Study design and data source
We conducted a retrospective observational study of abortion-related ED visits using data from the Nationwide Emergency Department Sample (NEDS), a nationally representative sample of ED visits. NEDS is a database of ED visits from 947 to 964 hospitals across the U.S. per year. Annually it includes more than 30 million unweighted visits, which represent more than 135 million weighted visits. NEDS was developed for the Healthcare Cost and Utilization Project (HCUP) and is maintained by the Agency for Healthcare Research and Quality. Data are available from 2006 onward. For this study, we utilized the five most recent years of data available (2009–2013). The study was certified exempt by the institutional review board of the University of California, San Francisco.
Unweighted visits are data collected on actual visits, which are then weighted proportionately to the total number of ED visits in the country based on the sampling strategy. The NEDS is a stratified single-stage cluster sample of state-level ED data reported to HCUP. Using the American Hospital Association Annual Survey of Hospitals as the target universe, the available data are selected to approximate a 20% stratified sample of all U.S. hospital-based EDs. More details of the sampling of hospitals can be found on the HCUP website [8, 9]. The characteristics used for sample stratification in the NEDS are U.S. region, urban or rural location, teaching status, ownership, and trauma level (see Fig. 1 for region definitions and states contributing data).
The NEDS includes patient-level and hospital-level information. Each ED visit has patient-level demographic characteristics including age, sex, primary and secondary payment source, and zip code-based urbanicity and income quartile. Each ED visit also has clinical characteristics, including up to 15 diagnoses (International Classification of Diseases, 9th Revision [ICD-9] codes), up to 15 procedures or treatments (Healthcare Common Procedure Coding System [HCPCS] and Current Procedural Terminology [CPT] codes), injury codes, admission and discharge status, diagnosis and treatment codes for inpatient care if admitted to the same hospital, and total charges. Each visit also has the corresponding hospital code, and hospital characteristics such as region, trauma level, urban-rural location, and teaching status. In this dataset, 5 of 13 states in the West, 11 of 12 states in the Midwest, 8 of 16 states in the South, and 8 of 9 states in the Northeast were represented in the data. Midwest hospitals were represented the most. The trauma level of a hospital refers to how well equipped it is to provide care to patients with traumatic injuries. Trauma level influences patient composition and was key to sample stratification in the dataset. Hospital ownership was categorized by the data distributor according to information reported in the American Hospital Association Annual Survey Database. Ownership could be governmental, private non-profit, or private for-profit. Hospitals with religious affiliations, including Catholic hospitals, are included, but are not distinguished as such and may fall into private for-profit or private non-profit categories. Federal hospitals (Veterans Affairs and Department of Defense) were not included in the sample. Patient-level and hospital-level weights were also provided to generate nationally representative estimates. HCUP provided weights for the NEDS data and these were calculated at the hospital level after sampling by hospital strata. Patient weights were calculated first by stratifying the data by hospital characteristics (region, urbanicity, trauma level, teaching status, and ownership). Within each of these strata, a weight was generated by dividing the total number of ED visits in the U.S. in that year for that stratum (from American Hospital Association data) by the number of ED visits for that stratum in the NEDS data. Weighted data thus represent all ED visits in the U.S. for a year.
Data preparation
We identified all ED visits that had an ICD-9 diagnosis code for abortion (ICD-9 diagnosis codes 635, 636, 637, and 638). We categorized ED visits by abortion relatedness and treatments received, and assessed whether the visit was for a major or minor incident. Visit categorization was based on the previously developed Procedural Abortion Incident Reporting and Surveillance (PAIRS) Framework [10]. Based on procedure codes, visits that were for observation care, repeat procedures (codes present in the dataset were CPT/HCPCS procedure codes 59812, 59820, 59821, 59840, 59841, 59851, 59855, and 59856; and ICD-9 procedure codes 6901, 6902, 6909, 6951, 6952, 6959, 734, and 750), blood transfusions (CPT/HCPCS procedure codes 36430 and P9021; and ICD-9 procedure codes 9903–9905), and abortion-related surgeries (CPT/HCPCS procedure codes 49320, 58662, 58999, 59300, and 59898; and ICD-9 procedure codes 680, 6831, 6839, 6849, 6851, 6859, 6869, and 7491) were systematically coded without individual visit review. Where procedure and diagnosis codes for a visit did not fall into one of these categories, several authors who are emergency medicine physicians or students with physician supervision provided an individual review of each visit following a modified version of the PAIRS framework (see Fig. 2). After a joint review of 100 visits with refinement of the decision rules, the physicians reviewed the remaining uncategorized visits, resulting in 1642 reviewed ED visits in total.
