanic or Latino origin.
dIncludes American Indians or Alaska Natives; Asians, Native Hawaiians, or Other
Pacific Islanders; and persons of two or more races.
Sources: Bureau of Justice Statistics, Deaths in Custody Reporting Program (DCRP),
2007–2010; and National Center for Health Statistics, National Death Index (NDI),
2007–2010.

The match rate did not show much variation by facility
type, decedent’s sex, race or Hispanic origin, age, or death
year. Even when differences were statistically significant
between groups (e.g., prison compared to jail deaths and
male compared to female decedents), they amounted to about
2 to 4 percentage points.

Comparing data coverage for cause of death
About 2% of death records in 2007–10 were missing COD in
the DCRP, compared to 6% in the NDI (table 3). COD data
may be missing in the NDI for a DCRP record for two reasons:
(1) no matched record was found in the NDI, or (2) a matched
record was found but the NDI had no COD information for
it. In this comparison, all missing COD in the NDI was due
to unmatched records. All matched records had COD data in
the NDI. Out of the 336 cases with missing COD in the DCRP,
COD data was available for 302 cases by linking with the NDI,
leaving 34 cases with no COD information (not shown).
The rate of missing data was higher for jail deaths than
prison deaths, particularly in the DCRP (6% missing in jail
compared to 1% in prison). This was due to a data anomaly
in 2008, when 21% of the DCRP jail deaths (n=202) were
missing COD information.6 By linking to the NDI database,
COD information was obtained for 185 out of those 202 cases,
providing a resolution for 92% of the cases (not shown).
6Excluding

2008, 1% of jail deaths were missing COD.

TABLE 3
Percent missing cause of death data in the DCRP and
NDI, by facility type, 2007–2010
Total
State prison
2007
2008
2009
2010
Local jail
2007
2008
2009
2010

DCRP
1.9%
0.7%
1.5
0.0
1.0
0.1
6.3%
1.8
21.2
2.1
0.4

NDI*
6.3%
5.8%
6.8
6.4
4.6
5.3
8.1%
8.8
8.0
7.8
7.6

*Missing cause of death in the NDI is due to unmatched records. All matched
records in the NDI have cause of death information.
Sources: Bureau of Justice Statistics, Deaths in Custody Reporting Program (DCRP),
2007–2010; and National Center for Health Statistics, National Death Index (NDI),
2007–2010.

Assessing Inmate Cause of Death: Deaths in Custody Reporting Program and National Death Index | April 2016

The NDI provided more ICD-10 codes than the DCRP. On
average, each subject received 1.06 ICD codes in the DCRP
(including unmatched cases), while each matched record
received an average of 2.69 codes in the NDI (table 4). This
large difference occurred in part because the DCRP asks
respondents for final cause of death, and the cause of death
question on the U.S. standard certificate of death requests
multiple causes. In addition, ICD-10 codes were not assigned
to unnatural deaths in the DCRP.
Prison deaths were assigned more ICD codes than jail deaths
both in the DCRP (1.17 for prison deaths compared to 0.68 for
jail deaths) and in the NDI (2.74 for prison deaths compared
to 2.54 for jail deaths). Large differences also existed by COD
category. The DCRP did not assign ICD codes for unnatural
deaths, such as suicides, homicides, intoxications, and
accidents. In the NDI, deaths from unnatural causes received
more ICD codes than deaths from illness.

Comparing underlying cause of death
In this report, DCRP and NDI data are compared to determine
the accuracy and completeness of CODs reported by DCRP
respondents. COD data were compared in three ways. The
distributions of underlying COD in the two datasets were
compared at the aggregate level. Also, the correspondence
of underlying COD between the two datasets was examined
at the individual level, and the reasons for mismatch were
explored. The first two analyses used the major category
of death, which classified COD into 11 groups. Finally, the
analysis compared the correspondence of multiple causes
between the two datasets at the individual level, using all
available ICD-10 codes. This analysis helped explain the
mismatch of the underlying COD in the two datasets.

TABLE 4
Mean number of ICD codes available in the DCRP and
NDI, 2007–2010

Total
Facility type
State prison
Local jail
Status codea
0
1
Sex
Male
Female
Year
2007
2008
2009
2010
Cause of death
Heart disease
AIDS-related
Cancer
Liver disease
Respiratory disease
Other illnessb
Suicide
Homicide
Drug/alcohol intoxication
Accident
Other/unknownc

DCRP
Standard
Mean
deviation
1.06
0.95

NDI
Standard
Mean
deviation
2.69
1.63

1.17
0.68

0.93
0.95

2.74
2.54

1.69
1.40

1.08
1.06

0.97
0.95

2.83
2.67

1.65
1.63

1.07
0.97

0.95
0.99

2.69
2.78

1.63
1.70

1.00
1.01
1.10
1.15

0.93
0.93
0.96
0.99

2.67
2.68
2.74
2.69

1.60
1.61
1.68
1.66

1.50
0.49
1.35
1.57
1.34
1.07
:
:
:
:
:

0.90
1.14
0.78
0.93
0.72
0.84
:
:
:
:
:

2.72
2.18
2.34
2.76
2.80
3.00
2.34
3.17
3.65
3.60
3.62

1.62
1.58
1.60
1.64
1.64
1.81
0.86
1.47
1.49
1.78
1.69

Note: The mean number of International Classification of Diseases (ICD-10) codes
in the DCRP is calculated by the DCRP cause of death category, while the mean
number in the NDI is calculated by the NDI cause of death category.
: Not calculated. Unnatural deaths in the DCRP were not assigned ICD-10 codes.
aThe NDI assigns a status code for each matched record. A status code of 1 indicates
a high probabilistic match score, and a code of 0 indicates otherwise.
bIncludes illnesses such as cerebrovascular disease, diabetes, and other nonleading
natural causes of death.
cIncludes cases where cause of death was provided but the manner was not known,
cases where an autopsy was inconclusive, and cases where the information listed
was a symptom rather than a cause or manner of death.
Sources: Bureau of Justice Statistics, Deaths in Custody Reporting Program (DCRP),
2007–2010; and National Center for Health Statistics, National Death Index (NDI),
2007–2010.

