A Nonparametric Imputation Approach for Dealing With Missing Variables in SHR Data

• Author: Robert L. Flewelling
• Published date: 01 August 2004
• Date: 01 August 2004
• DOI: http://doi.org/10.1177/1088767904265361
• Subject Matter: Articles

10.1177/1088767904265361
HOMICIDE STUDIES / August 2004
Flewelling / ANONPARAMETRIC IMPUTATION APPROACH
A Nonparametric Imputation
Approach for Dealing With
Missing Variables in SHR Data
ROBERT L. FLEWELLING
Pacific Institute for Research and Evaluation
Despite the considerable percentage of homicides recorded in the Supplementary Homi-
cide Reports (SHR) that have missing information, there is no standard approach used by
researchers to adjust homicide rates to accommodate missing data. Yet, rates that are
defined on the basis of offender characteristics or victim-offender relationship will be
severely underestimated if no adjustments, or imputations, are made. Several approaches
to imputation of SHR data have been described in the literature. This research note
describes an imputation strategy that appears to offer a viable alternative approach,
although further development and testing of this and other techniques are needed before
an optimal strategy can be determined.
Keywords: Supplementary Homicide Reports; missing data; imputation
It has now been more than 15 years since Kirk Williams and I pub-
lished an article in Criminology (Williams & Flewelling, 1987)
describing procedures we used to address missing data problems
in the Supplementary Homicide Reports (SHR). That effort was
part of a broader study to examine structural correlates of disag-
gregated homicide rates in American cities. A central theme of
that research, reiterated by Flewelling and Williams (1999), was
AUTHOR’S NOTE: This article was first submitted as a brief research note and com-
mentary but expanded into a longer piece at the suggestion of the editors for this special
issue. The approach described still warrants further development and validation, but it is
hoped that the article will stimulate consideration of this and other alternative approaches
for addressing missing data in the SHR. The author thanks Marc Riedel and Wendy
Regoeczi for their encouragement concerning this submission, and Kirk Williams for guid-
ing early work in this area. The development of the methods described herein was funded
in part by a grant (95-CX-0114) from the National Institute of Justice.
HOMICIDE STUDIES, Vol. 8 No. 3, August 2004 255-266
DOI: 10.1177/1088767904265361
© 2004 Sage Publications
255

---

256
HOMICIDE STUDIES / August 2004
the critical importance of disaggregating homicides into concep-
tually meaningful subgroups to advance understanding through
comparative research. The 1987 article focused on two problems
inherent in SHR data: underreporting of homicide incidents and
missing fields of information within reported incidents. Missing
information typically involves offender characteristics and the
victim-offender relationship and now affects more than one third
of all SHR records. Disaggregated homicide rates defined on the
basis of offender characteristics, therefore, are subject to signifi-
cant underestimation of varying degrees. The adjustment proce-
dures described in the article were designed to improve the accu-
racy of estimated homicide rates over calculations that ignored
missing data or that made simple proportionate adjustments
without using additional information that was available in the
homicide record data.
Although my own areas of research have since been refocused
onto other topics, I have always believed that those initial, and
admittedly still crude, adjustment procedures developed in the
mid-1980s could be further developed and substantially
enhanced. What we designed then served the purposes of the
underlying study in which we were engaged. At the same time,
we stressed the importance of further efforts to assess and refine
these procedures, which would include identifying both their
strengths and their limitations. Indeed, other investigators have
identified concerns in applying our strategy to other research
questions. For example, Langford, Isaac, and Kabat (1998)
reported significant overestimation of intimate partner victimiza-
tions when applying the procedures to SHR data from the state of
Massachusetts.
Although the percentage of homicides that have missing infor-
mation has increased significantly over the past two decades, a
cursory review of research published in Homicide Studies suggests
that until very recently there had not been substantial further
progress in developing techniques to compensate for missing
information through adjustment or imputation strategies. One
notable exception is the procedure designed by Fox (1997, 2001)
that employs a weighting technique that essentially replaces
records with missing offender information with incidents (or,

---

Flewelling / A NONPARAMETRIC IMPUTATION APPROACH
257
more precisely, weighted averages of incidents) from within the
same state and year that are similar with respect to victim charac-
teristics. This approach is noteworthy because it uses additional
information (victim characteristics) known about the incidents in
question to help infer the values of missing data elements. It also
provides users of public use files with a standard approach to
missing data adjustment that is easy to apply.
Two very recently published articles describe the use of formal
statistical modeling techniques to impute the values of missing
victim-offender relationship categories. Regoeczi and Riedel
(2003) employed an expectation-maximization (E-M) algorithm
in imputing the missing values for homicides recorded in Chi-
cago and Los Angeles, and Messner, Deane, and Beaulieu (2002)
used a log-multiplicative modeling approach with data from
the 1996 and 1997 SHR files. These methods employ statistical
models based on the relationships between victim-offender rela-
tionship and other variables in the homicide records that are
much less likely to be missing to estimate the value of the victim-
offender relationship for incidents in which it is not reported.
Using statistically related information from records that have
missing information on offender characteristics (including
victim- offender relationship) to help impute those characteristics
makes a great deal of sense, and there are a number of potentially
viable approaches for doing so.
...