Abstract
Objective. To examine risk factors for drug overdose by sex reflecting differing patterns of opioid and other drug use.
Design. National privately insured cohort.
Subjects. 206,869 subjects filling ≥2 opioid prescriptions from January 2009 through July 2012.
Methods. Sex-specific prediction models for future drug overdose developed and validated using variables measured within 6 months after starting opioids: demographics, substance use, comorbidities, opioid dose, and psychoactive drugs. Logistic regression and split-sample validation were used.
Results. Area under the receiver operating curves (AUCs) for both sex-specific risk models (0.80) were higher (P < 0.001) than for daily opioid dose alone. Risk factors for drug overdose were similar by sex but effects differed. For both sexes, substance use was the strongest predictor but the adjusted odds ratio (AOR) [95% CI] was 5.95 [4.33, 8.06] for women vs. 4.69 [3.24, 6.68] for men. AORs for daily opioid dose rose monotonically in men to 2.42 [1.76, 3.28] for high vs. low dose but were non-monotonic in women with 1.79 [1.35, 2.35] for high dose. AOR for 1–60 days of antidepressants vs. none was significant only in men (1.98 [1.32, 2.9]). AOR for benzodiazepine use was higher in men than women (2.75 vs 2.35, respectively). Zolpidem use was significant only in women. AUCs for sex-specific models were lower for the opposite sex and significantly lower for the men’s model in the women’s derivation dataset.
Conclusions. These models reveal similar risk factors by sex for drug overdose in opioid users but significant differences in effects that, if validated in other cohorts, may inform differing risk management strategies.
Keywords: Chronic Pain, Drug Overdose, Sex Characteristics, Narcotics, Polypharmacy, Benzodiazepines
Introduction
Despite significant changes in public policy and clinical practice, mortality from drug overdose is the leading cause of death from injury in the United States, with opioids contributing to the majority of these events [1–2]. The risk of death from drug overdose is greater for men than women but the gap between the sexes diminished between 1999 and 2010 due to a 415% increase in deaths for women and 265% for men [3]. Most drug overdose deaths have been due to prescription drug use and, among these, multiple types of drugs were involved: opioids (75%), benzodiazepines (29%), antidepressants (18%), and antiepileptic/antiparkinsonism drugs (8%) [4]. Among opioid-overdose deaths, other drugs have been commonly involved including benzodiazepines (30%) followed by antidepressants (13%) [4]. In 2010 alone, women made nearly 1 million visits to emergency departments for drug abuse and misuse [3]. Other studies have found that women are more likely to be prescribed opioids than men and to continue these drugs long term [5]. A recent Canadian study showed that men had a higher risk for escalation to high-dose opioid therapy than women and were twice as likely to die of opioid-related causes [6]. These data testify to the significant impact of drug abuse and related overdose for women as well as men.
One of the priorities identified by the U.S. Department of Health and Human Services’ Prescription Drug Abuse Subcommittee is the development of tools to reduce the risk of abuse and overdose that can be integrated into clinical practice [7]. Several tools have been developed to assess risk of aberrant behaviors such as the Opioid Risk Tool and SOAPP-R [8–9]. However, these models focus on risk of misuse and require patients to self-report. Additionally, performance of these models in women versus men has had little attention. A longitudinal study of 275 men and 335 women with chronic pain and prescribed opioids for pain showed that women were at greater risk of opioid misuse which was attributed to emotional issues and affective distress while men appeared to misuse opioids because of legal and behavioral problems [10]. In a study of 120 men and 42 women with chronic pain and diagnosed with opioid abuse, aberrant behaviors appeared to more prevalent in men while physical and psychological effects of pain were more common in women [11].
We aimed to identify and compare risk factors for drug overdose in women and men using predictors measured within 6 months after the first opioid prescription. To accomplish this objective, we used a national database of claims for over 200,000 privately insured persons filling two or more prescriptions for Schedule II or III opioids. Predictive models for drug overdose were developed and validated in women and men, separately, in an analysis that considered demographics, opioid therapy, use of psychoactive drugs, prior diagnosis of substance use disorder, and diagnosis of pain-related conditions within 6 months after the first opioid prescription. We hypothesized that the sex-specific risk prediction models would: 1) offer greater discrimination than average daily opioid dose, which is a standard, readily available metric of drug overdose risk [12–13]; and 2) perform less well when used for the opposite sex. Our goal was to develop models that could be applied to individual- or population-based data by policymakers, administrators, and clinicians to evaluate patterns of risk for drug overdose in patients with non-cancer pain who are treated with opioids.
Methods
Study Sample
This study cohort included 206,869 privately insured enrollees from the Aetna Health Maintenance Program, which provides comprehensive, full-service care to approximately 2.1 million people nationally. The details of the derivation of the study sample have been described in previous publications on a cohort of enrollees in the Aetna program [14,15]. Briefly, subjects aged 18 to 64 years with non-cancer pain who filled at least two Schedule II or III prescriptions for non-injectable opioid analgesics were identified from enrollment, claims for services, and prescriptions databases from January 2009 through July 2012. Eligible subjects were continuously enrolled at least 12 months in the Aetna program and had claims for service utilization at least 6 months before the first drug overdose event (Appendix A: ICD-9-CM codes for diagnosis of drug overdose). Subjects were excluded due to: incomplete opioid prescription data such as missing days’ supply, missing diagnostic data, or a cancer diagnosis other than basal cell cancer. This study was approved by the Institutional Review Board at the University of Texas Health Science Center at San Antonio.