Assessing Inmate Cause of Death: Deaths in Custody Reporting Program and National Death Index | April 2016

The marginal distributions of the underlying COD from the
two datasets were similar. The leading CODs in the DCRP
were heart disease (25%), cancer (22%), and suicide (11%),
while those in the NDI were cancer (23%), heart disease
(22%), and suicide (11%) (table 5). Illness, or natural, deaths
were far more prevalent among prisoners (89% of prison
deaths in the DCRP and the NDI) than among jail inmates
(55% of jail deaths in the DCRP and 59% in the NDI). About
30% of jail deaths were due to suicide, compared to 6% of
prison deaths.
The leading CODs were the same for male and female inmates
(i.e., heart disease, cancer, and suicide) with small differences:
Cancer and heart disease accounted for higher percentages of
male deaths than female deaths, while suicide accounted for a
higher percentage of female deaths than male deaths.7 These
patterns were evident in both the DCRP and the NDI.
To examine how COD data in the NDI correspond to data in
the DCRP at the individual level, the DCRP underlying COD
was cross-tabulated by the NDI underlying COD for 15,963
respondents with known COD in both datasets. This reveals
how frequently the COD categorization of NDI agreed with
that of DCRP (table 6 and table 7).
Two statistics are used to measure the congruence of
DCRP and NDI COD data (i.e., intercoder reliability): the
agreement rate and the kappa statistic. The agreement rate
is the percentage of records classified into the same category
by the two datasets. The kappa statistic was first proposed by

7Although

suicides accounted for a larger percentage of female inmate
deaths, males accounted for the majority (91% in jail and 95% in prison)
of suicides in correctional facilities. See Mortality in Local Jails and State
Prisons, 2000–2013 - Statistical Tables (NCJ 248756, BJS web, August 2015).

Cohen.8 It measures intercoder reliability while taking into
account agreement that is expected purely by chance—that is,
it indicates the level of agreement achieved beyond chance.
When the raters (i.e., the DCRP and NDI) are in complete
agreement, kappa takes on the value of 1. Landis and Koch
suggested characterizing values less than 0 as indicating no
agreement, 0 to 0.20 as slight, 0.21 to 0.40 as fair, 0.41 to 0.60
as moderate, 0.61 to 0.80 as substantial, and 0.81 to 1 as almost
perfect agreement.9 (See Methodology for more detail.)
The overall agreement rate was 70%, meaning that 70% of
the deaths were classified into the same major category of
death by the DCRP and the NDI. The kappa statistic was 0.64,
which indicates a substantial level of agreement following
Landis and Koch’s guidelines. The percentage of records that
received identical classification in both collections varied
by COD category, ranging from less than 40% to more than
90%. The categories in the DCRP with the highest agreement
rates were suicide (93%), cancer (87%), homicide (86%), and
AIDS-related (79%). In the DCRP, deaths due to liver disease
(38%) and respiratory disease (48%) had the lowest agreement
rates. This low agreement rate was likely due to how DCRP
respondents recorded these deaths. For example, a death due to
hepatitis (an inflammation of the liver caused by the hepatitis
virus) would be recorded accurately as an infectious disease
by the NDI, but may have been recorded as a liver-related
death by the DCRP. Similarly, respiratory deaths are frequently
coupled with other conditions and diseases (such as AIDS,
cancer, or heart disease), and it is likely that DCRP respondents
listed the causes that they deemed more serious than others.
8 Cohen,

J. (1960). A coefficient of agreement for nominal scales.
Educational and Psychological Measurement, 20, 37-46.
9 Landis,

J. R. & Koch, G. G. (1977). The measurement of observer
agreement for categorical data. Biometrics, 33(1), 159-174.

TABLE 5
Distributions of cause of death in the DCRP and NDI, by facility type and sex of decedent, 2007–10
All
All causes
Illness
Heart disease
AIDS-related
Cancer
Liver disease
Respiratory disease
Other illnessa
Unnatural deaths
Suicide
Homicide
Drug/alcohol intoxication
Accident
Other/unknownb

DCRP
100%
81.8%
24.5
3.0
21.8
8.1
5.9
18.5
16.8%
11.4
1.8
2.5
1.1
1.4%

NDI
100%
82.8%
22.3
3.7
23.2
5.1
5.1
23.3
16.6%
10.9
1.9
2.3
1.6
0.7%

State prison
DCRP
NDI
100%
100%
88.9%
89.0%
24.9
22.0
2.8
3.9
26.6
28.0
9.3
5.6
6.5
5.6
18.8
24.0
10.1%
10.4%
6.2
5.9
1.7
1.8
1.4
1.4
0.9
1.3
1.0%
0.6%

Local jail
DCRP
NDI
100%
100%
55.2%
59.3%
23.3
23.4
3.4
3.3
4.0
5.0
3.7
3.3
3.6
3.5
17.3
20.8
41.6%
39.6%
31.0
29.6
2.0
1.9
6.5
5.5
2.1
2.7
3.2%
1.1%

Male
DCRP
100%
82.1%
24.8
2.9
22.3
8.3
5.8
18.0
16.6%
11.2
1.9
2.4
1.1
1.4%

NDI
100%
83.0%
22.5
3.7
23.5
5.2
5.1
22.9
16.4%
10.7
2.0
2.1
1.6
0.7%

Female
DCRP
NDI
100%
100%
77.9%
79.4%
20.3
17.6
4.7
4.5
15.0
17.1
4.7
4.8
7.4
5.4
25.9
30.0
20.2%
19.6%
14.1
13.1
0.1
0.1
4.7
4.4
1.3
2.0
2.0%
1.0%

Note: Includes only records with a known underlying cause of death in both the DCRP and NDI (n=15,963).
aIncludes illnesses such as cerebrovascular disease, diabetes, and other nonleading natural causes of death.
bIncludes cases where cause of death was provided but the manner was not known, cases where an autopsy was inconclusive, and cases where the information listed was a
symptom rather than a cause or manner of death.
Sources: Bureau of Justice Statistics, Deaths in Custody Reporting Program (DCRP), 2007–2010; and National Center for Health Statistics, National Death Index (NDI),
2007–2010.

