Pike M C, Hill A P, Smith P G
Int J Epidemiol. 1980 Mar;9(1):89-95. doi: 10.1093/ije/9.1.89.
Two approaches to the estimation of the parameters in the logistic model applied to the analysis of stratum matched case-control studies are compared. It is shown that the "unconditional likelihood method' gives estimates of odds ratios which may be severely exaggerated if the stratum sizes are not large; however, if stratum sizes are not small the "conditional likelihood method' is computationally prohibitive. The use of small stratum sizes (e.g., 1-to-1 matching) leads to some loss of efficiency if the matching is irrelevant but it is shown that the loss is small for odds ratios up to at least 3. There would seem to be strong arguments for the routine use of matched designs as the reduction in confounding bias produced by such matching may be considerable whereas there is only a small loss in efficiency compared with unmatched designs.
比较了将逻辑模型应用于分层匹配病例对照研究分析时估计参数的两种方法。结果表明，如果层大小不大，“无条件似然法”给出的优势比估计可能会被严重夸大；然而，如果层大小不小，“条件似然法”在计算上是令人望而却步的。如果匹配不相关，使用小的层大小（例如1:1匹配）会导致一定的效率损失，但结果表明，对于至少高达3的优势比，损失很小。对于常规使用匹配设计似乎有充分的理由，因为这种匹配产生的混杂偏倚的减少可能相当可观，而与未匹配设计相比，效率损失很小。