Dependent Variable
The outcome is a drug overdose diagnosed from an inpatient or outpatient clinical encounter using discharge diagnoses after the first filled opioid analgesic prescription until the end of the study timeframe (3.5 years). We could not restrict the analysis to opioid-related overdose because the specific drugs contributing to an overdose were frequently not identified in this database.
Independent Variables
Built upon our previous findings on risk predictors for overdose using this database [14,15], we selected demographic, clinical, and treatment variables that we hypothesize may affect the risk of drug overdose. Data were collected on age as of July 2012, sex, and U.S. region in four categories defined by the Centers for Disease Control and Prevention. Clinical conditions and medications were measured between the first filled opioid analgesic prescription and the following 6 months. For all patients, data were available for at least 6 months prior to the overdose event. Pain-related conditions were identified from diagnosis codes for inpatient and outpatient encounters, including: back pain, large joint arthritis/other musculoskeletal disorders, neuropathic pain, chronic pain unspecified, and chronic headache. Mental health/substance use disorders were also identified as follows: anxiety or posttraumatic stress disorder (PTSD), depression, psychosis, drug use, and alcohol abuse. The ICD-9-CM codes used to identify clinical conditions are available upon request.
As reported previously, total morphine equivalent dose was calculated from all filled opioid prescriptions over 6 months as well as average daily dose over this timeframe [14,15]. As in other studies, average daily opioid dose was examined in four categories: <20, 20–49, 50–99, and ≥100 mg [12,14–17]; and total opioid dose was examined in four categories: ≤190, 191–450, 451–1830, and >1830 mg [14] [16]. We created a three-level opioid dose factor (High if daily dose ≥ 100 mg; Medium if daily dose 50–99 mg and total dose > 1830 mg; Low for all others). These categories are based on research showing that, compared with no opioid therapy, the hazard for drug overdose was highest for an average daily opioid dose ≥100 mg of any duration or 50–99 mg with a high total dose (>1830 mg), the latter being roughly equivalent to at least 4–6 weeks of opioid therapy within a 6-month period [14].
Duration of filled prescriptions for psychoactive medications was also measured within 6 months of starting opioid therapy and categorized as in other studies as: benzodiazepines (0, 1–30, 31–90, 91–180 days), zolpidem (0, 1–30, 31–90, 91–180 days), and antidepressants (0, 1–60, 61–180 days) [14,16]. Because 6–8 weeks can elapse before clinical benefit on depressive symptoms from antidepressant therapy is apparent, these categories differed slightly. All variables were measured before a drug overdose event including data prior to initiation of opioids if the event was <6 months after initiating opioid therapy.
Statistical Analyses
Differences in subjects’ characteristics by sex were tested using chi-square test for categorical variables and two-sample t-test with an unequal variance assumption for continuous variables. The incidence of drug overdose was calculated separately for women and men and sex differences were compared using the chi-square test and stratified by demographic, clinical conditions, and filled prescriptions variables.
For each sex, logistic regression models with drug overdose as the outcome were developed with a randomly selected derivation dataset comprised of half of the study cohort. The remaining half of the sample was used for validation. Subjects’ characteristics in the derivation and validation datasets were statistically equivalent. In the derivation sample for each sex, the full model predicting drug overdose included all study variables (i.e., demographics, comorbidities, opioids, other risky drugs, and length of follow up). A backward model selection procedure was employed. The final sex-specific risk prediction models included all clinically important and/or statistically significant variables. Subcategories for medications were collapsed when they had no significant difference and different subcategories were used, if needed, due to sex differences (i.e., duration of benzodiazepine use, duration of zolpidem use). The contribution of each variable to the final model’s explanatory power was examined using Pseudo R2. In principle, the more important a variable is for the model, the greater the decrease in the Pseudo R2 when that variable is removed from the model.
In the derivation sample for each sex, the area under the receiver operating characteristic (ROC) curve (AUC) was computed to measure the overall diagnostic performance for three models: 1) the final sex-specific risk prediction model, 2) a model with a 4-level daily opioid dose (<20, 20–49, 50–99, and ≥100 mg) as the only predictor, and 3) a model with a binary daily opioid dose (<100 mg vs. ≥100 mg). Differences in the AUCs for these three risk prediction models were tested using the nonparametric U-statistic [18,19]. The final model for each sex was also validated in the corresponding validation dataset, and AUCs were compared using the approach described above.
Because the final risk prediction models differed by sex, a secondary analysis examined the sex-specific risk prediction model in the opposite sex using both the derivation and validation datasets. The difference in AUC between women’s model and men’s model was examined in both the derivation sample and the validation sample. All statistical tests were performed with a two-sided significance level of 0.05 and analyses conducted using R (Version 3.1.2, The R Foundation for Statistical Computing). The funding source had no role in the study.
Results
Among 206,869 persons filling multiple opioid prescriptions, 57% were women and all patient characteristics, including demographics, comorbidities, and concurrent filled prescriptions, differed significantly between women and men (all P < 0.01). In particular, women were younger and more likely than men to be diagnosed with headache, anxiety or posttraumatic stress disorder (PTSD), and depression or psychotic disorder. Women received a lower opioid dose but were more likely to take antidepressants, benzodiazepines, or zolpidem than men (Table 1).
Table 1.