Assessing Inmate Cause of Death: Deaths in Custody Reporting Program and National Death Index | April 2016

TABLE 6
Number of NDI causes of death, by DCRP cause of death category, 2007–2010
DCRP cause of death
All causes
Heart disease
AIDS-related
Cancer
Liver disease
Respiratory disease
Other illnessa
Suicide
Homicide
Drug/alcohol intoxication
Accident
Other/unknownb

Total
15,963
3,917
471
3,487
1,296
943
2,946
1,816
279
398
181
229

Heart
disease
3,553
2,745
22
101
38
99
447
12
9
26
6
48

AIDSrelated
597
27
371
33
34
38
87
3
0
0
0
4

Cancer
3,696
146
15
3,046
65
83
290
11
6
3
5
26

Liver
Respiratory
disease disease
820
821
59
148
9
8
58
42
497
12
35
449
144
130
2
10
0
2
2
4
1
1
13
15

Other
illnessa
3,724
722
41
177
630
218
1,739
46
9
41
18
83

Drug/alcohol
Other/
Suicide Homicide intoxication Accident unknownb
1,732
295
361
256
108
7
9
21
24
9
0
0
1
4
0
8
4
6
10
2
2
0
6
8
4
0
3
2
11
5
8
21
24
40
16
1,679
9
11
25
8
0
240
0
5
8
12
3
259
4
44
3
5
19
116
7
13
1
12
9
5

TABLE 7
Percent of NDI causes of death, by DCRP cause of death category, 2007–2010
DCRP cause of death
All causes
Heart disease
AIDS-related
Cancer
Liver disease
Respiratory disease
Other illness
Suicide
Homicide
Drug/alcohol intoxication
Accident
Other/unknownb

Total
100%
100%
100%
100%
100%
100%
100%
100%
100%
100%
100%
100%

Heart
disease
22.3%
70.1
4.7
2.9
2.9
10.5
15.2
0.7
3.2
6.5
3.3
21.0

AIDSrelated
3.7%
0.7
78.8
1.0
2.6
4.0
3.0
0.2
0.0
0.0
0.0
1.8

Cancer
23.2%
3.7
3.2
87.4
5.0
8.8
9.8
0.6
2.2
0.8
2.8
11.4

Liver
Respiratory Other
disease disease
illnessa
5.1%
5.1%
23.3%
1.5
3.8
18.4
1.9
1.7
8.7
1.7
1.2
5.1
38.4
0.9
48.6
3.7
47.6
23.1
4.9
4.4
59.0
0.1
0.6
2.5
0.0
0.7
3.2
0.5
1.0
10.3
0.6
0.6
9.9
5.7
6.6
36.2

Drug/alcohol
Other/
Suicide Homicide intoxication Accident unknownb
10.9%
1.9%
2.3%
1.6%
0.7%
0.2
0.2
0.5
0.6
0.2
0.0
0.0
0.2
0.9
0.0
0.2
0.1
0.2
0.3
0.1
0.2
0.0
0.5
0.6
0.3
0.0
0.3
0.2
1.2
0.5
0.3
0.7
0.8
1.4
0.5
92.5
0.5
0.6
1.4
0.4
0.0
86.0
0.0
1.8
2.9
3.0
0.8
65.1
1.0
11.1
1.7
2.8
10.5
64.1
3.9
5.7
0.4
5.2
3.9
2.2

Assessing Inmate Cause of Death: Deaths in Custody Reporting Program and National Death Index | April 2016

The NDI attributed 18% of the 3,917 DCRP deaths due to
heart disease (n=722) to some undefined illness categories,
4% to respiratory disease (n=148), and 4% to cancer (n=146).
In total, the NDI attributed 3,553 deaths to heart disease,
9% less than the DCRP. The NDI also attributed 14% of DCRP
homicides to other causes. Specifically, 9% were classified as
natural deaths in the NDI—heart disease (3%), cancer (2%),
respiratory or other illness (4%)—and the remainder were
classified as accidental or other.
Close to half (49%) of deaths due to liver disease in the DCRP
were attributed to other illness in the NDI. This included
liver-related deaths caused by the hepatitis virus because of
its infectious nature. The NDI also attributed 11% of DCRP
deaths from respiratory disease to heart disease, 9% to cancer,
and 23% to other illness.
When the NDI classified a record differently from the DCRP,
it tended to place it in the same broad category of COD as in
the DCRP (either illness or unnatural death). In total, 68% of
the illness deaths in the DCRP were classified into the same
major category of death in the NDI (table 8). Thirty percent of
the illness deaths were attributed to a different illness category,
while only 2% were reported as unnatural deaths. Eighty-six
percent of the unnatural deaths in the DCRP received the same
classification in the NDI, 8% were attributed to illness deaths,
and 4% to other causes.