For each reviewed visit, the emergency medicine physicians assigned the reason for the patient’s ED visit to one of five categories: abortion-related, concurrent condition, pre-existing condition, not abortion-related, or cannot determine. They classified an ED visit as abortion-related based on the constellation of diagnosis and procedure codes for that visit. Abortion-related visits included adverse events, such as hemorrhage or infection, and symptoms directly related to the procedure, such as abdominal pain and vomiting. Concurrent conditions were defined as conditions that may have been noticed during or exacerbated by abortion, but were not directly caused by the abortion, such as ovarian cysts, vaginitis, urinary tract infection, or anxiety/depression. Pre-existing conditions were defined as chronic conditions such as hypertension or diabetes. The data were also categorized with regards to treatments received for abortion-related complaints. Categories of treatment included uterine reaspiration (which involves suction, not an incision, and does not meet the criteria for surgery), intravenous (IV) and non-IV medication, repeat abortion, blood transfusion, surgery, observation care, or other. The medication category excluded codes for injections or infusion of a therapeutic substance if no accompanying medication was listed. The observation care category included women who had routine testing for their symptoms but did not receive any medications or other treatments. This included women who had IV fluids, blood work, testing for sexually transmitted infections, and diagnostic imaging studies, but no treatments. The NEDS dataset also included information on whether patients were admitted as an inpatient to the same hospital, discharged home, transferred to another facility, or left against medical advice. Among women who were admitted, the reason for visit and treatment information in their chart was used to determine if the admission was likely abortion-related.
Major and minor incidents were systematically coded, with major incidents defined as those requiring an overnight inpatient stay, blood transfusion, or surgery. Minor incidents were defined as all other incidents that involved an abortion-related diagnosis or treatment. All overnight stays were further reviewed by physicians, and a group of treatments for an abortion-related diagnosis that together required the patient to stay overnight qualified as a major incident. Visits that involved concurrent conditions, pre-existing conditions, or visits that were not abortion-related were categorized as no incident. We identified the prevalence of comorbid conditions based on diagnosis codes. These included three key conditions hypothesized to be associated with abortion-related adverse events: diabetes, hypertension, and having an overweight/obese body mass index (BMI) [11,12,13,14].
ED visits were additionally categorized as being potentially indicative of a self-induced abortion. The physician team individually reviewed all abortion-related visits that had diagnosis codes of illegal abortion, failed attempted abortion, and certain injury codes (including poisonings and indications of self-harm). They looked at all of the diagnosis codes for that case and made a clinical judgement based on the group of codes together and their ED experience. Visits that were coded as illegal abortions, particularly those that included injury codes, were more often considered potentially self-induced. In addition, visits that included injury codes consistent with self-induction were categorized as such. Cases which were unlikely to have been abortion-related were removed.
Statistical analysis
We estimated the number of abortion-related ED visits annually in the U.S. and the proportion of ED visits among women of reproductive age (15–49) that were for abortion-related care. We examined the characteristics of the sample and compared these to published estimates of the characteristics of abortion patients [15]. We also described outcomes including treatments received and discharge status. Based on treatments received, ED visits were categorized as being for a major incident, a minor incident, no incident, or that the incident type could not be determined. We then built multivariable logistic regression models to examine the factors associated with major incidents and observation care, controlling for sociodemographic characteristics, comorbidities, and hospital characteristics with the ED visit as the unit of analysis. Per the recommendation of the HCUP, the organization which oversees the NEDS database, sample weights were not used in the multivariable models [16]. We also estimated the proportion of ED visits that were potentially due to attempted self-induced abortion. Because the weighted estimates are nationally representative, we were also able to use published national estimates of abortion incidence [17] to estimate the major incident rate for abortion in the U.S. during the study period. This assumes that the vast majority of major incidents go through an ED evaluation (although we acknowledge that a small percentage do not). For all analyses, we report weighted results unless otherwise specified. Where data were missing, a missing category was retained for all analyses. Statistical significance was set at P < 0.05 for all chi-squared tests and adjusted odds ratios (AORs). We used STATA 14 for all analyses.