To quantify how adherence mismeasurement affects the estimated impact of adherence on inpatient costs among patients with serious mental illness (SMI).

Proportion of days covered (PDC) is a common claims-based measure of medication adherence. Because PDC does not measure medication ingestion, however, it may inaccurately measure adherence. We derived a formula to correct the bias that occurs in adherence-utilization studies resulting from errors in claims-based measures of adherence. 

We conducted a literature review to identify the correlation between gold-standard and claims-based adherence measures. We derived a bias-correction methodology to address claims-based medication adherence measurement error. We then applied this methodology to a case study of patients with SMI who initiated atypical antipsychotics in 2 large claims databases.

 Our literature review identified 6 studies of interest. The 4 most relevant ones measured correlations between 0.38 and 0.91. Our preferred estimate implies that the effect of adherence on inpatient spending estimated from claims data would understate the true effect by a factor of 5.3, if there were no other sources of bias. Although our procedure corrects for measurement error, such error also may amplify or mitigate other potential biases. For instance, if adherent patients are healthier than nonadherent ones, measurement error makes the resulting bias worse. On the other hand, if adherent patients are sicker, measurement error mitigates the other bias. 

Measurement error due to claims-based adherence measures is worth addressing, alongside other more widely emphasized sources of bias in inference. 

Using pharmacy claims data on patients with serious mental illness (SMI), this study demonstrates that increased medication adherence correlates with lower inpatient costs. A bias-correction formula was used to show that measurement error in claims-based adherence measures results in the effects of adherence being underestimated by a factor of 5.3. 

		Medication nonadherence is a burdensome problem in SMI, substantially contributing to healthcare costs. 

		Few studies address errors in adherence measurement, which can introduce bias when estimating the impact of adherence on inpatient costs. 

		Using a reliable estimate of the correlation between true adherence and claims-based adherence can identify the effects of measurement error on the estimated relationship between adherence and cost.

Nonadherence to medication is a prevalent and burdensome problem among patients with serious mental illness (SMI).

 Estimates of adherence to antipsychotics among patients with schizophrenia-spectrum disorders, for instance, range from 47% to 95%.

 The consequences of poor adherence include suboptimal health outcomes and higher avoidable healthcare costs.

 In patients with schizophrenia, medication nonadherence impedes recovery,

 and extends the length of in-hospital stays.

 Overall, hospitalizations due to medication nonadherence have been estimated to cost more than $100 billion annually in the United States,

 and hospitalization costs due to antipsychotic nonadherence specifically have been estimated at $1.5 billion annually.

CMS includes medication adherence as a rating measure when determining healthcare quality.

 A common method to indirectly assess adherence is the proportion of days covered (PDC),

 which typically uses prescription claims data and is calculated as the proportion of days in the measurement period, usually 1 year, for which the patient has medication on hand. The International Society for Pharmacoeconomics and Outcomes Research (ISPOR) has determined that PDC is one of the preferred methods to calculate medication adherence.

 PDC also is used to measure adherence to antipsychotic medications as part of the Healthcare Effectiveness Data and Information Set quality measures.

Although PDC is widely used, it suffers from 2 key shortcomings. First, PDC underestimates adherence when patients pay cash for medication or use other coverage options that fail to result in a recorded insurance claim. Second, PDC overstates adherence when patients purchase but do not take a given medication. New technologies, such as electronic pillboxes, smart caps, or ingestible sensors, may provide more accurate adherence measurements, but currently, payers and providers rarely use these technologies to monitor adherence. 

Neither the magnitude nor the direction of the bias associated with PDC adherence estimates are widely discussed or incorporated into adherence analyses. This study aimed to quantify how adherence mismeasurement affects the estimated impact of adherence on inpatient costs among patients with SMI. 

Our methodology relies on a 3-step approach to estimate the potential impact of measurement error on inpatient spending. We mathematically derived a formula for a bias-correction factor. As this factor depends principally on the link between true and measured adherence, we next conducted a review of the literature to identify studies that measured the relationship between a “gold standard” measure of adherence (eg, Medication Event Monitoring System [MEMS] caps, electronic pill counts) and adherence measured in claims data. Finally, we applied the bias-correction factor to a case study of patients with SMI who initiated therapy with an atypical antipsychotic. 

Consider the case where a researcher wants to measure the relationship between patient adherence and inpatient spending. One commonly used approach is an ordinary least squares (OLS) regression, such as the following: 

 represents inpatient spending for patient 

 represents patient adherence to atypical antipsychotics, and the vector W

' contains other patient covariates of interest. The coefficient on adherence, 

Mismeasurement in PDC biases the estimated effect of adherence on inpatient spending (ie, 

 Ì‚) toward 0. However, the true effect of adherence on inpatient costs can be derived if the relationship between measured adherence in claims data and true adherence is known. As shown in eAppendix 1 (

), one can correct for measurement error bias using the following formulation:

  is the true effect of adherence on inpatient spending after adjusting for adherence mismeasurement, 

 Ì‚is the OLS estimator from the regression in equation 1, and (1-

) is the correction factor for the mismeasurement in 

 is the R-squared value of linear regression of measured adherence (

 is the square of the correlation between measured and true adherence. 

The true effect of adherence is feasible to estimate, but problematic. First, one must assume that there cannot be any unobserved patient characteristics that affect both medication adherence and inpatient spending. For instance, patients with less severe forms of SMI may be more likely to be adherent to their medication and have lower inpatient costs.

