St-Pierre N R
Department of Animal Sciences, The Ohio State University, Columbus 43210, USA.
J Dairy Sci. 2001 Apr;84(4):741-55. doi: 10.3168/jds.S0022-0302(01)74530-4.
In animal agriculture, the need to understand complex biological, environmental, and management relationships is increasing. In addition, as knowledge increases and profit margins shrink, our ability and desire to predict responses to various management decisions also increases. Therefore, the purpose of this review is to help show how improved mathematical and statistical tools and computer technology can help us gain more accurate information from published studies and improve future research. Researchers, in several recent reviews, have gathered data from multiple published studies and attempted to formulate a quantitative model that best explains the observations. In statistics, this process has been labeled meta-analysis. Generally, there are large differences between studies: e. g., different physiological status of the experimental units, different experimental design, different measurement methods, and laboratory technicians. From a statistical standpoint, studies are blocks and their effects must be considered random because the inference being sought is to future, unknown studies. Meta-analyses in the animal sciences have generally ignored the Study effect. Because data gathered across studies are unbalanced with respect to predictor variables, ignoring the Study effect has as a consequence that the estimation of parameters (slopes and intercept) of regression models can be severely biased. Additionally, variance estimates are biased upward, resulting in large type II errors when testing the effect of independent variables. Historically, the Study effect has been considered a fixed effect not because of a strong argument that such effect is indeed fixed but because of our prior inability to efficiently solve even modest-sized mixed models (those containing both fixed and random effects). Modern statistical software has, however, overcome this limitation. Consequently, meta-analyses should now incorporate the Study effect and its interaction effects as random components of a mixed model. This would result in better prediction equations of biological systems and a more accurate description of their prediction errors.
在畜牧业中，理解复杂的生物、环境和管理关系的需求日益增加。此外，随着知识的增长和利润率的缩小，我们预测各种管理决策反应的能力和愿望也在增强。因此，本综述的目的是帮助展示改进的数学和统计工具以及计算机技术如何能够帮助我们从已发表的研究中获取更准确的信息，并改进未来的研究。在最近的几篇综述中，研究人员从多个已发表的研究中收集数据，并试图构建一个能最好地解释这些观察结果的定量模型。在统计学中，这个过程被称为荟萃分析。一般来说，不同研究之间存在很大差异，例如，实验单位的生理状态不同、实验设计不同、测量方法不同以及实验室技术人员不同。从统计学角度来看，研究是区组，其效应必须被视为随机的，因为我们寻求的推断是针对未来未知的研究。动物科学中的荟萃分析通常忽略了研究效应。由于跨研究收集的数据在预测变量方面不均衡，忽略研究效应的结果是回归模型参数（斜率和截距）的估计可能会出现严重偏差。此外，方差估计向上偏倚，导致在检验自变量效应时出现较大的II类错误。从历史上看，研究效应被视为固定效应，并非因为有强有力的论据表明这种效应确实是固定的，而是因为我们之前无法有效地求解哪怕是中等规模的混合模型（那些包含固定效应和随机效应的模型）。然而，现代统计软件已经克服了这一限制。因此，荟萃分析现在应该将研究效应及其交互效应纳入混合模型的随机成分中。这将产生更好的生物系统预测方程，并更准确地描述其预测误差。
