Commentary on Dowd <i>et al</i> .: Beyond single data sources—Consistency checking and evidence synthesis for prevalence estimation

New and improved methods are needed to estimate the prevalence of dependence on drugs including opioids—particularly where treatment data are partial and unlinked to mortality, health and criminal justice records. Confidence in new estimates and methods can be achieved through consistency checks with independent information on drug-related harms. Reliable estimates of the prevalence of opioid use disorder (OUD) or injecting drug use are particularly important in populations facing an opioid emergency and/or changing levels of drug-related harm. We agree with Dowd et al. [1] that prevalence estimation is challenging. Population surveys inevitably underestimate prevalence because of missing key populations, while all routine data sets are partial, such that any single one is likely to produce biased estimates. In the United States, national estimates of injecting drug use prevalence rely on unvalidated assumptions, including an estimate of the ratio of non-fatal to fatal overdose that is neither period nor area-specific [2]. We agree also that multiple systems estimation (MSE) or capture-recapture will often produce biased results [3] and, as noted by Dowd et al. [1], estimates that fluctuate widely between years - which suggests violation of one or more underlying assumptions. Critically, MSE is rarely validated against other information. We developed the multi-parameter estimation of prevalence (MPEP) approach [4-6], a type of multi-parameter evidence synthesis (MPES) [7, 8] model that uses linkage between opioid agonist therapy (OAT) and drug-related harms data to jointly estimate prevalence from multiple data sources. Importantly, this allows consistency of evidence across different sources to be assessed. Availability of linked data is variable and may also not be complete, as highlighted by Dowd et al. [1]. In the United States, fragmented payer systems mean MPEP estimates may be restricted to subsets of the population. For example, Doogan et al. [9] applied MPEP in 19 of 88 Ohio counties, but estimated prevalence only within the Medicaid beneficiary population. We welcome Dowd et al. [1]’s suggestion for estimating prevalence through extrapolation from Medicaid claims with recorded OUD when linked data are not available. Their approach adjusts for misclassification of OUD diagnoses (both false positives and false negatives) and for a proportion of individuals with OUD having no healthcare encounters within the period. A notable strength is the use of Monte Carlo simulation to propagate uncertainty through these adjustments, producing prevalence estimates with uncertainty intervals. The approach could be strengthened further by incorporating sampling uncertainty in the observed proportion of recorded OUD cases within the Medicaid population—particularly for smaller geographic areas. Dowd et al. [1] additionally show how prevalence estimates derived for the Medicaid population can be extrapolated to the total population, using information from the National Survey on Drug Use and Health (NSDUH) on population size and the ratio of reported OUD prevalence among non-Medicaid and Medicaid populations. Use of this ratio is a pragmatic approach: although unlikely to fully eliminate bias, ratios are likely to be more stable and transferable across settings than absolute prevalence estimates. This extrapolation could be extended to allow for potential variation in this ratio by demographic factors or county via a regression modelling structure. An important next step would be to assess consistency of the resulting prevalence estimates with drug-related harm data, such as opioid-related hospital admissions and deaths. Moreover, if there is any prior information—however uncertain—on plausible rates of drug-related harms within the population, all available data can be jointly modelled within a Bayesian MPES framework [7, 8]. Notably, MPES (unlike the more specific form, MPEP) does not require data linkage. This work also raises important questions about the target population—definition of which is also a core component of MPEP. Dowd et al. [1] aim to estimate the number of individuals meeting Diagnostic and Statistical Manual of Mental Disorders, fifth edition, criteria for OUD. However, it is worth considering whether this population is a mixture of individuals using illicit opioids, people in receipt of OAT and people in receipt of prescription opioids for non-cancer pain (potentially 5%–15% of people prescribed, as highlighted by a recent systematic review [10]). Several assumptions deserve scrutiny: whether diagnostic sensitivity and specificity are similar among these different populations, whether these vary by engagement with healthcare and whether geographic variation in diagnostic behaviour or access to services could be contributing to variation in prevalence estimates. An additional limitation is that NSDUH omits incarcerated and homeless populations, which vary geographically, and potentially complicates the extrapolation to non-Medicaid residents. Nonetheless, Dowd et al.'s approach [1] is an important building block toward improved prevalence estimates in settings without record linkage. Where treatment data are incomplete because of multiple payer systems, incorporating survey-based ratios within MPEP to produce estimates for the overall population also warrants consideration. Hayley E Jones: Conceputalization (equal); methodology (lead); writing—original draft preparation (lead); writing—review and editing (equal). Andreas Markoulidakis: Methodology (supporting); writing—original draft preparation (supporting); writing—review and editing (equal). Matthew Hickman: Conceputalization (equal); methodology (equal); writing—original draft preparation (supporting); writing—review and editing (equal). The authors acknowledge funding from the NIH (1R01DA059822-01A) and NIHR (NIHR160569 and NIHR200877). None. N/A.