Assessing the sources of disagreement between the
DCRP and NDI
Differences in the COD classification between the DCRP
and NDI are due to several factors. Mismatched records
are the first factor to consider. Two records may not be true
matches even if they have the same identifiers, such as last
name and birth and death dates. It is not possible to accurately
distinguish true matches from false matches, but the NDI
status code can be used as a proxy. As previously discussed, the
NDI matching process assigned a status code of 1—indicating
a high probabilistic score—to 82% of the matches, and a code
of 0—indicating a low probabilistic score—to 18% of matches.
The congruence in underlying COD between the DCRP and
NDI should be higher for matches with a status code of 1 than
for those with a status code of 0.
The second factor is data source. The information on the
death certificate may be different from that on the medical
examiner’s report, or the text may be the same, but the case

was assigned different ICD-10 codes by NDI and DCRP
nosologists. This explanation was examined by directly
comparing the sets of ICD codes from the DCRP and NDI for
matched records.
The two datasets use different methods to identify the
underlying COD when decedents have multiple health
conditions. Specifically, the NDI uses ACME to identify the
underlying COD from the list of ICD-coded conditions.
When using ACME, shuffling the order in which the codes are
listed can change the underlying COD. The DCRP instructs
respondents to report a final COD as determined by a medical
examiner. For multiple-cause deaths in the DCRP, the most
common cause is designated as the underlying COD. Due to
such procedural differences, even records with an identical
set of ICD-10 codes may be assigned to different underlying
COD categories in the two datasets. To understand how much
this factor affects the congruence of the underlying COD
between the DCRP and NDI, agreement rates were calculated
separately for records with an identical set of ICD codes in the
two datasets and those with different ICD codes. Agreement
rates were also calculated by the number of ICD codes. The
agreement rates should be higher for records with an identical
set of ICD codes and for those with fewer ICD codes.
TABLE 8
Percent of cases in the DCRP and NDI that match on
cause of death, 2007–10
Same
category
DCRP cause of death
as DCRP
All causes
69.8%
Illnessb
67.7%
Heart disease
70.1
AIDS-related
78.8
Cancer
87.4
Liver disease
38.3
Respiratory disease
47.6
Other illness
59.0
Unnatural
85.8%
Suicide
92.5
Homicide
86.0
Drug/alcohol intoxication
65.1
Accident
64.1

Different category from DCRP
Other/
Illness
Unnatural unknowna
27.3%
2.2%
0.7%
30.3%
1.7%
0.3%
28.1
1.6
0.2
20.2
1.1
0.0
11.8
0.8
0.1
60.1
1.2
0.3
50.2
1.7
0.5
37.3
3.2
0.5
8.1%
3.6%
2.5%
4.6
2.5
0.4
9.3
1.8
2.9
19.1
4.8
11.1
17.1
14.9
3.9

aIncludes cases where cause of death was provided but the manner was not known,
cases where an autopsy was inconclusive, and cases where the information listed
was a symptom rather than a cause or manner of death.
bIncludes illnesses such as cerebrovascular disease, diabetes, and other nonleading
natural causes of death.
Sources: Bureau of Justice Statistics, Deaths in Custody Reporting Program (DCRP),
2007–2010; and National Center for Health Statistics, National Death Index (NDI),
2007–2010.

Assessing Inmate Cause of Death: Deaths in Custody Reporting Program and National Death Index | April 2016

The congruence of the underlying COD between the DCRP
and the NDI was compared by status code and other factors
(table 9). The kappa statistic of 0.46 and the overall agreement
rate of 56% for records with a status code of 0 (i.e., matches
with lower probabilistic match scores) were significantly lower
than the kappa statistic of 0.66 and 72% for the corresponding
statistics for those with a status code of 1 (i.e., those with
higher probabilistic match scores). Except for accidental
and other or unknown COD, deaths in each COD category
by DCRP had a higher agreement rate for matches with a
status code of 1 than for matches with a status code of 0. For
example, 83% of the AIDS-related deaths in the DCRP with a
status code of 1 were also classified as AIDS-related deaths in
the NDI, compared to 57% of those with a status code of 0.

also categorized as respiratory disease in the NDI, compared
to 42% of the deaths attributed to respiratory disease in the
DCRP that also had other causes. Because the DCRP did not
assign ICD-10 codes to unnatural deaths, this analysis was
restricted to DCRP illness deaths only.
Jail deaths had better agreement than prison deaths, with a
kappa statistic of 0.68 (compared to 0.61 for prison deaths) and
an overall agreement rate of 74% (compared to 69%). This was
likely because nearly all jail deaths in the study group were sent
for autopsy (95%), while about two-thirds (65%) of prisoner
deaths were autopsied.
DCRP–NDI agreement improved over time. Within 4 years,
the kappa statistic increased from 0.60 in 2007 to 0.68 in 2010
and the overall agreement rate increased from 67% in 2007 to
74% in 2010.

The agreement rates were also compared by the number of
conditions (i.e., single-cause compared to multiple-cause
deaths), facility type, sex, and death year. In this context,
single-cause deaths were those that had only one ICD-10
code in both datasets, and multiple-cause deaths had more
than one ICD-10 code in either the NDI or DCRP. Because
no ambiguity was involved in identifying the underlying
COD for single-cause deaths, single-cause deaths had higher
intercoder reliability, with a kappa statistic of 0.77 and an
overall agreement rate of 83%, compared to a kappa statistic of
0.52 and an overall agreement rate of 63% for multiple-cause
illness deaths. The difference in agreement rates was also
substantial within each illness category. For example, 73% of
deaths attributed only to respiratory disease in the DCRP were

To further explore how DCRP–NDI agreement varies by
decedent demographics and circumstances of death, a
logistic regression was run with the 15,293 death records
with COD information in both datasets. The dependent
variable is binary, with 1 indicating that the record received
the same COD classification in the DCRP and the NDI, and
0 indicating otherwise. The independent variables include
facility type, sex, race or Hispanic origin, age group of
the deceased, year of death, DCRP COD category, match
quality (or NDI’s status code), and the number of conditions
(i.e., single compared to multiple).

