Study Design
Sampling based on Milestones (SMile): A potential alternative study design Ian Shrier* Ian Shrier Tibor Schuster Yi Li Zachary Vernec Russell Steele
Some authors suggest simulating data to evaluate potential biases prior to starting a study. We wanted to evaluate a current hypothesis that concussion causes symptoms through decreases in binocular vision near point convergence (NPC). Because symptom resolution ranges from days-months, sampling at fixed timepoints is inefficient. Rather, if the NPC vs. symptom relationship is time-independent, we expect no bias with sampling based on milestones (SMile design): time of concussion, 50% improved symptoms, and healed. We created Oracle data (comprehensive Monte Carlo simulation study) using the same healing rate and linear slope for NPC vs. symptom for each participant. Bias was the difference of the average slope and the Oracle data slope (= 1). There was never bias when there was no measurement error. With multiplicative measurement error up to 50%, the overall bias was <5%. Grouped by initial symptom score, the slope was underestimated by ~30% for the mild (low symptom score) group, and overestimated by ~8% for the severe group (high symptom score). Bias increased in the mild group when we excluded participants who healed quickly. Bias decreased when measurement was delayed by days after the milestone. All these biases completely depend on observed error measurement. The SMile design may be useful when, within the same individual, two linearly related characteristics (ie. NPC and symptom scores) can be precisely measured over time. Further work will explore other assumptions.