Patient characteristics and incidence of overdose
Characteristics1 | Women |
Men |
Total |
||||
---|---|---|---|---|---|---|---|
N (Column %) | Drug overdose2 (%) | N (Column %) | Drug overdose2 (%) | N (Column %) | Drug overdose2 (%) | P-value3 | |
All subjects | 117,472 | 0.76 | 89,397 | 0.56 | 206,869 | 0.67 | <0.001 |
Demographics | |||||||
Age (year) | |||||||
18–34 | 29765 (25.34) | 0.76 | 20388 (22.81) | 0.87 | 50153 (24.24) | 0.81 | 0.19 |
35–44 | 27902 (23.75) | 0.79 | 19522 (21.84) | 0.45 | 47424 (22.92) | 0.65 | <0.001 |
45–54 yr | 34277 (29.18) | 0.79 | 27514 (30.78) | 0.51 | 61791 (29.87) | 0.67 | <0.001 |
55–64 yr | 25528 (21.73) | 0.66 | 21973 (24.58) | 0.43 | 47501 (22.96) | 0.55 | 0.001 |
U.S. region | |||||||
Midwest | 6701 (5.7) | 0.81 | 5327 (5.96) | 0.45 | 12028 (5.81) | 0.65 | 0.016 |
Northeast | 32468 (27.64) | 0.79 | 28099 (31.43) | 0.59 | 60567 (29.28) | 0.70 | 0.005 |
South | 56429 (48.04) | 0.75 | 40643 (45.46) | 0.59 | 97072 (46.92) | 0.68 | 0.003 |
West | 21874 (18.62) | 0.72 | 15328 (17.15) | 0.45 | 37202 (17.98) | 0.61 | 0.001 |
Clinical Conditions | |||||||
Pain conditions | |||||||
Back pain-related | 29842 (25.4) | 1.13 | 23220 (25.97) | 0.82 | 53062 (25.65) | 0.99 | <0.001 |
Musculoskeletal conditions | 38004 (32.35) | 0.94 | 29987 (33.54) | 0.62 | 67991 (32.87) | 0.80 | <0.001 |
Neuropathy | 409 (0.35) | 1.71 | 410 (0.46) | 2.20 | 819 (0.4) | 1.95 | 0.62 |
Chronic pain | 3936 (3.35) | 2.59 | 2797 (3.13) | 2.15 | 6733 (3.25) | 2.41 | 0.239 |
Headache | 6349 (5.4) | 1.48 | 1414 (1.58) | 1.56 | 7763 (3.75) | 1.49 | 0.833 |
Anxiety or PTSD | 10455 (8.9) | 2.06 | 5077 (5.68) | 1.77 | 15532 (7.51) | 1.96 | 0.232 |
Depression or psychotic disorder | 11641 (9.91) | 3.09 | 4703 (5.26) | 2.55 | 16344 (7.9) | 2.93 | 0.064 |
Alcohol abuse or substance use disorder | 1512 (1.29) | 8.86 | 2092 (2.34) | 4.35 | 3604 (1.74) | 6.24 | <0.001 |
Filled Prescriptions | |||||||
Daily opioid dose (mg) | |||||||
<20 | 12183 (10.37) | 0.48 | 8050 (9) | 0.34 | 20233 (9.78) | 0.43 | 0.11 |
≥20 to < 50 | 68853 (58.61) | 0.69 | 49721 (55.62) | 0.43 | 118574 (57.32) | 0.58 | <0.001 |
≥50 to < 100 | 28744 (24.47) | 0.76 | 23577 (26.37) | 0.51 | 52321 (25.29) | 0.65 | <0.001 |
≥100 | 7692 (6.55) | 1.77 | 8049 (9) | 1.69 | 15741 (7.61) | 1.73 | 0.71 |
Total opioid dose (mg) | |||||||
≤190 | 29710 (25.29) | 0.51 | 18881 (21.12) | 0.48 | 48591 (23.49) | 0.5) | 0.56 |
>190 to ≤ 450 | 40058 (34.1) | 0.45 | 27735 (31.02) | 0.35 | 67793 (32.77) | 0.41 | 0.038 |
>450 to ≤ 1830 | 29467 (25.08) | 0.71 | 25016 (27.98) | 0.34 | 54483 (26.34) | 0.54 | <0.001 |
>1830 | 18237 (15.52) | 1.90 | 17765 (19.87) | 1.28 | 36002 (17.4) | 1.60 | <0.001 |
Opioid dose risk category4 | |||||||
Low | 104642 (89.08) | 0.62 | 76288 (85.34) | 0.4 | 180930 (87.46) | 0.53 | <0.001 |
Medium | 5138 (4.37) | 1.91 | 5060 (5.66) | 1.11 | 10198 (4.93) | 1.51 | 0.001 |
High | 7692 (6.55) | 1.77 | 8049 (9) | 1.69 | 15741 (7.61) | 1.73 | 0.71 |
Antidepressants (days) | |||||||
0 | 89214 (75.94) | 0.49 | 77841 (87.07) | 0.42 | 167055 (80.75) | 0.46 | 0.029 |
1–60 | 8558 (7.29) | 1.51 | 4062 (4.54) | 1.82 | 12620 (6.1) | 1.61 | 0.19 |
61–180 | 19700 (16.77) | 1.63 | 7494 (8.38) | 1.31 | 27194 (13.15) | 1.54 | 0.054 |
Benzodiazepines (days) | |||||||
0 | 90641 (77.16) | 0.45 | 75426 (84.37) | 0.37 | 166067 (80.28) | 0.41 | 0.013 |
1–30 | 11844 (10.08) | 0.98 | 5875 (6.57) | 0.94 | 17719 (8.57) | 0.97 | 0.78 |
31–90 | 6147 (5.23) | 2.15 | 3167 (3.54) | 1.55 | 9314 (4.5) | 1.94 | 0.047 |
91–180 | 8840 (7.53) | 2.66 | 4929 (5.51) | 2.35 | 13769 (6.66) | 2.55 | 0.28 |
Zolpidem (days) | |||||||
0 | 106009 (90.24) | 0.65 | 82673 (92.48) | 0.51 | 188682 (91.21) | 0.58 | <0.001 |
1–30 | 3656 (3.11) | 1.23 | 2340 (2.62) | 1.15 | 5996 (2.9) | 1.20 | 0.79 |
31–90 | 2788 (2.37) | 2.15 | 1673 (1.87) | 1.37 | 4461 (2.16) | 1.86 | 0.063 |
91–180 | 5019 (4.27) | 1.97 | 2711 (3.03) | 1.07 | 7730 (3.74) | 1.66 | 0.003 |
The difference between women and men are statistically significant for all patients’ characteristics (all P < 0.01).