 To address this, we derived the bias that would remain in the case where medication adherence was an endogenous variable in eAppendix 1. The formula we derived fully addresses the bias due to measurement error, even when adherence is endogenous, so that researchers will recover the same parameter that would have been estimated if they had had access to correctly measured adherence. To be clear, our approach does not simultaneously address the endogeneity itself; rather, it represents a full solution to the problem of measurement error in adherence due to the use of claims data. 

Second, one needs a reliable estimate of the correlation between true medication adherence and claims-based adherence. The following section describes our review of the literature that was used to identify this parameter. 

A targeted literature review was conducted in Google Scholar and PubMed to identify estimates of the relationship between gold-standard adherence measures and claims-based adherence measures. The search combined free text and medical subject headings (MeSHs; in PubMed only) that describe various measures of adherence (MEMS caps, direct observation, PDC, medication possession ration, self-report, prescription claims, lab tests, pill count, and physician estimate) and the search term “adherence accuracy.” Searches were conducted without disease specification, with the term “schizophrenia,” and with the term “serious mental illness.” After reviewing the results of this initial search, the search was conducted again with other central nervous system diseases, specifically multiple sclerosis, Alzheimer’s disease, and Parkinson’s disease. A maximum of 50 titles were screened in each search or until titles were no longer generally relevant to the research questions. 

Abstracts were screened from relevant titles, which were defined as papers that discussed or compared multiple measures of medication adherence. Titles were not relevant if they indicated an intervention to improve adherence, factors influencing adherence, or outcomes associated with adherence or nonadherence, without language suggesting relevance. Titles that were non-English, nonhuman, and did not focus on schizophrenia or another SMI were also not investigated. Full texts were assessed if their abstracts included adherence measures collected using different methodologies and these data were explicitly compared in the results or conclusions sections. Abstracts and full texts were not identified or screened more than once if they appeared as results from multiple searches or were duplicated in databases. Additional citations were identified through previous literature searches, forward reference searches of each manuscript, and the references used in each manuscript. A complete description of the search terms used, number of full texts, abstracts, and articles reviewed is contained in eAppendix 2. 

We used the Truven Health Analytics MarketScan (MarketScan) Commercial Claims and Encounters Database and the Medicaid Multi-State Database from October 1, 2007 through December 31, 2013 to identify patients with SMI. The commercial database included medical and pharmacy claims for individuals and their dependents who were covered by employer-sponsored private health insurance. The Medicaid database included medical and pharmacy claims of Medicaid beneficiaries from 11 deidentified but geographically dispersed states. We limited the sample to individuals aged 18 or older who had at least 1 inpatient or 2 outpatient claims with a diagnosis code for schizophrenia (

International Classification of Diseases, Ninth Revision, Clinical Modification

] diagnosis code: 295.x), bipolar disorder (

 diagnosis code: 296.0x—296.1x, 296.4x–296.8x), or major depressive disorder (

 diagnosis code: 296.2x—296.3x, 311.x). 

To measure adherence and healthcare utilization among patients initiating therapy, patients were required to have a new prescription for an antipsychotic and to be continuously enrolled for 6 or more months before and 12 or more months after the date they filled the new prescription. We required that a patient have an SMI diagnosis and no antipsychotic prescriptions during the 6-month “clean” period before the medication initiation date. Both atypical (eg, aripiprazole, asenapine, iloperidone, lurasidone, olanzapine, paliperidone, quetiapine, risperidone, ziprasidone) and typical (eg, chlorpromazine, fluphenazine haloperidol, perphenazine) oral antipsychotics were included in the analysis. Patients using clozapine were excluded from the sample, as clozapine is typically used only for patients who do not respond to at least 2 other antipsychotic medications.

 Also excluded were patients missing data on age or patients who received antipsychotics via mail order.

We applied an OLS regression to measure the effect of medication adherence on inpatient cost measured over the 365 days following the initiation of the antipsychotic medication regimen. Costs included payments made by primary payers (ie, commercial insurers or Medicaid), patient out-of-pocket payments, and other third-party payments. All costs were inflated to 2015 US dollars using the Consumer Price Index.

 The primary independent variable of interest was a patient’s PDC, which was calculated as the total days covered by all antipsychotics supplied across all claims in the 365 days after the initial medication fill divided by 365. As this study focused on nonadherence rather than overuse, PDC values were capped at 100%. 

To account for compositional differences between the populations of adherent and nonadherent patients, we included patient demographics and health status as explanatory variables in our regression analysis. The Charlson Comorbidity Index

 was used to measure patients’ overall health status.

 Mental health status was measured based on the presence of comorbid SMI conditions (ie, schizophrenia, bipolar disorder, or major depressive disorder), a diagnosis of alcoholism (

: 265.2, 291.1-291.3, 291.5-291.9, 303.0, 303.9, 305.0, 357.5, 425.5, 535.3, 571.0-571.3, 980.x, V11.3), or drug dependence or abuse (

 We also included an indicator for whether the patient was commercially insured or covered by Medicaid and a measure of inpatient spending levels during the 6 months before antipsychotic initiation, as a measure of disease severity. In addition to our baseline specification, we included interaction terms of PDC with indicator terms for specific SMI diagnoses (ie, schizophrenia, bipolar disorder, and major depressive disorder) to estimate the effect of adherence for patients with a specific type of SMI.

All statistical analyses were performed using Stata-MP version 13.0 (StataCorp LP; College Station, Texas).