TABLE 9
DCRP–NDI cause of death (COD) agreement, by status code, COD count, and facility type, 2007–10
Kappab
Agreement rate
All causes
Illness
Heart disease
AIDS-related
Cancer
Liver disease
Respiratory disease
Other illnessc
Unnatural
Suicide
Homicide
Drug/alcohol intoxication
Accident
Other/unknownd

All
0.64

Status codea
0
1
0.46
0.66

69.8%
67.7%
70.1
78.8
87.4
38.3
47.6
59.0
85.8%
92.5
86.0
65.1
64.1
2.2

56.0%
56.2%
60.7
57.1
72.0
23.8
36.5
50.1
61.2%
64.7
55.6
46.9
65.0
0.0

71.8%
69.5%
71.4
82.5
89.8
40.5
49.0
60.4
88.4%
95.3
90.5
66.7
64.0
2.5

Number of CODs
Single
Multiple
0.77
0.52
:
83.0%
84.7
92.5
94.3
63.2
72.5
65.3
:
:
:
:
:
:

:
62.6%
66.7
65.8
83.3
33.2
42.3
57.2
:
:
:
:
:
:

Facility type
State prison Local jail
0.61
0.68
68.7%
67.6%
68.9
79.5
87.6
37.3
46.2
58.3
84.6%
91.0
88.7
64.8
64.5
1.6

74.1%
68.5%
74.9
76.5
81.2
48.0
57.5
62.1
86.9%
93.6
77.6
65.3
63.4
2.8

Year
2007
0.60
66.8%
64.5%
68.9
78.2
85.5
36.4
45.3
54.7
85.8%
92.5
87.1
65.1
59.5
2.5

2008
0.61

2009
0.66

2010
0.68

67.6%
66.6%
70.2
76.9
88.3
35.5
48.3
56.0
84.8%
92.3
86.0
61.5
66.7
1.0

71.6%
69.5%
70.6
79.3
86.9
40.8
49.8
62.1
84.8%
92.6
83.6
65.4
56.6
0.0

73.5%
70.6%
70.6
81.3
88.5
40.9
47.2
65.1
87.4%
92.4
87.2
68.5
73.1
13.3

Note: Includes only records with a known cause of death in both the DCRP and NDI (n=15,963).
: Not calculated. Unnatural deaths in the DCRP were not assigned International Classification of Diseases (ICD-10) codes.
aThe NDI assigns a status code for each matched record. A status code of 1 indicates a high probabilistic match score, and a code of 0 indicates otherwise.
bKappa is a measure of interrater agreement for qualitative items. It is a more robust measure than percent for agreement, as it takes into account the agreement occurring
by chance.
cIncludes illnesses such as cerebrovascular disease, diabetes, and other nonleading natural causes of death.
dIncludes cases where cause of death was provided but the manner was not known, cases where an autopsy was inconclusive, and cases where the information listed was a
symptom rather than a cause or manner of death.
Sources: Bureau of Justice Statistics, Deaths in Custody Reporting Program (DCRP), 2007–2010; and National Center for Health Statistics, National Death Index (NDI), 2007–2010.

Assessing Inmate Cause of Death: Deaths in Custody Reporting Program and National Death Index | April 2016

The regression results confirm the relationships observed in
the previous analysis (table 10). Records with a status code
of 1 were more likely to receive the same categorization for
underlying COD in both datasets than those with a status
code of 0, with an odds ratio of 2.2. This effect was statistically
significant. Likewise, single-cause deaths were more likely to
be consistently classified than multiple-cause deaths, with an
odds ratio of 2.05.
Intercoder reliability also varied across COD categories.
Deaths due to suicide, homicide, cancer, and AIDS-related
problems had higher agreement rates than deaths by heart
disease, while liver disease, respiratory disease, and drug and
alcohol intoxication had lower agreement rates. Furthermore,
deaths that occurred in jail had higher agreement rates than

deaths that occurred in prison, with an odds ratio of 1.21. The
sex of the decedent made no difference on agreement rate.
Deaths of black decedents had lower agreement rates than
those of white decedents, with an odds ratio of 0.91. There
were no statistical differences among Hispanics, whites, and
persons of other races. Lastly, the agreement rates in 2009 and
2010 were significantly higher than in 2007.
To further understand the source of mismatch between the
underlying COD, detailed ICD-10 codes from the DCRP and
the NDI were analyzed for matched records. This analysis was
limited to 12,131 records with ICD codes from both records.
Thirty percent of the cases were excluded either because they
were classified as unnatural deaths, and therefore not assigned
ICD-10 codes in the DCRP, or because no matches were found

TABLE 10
Logistic regression of DCRP–NDI cause of death agreement, 2007–2010
Dependent variable: Cause of death in agreementa
Status code = 1b
Single-cause death (ref = multiple-cause death)
DCRP cause of death category (ref = heart disease)
AIDS-related
Cancer
Liver disease
Respiratory disease
Other illnessc
Suicide
Homicide
Drug/alcohol intoxication
Accident
Other/unknownd
Local jail (ref = prison)
Female (ref = male)
Race/Hispanic origin (ref = white)
Black/African Americane
Hispanic/Latino
Otherf
Death year (ref = 2007)
2008
2009
2010
Constant