% of overdose was reported for each row.
P values were calculated for comparing the incidence of rate between women and men for each row using Chi2 test.
Opioid dose risk category was high if daily opioid dose ≥100 mg; Medium if daily opioid dose 50–99 mg and total opioid dose >1830 mg; Low for the remaining subjects.
All study subjects were followed for an average of 1.6 years after the first opioid prescription (SD = 0.88, range = 0.5–3.6). Over the study timeframe, 1,386 (.67%) subjects were diagnosed with a drug overdose but about half of these events occurred within 1 year after the first opioid prescription, with 33% occurring from years 1 to 2, 13% from years 2 to 3, and <4% from years 3 and on. Overall, the incidence of drug overdose was significantly greater for women than men (0.76 vs. 0.56%, Table 1). However, the median of the average morphine equivalent daily dose (mg) was significantly higher for men than women (median [first quartile, third quartile]: 40 [29.2, 59.6] versus 37.5 [27, 53.6] respectively, P < 0.001) as was the total dose (450 [225, 1275] versus 365.5 [190, 900], respectively, P < 0.001).
Women with the following characteristics were significantly more likely than men to have a drug overdose: older age (≥35 year), back pain or musculoskeletal conditions, alcohol abuse or substance use disorder, average daily opioid dose of 20–99 mg, total dose >190 mg within 6 months, and treatment with other risky drugs (Table 1).
Tables 2 and 3 summarize the final variables in the sex-specific risk prediction models using the derivation dataset. The adjusted odds ratios (AOR) for drug overdose for women and men were similar for age (i.e., 2% lower odds per one year of age) and depression or psychotic disorder (i.e., approximately three-fold higher odds versus neither condition). Alcohol abuse or other substance use disorder was associated with higher odds of overdose for women (AOR = 5.95 [4.33, 8.06]) than men (AOR = 4.69 [3.24, 6.68]). The association of opioid dose with drug overdose was stronger and monotonic in men (AOR = 2.02 [1.29, 3.03] for medium vs. low; AOR = 2.42 [1.76, 3.28] for high vs. low) but non-monotonic in women (AOR = 1.96 [1.41, 2.66] for medium vs. low. AOR = 1.79 [1.35, 2.35] for high vs. low).
Table 2.
Overdose risk model for women
Characteristics | Full model | Reduced model | ||
---|---|---|---|---|
OR [95% CI] | P-value | OR [95% CI] | P-value | |
Demographics | ||||
Age (year) | Per 1-yr increase | |||
18–34 | 1 | 0.98 [0.98, 0.99] | <0.001 | |
35–44 | 0.87 [0.66, 1.14] | 0.32 | ||
45–54 | 0.7 [0.54, 0.92] | 0.01 | ||
55–64 | 0.63 [0.46, 0.85] | 0.003 | ||
U.S. Region | ||||
Midwest | 1 | |||
Northeast | 0.92 [0.6, 1.46] | 0.7 | ||
South | 0.93 [0.62, 1.45] | 0.73 | ||
West | 1 [0.64, 1.62] | 1 | ||
Clinical Conditions | ||||
Pain conditions | ||||
Back pain-related | 1.35 [1.09, 1.68] | 0.01 | ||
Musculoskeletal conditions | 0.87 [0.7, 1.08] | 0.2 | ||
Neuropathy | 1.8 [0.53, 4.55] | 0.27 | ||
Chronic pain | 1.33 [0.93, 1.85] | 0.11 | ||
Headache | 1.22 [0.87, 1.67] | 0.23 | ||
Anxiety or PTSD | 1.26 [0.97, 1.61] | 0.08 | ||
Depression or psychotic disorder | 2.77 [2.18, 3.51] | <0.001 | 3.04 [2.41, 3.82] | <0.001 |
Alcohol abuse or substance use disorder | 5.55 [4.01, 7.55] | <0.001 | 5.95 [4.33, 8.06] | <0.001 |
Filled Prescriptions | ||||
Opioid dose risk category | ||||
Low | 1 | 1 | ||
Medium | 1.72 [1.23, 2.36] | 0.001 | 1.96 [1.41, 2.66] | <0.001 |
High | 1.66 [1.24, 2.19] | <0.001 | 1.79 [1.35, 2.35] | <0.001 |
Antidepressants (days) | ||||
0d | 1 | 1 | ||
1–60d | 1.25 [0.9, 1.69] | 0.17 | 1.36 [0.99, 1.84] | 0.05 |
61–180d | 1.33 [1.04, 1.7] | 0.02 | 1.41 [1.11, 1.8] | 0.01 |
Benzodiazepines (days) | ||||
0 | 1 | 1 | ||
1–30 | 1.31 [0.95, 1.78] | 0.09 | ||
31–90 | 2.38 [1.73, 3.21] | <0.001 | 2.35 [1.88, 2.93] | <0.001 |
91–180 | 2.19 [1.66, 2.87] | <0.001 | ||
Zolpidem (days) | ||||
0 | 1 | 1 | ||
1–30 | 1.33 [0.84, 2.01] | 0.19 | ||
31–90 | 1.52 [0.96, 2.29] | 0.06 | ||
91–180 | 1.77 [1.27, 2.41] | <0.001 | 1.74 [1.26, 2.35] | 0.001 |
Length of follow-up (yr) | 1.58 [1.43, 1.75] | <0.001 | 1.57 [1.42, 1.74] | <0.001 |
Table 3.
