Evaluating Misclassification Effects on Single Sequential Treatment in Sequential Multiple Assignment Randomized Trial (SMART) Designs

Sequential multiple assignment randomized trial designs tailor individual treatment by rerandomizing participants to subsequent therapies based on their response to initial treatment. Misclassification of participant responses to initial treatment can lead to inappropriate treatment assignment and thus impact the final outcome. The aim of this study is to derive a series of formulas for quantifying potential misclassification effects on the mean, variance, and statistical inference of a single sequential treatment (SST) effect with continuous outcome. Relative bias is expressed as a function of sensitivity, specificity, and the probability of being true responders. Results show that misclassification can introduce bias to the estimated treatment effect. Though the magnitude of bias varies, there are a few general conclusions: (1) for any fixed sensitivity (or specificity) the relative bias of the mean of responders (or nonresponders) always approaches 0 in a monotonic nonlinear pattern as specificity (or sensitivity) increases; (2) the relative bias of SST variance always has nonmonotone nonlinear relationship with sensitivity or specificity; (3) the SST variance under misclassification is always over-estimated. Furthermore, the results show that misclassification can affect statistical inference, with power exhibiting either monotonic or nonmonotonic patterns and resulting in either under- or over-estimation.

序贯多重分配随机试验（sequential multiple assignment randomized trial）设计会根据受试者对初始治疗的应答情况，将其再随机分配至后续治疗方案，从而实现个体化治疗。受试者对初始治疗应答的分类错误，可能导致治疗分配不当，进而对最终试验结局产生负面影响。本研究旨在推导一系列公式，以量化分类错误对连续结局下单序贯治疗（single sequential treatment, SST）效应的均值、方差及统计推断的潜在影响。相对偏倚可表示为灵敏度、特异度以及真正应答者概率的函数。研究结果表明，分类错误会对治疗效应的估计值引入偏倚。尽管偏倚的大小存在差异，但可得出若干一般性结论：（1）对于固定的灵敏度（或特异度），应答者（或非应答者）均值的相对偏倚，会随特异度（或灵敏度）的提升以单调非线性模式逐渐趋近于0；（2）单序贯治疗方差的相对偏倚，与灵敏度或特异度始终呈现非单调非线性关系；（3）存在分类错误时，单序贯治疗的方差始终被高估。此外，研究结果还显示，分类错误会影响统计推断，检验效能可呈现单调或非单调模式，并可能导致估计值出现低估或高估的情况。