Odds ratio
2.20*
2.05*

Coefficient
0.12*
0.11*

Standard error
0.120
0.106

1.39*
2.82*
0.26*
0.38*
0.61*
5.29*
2.94*
0.76*
0.80
0.01*
1.21*
0.98

0.17*
0.18*
0.02*
0.03*
0.03*
0.52*
0.52*
0.09*
0.13
0.00*
0.06*
0.08

0.170
0.179
0.018
0.028
0.032
0.521
0.522
0.087
0.130
0.004
0.063
0.081

0.91*
0.90
0.83

0.04*
0.06
0.12

0.039
0.056
0.117

1.07
1.16*
1.23*
0.96

0.06
0.06*
0.07*
0.07

0.057
0.062
0.067
0.068

Observations

15,923

Note: The reference group is the baseline used in comparisons. For example, an effect of 2.05 for single-cause deaths means that the odds of having a cause of death in
agreement between the DCRP and the NDI were twice as high for single-cause deaths than for the reference group multiple-cause deaths.
*Indicates statistical significance at 0.05 level.
aThe dependent variable is an indicator that equals 1 when DCRP and NDI major categories of death are in agreement and 0 when they are not.
bThe NDI assigns a status code for each matched record. A status code of 1 indicates a high probabilistic match score, and a code of 0 indicates otherwise.
cIncludes illnesses such as cerebrovascular disease, diabetes, and other nonleading natural causes of death.
dIncludes cases where cause of death was provided but the manner was not known, cases where an autopsy was inconclusive, and cases where the information listed was a
symptom rather than a cause or manner of death.
eExcludes persons of Hispanic or Latino origin.
fIncludes American Indian or Alaska Native; Asian, Native Hawaiian, or Other Pacific Islander; and persons of two or more races.
Sources: Bureau of Justice Statistics, Deaths in Custody Reporting Program (DCRP), 2007–2010; and National Center for Health Statistics, National Death Index (NDI),
2007–2010.

Assessing Inmate Cause of Death: Deaths in Custody Reporting Program and National Death Index | April 2016

in the NDI. The 12,131 records were classified into 5 groups
based on the relationship of their ICD codes reported in the
DCRP and NDI (table 11).
Twenty percent of the cases (2,402) received the same set of
ICD codes in the DCRP and NDI. In 3,371 cases (28%), the
DCRP ICD codes were a subset of the NDI codes. That is, all
DCRP codes were also reported on the matching NDI record,
but the latter also had one or more codes not reported in the
DCRP. The NDI codes were a subset of the DCRP codes in
342 cases (3%). Because the DCRP provides up to 5 ICD codes
for each illness death, while the NDI provides up to 14 codes
for each death in this dataset, it is more likely that the COD
codes assigned by the DCRP were a subset of those assigned
by the NDI than the reverse. In 1,129 cases (9%), the ICD-10
codes from the two datasets were different with some overlap.
In 4,887 cases (40%), the ICD-10 codes from the two datasets
were different with no overlap. The proportions of the last two
groups indicate substantial mismatch in the all-cause ICD
codes between the two datasets.
Ninety-four percent of records with the same set of
ICD-10 codes in the DCRP and the NDI were assigned to
the same major category of death by the two datasets in
the categorization of the underlying COD (table 12). The
agreement rate was 78% for records with DCRP codes being
a subset of NDI codes, and 72% for records with NDI codes
being a subset of DCRP codes. The agreement rate was lower
for records with different sets of ICD codes with overlap (62%)
and without overlap (52%).
Six percent of the cases with identical ICD-10 codes in the
DCRP and the NDI ended up in different major categories of
underlying COD. This could be due to the order in which the
ICD-10 codes were entered into the system, as order can affect

the final COD determination in ACME. For cases with a single
identical ICD-10 code in both datasets, the agreement rate was
almost 100%. As the number of ICD-10 codes increased to
5, the agreement rate dropped to 57%. The same pattern was
observed for groups B and C. As the number of ICD-10 codes
in the DCRP increased, the agreement rate dropped from 79%
to 58% for group B, and from 84% to 57% for group C. For
groups D and E, where DCRP and NDI codes were different
with or without overlapping, the agreement rate was relatively
low, regardless of the number of ICD-10 codes in the DCRP.
TABLE 11
Comparison of DCRP and NDI ICD-10 codes, 2007–10
Agreement of DCRP and NDI ICD-10 codes
Total
A. Identical codes in the DCRP and NDIa
B. DCRP codes were a subset of NDI codesb
C. NDI codes were a subset of DCRP codesc
D. NDI and DCRP codes overlappedd
E. No overlap between DCRP and NDI codese

Number
12,131
2,402
3,371
342
1,129
4,887

Percent
100%
19.8
27.8
2.8
9.3
40.3

Note: Includes only matched records that had International Classification of
Diseases (ICD-10) codes in both the DCRP and NDI datasets (n=12,131). The DCRP
provides up to 5 ICD codes for illness deaths, while the NDI provides up to 14 codes
for each matched case.
aThis group received an identical set of ICD codes from the DCRP and NDI.
bThis group received more ICD codes from the NDI than from the DCRP, and the
DCRP codes were a subset of the NDI codes.
cThis group received more ICD codes from the DCRP than from the NDI, and the NDI
codes were a subset of the DCRP codes.
dThe codes differed between DCRP record and its NDI match, but there was some
overlap in the reported codes.
eThe codes in the DCRP record and the NDI matched record were different and had
no overlap.
Sources: Bureau of Justice Statistics, Deaths in Custody Reporting Program (DCRP),
2007–2010; and National Center for Health Statistics, National Death Index (NDI),
2007–2010.

