Walker Ian, Milne Sarah
Department of Psychology, University of Bath, Bath, England.
Behav Res Methods. 2005 Feb;37(1):23-36. doi: 10.3758/bf03206395.
Most forms of regression analysis make assumptions about the relationships between the variables being modeled. As a consequence, it can be difficult to know which form of analysis is most appropriate for a given data set. In this article, we explore the idea that function estimators might provide a better alternative in many situations. Function estimators discover the best function to link dependent and independent variables, no matter what form this takes. Four studies demonstrate that one type of function estimator (a neural network) not only performs the same tasks as linear regression and nonlinear regression, but often performs these tasks better and with more flexibility. Moreover, neural networks allow a useful secondary analysis in which useful groups of people can be identified. We recommend that function estimators be used in preference to regression-based techniques for many analyses. The Matlab script used to write this article may be downloaded from www.psychonomic.org/archive/.
大多数回归分析形式都对所建模的变量之间的关系做出假设。因此，很难知道哪种分析形式最适合给定的数据集。在本文中，我们探讨了函数估计器在许多情况下可能提供更好替代方案的观点。函数估计器会发现连接因变量和自变量的最佳函数，无论其形式如何。四项研究表明，一种函数估计器（神经网络）不仅能执行与线性回归和非线性回归相同的任务，而且通常能更好地执行这些任务，并且具有更大的灵活性。此外，神经网络允许进行有用的二次分析，从中可以识别出有用的人群组。我们建议在许多分析中优先使用函数估计器而非基于回归的技术。用于撰写本文的Matlab脚本可从www.psychonomic.org/archive/下载。
