Association Between Migraine and Suicidal Behaviors

A Nationwide Study in the USA

Lauren E. Friedman, PhD; Qiu-Yue Zhong, MD, ScM; Bizu Gelaye, PhD, MPH; Michelle A. Williams, ScD; B. Lee Peterlin, DO


Headache. 2018;58(3):371-380. 

In This Article


Study Sample

Our study included adult participants, 18 years of age or older, with hospital discharges in 2007–2012 from the Nationwide Inpatient Sample (NIS) database. The NIS is a database of hospital inpatient stays compiled from billing data across the USA as part of the Healthcare Cost and Utilization Project (HCUP). The HCUP is the largest inpatient health care utilization database in the USA and includes approximately 20% of discharges from all nonfederal, acute care hospitals. All hospitals participating in HCUP are included in the sample. Hospital characteristics, including geographic region, ownership, location, teaching status, and bed size, were used for weighting to account for the survey structure and to create a sample that is representative of hospital admissions.[19]

Standard Protocol Approvals, Registrations, and Patient Consents

As the NIS data are publicly available and do not contain personal identifiers, this study was exempt from review by the Office of Human Research Administration at the Harvard T.H. Chan School of Public Health, Boston, MA.

Demographic Characteristics

Sociodemographic variables including age, sex, race, median household income quartiles, length of stay, and total hospital charges were evaluated. Smoking was identified using International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) diagnosis codes (305.1X, 649.0X, and V15.82).


Adult onset migraine diagnoses were based on the ICD-9-CM diagnosis codes (346. XX). Hospitalizations without migraine were composed of hospitalizations without a diagnosis of migraine (ICD-9-CM: 346.XX), tension-type headache (ICD-9-CM: 307.81), and headache (ICD-9-CM: 784.0X and 339.XX).[20]

Suicidal Behaviors

Suicidal behaviors were characterized based on ICD-9-CM diagnosis codes and were used to identify hospitalizations with suicidal behaviors, including suicidal ideation (V62.84) and suicide and self-inflicted injury (E950.XX-E959.XX). We did not utilize datasets prior to 2007 as the diagnosis codes for suicidal ideation were not used prior to October 2005.

Psychiatric Disorders

ICD-9-CM diagnosis codes were used to diagnose psychiatric disorders, including non-psychotic depression (296.82, 301.12, 309.28, 296.2X, 296.3X, 300.4X, 309.0X, 309.1X, 311.XX), anxiety (300.00, 300.01, 300.02, 300.21, 300.22, 300.23, 300.29, 309.21, 300.81, 300.2X, 300.3X, 300.7X, 308.2X, 308.3X), posttraumatic stress disorder (PTSD; 309.81), alcohol or substance abuse (291.XX, 292.XX, 303.XX, 304.XX, 305.XX, 648.3X, 655.5X, 965.0X, V65.42), psychosis (295.XX, 296.XX, 297.XX, 298.XX), and any psychiatric disorder from NIS hospitalization records[21] (Supporting Information Table 1).

Statistical Analysis

Our primary unit of analysis was per hospitalization. Repeated hospital discharges are not linked in the NIS database; therefore, a patient who was admitted to the hospital multiple times in one year would be counted each time as a separate hospitalization. In all analyses, we used discharge-level sampling weights based on the sampling scheme and provided by the datasets to report national estimates from all USA community hospitals. Adjusted discharge-level sampling weights were applied to each year and the estimates for each year were then merged. We compared the distributions of sociodemographic, psychiatric, and hospital characteristics between hospitalizations with and without migraine by performing Wald chi-square and t-tests. We calculated odds ratios (ORs) and 95% confidence intervals (CIs) using multivariable logistic regressions. We also calculated adjusted ORs and 95% CIs with a priori confounders including age (continuous), race/ethnicity, median household income quartiles for patient ZIP code, hospital region (Northeast, Midwest, South, West), hospital location (rural or urban), year, and age-adjusted Charlson Comorbidity Index, which assigns disease weights. Diseases with a weight of one include myocardial infarction, congestive heart failure, peripheral vascular disease, cerebrovascular disease, dementia, chronic pulmonary disease, ulcer disease, mild liver disease, and diabetes, and diseases with a weight of two include diabetes with end organ damage, any tumor, leukemia, and lymphoma. Moderate or severe liver disease has a weight of three, and metastatic solid tumor and AIDS have a weight of six. All variables had < 5% missing data except for race/ethnicity. We created missing indicator variables to address missing data for race/ethnicity and median household income quartiles. Total hospitalization charges were adjusted for inflation to reflect 2012 US dollars.[22] An age-unadjusted Charlson Comorbidity Index was calculated by adding separate disease weights, with a Charlson Comorbidity Index score of 0 as the lowest risk attributable to comorbid disease. For calculating an age-adjusted Charlson Comorbid Index, a patient aged ≤ 40 years old is assumed to have the lowest risk of comorbid death attributable to age and each decade of age over 40 adds 1 point to risk (eg, 50–59 years, 1 point; 60–69 years, 2 points; 70–79 years, 3 points). These points for age are added to the score from the age-adjusted Charlson Comorbidity Index (eg, 0, 1, 2, 3, etc.).[23–25] Prior studies have reported migraine is modified by age[26] and sex.[27] Given this, we repeated the primary analysis stratifying by age categories (<50 vs ≥50 years old) and sex (male vs female; Supporting Information Table 3). Given previous studies have shown psychiatric comorbidities modify the association between migraine and suicidal behaviors,[3–7] we repeated the analyses after stratifying participants according to depression, anxiety, or PTSD, respectively. We performed analyses for suicidal ideation and suicide and self-inflicted injury separately and the associations of migraine with suicidal ideation/suicide and self-inflicted injury were similar (data not shown).

All analyses were conducted using SAS 9.4 (SAS Institute, Cary, NC, USA) and SAS-callable SUDAAN software (version 11.0.1, RTI International, Research Triangle, NC, USA). Statistical significance was set at a two-sided P < .05. Computations were run on the Odyssey cluster supported by the Faculty of Arts & Sciences Division of Science, Research Computing Group at Harvard University.