Interval-cohort designs and bias in the estimation of per-protocol effects: a simulation study