Estimates of the correlation between gold-standard medication adherence and other adherence measures varied widely in the literature. As shown in 

, only 2 studies measured the correlation between electronic monitoring systems and prescription refill records, with correlation values of 0.32 and 0.48, respectively.

 Two additional studies calculated the correlation between electronic monitoring systems and non-claims measures.

 When we broadened our definition of gold standard to include pill counts, we found 2 additional studies.

 Ignoring the relationship between adherence and patient characteristics, these figures indicate that the effect of adherence on inpatient spending could be underestimated by a factor ranging from 1.2 to 9.8, due to measurement error alone.

In our empirical case study, 145,235 patients in the commercially insured population and 86,321 patients in the Medicaid population had an SMI and initiated an antipsychotic medication regimen (

). Descriptive statistics for patient characteristics are shown in 

. Patient age was concentrated between 28 and 55 years, and 65.3% of patients in our sample were women. Of these patients, 17.7% had an alcohol dependence problem and 9.8% suffered from drug abuse. Among patients with an SMI, 16.4%, 47.8%, and 75.6% had schizophrenia, bipolar disorder, or major depressive disorder, respectively. The average patient had a PDC of 48.0%. 

Results from the regression analysis indicated that patients with higher adherence had lower levels of inpatient spending (

). In our baseline approach (model 1), a 10-percentage-point increase in PDC was correlated with a change in annual inpatient cost of —$41 (95% confidence interval [CI], –$65 to –$16; 

= .001) per patient with an SMI. When allowing for PDC to interact with each of the 3 SMIs of interest (model 2), schizophrenia patients with a 10-percentage-point higher PDC had a higher inpatient cost of $86 (95% CI, —$152 to –$20; 

= .011). The corresponding amounts for major depressive disorder and bipolar disorder were $67 (95% CI, —$106 to –$27; 

When we corrected for adherence mismeasurement using the bias-correction formula above and our preferred estimate for the correlation between claims-based and true adherence, the impact of medication on inpatient spending increased by a factor of 5.3. This estimate is calculated from the correlation from Hansen et al

 between the PDC and patient covariates: 

When we applied this bias-correction factor to our case study, patients with a 10-percentage-point greater PDC had $217 (95% CI, —$347 to –$87) higher inpatient cost per patient (

). Even if we applied the most conservative adjustment factor from Bruce et al,

 that would still inflate the relationship by 20%:

Articles

WEB EXCLUSIVE

This study demonstrates that common pharmacy claims-based measures underestimate the effect of actual adherence on inpatient costs among patients with serious mental illness.

The Cost of Adherence Mismeasurement in Serious Mental Illness: A Claims-Based Analysis

 To establish a prescription opioid registry protocol in a large health system and to describe algorithms to characterize individuals using prescription opioids, opioid use episodes, and concurrent use of sedative/hypnotics. 

Protocol development and retrospective cohort study.

 Using Kaiser Permanente Northern California (KPNC) electronic health record data, we selected patients using prescription opioids in 2011. Opioid and sedative/hypnotic fills, and physical and psychiatric comorbidity diagnoses, were extracted for years 2008 to 2014. Algorithms were developed to identify each patient’s daily opioid and sedative/hypnotic use, and morphine daily-dose equivalent. Opioid episodes were classified as long-term, episodic, or acute. Logistic regression was used to predict characteristics associated with becoming a long-term opioid user. 

In 2011, 18% of KPNC adult members filled at least 1 opioid prescription. Among those patients, 25% used opioids long term and their average duration of use was more than 4 years. Sedative/hypnotics were used by 76% of long-term users. Being older, white, living in a more deprived neighborhood, having a chronic pain diagnosis, and use of sedative/hypnotics were predictors of initiating long-term opioid use. 

This study established a population-based opioid registry that is flexible and can be used to address important questions of prescription opioid use. It will be used in future studies to answer a broad range of other critical public health issues relating to prescription opioid use.

We describe a protocol for developing a prescription opioid registry using electronic health record data and algorithms to characterize individuals using prescription opioids, opioid use episodes, and concurrent use of sedative/hypnotics. 

		In 2011, 18% of adult members of a large integrated health plan filled at least 1 opioid prescription. 

		Among these patients, 25% used opioids long term, and the average duration of use was more than 4 years. Sedative/hypnotics were used by 76% of long-term users. 

		Being older and white; living in a more deprived neighborhood; having chronic pain, arthritis, chronic obstructive pulmonary disease, hypertension, or a psychiatric disorder diagnosis; and a history of sedative/hypnotic use were predictors of becoming a long-term opioid user.

The use of prescription opioids has increased dramatically in the past 2 decades, with associated increases in opioid misuse/abuse and opioid overdose. These are among the most commonly prescribed medications,

 with 259 million prescriptions written for opioid pain relievers in the United States in 2012.

 However, effectiveness of these medications for long-term use has not been established,

 and the risk of opioid-related abuse and overdose has led to a prescription opioid epidemic.

 Close to 2 million Americans abused or were dependent on opioids in 2014.

 From 2003 to 2013, the proportion of drug-abuse treatment admissions for nonheroin opiates tripled.

 There were more than 18,000 fatal overdoses in 2014 related to prescription opioids, more than 4 times the number in 1999.

 Sedative/hypnotics are frequently involved in overdose deaths, and their concurrent use with opioids is of high concern.

The goal of the overall project was to use electronic health record (EHR) data from Kaiser Permanente Northern California (KPNC) to develop a patient prescription opioid registry with the potential to ultimately address several research inquiries: the characterization of opioid use and opioid users, identification of prescription opioid misuse, predictors of opioid overdose, and to describe patients’ services utilization and costs. It drew on our, and others’, previous research.

