Centre for Biostatistics, Institute of Population Health, University of Manchester, 1st Floor, Jean McFarlane Building, Manchester M13 9PL, UK.
BMC Med Res Methodol. 2013 Jun 6;13:73. doi: 10.1186/1471-2288-13-73.
IVF treatments for infertility involve the transfer of multiple embryos in any one treatment cycle. When data is available on individual embryos the outcomes of each embryo are only partially observed, as treatment outcome (live birth) is assessed at the patient level. Two-level Embryo-Uterus (EU) models have been developed which assume a biologically plausible mechanism and assume that effects are mediated directly through the embryo (E) and also through the uterine environment (U), represented by two sub-models. This approach potentially allows inference as to the association of patient variables with outcome. However, when the variable is measured at the patient level either additional decisions have to be made in the modelling process as to in which sub-model the variable should be included or some model selection algorithm has to be invoked. These uncertainties have limited the practical application of these models.
We have conducted simulation studies based around realistic parameter values of situations where a putative patient-level variable is being considered for inclusion in an EU model and/or the mechanistic interpretation from the sub-model assignment is of interest. Firstly we explore various strategies for inference for a variable of interest where the sub-model is either pre-specified or considered unknown. Secondly we explore the use of information criteria to select the appropriate sub-model and the strength of evidence for that assignment. These are demonstrated in a reanalysis of a previously published dataset.
In the absence of prior evidence for potential prognostic factors measured at the patient level, two single degree-of-freedom likelihood ratio tests with a Bonferroni correction including the variable of interest in first the E then the U sub-model performs well as a statistical test for association with outcome. For model building the information criteria can be used, but large differences are required (≥ 6) to provide reasonable evidence of sub-model assignment. Previous interpretations have been over-optimistic.
These results suggest simple strategies and should enable these models to be used more confidently in practical applications.
体外受精（IVF）治疗不孕涉及在一个治疗周期中转移多个胚胎。当有关于单个胚胎的数据时，每个胚胎的结果仅部分观察到，因为治疗结果（活产）是在患者层面评估的。已经开发了两级胚胎-子宫（EU）模型，这些模型假设了一种生物学上合理的机制，并假设效应是通过胚胎（E）直接介导的，也通过子宫环境（U）介导，由两个子模型表示。这种方法有可能推断出患者变量与结果的关联。然而，当变量在患者层面测量时，在建模过程中必须做出额外的决策，以确定该变量应该包含在哪个子模型中，或者必须调用一些模型选择算法。这些不确定性限制了这些模型的实际应用。
我们根据实际的参数值进行了模拟研究，这些参数值涉及到正在考虑将一个假定的患者层面变量纳入 EU 模型的情况，以及子模型分配的机制解释是否相关。首先，我们探索了各种策略，用于推断感兴趣的变量，其中子模型要么是预先指定的，要么被认为是未知的。其次，我们探索了使用信息准则来选择适当的子模型以及该分配的证据强度。这些在对先前发表的数据集的重新分析中得到了证明。
在没有预先存在的患者层面潜在预后因素的证据的情况下，对于感兴趣的变量，首先将其包含在 E 子模型中，然后在 U 子模型中进行两次单自由度似然比检验，并进行 Bonferroni 校正，作为与结果关联的统计检验，效果良好。对于模型构建，可以使用信息准则，但需要较大的差异（≥6）才能提供子模型分配的合理证据。以前的解释过于乐观。
这些结果表明了简单的策略，应该使这些模型在实际应用中更有信心地使用。
