OBJECTIVES:

We demonstrate how estimates of retention on OUD treatment in methadone clinics (MC) vs office-based buprenorphine (OB) in environments where treatments are mandated by protocol (MC=82.6%; OB=85.3%) are generally optimistic relative to results to be expected when patients are offered choices of treatments.

METHODS:

Manski called this estimation problem a “mixing problem.” Various solutions are possible; most yield only bounds on estimates, the simplest being that for which no assumptions are employed. The method employs a counterfactual argument, using “potential” outcomes of treatment assignments, only one of which is observed. It suggests that there are 4 fractions of patients, two for which treatment is inconsequential (patients succeed or fail regardless of treatment) and two for which only one or the other of the treatments is successful. Speculation on the sizes of these fractions consistent with the observed trial outcomes with mandated treatments yields bounds on retention for a scenario where patients choose which of the two treatments they receive.

RESULTS:

If outcomes are ordered (e.g. for all patients, OB > MC) then the success percentage bounds correspond to those observed when mandated [82.6%, 85.3%]. This is also the case if treatment is independent of outcomes. However, when employing no assumptions (no prior information), bounds on retention percentages when patients have choices are shown to be wider [67.9%, 100%] than the observed range in the trial.

CONCLUSIONS:

Outside trials, patients generally have treatment choices and outcomes will generally differ from environments with mandated treatments. The wider bounds result from patients choosing treatments at the extremes (optimizing or minimizing their success), choosing treatments that work (or don’t work) for them. Perfect foresight being rare, predicting outcomes must, absent assumptions, allow for both extremes – and anywhere in between. Other ranges are possible with various assumptions on prior outcomes or treatment policies.