 Our objective here was to describe our protocol to develop the registry and to address 3 initial research questions: 1) characterize all individuals who used prescription opioids in 2011; 2) analyze their opioid use and concurrent use of sedative/hypnotics; and 3) identify predictors of becoming a new long-term user of opioids. We provide detail and context for our methodological approach, which can be a foundation for future analyses and, we hope, a methodological resource for other research teams addressing these questions. 

KPNC is a nonprofit integrated healthcare delivery system providing comprehensive health services to approximately 3.8 million members in Northern California. The membership reflects the region’s general population, although it underrepresents individuals with very low levels of education and income.

 Membership includes enrollees from Medicare, Medicaid, and the State Health Insurance Assistance Program. 

Data SourcesMembership, outpatient pharmacy, and medical encounter data are archived in KPNC’s EHR. Demographic (eg, age, race/ethnicity, sex, address), membership status, health services, diagnostic, and pharmacy data are regularly extracted from the EHR and stored in a Virtual Data Warehouse (VDW).

 The VDW is a distributed data model that includes EHR data and other data, such as mortality and Census data. Although the current study only uses KPNC data, the VDW is a feature of the Health Care Systems Research Network, whose data are harmonized across 19 health systems.

KPNC pharmacy data in the VDW include generic name, strength, directions for use, date dispensed, quantity dispensed, days’ supply, prescriber identification number, and National Drug Code. Surveys have found that more than 90% of members obtain all or almost all of their prescription medications through KPNC pharmacies.

 KPNC does not have policies that impose mandatory opioid dosage limits or prior authorization.

Clinical diagnostic and health services utilization data include hospitalizations, and emergency department (ED) and outpatient clinic visits (primary and specialty care). Mortality data, including date of death and underlying cause of death, are created from the EHR and death certificates. The VDW tumor registry contains information on all new cancers for KPNC members diagnosed after January 1, 1997. 

We extracted all opioid fills made at KPNC outpatient pharmacies during 2011 (eAppendix Table 1 [

 we focused on formulations with higher likelihood of abuse, and those used primarily to treat pain. Thus, we excluded opioid formulations used primarily as antitussives, anesthetics, antihistamines, antidiarrheals, or opioid agonists/antagonists. Eligible registry participants were aged 19 years or older (on January 1, 2011) with at least 1 opioid fill during 2011, and without a cancer diagnosis between January 1, 1997, and December 31, 2014. 

To compare the demographics of individuals with and without opioid use, we selected a comparison cohort of all individuals who were: KPNC members during 2011; aged 19 years or older on January 1, 2011; without a cancer diagnosis between 1997 and 2014; and without an opioid fill in 2011. 

For each individual, we extracted gender, race/ethnicity, date of birth, and home address as of January 1, 2011. Using members’ home addresses in combination with the 2006 to 2010 American Community Survey collected by the US Census Bureau,

 a neighborhood deprivation index (NDI) was generated at the Census tract level.

 tests were used to identify differences (at 

For the individuals who used prescription opioids, we extracted all outpatient opioid fills between January 1, 2008, and December 31, 2014. This allowed at least 3 years of opioid use before and after 2011, and facilitated analyses related to history and subsequent course of opioid use. We constructed a dataset of daily use: 1 record per person per day for each day of those 7 years; for each day, variables indicated if the person was considered to be using opioids, the opioid type, and the morphine equivalent of milligrams used. To identify opioid use on any given day, we assumed: 1) individuals used opioids according to the days’ supply variable (ie, according to provider instructions); 2) if a person had 7 or more days remaining on a prior fill at dispensation, the new fill was assumed to be used concurrently with the prior fill; otherwise, the new fill was assumed to be used consecutively; and 3) stock from 3 fills could be used concurrently, or be held for future use. 

Opioid Episode Definition and Classification 

Using the daily use dataset, we created episodes of opioid use, defined as the period from the start of any opioid use until a gap in use of more than 180 days.

 For each episode, we calculated: 1) episode duration; 2) number of fills; 3) prescribed days’ supply of opioids filled during the episode; 4) mean morphine daily-dose equivalent (MDDE) across the duration of the episode, by opioid type; 5) MDDE as prescribed, by opioid type (where the denominator is prescribed days’ supply). Because episodes can include gaps in use, MDDE across episodes will typically be lower than MDDE as prescribed, since the latter’s denominator does not include gaps in use. 

Episodes were classified into 3 mutually exclusive types: 1) acute, 2) episodic, and 3) long-term, as in previous research.

 Acute episodes were those lasting less than 90 days. Episodic episodes lasted 90 days or longer, during which the individual was dispensed less than 120 days’ supply of opioids, and during which there were fewer than 10 opioid fills. Long-term episodes lasted 90 days or more with either 120 days’ supply or more of opioids, or at least 10 opioid fills. When describing episodes spanning any part of 2011, we retained only the most severe episode per individual and classified the individual according to that episode type. Thus, all individuals who had a long-term opioid episode covering any part of 2011 were classified as “persons with long-term opioid use.” Among those remaining, individuals with at least 1 episodic episode were “persons with episodic opioid use,” and the remaining individuals were “persons with acute opioid use.” 

