Identifying compliant participants through data matching improved estimation of intervention efficacy: randomized...

We propose a data-matching approach to estimate intervention efficacy for randomized controlled trials (RCTs) when there is noncompliance to the allocated treatment with induced selection bias.

We considered a large RCT to compare health care costs and hospital length of stay 12 months after randomization. Participants allocated to the intervention group were eligible to receive health-coaching and disease-management services. An opt-out approach was adopted for recruitment. Control-group participants received usual care but were allowed to opt-in to receive the intervention. Using "nearest-neighbor"-matched data, we identified compliant participants in both arms to estimate intervention efficacy. Results were compared with intention-to-treat (ITT), instrumental-variable-adjusted ITT, per-protocol (PP), and as-treated (AT) analyses.

The ITT estimated an intervention effect of a 1.5% reduction in cost, but 56.7% of intervention-group participants did not receive health coaching. The PP and AT found an increase in cost of 9.4% and 17.1%, respectively. The matching method estimated a 12.3% reduction in cost. After adjustment for baseline covariates, the intervention group had lower same-day admission cost (13.6%; 95% CI: 7.3%-20.0%; P < 0.001) and shorter hospital stay (11.2%; 95% CI: 2.6%-19.9%; P = 0.021).

Opt-in/opt-out strategies in RCTs misled intervention comparisons and the matching approach improved estimation of intervention efficacy.