Overdose risk model for men
Characteristics | Full model | Reduced model | ||
---|---|---|---|---|
OR [95% CI] | P-value | OR [95% CI] | P-value | |
Demographics | ||||
Age (year) | 1-yr increase | |||
18–34 | 1 | 0.98 [0.97, 0.99] | <0.001 | |
35–44 | 0.57 [0.39, 0.83] | 0.003 | ||
45–54 | 0.55 [0.39, 0.77] | <0.001 | ||
55–64 | 0.45 [0.3, 0.67] | <0.001 | ||
U.S. Region | ||||
Midwest | 1 | |||
Northeast | 1.12 [0.64, 2.13] | 0.71 | ||
South | 0.91 [0.52, 1.71] | 0.75 | ||
West | 0.74 [0.39, 1.5] | 0.38 | ||
Clinical Conditions | ||||
Pain conditions | ||||
Back pain-related | 1.4 [1.05, 1.86] | 0.02 | ||
Musculoskeletal conditions | 0.75 [0.56, 1] | 0.05 | ||
Neuropathy | 4.04 [1.36, 9.52] | 0.004 | ||
Chronic pain | 1.7 [1.07, 2.61] | 0.02 | ||
Headache | 1.45 [0.67, 2.77] | 0.30 | ||
Anxiety or PTSD | 1.02 [0.69, 1.48] | 0.91 | ||
Depression or psychotic disorder | 3.05 [2.13, 4.32] | <0.001 | 3.23 [2.27, 4.54] | <0.001 |
Alcohol abuse or substance use disorder | 4.62 [3.17, 6.61] | <0.001 | 4.69 [3.24, 6.68] | <0.001 |
Filled Prescriptions | ||||
Opioid dose risk category | ||||
Low | 1 | 1 | ||
Medium | 1.8 [1.15, 2.74] | 0.01 | 2.02 [1.29, 3.03] | 0.001 |
High | 2.14 [1.55, 2.93] | <0.001 | 2.42 [1.76, 3.28] | <0.001 |
Antidepressants (days) | ||||
0 | 1 | 1 | ||
1–60 | 1.98 [1.31, 2.91] | 0.001 | 1.98 [1.32, 2.9] | 0.001 |
61–180 | 1.13 [0.75, 1.66] | 0.55 | 1.2 [0.8, 1.76] | 0.36 |
Benzodiazepines (days) | ||||
0 | 1 | 1 | ||
1–30 | 2.34 [1.55, 3.44] | <0.001 | 2.75 [2.07, 3.64] | <0.001 |
31–90 | 2.05 [1.23, 3.29] | 0.004 | ||
91–180 | 3.14 [2.15, 4.53] | <0.001 | ||
Zolpidem (days) | ||||
0 | 1 | |||
1–30 | 1.47 [0.79, 2.53] | 0.19 | ||
31–90 | 0.5 [0.15, 1.2] | 0.18 | ||
91–180 | 1.35 [0.78, 2.2] | 0.26 | ||
Length of follow-up (yr) | 1.37 [1.19, 1.57] | <0.001 | 1.38 [1.2, 1.58] | <0.001 |
Antidepressant use for 1–60 days by men was associated with two-folder higher adjusted odds of drug overdose but not significant for longer (90–180 days) treatment while the AORs for antidepressant therapy were slightly but significantly increased in women regardless of duration. Similarly, benzodiazepine use for 1–30 days had nearly three-fold higher AOR for drug overdose in men but only slightly increased in women. Treatment with zolpidem for at least 90 days significantly increased odds of drug overdose by 74% in women but had no significant effect in men.
In the derivation sample for women, daily opioid dose models had low Pseudo R2 values: 0.8% for four-level categories and 1.6% for two-level categories (Appendix B). In contrast, the sex-specific full model in women had a Pseudo R2 of 11%. In women, alcohol abuse or substance use disorder contributed the most to predicting overdose, reflecting greatest drop in Pseudo R2 after removal from the model while antidepressant therapy contributed the least (Appendix B). In men, the daily opioid dose models performed somewhat better (i.e., Pseudo R2 = 2.4% for four-level categories; 2.7% for two-level categories) but the sex-specific model performed the best (Pseudo R2 = 12.2%) (Appendix B). A diagnosis of alcohol abuse or substance use disorder and benzodiazepine therapy contributed the most to the risk model in men.
As shown by ROC curves, the sex-specific models performed significantly better than the opioid dose categories in both the derivation and validation datasets for both women and men (Figure 1). For example, in the derivation sample for women, the AUC was 0.8 for the sex-specific model versus 0.56 for the four-level daily dose model versus 0.54 for the two-level daily dose model (all P < 0.001).
