Concordance of data about sex from electronic health records and the National Death Index: Implications for transgender populations

Abstract

Background:

Transgender individuals have greater health risks than cisgender individuals, which may bode for greater mortality. However, research is limited by lack of gender identity information at the time of death. Novel opportunities to combine administrative data with National Death Index (NDI) data may facilitate mortality research about transgender populations, but binary measures of sex and gender may pose problems for analyses. This study explored differences in sex recorded in Veterans Health Administration (VHA) electronic health record (EHR) and NDI data between transgender and cisgender decedents.

Methods:

We used VHA EHR data from fiscal years 2000-2016 to identify deaths among a sample of transgender and cisgender patients. We cross-tabulated sex recorded in the NDI with EHR-based sex from VHA EHR data. We extracted data in 2018 and conducted analyses in 2020.

Results:

Death occurred for 1,109 transgender patients and 7,757 cisgender patients. For cisgender decedents, EHR-based sex and NDI-based sex were 100% concordant. For transgender decedents, 46 (4%) were discordant between data sources. Of transgender decedents with female EHR-based sex (n = 259), 17% were indicated as male in NDI data; of those with male EHR-based sex (n = 850), 0.2% were indicated as female in NDI data.

Conclusions:

Data linkage between EHR and the NDI can facilitate transgender mortality research, but examining mortality specific to various transgender identities remains difficult. Improved documentation of sex and gender is needed within US mortality surveillance.

INTRODUCTION

Transgender individuals have greater risks to their health (e.g., homelessness, violence) than cisgender individuals,¹ which bodes for greater risk of death.² For example, homicide among transgender individuals (specifically transgender women of color), prompted the American Medical Association to issue a policy statement advocating for infrastructure, research, and practice to reduce violence against people who are transgender.³ However, mortality studies are strongly limited by lack of gender identity information at the time of death.⁴ Currently, only binary data about sex (male or female) are available in the National Death Index (NDI).

Matching electronic health record (EHR) with NDI data presents opportunities to learn about mortality among transgender populations,⁵ ideally if EHR data contain self-identified gender identity data. Unfortunately, collection of gender identity data in EHR is currently rare.⁶ International Classification of Disease (ICD) codes for gender identity disorder (GID), which is often (but not always) required for transgender individuals to access gender affirming therapies, can identify a subset of transgender populations (i.e., those with GID diagnosis codes).⁷

The Veterans Health Administration (VHA) linked EHR and NDI data through 2017 for basic mortality data, including sex documented on the death certificate. However, it is unclear how something often construed as a basic demographic – the sex of an individual – becomes complex for transgender individuals when cross-referenced with administrative data. The objective of this brief report was to examine concordance of sex denoted between VHA EHR and NDI. We hypothesize that cisgender patients would be 100% concordant between the two data sources, but transgender patients would be less than 100% concordant.

METHODS

This secondary analysis used data from a previous study that included all VHA patients between fiscal years 2000-2016 who ever had one of four ICD 9 (e.g., 302.85 GID in adolescents or adults) or one of seven 10 GID-related codes (e.g., F64.9 GID, unspecified), creating the group of transgender patients (n=8,977). A complete list of these ICD codes has been previously published.⁷ In the absence of self-reported gender identity in VHA data,⁸ ICD codes have been used to define transgender status with documented validity from clinical text notes. For example, among patients with a GID-related code, nearly 90% had clinical text notes that included text suggesting the patient’s gender identity (e.g., “patient is transgender man”).⁷ For each transgender patient, three patients were randomly chosen based on not having GID-related ICD codes and, to account for facility and temporal factors, had to have an outpatient visit in the same facility within +/− 5 days from the date of the transgender patient’s GID diagnosis. This group was referred to as cisgender patients (n=26,907). Data older than 5 years from the time of the study were necessary due to the rarity of GID-related codes and the need to create the most comprehensive sample of transgender patients. From this existing study data, we focused the present analysis only on individuals who had died during the observation period pursuant to this brief report’s focus on mortality.

NDI is considered the gold standard for mortality data in the United States,⁹ aggregating information about cause and date of death gathered from offices of vital statistics from all US states, Washington DC, and US territories. Sex data in the NDI are compiled by the National Center for Health Statistics from death certificate data reported from 57 jurisdictions (e.g., U.S. states, the District of Columbia, Puerto Rico); death certificate data typically include information input from medical professionals, funeral directors, and next-of-kin contacts.¹⁰

VHA contracts with the NDI to conduct probabilistic matching on all patients to ascertain vital status and mortality data for deceased patients currently through 31 December 2017 and includes information such as date, cause of death, and sex documented on the death certificate. We categorized deaths according to National Vital Statistics System (NVSS) conventions.⁸ At the time of the study, NDI data for VHA patients were available only through 2017.

VHA EHR include a single data field for sex of the patient, which can only be populated as either male or female (i.e., EHR-based sex). VHA has a formal process through which patients may request changes to their EHR-based sex.⁸ Thus, it is unclear if EHR-based sex represents a patient’s sex designated at birth or indicates their current gender identity. For all patients in our cohort, we extracted EHR-based sex from their most currently available patient data as the closest approximation of sex prior to death.