Comorbid Health Conditions, Mortality, and Sedative/Hypnotic Use 

To examine comorbidities, we extracted from the VDW all diagnoses (

International Classification of Diseases, 9th Revision, Clinical Modification

) associated with healthcare encounters. As in prior research with other complex patient populations,

 we identified whether individuals received a diagnosis for 1 or more of 13 chronic conditions: arthritis, asthma, congestive heart failure, chronic obstructive pulmonary disease, chronic pain, diabetes, epilepsy, end-stage renal disease, hypertension, ischemic heart disease, osteoporosis, Parkinson’s disease, and stroke. We also identified individuals with the following psychiatric and substance use disorders

: attention-deficit disorders, anxiety disorder, autism, bipolar disorder, dementia, depression, other psychoses, personality disorder, schizophrenia, opioid abuse/dependence, and nonopioid substance use disorders (excluding tobacco). Date of death and underlying cause of death were also extracted, as was KPNC membership information for each month from 2008 to 2014. 

For individuals with opioid use, we extracted all KPNC pharmacy fills for sedative/hypnotic medications (eAppendix Table 2) during 2008 to 2014. Using days’ supply and the same approach as for opioid fills (without calculating dose), we identified daily use of sedative/hypnotics for each person. These records were merged with the opioid episodes to calculate the number of days of sedative/hypnotic use during each opioid episode. 

Predictors of Initiating Long-Term Opioid Use

Individuals who use opioids long term are at higher risk of adverse events (AEs) compared with individuals who use opioids over a shorter term.

 To identify predictors of becoming a long-term user within 3 to 4 years of starting opioid use, we identified “opioid-naïve” individuals as those without an opioid fill prior to their first fill of 2011—in other words, having had at least 3 years without prior use of prescription opioids. We required continuous KPNC membership from 2008 to 2014, or death, subsequent to their first 2011 fill (allowing gaps in membership of ≤3 months). In the year prior to their first opioid fill, we examined medical diagnoses, sedative/hypnotic use, and 4 measures of utilization as proxies for severity and propensity to use resources: 1) number of inpatient days, 2) number of ED visits, 3) number of office visits, and 4) number of nonopioid prescription fills. 

We conducted logistic regression analysis with the dependent variable indicating whether or not the person subsequently had a long-term opioid episode, from initial fill to December 31, 2014. In addition to sex, age (in 7 groups), race/ethnicity, and NDI quartile, we included the following covariates measured in the year before the first opioid fill: dichotomous indicators for each of 13 chronic medical conditions; dichotomous indicators for any psychiatric disorder, opioid abuse/dependence, and nonopioid abuse/dependence; a dichotomous indicator of sedative/hypnotic use; the 4 utilization measures; and a class variable indicating the KPNC clinic where the patient received most of their care. 

Among 2,480,030 adult KPNC members in 2011, 455,693 (18.3%) had at least 1 opioid fill at a KPNC pharmacy. Individuals with opioid use were different from individuals without opioid use on every characteristic, including being more likely to be female, older, white, have a chronic medical or psychiatric condition, and to have a diagnosis of opioid abuse/dependence or nonopioid substance use disorder (

The 455,693 individuals with opioid use had 474,045 unique episodes occurring in some part of 2011. Due to the 180-day gaps used when creating episodes, few individuals had more than 1 episode in 2011: 18,352 individuals had 2 episodes covering any part of 2011, and the rest had 1. After retaining the most severe episode per person, there were 112,089 long-term opioid episodes, 71,011 episodic episodes, and 272,593 acute episodes (

). Long-term users were, on average, aged 56 years, and 61% were female. Forty percent of individuals with long-term opioid use received at least 1 diagnosis for a psychiatric disorder in 2011, 3% were diagnosed with opioid abuse/dependence, and 7% were diagnosed with a nonopioid substance use disorder. 

On average, long-term episodes were 1609 days long and included 54 opioid fills (

). Because the opioid data spanned 2008 to 2014, it is possible that some episodes began prior to 2008 or continued after 2014; thus, episode durations may be underestimated. Among long-term opioid episodes, 27% began prior to July 1, 2008, and ended after June 30, 2014, and may be both left and right truncated. Another 17% of long-term episodes may have been left truncated only, and 24% may have been right truncated only.

Among individuals with long-term use, the mean MDDE was 38.98 mg, with the highest MDDE for long-acting Schedule II opioids. However, use of those opioids was highly skewed (median MDDE was 0), with only 28% of long-term users using any long-acting Schedule II opioids. Among individuals with episodic and acute use, on the other hand, Schedule III opioids had the highest mean MDDE. Individuals with episodic use used at lower levels than long-term users, and tended to have substantial gaps between fills; therefore, they had a much lower mean MDDE (5.37 mg). Individuals with long-term use also had higher mean MDDE than episodic or acute users. 

Sedative/hypnotics were used by 76% of individuals with long-term opioid use during their episodes, and for an average of 34% of episode days. Among individuals with acute use, 16% used sedative/hypnotics during their (much shorter) acute episode.

Among all individuals using prescription opioids in 2011, 175,558 (39%) were opioid naïve. Of these, 85,305 had continuous KPNC membership from 2008 to 2014 (n = 81,809), or until death (n = 3496), and were the analytic sample for initiating long-term opioid use (

). Multivariate analysis indicated that individuals at least 80 years of age were more likely to become long-term users than individuals younger than 50 years of age. Compared with whites, Asians and Hispanics were less likely to become long-term users (odds ratio [OR], 0.41; 95% confidence interval [CI], 0.35-0.47, and OR, 0.67; 95% CI, 0.60-0.75, respectively) (

). Individuals in more-deprived neighborhoods were more likely to become long-term users than those in the least-deprived neighborhoods (most-deprived neighborhood: OR, 1.26; 95% CI, 1.12-1.43). 