Figure 1.
Diagnostic performance of sex-specific risk prediction models with comparison to other strategies.
Model 1 (purple): sex-specific final reduced risk prediction model (see Tables 2 and 3). Model 2 (blue): model with a four-level daily MED dose only (i.e., <20, 20–49, 50–99, and ≥100 mg). Model 3 (orange): model with a two-level daily MED dose only (i.e., daily MED≥100 mg vs. not).
Because the final risk prediction models differed by sex, a secondary analysis examined the performance of the sex-specific models in the opposite sex (Table 4). The AUC for sex-specific model applied to the opposite sex was lower but not significant. However, although it is difficult to find a significant improvement in the AUC [20–22], the women’s model performed significantly better than the men’s model in women’s derivation dataset (AUC 0.804 versus 0.785, respectively, P < 0.001).
Table 4.
Comparison of the final sex-specific models in different samples
Sex | Sample | Model | AUC | P-value |
---|---|---|---|---|
Women | Derivation | Women’s model | 0.804 | <0.001 |
Men’s model | 0.785 | |||
Validation | Women’s model | 0.801 | 0.85 | |
Men’s model | 0.800 | |||
Men | Derivation | Women’s model | 0.777 | 0.08 |
Men’s model | 0.791 | |||
Validation | Women’s model | 0.795 | 0.41 | |
Men’s model | 0.801 |
Discussion
In a large national cohort of privately insured adults with non-cancer pain who filled prescriptions for more potent opioid analgesics, women were significantly more likely than men to have a drug overdose over a 3.5-year timeframe despite their receiving a significantly lower mean opioid dose. In another privately insured cohort on opioids, the adjusted odds of high dose opioid therapy (>100 mg morphine equivalent dose per day) were 68% greater for men than women [23]. Yet nationally, women are more likely to be prescribed opioids than men (7.2% versus 6.3%, respectively) according to the 2010–2012 National Health and Nutrition Examination Survey (NHANES) [24]. These data paint a complicated picture of risk factors for drug overdose in women compared with men that requires targeted analyses of specific risk factors to guide management and risk reduction.
In both the derivation and validation cohort analyses, there were important similarities but significant differences in risk factors for drug overdose by sex. First, the overall goodness-of-fit (i.e., Pseudo R2) for the reduced model for both sexes was relatively low but much higher than for a commonly used metric, the average daily opioid dose. In regard to similarities, the adjusted odds of overdose decreased by 2% for each additional year of age for both sexes; this effect is similar in magnitude to that observed in other studies [25]. For both sexes, a prior diagnosis of alcohol abuse and/or substance use disorder contributed the most to the fit of the risk prediction models. A high risk of drug overdose death for persons with substance use disorders has been widely reported in both veteran and civilian populations [12,17]. However, the adjusted odds of drug overdose for alcohol abuse or substance use disorder were increased by nearly six-fold in women and only 4.7-fold in men versus persons with neither diagnosis. Benzodiazepine use was associated with a higher risk among men than women. Possibly reflecting this difference, treatment with benzodiazepines was the second most important variable in the reduced model for men but fourth for women. In a national cohort study from 2004 to 2009 of predominantly male veterans treated with opioids, benzodiazepines were associated with an adjusted hazard of 3.86 for drug overdose but the risk varied by dose. However, our study focused instead on duration of benzodiazepine therapy [26].
In women, depression or psychosis was the second most important factor in the reduced model for drug overdose, but the likelihood of overdose was similarly increased by about three-fold for both sexes versus persons with neither diagnosis. In both genders, opioid dose based on categories incorporating both daily and total dose were significantly associated with future risk of drug overdose [14]. Interestingly, high dose opioid therapy (>100 mg/day) had a stronger association for men than women but risks were similar for moderate dose (50–99 mg/day plus a total dose >1830).
All variables in the final models were the same for both sexes except zolpidem use, which was a significant risk factor for drug overdose in only women. Other studies have not compared outcomes for zolpidem use by sex but, nationally from 2009 to 2011, this drug contributed to more emergency department visits for adverse events (N = 10,210) than any other psychiatric medication [27]. Among over 2,000 drug-related fatalities in 2013, zolpidem was involved in more cases for women than men [28].
The ROC curves for the reduced model in the derivation and validation datasets for men and women showed that they performed equally well and much better than risk classification using opioid dose alone. Because the variables in the models for both sexes were largely the same, it was interesting to find that the model developed for women performed significantly better than the model developed for men when applied to women. The AUCs appear to be quite similar across the different models (range 0.777–0.804), but the AUC is a very robust statistic so it is difficult to observe a significant improvement [21,22]. For example, a study of multiple biomarkers for predicting the risk of cardiovascular events found that AUC increased from 0.795 to 0.816 from the addition of the five strongly predictive biomarkers to the standard risk factors [29].
Several limitations should be acknowledged. First, it is difficult to compare our results with those of other cohorts because some studies only examine deaths while others examine only opioid-related adverse events. We used a broad range of diagnoses to identify drug overdose because coded diagnoses may not distinguish whether opioids were directly contributing to the drug overdose and opioid-related overdose appears to be under-reported in this database (17% of overdose events).