TABLE 12
Percent of records with same underlying cause of death, by International Classification of Diseases (ICD-10) codes,
2007–10
Agreement of DCRP and NDI ICD-10 codes
Total
A. Identical codes in the DCRP and NDIa
B. DCRP codes were a subset of NDI codesb
C. NDI codes were a subset of DCRP codesc
D. NDI and DCRP codes overlappedd
E. No overlap between DCRP and NDI codese

Total
69.0%
93.5
78.1
71.9
61.9
52.0

One
70.6%
99.7
78.7
~
~
52.1

Number of ICD-10 codes in the DCRP
Two
Three
Four
68.2%
61.7%
53.4%
78.4
76.3
65.2
76.3
69.0
58.3
83.6
60.2
48.8
65.4
61.5
52.0
53.2
46.7
43.4

Five
56.7%
57.1
~
56.5
58.3
47.6

Note: Includes only matched records that had International Classification of Diseases (ICD-10) codes in both the DCRP and NDI datasets (n=12,131). The DCRP provides up to 5
ICD codes for illness deaths, while the NDI provides up to 14 codes for each matched case.
~Not applicable.
aThis group received an identical set of ICD codes from the DCRP and NDI.
bThis group received more ICD codes from the NDI than from the DCRP, and the DCRP codes were a subset of the NDI codes.
cThis group received more ICD codes from the DCRP than from the NDI, and the NDI codes were a subset of the DCRP codes.
dThe codes differed between DCRP record and its NDI match, but there was some overlap in the reported codes.
eThe codes in the DCRP record and the NDI matched record were different and had no overlap.
Sources: Bureau of Justice Statistics, Deaths in Custody Reporting Program (DCRP), 2007–2010; and National Center for Health Statistics, National Death Index (NDI),
2007–2010.

Assessing Inmate Cause of Death: Deaths in Custody Reporting Program and National Death Index | April 2016

These results confirm that disagreement in the underlying
COD between the DCRP and the NDI was due partly to
differences in the source data and partly to the different
procedures for identifying the underlying COD. Forty percent
of the cases received completely different ICD codes in the
DCRP and the NDI. This group had a much lower agreement
rate in underlying COD than records that received identical
or similar ICD codes in the two datasets. Even among records
that received the same ICD codes, the agreement rate declined
as the number of conditions increased, suggesting that the
different procedures for identifying the underlying COD in
the two datasets can lead to different classifications of major
categories of deaths.

Summary and implications
This study evaluated the feasibility of linking the DCRP data
to the NDI database and assessed whether the DCRP was
accurately capturing and measuring COD, whether the NDI
could provide COD for DCRP cases missing this data, and
whether the NDI could provide additional detail to illness
deaths. Matches were found for 94% of the death records sent
to the NDI, and 82% of the matches were considered a true
match by the NDI matching program. Notably, the percentage
of cases that were unable to be matched declined over time,
from 7.3% in 2007 to 5.8% in 2010. To further improve the
matching effectiveness of NDI search, BJS may consider
the following:
Create and use a unique subject ID variable for DCRP

records sent to the NDI. A subject ID would allow the study
to include persons who share the same last name, date of
birth, and date of death as unique individuals with shared
traits. Without the identifier, the distinction between unique
persons and potential duplicate records could not be made,
and these cases were excluded.

Send multiple records with all known aliases to the NDI to

improve matching. All aliases for one individual should be
identified by the same subject ID. This would be applicable
if the data providers included information about aliases by
which an inmate is known (e.g., Sammy or Sam). If more
than one match was returned, the match with the highest
probabilistic score could then be used as the true match.

Use decedents’ SSNs in matching. The DCRP does not

collect SSNs; however, BJS may be able to obtain them
for decedents by matching DCRP records to other BJS
administrative databases that include them.

The NDI linkage provides reliable, accurate, and
comprehensive information of COD on jail and prison
inmates. NDI data are based on death certificates and
processed with standard National Center for Health Statistics
(NCHS) codes. The NDI uses a reproducible automated
algorithm (i.e., ACME) to identify the underlying COD. It
allows comparisons with national data on noninstitutionalized
populations because the same algorithms are used for ranking
COD, determining underlying COD, and rolling up the CODs
into broad categories of death. These algorithms are stable over
time and systematic, and they are easily understood by public
health and medical audiences.
The NDI provided more than twice as many ICD-10 codes per
subject than did the DCRP (2.69 for the NDI compared to 1.06
for the DCRP). This was likely a result of the way CODs were
recorded in death certificates. BJS may consider adopting the
version found in standard death certificates.
With more accurate and comprehensive COD data from
the NDI, BJS may produce statistics on specific diseases
in correctional populations. Prison administrators and
correctional health professionals may use detailed NDI trend
data to plan and budget for treating chronic health conditions
among long-term inmates. For example, hepatitis C has
received a lot of attention in correctional health circles recently,
in part because new pharmacotherapy to treat hepatitis C
is expensive and the older standard course of treatment is
complicated and hard to implement among inmates.10 National
statistics could be used by administrators and health care
professionals to justify the funding and resources needed to
mitigate the high rates of hepatitis C found in some facilities.
NDI search is a viable solution if BJS wants to reduce missing
COD data or improve data quality. NDI search provided COD
information for 90% of the cases with missing data in the
DCRP. It takes approximately 12 months after the end of the
calendar year for deaths to be included in the NDI database,
which is 3 months longer than the DCRP. However, relying
exclusively on the NDI for COD is not recommended because
a subset of DCRP cases (6% in 2007–10) cannot be matched
with NDI data, and some matches may not be correct.
10Keller,

J. E. (2014). Hepatitis C treatment: Between a rock and a
hard place. Correct Care, 28(2), 8-9, 20. http://www.ncchc.org/ filebin/
CorrectCare/28-2.pdf; and Mortality in Local Jails and State Prisons,
2000–2013 - Statistical Tables, NCJ 248756, BJS web, August 2015.