Numerous conditions were associated with long-term opioid use, including chronic pain (OR, 2.57; 95% CI, 2.26-2.93), nonopioid substance use disorders (OR, 2.25; 95% CI, 1.89-2.69), psychiatric disorders (OR, 1.22; 95% CI, 1.12-1.33), and arthritis (OR, 1.41; 95% CI, 1.31-1.52). Use of sedatives/hypnotics was associated with increased odds of becoming a long-term user (OR, 1.67; 95% CI, 1.54-1.81, vs no use). Even after adjusting for diagnosed conditions, inpatient hospital days and use of nonopioid medications in the prior year remained predictive of the long-term user. On the other hand, outpatient office visits in the prior year were associated with lower odds of becoming a long-term user. 

This study developed a protocol for an EHR-based prescription opioid registry that can be used to address important research questions about prescription opioid use in noncancer patients on a population level. The current paper also addressed initial questions about the characteristics of prescription opioid users and what predicts initiating long-term use.

 individuals who used opioids were older, more likely to be white, and were more clinically complex patients, with more medical and psychiatric conditions and substance use disorders than individuals not using opioids. Further, a considerable portion of patients was using opioids long term. Patients using opioids long term are especially important to identify, because duration of use is associated with abuse, overdose, and other AEs.

 we found individuals with long-term opioid use to be more likely than those with shorter-term use to have higher daily dosages, chronic medical or psychiatric conditions, and opioid or other substance use disorders. Given the current epidemic of misuse and overdose, identifying long-term users with population-based data can help health systems identify patients early, monitor them, and refer them to specialty services (eg, substance use treatment, pain management) as needed. 

Our analysis of “opioid-naïve” users indicated that only 4.2% went on to long-term use within 3 years, although at any given time, the percentage of long-term opioid users is quite high (25%). Although individual risk is low, at a population level, this is consistent with the high level of AEs observed in recent years. 

Predictors of developing long-term use included chronic pain, sedative/hypnotic use, psychiatric disorders, and nonopioid substance use disorders. Concurrent use of sedative/hypnotics and opioids has been shown to be associated with a substantial increased risk of death from drug overdose.

 Federal and health-system guidelines have focused on reducing high daily dosages, and also on restricting concurrent opioid and sedative/hypnotic use.

 Individuals who lived in more-deprived neighborhoods were also more likely to develop long-term opioid use; to our knowledge, this is a relationship not previously identified in other research studies. Our data do not contain information on pain severity or control. However, findings may suggest that individuals residing in more-deprived neighborhoods (which may also be a proxy for individual deprivation) have more complex health status, or fewer nonmedication treatment alternatives available—these hypotheses deserve further study.

There is increased interest in using registries to address critical clinical and policy questions.

 A goal of this project was to develop a protocol that can serve as a reference for other clinical and research teams addressing similar questions. Study algorithms can be used in health systems with pharmacy dispensation data and encounter data. For example, because our approach used the VDW, investigators from 19 other health systems in the Health Care Systems Research Network can also use the VDW to similarly address important questions about prescription opioid use. We recognize this is not without challenges, and would require adaptations, particularly for systems that have dissimilar EHR data elements or claims data. However, by sharing details about our methodology, we hope to contribute to developing harmonized approaches across systems to address the opioid epidemic. 

Our measures of opioid and sedative/hypnotic use depend on pharmacy dispensation data, which is commonly used in the literature, and which we consider a reasonable proxy for use. Uncertainty also exists about calculating use for overlapping fills. However, in contrast to some prior studies, we make explicit our assumptions for overlapping fills. The vast majority of KPNC members fill prescriptions at KPNC pharmacies,

 but we miss potential non-KPNC pharmacy fills. Although all registry members filled an opioid prescription at KPNC, it is possible that individuals using opioids may be more likely to seek opioid prescriptions externally. Identification of medical and psychiatric conditions, and substance use disorders, is based on diagnoses recorded in the EHR as part of routine care; thus, individuals with more visits may have more opportunity to receive a diagnosis. Also, there can be truncation of episodes that began prior to 2008 or continued post 2014 and, therefore, possible underestimation of long-term episode duration. These limitations are similar to those of other EHR-data—based studies. Finally, generalizability to other systems may be limited, although study algorithms can be adapted. 

This study established a population-based opioid registry that is flexible, and can be used to address important questions of prescription opioid use. Future analyses will leverage the prescription opioid registry and its algorithms to examine prescription opioid misuse, fatal and nonfatal overdose, and health service utilization and cost. Thus, with this same registry, we will be able to address a broad range of critical public health issues relating to prescription opioid use. 

A flexible population-based prescription opioid registry was established for addressing a broad range of critical public health questions relating to prescription opioid use.

Prescription Opioid Registry Protocol in an Integrated Health System

Adverse events (AEs) are common, estimated to occur in around 10% internationally. Although preventable harm can be minimized, when AEs occur it is important that they be managed appropriately. AEs can be traumatic not only for patients, their friends, and relatives, but also for the involved clinicians, who have been referred to as “second victims” in a growing body of international research. Despite the frequency with which AEs occur, organizational mechanisms for supporting staff in these circumstances are not routinely embedded in healthcare settings. Critical Incident Stress Debriefing (CISD) has long been provided for professionals, such as disaster workers, who are exposed to traumatic and high-stress events. CISD is considered an effective strategy to promote resilience and recovery. We explore the potential value of providing CISD for health professionals involved in patient safety—related AEs and discuss the instances in which this could be routinely implemented.