By comparison, in an analysis of death from opioid overdose from the National Vital Statistics System from 2010, men had 55% greater odds than women but this differential decreased substantially over the prior decade [3]. However, the rate of intentional death from drug overdose was similar by sex and women and men were equally to visit the emergency department for opioid misuse or abuse. Other studies have found that women are more likely to be hospitalized for drug overdose [3]. Our data add to concerns about the major risk of opioids in alcohol or substance abusing women and men.
Another limitation of the study is our lack of patient-reported information from risk measures such as the Opioid Risk Tool and the SOAPP-R [8,9]. The AUC for the SOAPP-R in predicting risk of aberrant behaviors among 221 opioid-treated persons with chronic pain was 0.74 compared with 0.79 and 0.8 for our models predicting opioid overdose. But these and other studies do not specifically address differences by sex [30].
Third, although ours is a national cohort, it only involves a commercially insured population. Fourth, days’ supply data were used to compute the average daily opioid dose as in other studies [14–16], [31,32], but may not reflect the actual days that the patient was taking the medication. Fifth, our study could not distinguish prevalent opioid users from new users, but studies of drug overdose find that the risk for overdose continues to be increased even among persons who have been taking these drugs for years [33]. Sixth, we considered only predictors measured within 6 months after the first filled opioid prescription because it is important to distinguish risks early after starting therapy. Lastly, due to restricted availability of clinical variables for this dataset, we could not examine interactions with other prescribed drugs or clinical variables such as body mass index.
Clinical Significance
This study offers the first sex-specific models assessing risk for drug overdose after starting opioid therapy. The clinical implications are significant. Although all persons with a history of alcohol abuse or substance use disorder have significantly increased risk for drug overdose, this risk appears to be even greater for women. Even though women were more likely to fill prescriptions for multiple risky drugs than men, men had a greater risk of drug overdose from receiving multiple drugs including benzodiazepines and high dose opioids as well as short-term after initiating antidepressant therapy. On the other hand, zolpidem therapy was more risky for women than men.
Conclusions
Risk factors for drug overdose were similar by sex but effects differed. If validated in other cohorts, these sex-specific models may be useful for policymakers, administrators, and clinicians to modify management and offer greater support to persons at increased risk of serious adverse consequences of treatment for non-cancer pain.
Acknowledgments
The authors would like to thank Aetna, Inc. for the data and Benjamin Ehler for data cleaning.
Appendix A: ICD-9-CM Code for Diagnosis of Drug Overdose
ICD-9-CM code, 965.0, 965.00, 965.02, 965.09, 965.1, 965.4, 965.61, 965.69, 965.8, 965.9, 967.6, 967.8, 967.9, 969.4, 977.9, E850.1-E850.6, E850.8, E850.9, E852.8, E852.9, E853.2, E950.0 or E950.2
Appendix B: Variable Importance Based on Logistic Regressions
Gender | Model | Pseudo R2 | Decrease in R2* |
---|---|---|---|
Women | Logistic regression (reduced) | 0.1101 | |
− Alcohol abuse or substance use disorder | 0.0919 | 0.0183 | |
− Depression or psychotic disorder | 0.0938 | 0.0163 | |
− Length of follow-up | 0.0959 | 0.0142 | |
− Benzodiazepines | 0.0995 | 0.0106 | |
− Opioid dose risk category | 0.1049 | 0.0052 | |
− Age | 0.1071 | 0.0030 | |
− Zolpidem | 0.1080 | 0.0021 | |
− Antidepressants | 0.1084 | 0.0017 | |
− Four-level daily MED only | 0.0083 | ||
− Two-level daily MED only | 0.0164 | ||
Men | Logistic regression (reduced) | 0.1219 | |
− Alcohol abuse or substance use disorder | 0.1032 | 0.0187 | |
− Benzodiazepines | 0.1063 | 0.0156 | |
− Depression or psychotic disorder | 0.1085 | 0.0134 | |
− Opioid dose risk category | 0.1109 | 0.0110 | |
− Length of follow-up | 0.1153 | 0.0066 | |
− Age | 0.1161 | 0.0058 | |
− Antidepressants | 0.1184 | 0.0035 | |
− four-level daily MED only | 0.0237 | ||
− two-level daily MED only | 0.0269 |
References
- 1. McCarthy M. Drug overdose has become leading cause of death from injury in US. BMJ 2015;350: h3328.. [DOI] [PubMed] [Google Scholar]
- 2.Warner M, Chen LH, Makuc DM, Anderson RN, Miniño AM. Drug poisoning deaths in the United States, 1980-2008. NCHS data brief, no 81. Hyattsville, MD: National Center for Health Statistics. 2011. Available at: http://www.cdc.gov/nchs/data/databriefs/db81.htm (cited Sept 21 2015). [PubMed]
- 3. Centers for Disease Control and Prevention. Vital signs: Overdoses of prescription opioid pain relievers and other drugs among women — United States, 1999–2010. MMWR 2013;62:537–42. [PMC free article] [PubMed] [Google Scholar]
- 4. Jones CM, Mack KA, Paulozzi LJ. Pharmaceutical overdose deaths, United States, 2010. JAMA 2013;309:657–9. [DOI] [PubMed] [Google Scholar]
- 5. Campbell CI, Weisner C, Leresche L, et al. Age and gender trends in long-term opioid analgesic use for noncancer pain. Am J Public Health 2010;100: 2541–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Kaplovitch E, Gomes T, Camacho X, et al. Sex differences in dose escalation and overdose death during chronic opioid therapy: A population-based cohort study. PLoS One 2015;10(8):e0134550.. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. U.S. Department of Health and Human Services. Addressing prescription drug abuse in the United States: Current activities and future opportunities. 2015 Sept. Available at: http://www.cdc.gov/drugoverdose/pdf/hhs_prescription_drug_abuse_report_09.2013.pdf (accessed Sept 13 2015).
