# Identifying opioid agonist treatment prescriber networks from health administrative data: A validation study

## TL;DR

This study shows how social network analysis can identify clinics prescribing opioid agonist treatment using patient data, even without direct clinic identifiers.

## Contribution

The novel use of social network analysis to identify OAT clinics from administrative data without clinic-level identifiers.

## Key findings

- Social network analysis successfully identified clinics with high specificity and negative predictive value.
- Modularity maximization provided the best performance among tested algorithms.
- Network-based identification can facilitate clinic-level outcome comparisons in opioid treatment.

## Abstract

Given the growth of collaborative care strategies for people with opioid use disorder and the changing composition of the illicit drug supply, there is a need to identify and analyze clinic-level outcomes for centers prescribing opioid agonist treatment (OAT). We aimed to determine and validate whether prescriber networks, constructed with administrative data, can successfully identify distinct clinical practice facilities in Ontario, Canada.

We executed a retrospective population-based cohort study using OAT prescription records from the Canadian Addiction Treatment Centres in Ontario, Canada between 01/01/2013 and 12/31/2020. Social network analysis was utilized to create networks with connections between physicians based on their shared OAT clients. We defined connections two different ways, by including the number of clients shared or a relative threshold on the percentage of shared OAT clients per physician. Clinics were identified using modularity maximization, with sensitivity analyses applying Louvain, Walktrap, and Label Propagation algorithms. Concordance between network-identified facilities and the (gold standard) de-identified facility-level IDs was assessed using overall, positive and negative agreement, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV).

From 144 physicians at 105 clinics with 32,842 OAT clients, we assessed 250 different versions of the created networks. The three different detection algorithms had wide variation in concordance, with ranges on sensitivity from 0.02 to 0.88 and PPV from 0.06 to 0.97. The optimal result, derived from the modularity maximization method, achieved high specificity (0.98, 95% CI: 0.98, 0.98) and NPV (0.98, 95% CI: 0.97, 0.98) and moderate PPV (0.54, 95% CI: 0.52, 0.57) and sensitivity (0.45, 95% CI: 0.43, 0.47). This scenario had an overall agreement of 0.96, negative agreement of 0.98, and positive agreement of 0.49.

Social network analysis can be used to identify clinics prescribing OAT in the absence of clinic-level identifiers, thus facilitating construction and comparison of clinic-level caseloads and treatment outcomes.