How do we optimize message matching interventions? Identifying matching thresholds, and simultaneously matching to multiple characteristics