- 8. Webster LR, Webster RM. Predicting aberrant behaviors in opioid-treated patients: Preliminary validation of the Opioid Risk Tool. Pain Med 2005;6:432–42. [DOI] [PubMed] [Google Scholar]
- 9. Butler SF, Budman SH, Fernandez KC, Fanciullo GJ, Jamison RN. Cross-validation of a screener to predict opioid misuse in chronic pain patients (SOAPP-R). J Addict Med 2009;3:66–73. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Jamison RN, Butler SF, Budman SH, Edwards RR, Wasan AD. Gender differences in risk factors for aberrant prescription opioid use. J Pain 2010;11(4): 312–20. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Manubay J, Davidson J, Vosburg S, et al. Sex differences among opioid-abusing patients with chronic pain in a clinical trial. J Addict Med 2015;9(1):46–52. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Bohnert AS, Valenstein M, Bair MJ, et al. Association between opioid prescribing patterns and opioid overdose-related deaths. JAMA 2011;305:1315–21. [DOI] [PubMed] [Google Scholar]
- 13. Berland D, Rodgers P. Rational use of opioids for management of chronic nonterminal pain. Am Fam Physician 2012;86:252–8. [PubMed] [Google Scholar]
- 14. Liang Y, Turner BJ. Assessing risk for drug overdose in a national cohort: Role for both daily and total opioid dose? J Pain 2015;16:318–25. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Turner BJ, Liang Y. Drug overdose in a retrospective cohort with non-cancer pain treated with opioids, antidepressants, and/or sedative-hypnotics: Interactions with mental health disorders. J Gen Intern Med 2015;30:1081–96. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Liang Y, Turner BJ. National cohort study of opioid analgesic dose and risk of future hospitalization. J Hosp Med 2015;10:42–431. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Dunn KM, Saunders KW, Rutter CM, et al. Opioid prescriptions for chronic pain and overdose: A cohort study. Ann Intern Med 2010;152:85–92. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Hanley JA, McNeil BJ. A method of comparing the areas under receiver operating characteristic curves derived from the same cases. Radiology 1983;148:839–43. [DOI] [PubMed] [Google Scholar]
- 19. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics 1988;44:837–45. [PubMed] [Google Scholar]
- 20. Liang Y, Ankerst DP, Sanchez M, Leach RJ, Thompson IM. Body mass index adjusted prostate-specific antigen and its application for prostate cancer screening. Urology 2010;76:1268.e1–6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Ware JH. The limitations of risk factors as prognostic tools. N Engl J Med 2006;355(25):2615–7. [DOI] [PubMed] [Google Scholar]
- 22. Baker SG, Schuit E, Steyerberg EW, et al. How to interpret a small increase in AUC with an additional risk prediction marker: Decision analysis comes through. Stat Med 2014;33(22):3946–59. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Kobus AM, Smith DH, Morasco BJ, et al. Correlates of higher-dose opioid medication use for low back pain in primary care. J Pain 2012;13:1131–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Frenk SM, Porter KS, Paulozzi LJ. Prescription opioid analgesic use among adults: United States, 1999–2012. NCHS data brief, no 189. Hyattsville, MD: National Center for Health Statistics; 2015. Available at: http://www.cdc.gov/nchs/data/databriefs/db189.htm (accessed Sept 21 2015). [PubMed]
- 25. Paulozzi LJ, Kilbourne EM, Shah NG, et al. A history of being prescribed controlled substances and risk of drug overdose death. Pain Med 2012;13:87–95. [DOI] [PubMed] [Google Scholar]
- 26. Park TW, Saitz R, Ganoczy D, Ilgen MA, Bohnert ASB. Benzodiazepine prescribing patterns and deaths from drug overdose among US veterans receiving opioid analgesics: Case-cohort study. BMJ 2015;350:h2698.. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Hampton LM, Daubresse M, Chang HY, Alexander GC, Budnitz DS. Emergency department visits by adults for psychiatric medication adverse events. JAMA Psychiatry 2014;71:1006–14. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Mowry JB, Spyker DA, Cantilena LR, Jr, McMillan N, Ford M. 2013 Annual report of the American Association of Poison Control Centers' National Poison Data System (NPDS): 31st annual report. Clin Toxicol 2014;52:1032–283. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Wang TJ, Gona P, Larson MG, et al. Multiple biomarkers for the prediction of first major cardiovascular events and death. N Engl J Med 2006;355(25):2631–9. [DOI] [PubMed] [Google Scholar]
- 30. Jones T, Moore T, Levy JL, et al. A comparison of various risk screening methods in predicting discharge from opioid treatment. Clin J Pain 2012;28:93–100. [DOI] [PubMed] [Google Scholar]
- 31. Braden JB, Russo J, Fan MY, et al. Emergency department visits among recipients of chronic opioid therapy. Arch Intern Med 2010;170:1425–32. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32. Edlund MJ, Martin BC, Fan MY, et al. An analysis of heavy utilizers of opioids for chronic noncancer pain in the TROUP study. J Pain Symptom Manage 2010;40:279–89. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33. Darke S. Opioid overdose and the power of old myths: What we thought we knew, what we do know and why it matters. Drug Alcohol Rev 2014;33:109–14. [DOI] [PubMed] [Google Scholar]