		Health professionals can suffer negative professional and personal effects after involvement in patient safety—related adverse events (AEs). 

		The quality and safety of patient care can be compromised when health professionals experience these negative effects. 

		Most organizations do not provide any support for staff in relation to involvement in adverse events. 

		Critical incident stress debriefing has long been provided for professionals exposed to traumatic and high-stress events and may be a useful strategy for those who experience AEs.

It has long been recognized that clinical work carries an emotional burden. Healthcare workers are regularly exposed to loss, injury, traumatic stimuli, and undesirable patient outcomes. Adverse events (AEs) are common in medicine, with estimates that as many as 1 in 9 hospitalized patients are harmed by the healthcare they receive.

 Although preventable harm can be minimized, it is important to appropriately manage AEs when they occur.

 AEs can be traumatic not only for patients, their friends, and relatives, but also for the clinicians involved—referred to as second victims.

 Critical incident stress debriefing (CISD) has long been provided for professionals, such as disaster workers, who are exposed to traumatic and high-stress events; it is considered an effective strategy to promote resilience and recovery.

 This paper explores the potential value of providing CISD for health professionals involved in patient safety—related AEs and the instances in which this could be routinely implemented.

Clinicians who witness or contribute to the occurrence of an AE can experience psychological effects that disrupt their professional and personal lives and their ability to deliver high-quality, safe care.

Anxiety, depression, and sleep disturbances are consistently reported, in addition to shame, guilt, loss of self-confidence, and feelings of incompetence and worthlessness.

 The severity of these effects is related to the degree of harm to the patient and the clinician’s experience of the investigation process.

These symptoms have ripple effects, with adverse consequences for patients, clinicians, and the wider healthcare system. Patient safety can be threatened in the immediate aftermath of an incident, when a clinician’s ability to manage other patients may be impaired.

 In extreme cases, clinicians may change careers or leave the profession.

 In the longer term, the ability for individuals and organizations to learn from AEs can be diminished if clinicians are reluctant to report incidents.

A survey of UK physicians reported that most (83%) had been involved in an AE and that 76% had experienced personal or professional distress that required support.

 Despite this, progress in providing effective support for these professionals has been slow. At present, health professionals report difficulty in identifying appropriate sources of support after their involvement in an AE. CISD has been utilized in other healthcare contexts, but rarely in the context of patient safety—related AEs.

CISD is a structured, small-group, supportive crisis intervention process that proceeds in 7 phases 

 and is followed by individual sessions and follow-up engagement with other support services. It is one of many crisis intervention techniques that are included under the umbrella of a Critical Incident Stress Management (CISM) program.

 There is evidence to suggest that CISM approaches are effective in reducing the negative psychological aftermath of a wide variety of critical incidents.

CISD is embedded within a web of crisis support services, such as pre-event education, follow-up services, and referral to professional care and postincident education programs.

 Utilizing CISD in the context of AEs offers the benefit of an established method in which many counselors are trained. Debriefings are designed for small, homogeneous groups who are unified by their encounter of a traumatic event, such as those involved in a patient safety—related AE. They are a form of psychological first-aid in the immediate aftermath of the event that aims to reduce feelings of distress and to restore group cohesion and performance.

Each health professional has a unique set of needs in relation to his or her experience of an AE. Health professionals involved in some AEs, such as those that did not involve an error or did not lead to adverse consequences for the patients, may not feel that they warrant CISD. By offering CISD at an organizational level, health professionals can choose to access this when they consider it necessary or helpful to provide immediate support after an event, and access to resources and individualized support in the following days, weeks, and months. CISD may then be routinely implemented at a local level for specific types of AEs associated with greater psychological distress (eg, the unexpected death of a child). Selecting the AEs for which CISD is routinely provided is challenging, given that the level of distress experienced following an event is subjective. Evidence of those events that lead to the greatest distress is limited, but studies indicate that those in which there is patient harm or harm that is severe are more damaging to the health professionals involved.

 Routinely providing CISD for these events may be an appropriate first step, and it may be possible to achieve consensus on a list of “must support” events.

Lessons From Current Support Programs for Clinicians Involved in AEs 

Evidence of the need for psychological first-aid in the days and weeks that follow the event, in addition to a broader infrastructure of support and guidance, has provided the foundation for in-house interventions, such as the forYOU program at the University of Missouri Health System, the peer-support program at Brigham and Women’s Hospital, and the Resilience in Stressful Events (RISE) program at Johns Hopkins Hospital.

 Psychological first-aid used in these programs presents an adapted form of CISD, although evaluation is currently limited and a full CISD process has not yet been attempted. The ethical challenges associated with evaluating the effectiveness of such programs has been a key barrier. 

Data on the development of such programs provide some guidance around the level of demand, as well as the organizational infrastructure and training needed to operate support programs. Data collected in the first year of the forYOU program at the University of Missouri Health System reported 30% of health professionals surveyed as being adversely affected by involvement in an AE over a 12-month period, with 15% experiencing serious professional implications (eg, considering leaving their profession).

 At Johns Hopkins, nearly 50% reported experiencing such an event at some point during their career.