Sex recorded in NDI from the death certificate was cross-tabulated with the EHR-based sex from VHA administrative data for overall causes of death. The institutional review board of the VA Pittsburgh Healthcare System approved this study.

RESULTS

Over the observation period, there were 1,109 deaths among transgender patients and 7,757 deaths among non-transgender patients. For cisgender decedents, there was 100% concordance between the EHR-based sex and the sex documented in NDI data (Table). For transgender decedents, 46 (4%) were discordant between data sources. Among transgender decedents whose EHR-based sex indicated female (n=259), 17% were indicated as male in NDI data; among those whose EHR-based sex indicated male (n=850), 0.2% were indicated as female in NDI data.

Table.

Comparisons of sex recorded at last medical visit and sex recorded in the National Death Index for all causes of death

	Decedents without GID ICD code (i.e., cisgender) (n=7,757)		Decedents with GID ICD code (i.e., transgender) (n=1,109)
	EHR-based Sex a
	Male	Female	Male	Female
	(n=7,527)	(n=230)	(n=850)	(n=259)
NDI-based Sex	n (%)	n (%)	n (%)	n (%)
Male	7,527 (100)	0 (0)	848 (99.8)	44 (17)
Female	0 (0)	230 (100)	2 (0.2)	215 (83)

^aSex documented in patients’ most recently available medical data

EHR=electronic health record; GID=gender identity disorder; ICD=International Classification of Disease; NDI=National Death Index

DISCUSSION

Because US mortality surveillance does not include gender identity⁴ and few EHR systems include gender identity information,⁶ mortality research about transgender individuals currently requires proxy-based procedures. However, as this first study illustrated with linked mortality and EHR data for a subsample of transgender individuals (i.e., those with GID diagnoses), discordant information emerged about the decedent’s sex, which supported the study hypotheses.

The discordance among transgender patients is difficult to interpret. For example, among the 44 transgender decedents with female EHR-based sex who were documented as male in NDI, it is unclear if these decedents may be misgendered transgender women (EHR valid, NDI invalid), misgendered transgender men (EHR invalid, NDI valid), or some combination. Thus, although data linkage can facilitate mortality research for transgender people generally, challenges remain for learning about transgender women, transgender men, and especially non-binary individuals who may specifically be lost in solely binary sex and gender data.

This brief report was a first step in exploring this heretofore undocumented issue in data systems. Further research should explore if other structured data (e.g., Current Procedural Terminology [CPT] codes) or unstructured data (e.g., clinical text notes; chart reviews) of gender affirming surgeries or hormones, which have been used in recent studies to try to determine sex assigned at birth, ^11,12 may clarify gender identity for mortality studies. However, because VHA currently does not cover gender affirming surgeries⁸ and only 58% of transgender veterans receive hormones through VA,¹³ data may be limited compared to other systems that do cover affirming surgeries (e.g., Medicare), which may have implications for deaths involving autopsies and consequently sex data recorded on death certificates.

Because mortality is a key metric for health services research, the ramifications of these results for transgender populations require bold future directions. The breadth of NVSS data facilitates identification of disparities by socio-demographics, such as race, ethnicity, age, and sex.¹⁴ Transgender populations are invisible in these data. Piecemeal efforts (e.g., EHR linkage) cannot replace surveillance powered by the >2.5 million death records processed annually in the NVSS.¹⁴ Widespread implementation of self-identified gender identity in EHR may help but remains vexed by disparities in access to care for transgender individuals.¹⁵ There are efforts to improve data quality for the National Violent Death Reporting System (NVDRS) by training death investigators to document decedent gender identity as they do any other demographic characteristic.⁴ However this effort is predicated on the existence of a NVDRS data field for “transgender status,” which is, itself, problematic but is a start.¹⁶ The NDI does not include a data field for gender identity. Expanding data fields on death certificates is no small task, including variability in professionals involved in completing them (e.g., medical examiners, coroners, funeral directors). However, it has been done before, as when Hispanic ethnicity and educational attainment were added in 1989.¹⁷

We note several limitations. Using ICD codes to define transgender status invites misclassification bias; not all patients who self-identify as transgender need or have a GID diagnosis and not all patients with a GID diagnosis may self-identify as transgender. The sample of veterans utilizing VHA care may limit generalizability to non-veteran populations or veterans who do not utilize VHA care.

Transgender individuals face several challenges jeopardizing their health (e.g., violence, homelessness, discrimination),¹ and without mortality data, research cannot identify patterns and correlates to guide prevention. Although linking EHR and NDI data provide opportunities to learn about transgender mortality,¹⁸ the limitations of these methods – and the fact that such data gymnastics are needed in the first place – reify the very disparity of transgender communities, i.e., second-rate research methods for a population treated like second-class citizens.¹⁵ This is not to say progress cannot happen with creative methods with administrative data; indeed it has, but incrementally. To identify, understand, and ultimately address health disparities for transgender communities, US mortality surveillance must reckon with the nuances, challenges, and improvements around sex and gender identity data.