Assessing Inmate Cause of Death: Deaths in Custody Reporting Program and National Death Index | April 2016

At the individual level, 70% of the records were classified into
the same major category of death by the two datasets. The
discrepancies were due to several factors: (1) the matched
records may not be true matches, (2) most cases were not
assigned the same set of ICD-10 codes by the two datasets,
and (3) the process for identifying the underlying COD from
multiple ICD codes was different between the two datasets
such that even records with identical ICD codes may be
classified to different major categories of death. Deaths with
more ICD-10 codes were less likely to be categorized as the
same underlying COD in the two datasets.
BJS publishes annual statistical tables on the distribution of
COD over 11 major categories. When detailed NDI data were
rolled into the same categories, the overall distributions from
the two datasets were similar. In other words, the current
DCRP method for obtaining COD produces aggregate statistics
that are roughly comparable to what may be produced using
NDI data.
Overall, the NDI linkage provides significant improvement
in the quality of COD data in the DCRP. BJS will report COD
statistics with increased confidence when DCRP COD data
align with NDI data.

Methodology
Data source: Deaths in Custody Reporting Program
The Deaths in Custody Reporting Program (DCRP) is an
annual Bureau of Justice Statistics (BJS) data collection.
The DCRP collects national, state, and incident-level data
on persons who died while in the physical custody of the
50 state departments of corrections or the approximately 2,800
local adult jail jurisdictions nationwide. The DCRP began
in 2000 under the Death in Custody Reporting Act of 2000
(P.L. 106-297), and it is the only national statistical collection
to obtain comprehensive information about deaths in adult
correctional facilities. BJS uses DCRP data to track national
trends in the number and causes (or manners) of deaths
occurring in state prison or local jail custody.
The DCRP collects data about the characteristics of the
decedents as well as circumstances surrounding the death,
including the cause, time and location where the death occurred,
and information on whether an autopsy was conducted and the
availability of results to the respondent. Data on executions are
excluded from this report but are accessible on the BJS website
along with the DCRP mortality data.
The DCRP data collection instruments are administered
annually to both state prisons and local jails. In addition to
individual death records, respondents provide an aggregate
count of the number of deaths that occurred during the
referenced calendar year.

BJS obtains a separate report describing the decedent’s
characteristics and the circumstances surrounding the death
for each death that occurred in a state prison or local jail. State
prison and local jail respondents can submit individual records
on decedents at any time during a collection cycle through a
BJS web-based collection system.
Data source: The National Death Index
The National Death Index (NDI) is a centralized database of
death certificates filed by state vital statistics offices. The NDI
was established by the National Center for Health Statistics
(NCHS) established the NDI as a mortality register for
epidemiologists and other professional medical investigators
and death researchers. Like the DCRP, the NDI is available
for statistical purposes only, which precludes its use for legal,
administrative, private or genological purposes.
The NDI is considered the “gold standard” for death
identification because both the source, i.e., death certificates,
and coverage is better than other mortality death files, such as
the Social Security Administration’s Death Master File.11
Nonresponse
All state department of corrections and more than 99% of
jails participated for data years 2007 through 2010. However,
there is varying degree of item nonresponse, particularly for
the cause of death item. BJS changed data collection agents in
2009. As a result, the follow-up period for data year 2008 was
cut short, and 21% of jail deaths and 3% of prison deaths were
missing cause of death information. Aside from data year 2008,
missingness affects a small percentage of cases—less than 1% of
prison and 3% of jail deaths—annually.
Reported statistics
The kappa statistic was first proposed by Cohen.12 It measures
intercoder reliability while taking into account agreement
that is expected purely by chance—that is, it indicates the
level of agreement achieved beyond chance. For reliability
between two raters (i.e., the DCRP and NDI), kappa is
calculated as (po – pe) / (1 – pe) , where po is the observed
proportionate agreement among raters and pe is the expected
proportionate agreement by chance. When the raters are in
complete agreement, kappa is 1. Landis and Koch suggested
characterizing values less than 0 as indicating no agreement,
0 to 0.20 as slight, 0.21 to 0.40 as fair, 0.41 to 0.60 as moderate,
0.61 to 0.80 as substantial, and 0.81 to 1 as almost perfect
agreement.13
11Fillenbaum,

G.G., Burchett, B.M. & Blazer, D.G. (2009). Identifying a
National Death Index Match. American Journal of Epidemiology, 170(4),
515–518.
12Cohen,

J. (1960). A coefficient of agreement for nominal scales.
Educational and Psychological Measurement, 20, 37–46.
13Landis,

J. R. & Koch, G. G. (1977). The measurement of observer
agreement for categorical data. Biometrics, 33(1), 159–174.

Assessing Inmate Cause of Death: Deaths in Custody Reporting Program and National Death Index | April 2016

The Bureau of Justice Statistics of the U.S. Department of Justice is the
principal federal agency responsible for measuring crime, criminal
victimization, criminal offenders, victims of crime, correlates of crime,
and the operation of criminal and civil justice systems at the federal, state,
tribal, and local levels. BJS collects, analyzes, and disseminates reliable and
valid statistics on crime and justice systems in the United States, supports
improvements to state and local criminal justice information systems,
and participates with national and international organizations to develop
and recommend national standards for justice statistics. Jeri M. Mulrow is
acting director.
This report was written by Zhen Zeng, Margaret Noonan, and E. Ann Carson,
BJS Statisticians; Ingrid Binswanger, BJS Visiting Fellow; Patrick Blatchford,
Colorado School of Public Health; and Hope Smiley-McDonald, and Chris
Ellis, RTI International. Anastasios Tsoutis and Emmaline Mitchell verified
the report.
Lynne McConnell, Morgan Young, and Jill Thomas edited the report.
Tina Dorsey and Barbara Quinn produced the report.
April 2016, NCJ 249568

NCJ249568

Office of Justice Programs
Innovation • Partnerships • Safer Neighborhoods
www.ojp.usdoj.gov