Identifying data sources
With respect to exposures, we distinguished between those that heighten overdose risk, with a focus on published relationships between opioid analgesic dose [
5], concomitant opioid and benzodiazepine prescriptions [
6], recent release from incarceration [
7], or cessation of treatment for OUD [
8], and those that may reduce overdose risk, such as use of sublingual buprenorphine products and methadone for the treatment of OUD [
9]. In Connecticut, controlled substance prescription data (including buprenorphine products) is collected and managed by the Department of Consumer Protection (DCP). Data related to methadone treatment in the state is managed by the Department of Mental Health and Addiction Services (DMHAS) (see Table
1).
Table 1
Data sources relevant to opioid-related overdoses
Exposures | Controlled substance prescriptions (opioid analgesics, methadone, and benzodiazepines) | Connecticut Prescription Monitoring and Reporting System (CPMRS) | Department of Consumer Protection |
Buprenorphine for opioid use disorder (OUD) | CPMRS | Department of Consumer Protection |
OUD treatment, including methadone from opiate treatment programs | | Department of Mental Health and Addiction Services |
Incarceration periods, methadone for OUD | | Department of Correction |
Outcomes | Fatal opioid overdoses | | Office of the Chief Medical Examiner |
Fatal and nonfatal opioid overdoses | ChimeData | Connecticut Hospital Association |
Out of hospital opioid overdoses | | Department of Public Health/Department of Emergency Services and Public Protection |
With respect to outcomes, we focused on the collection of data on both fatal and non-fatal cases. Compared to data pertaining to non-fatal cases, fatal overdose information is typically more accurate because of the involvement of the Office of the Chief Medical Examiner (OCME), which performs an autopsy and toxicological testing when indicated in all unnatural deaths. Because of the detailed fatal case reports, we differentiated between single-, poly-opioid, and poly-substance related events, and whether the overdose was intentional or unintentional.
For non-fatal overdoses, the source of information and level of detail depends on where the victim is encountered. Overdose reversals with naloxone are often performed by first responders, which can include emergency medical service (EMS) personnel, local police and firefighters, and, in the cases of many of the smaller Connecticut towns, the state police. Uniformly detailed reporting on these cases is not mandatory. The Department of Public Health (DPH) maintains the database for EMS reversals collected through National EMS Information Systems forms. The Department of Emergency Services and Public Protection (DESPP) collects data from the state police. Reversals can also be performed by laypeople in the community, but there is very little reliable information on naloxone administration unless individuals performing the reversal engage with the EMS or healthcare systems at the time of the event or subsequently report it to a community program that supplied the naloxone. The Connecticut Hospital Association’s ChimeData collects and maintains administrative discharge (UB-04 claims-based) data from inpatient admissions, hospital-based outpatient surgery, and emergency department (ED) non-admissions and provides this to the DPH.
Accessing data
Of these datasets, the one most readily obtained was from the OCME, aided by the fact that two members of the CORE team (LEG, RH) have had long-standing relationships with the OCME working on similar projects [
10]. The CORE team reviews autopsy reporting for each opioid-involved fatality to determine if it was unintentional or intentional and if the final immediate cause of death involved an opioid (± other substances) as determined by either the medical examiner and/or toxicological evidence. Research involving data on deceased individuals is IRB-exempt and some information is publicly available, but an agreement had to be established between the OCME and the research team as most of the detailed decedent-level data are not publicly available. Department of Correction (DOC) data are publicly available, though, similar to OCME data, specific elements of most interest to our work (e.g. receipt of methadone during incarceration) are not.
In contrast, almost all other data sought by the CORE team are considered protected personal health information, and as a consequence, the agencies that maintain the databases have identified HIPAA and other barriers that create legal challenges to sharing data within state agencies and with the CORE team. Overall, developing a satisfactory protocol and obtaining the approvals required over a year’s time and review by multiple attorneys. A key factor in our eventual success was that identifiers were only linked between state agencies by a designated member of our team, a UConn researcher who had the credentials required to view confidential data from two of the state agencies and who worked closely with other agency employees to conduct the linkage in a manner consistent with state protocols for data security.
Data merging and matching
We used a public domain software program that integrates both probabilistic and deterministic matching algorithms (The Link King V9,
www.the-link-king.com) to identify and match individual records across multiple agencies using all available demographic identifiers and geographic information, such as residential address and zip code [
11]. The deterministic protocol ascertains whether record pairs matched or did not match on a set of established criteria; the Link King application employs a complex deterministic protocol that allows some discrepancy on the record elements through “fuzzy” equivalence algorithms. It also includes an array of probabilistic algorithms such as phonetic name matching, approximate string matching and spelling distance, and calculation of distance between the geographic centers of zip codes. These probabilistic procedures use statistical formulae to calculate an overall similarity score between data elements for each record pair and cut points to determine if the records were from the same individual. In a study using similar data elements, the Link King application was shown to have high accuracy for records linkage, with sensitivity at 96.6% and positive predictive at 96.1% [
12].
The record matching was carried out in two phases. The first phase involved linking records across data extracts from administrative databases concerning hospital care, emergency medical treatment, medical examiner reports, incarceration dates, and substance use disorder treatment through the state secured data exchange protocol. A randomly generated unique identifier was assigned to each matched individual. Once datasets were linked through the unique identifier, all other personal identifiers were removed from these datasets, except the master individual list file. In the second phase, the master file was transported via a state securely encrypted laptop, to the state agency—the DCP—that administers the state’s PDMP. Because the DCP does not allow identified data to leave its premises, the record matching was performed on site and all identifiers were removed prior to departure. After the second phase matching was completed, individual identifiers were stripped and only the unique identifier and de-identified data were saved in the laptop to be merged with the de-identified data extracts from the first phase.
As above, we identified both exposure and outcome datasets with personal identifiers—including sex, age, race/ethnicity and census tract allowing for inference of socioeconomic status—with sufficient detail for project goals. For exposure ascertainment, we included CPMRS and DMHAS treatment datasets, DOC data, ChimeData, and, for outcome ascertainment, we included OCME cause-of-death files (fatal opioid-related events), ChimeData (fatal and non-fatal overdose events), EMS and DESPP data (non-fatal overdose events).
With exposure and outcome data matched, we generated a per-subject profile of the following: (1) prescription opioid and benzodiazepine receipt including products, dosages, duration of therapy, and timing in relation to outcome; (2) hospitalizations including duration, admitting and discharge diagnoses, controlled substances received, and timing in relation to outcome; (3) DMHAS- and DCP-captured OUD treatment episodes, categorized as detoxification episodes, longer-term abstinence-based treatment, and methadone and buprenorphine treatment episodes, and (4) DOC episodes including arrest charge, duration of incarceration, methadone treatment, and timing in relation to outcome. For DMHAS exposure data, elements included OUD diagnoses, OUD medication treatment received, start and end dates of treatment, and timing in relation to outcomes in the cohort.