The forYOU team consists of 51 members to support the 5300 faculty, staff, students, and volunteer employees, whereas the RISE program has 29 on-call responders for approximately 10,000 employees. In Missouri, most staff (60%) can gain the necessary support from those in their local environment if staff have an understanding of the issues facing “second victims.” An additional 30% of health professionals benefit from self-referral to organizational peer support, and a further 10% require follow-up with external support services. In the first 10 months of operation, an average of 5 peer support sessions were requested per month, from which 13 employees were referred to external support services. One team debriefing was required every other month.

Data from the development of the peer-support program at Brigham and Women’s Hospital and a cross-country study of UK and US health professionals suggest that those seeking peer support locally or from trained peer supporters indicate a preference for support from members of their own profession who are at the same or slightly senior level, but that there are instances in which people receive support from those in other professions.

The application of CISD beyond AEs to a range of stressful events experienced by health professionals means that it is advantageous for health service providers to offer CISD for staff experiencing any type of traumatic event. Less resource-intensive mechanisms for providing support are particularly important in public health systems, such as the United Kingdom’s National Health Service. A further advantage of CISD is that there is a large reservoir of relief workers already trained and experienced in providing this kind of support. This is particularly useful for smaller organizations that do not already have institutional support services in place, such as employee assistance programs or formal support for “second victims.”

Current models of support for those affected by AEs reveal a number of practical considerations that must be considered in the utilization of CISD programs for this group. Although there is demand for support services, CISD will not be required every day or every week in relation to AEs, and the program structure must provide the necessary flexibility for this. Debriefings work best with small, homogeneous groups. If implemented at a local level, a team may be united by their involvement in the AE and their feelings of distress, but preference for peer support from the same profession must be taken into consideration.

There may also be tension between individuals and professional groups around attribution of responsibility for the AE.

 Support services require organizational resources and, therefore, buy-in. Organizations also have to invest in making frontline providers and service directors aware of the support available and mechanisms for accessing the CISD program, if implemented, in addition to developing human resource policies and guidance for those affected and for those providing counseling. 

Work is still needed to test the effectiveness of a full CISD program for "second victims" and to identify any modifications needed for healthcare workers experiencing this type of traumatic event. However, if successfully deployed, there are advantages for healthcare organizations. Routinely implementing a support process for staff following harmful AEs would demonstrate that organizations recognize patient safety—related AEs as a potentially distressing aspect of clinical work. Greater acknowledgement of the profound impact of AEs for those involved may facilitate the open discussion of AEs that are critical to constant learning and improvement. In these ways, CISD may be a useful strategy to support clinicians to promote resilience, recovery, and the continuous provision of safe patient care. Trial and evaluation of the applicability and effectiveness of CISD in the context of AEs is now needed.

 School of Public Health, University of New South Wales (RH), New South Wales, Australia; Johns Hopkins Bloomberg School of Public Health (AW), Baltimore, MD.

 The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article. 

 Concept and design (RH; AW); analysis and interpretation of data (RH; AW); drafting of the manuscript (RH; AW); critical revision of the manuscript for important intellectual content (RH; AW). 

 Reema Harrison, PhD, MSc, BSc, School of Public Health and Community Medicine, University of New South Wales, Rm 308 Samuels Bldg, New South Wales, Australia 2026. E-mail: reema.harrison@unsw.edu.au. 

1. Classen DC, Resar R, Griffin F, et al. Global trigger tool shows that adverse events in hospitals may be ten times greater than previously measured. 

. 2011;30(4):581-589. doi: 10.1377/hlthaff.2011.0190.

2. de Vries EN, Ramrattan MA, Smorenburg SM, Gouma DJ, Boermeester MA. The incidence and nature of in-hospital adverse events: a systematic review. 

. 2008;17(3):216-223. doi: 10.1136/qshc.2007.023622.

3. Conway JF, Federico F, Stewart K, Campbell MJ. Respectful management of serious clinical adverse events (second edition). Institute for Healthcare Improvement website. www.ihi.org/resources/Pages/IHIWhitePapers/RespectfulManagementSeriousClinicalAEsWhitePaper.aspx. Published 2011. Accessed March 16, 2017.

4. Wu AW. Medical error: the second victim. The doctor who makes the mistake needs help too. 

5. Scott SD, Hirschinger LE, Cox KR, McCoig M, Brandt, J, Hall LW. The natural history of recovery for the healthcare provider “second victim” after adverse patient events. 

. 2009;18(5):325-330. doi: 10.1136/qshc.2009.032870.

6. Sirriyeh R, Lawton R, Gardner P, Armitage G. Coping with medical error: a systematic review of papers to assess the effects of involvement in medical errors on healthcare professionals’ psychological well-being. 

. 2010;19(6):e43. doi: 10.1136/qshc.2009.035253.

The authors explore the potential value of providing critical incident stress debriefing for health professionals involved in adverse patient safety events and the instances in which this could be routinely implemented.

May 2017

The Cost of Adherence Mismeasurement in Serious Mental Illness: A Claims-Based Analysis

Prescription Opioid Registry Protocol in an Integrated Health System

Critical Incident Stress Debriefing After Adverse Patient Safety Events

Opioid Prescribing for Chronic Pain in a Community-Based Healthcare System

Disparities in Diabetes and Hypertension Care for Individuals With Serious Mental Illness

Assessing the Effect of the VHA PCMH Model on Utilization Patterns Among Veterans With PTSD

The Association of Mental Health Program Characteristics and Patient Satisfaction

State Prescription Drug Monitoring Programs and Fatal Drug Overdoses

Drivers of Excess Costs of Opioid Abuse Among a Commercially Insured Population

Medicaid Prior Authorization and Opioid Medication Abuse and Overdose