Sources of bias in single-trial normalization procedures
Sources of bias in single-trial normalization procedures
复制标题
DOI：
10.1111/ejn.13179
复制
发表时间：
2016-04-01
影响因子：
3.4
通讯作者：
Muresan, Raul C.
中科院分区：
文献类型：
--
作者：
Ciuparu, Andrei;Muresan, Raul C.
基线标准化程序对于分析大脑活动是必不可少的。这些使用参考（基线）期的统计数据对整个试验（基线期和刺激期）的沿着数据进行标准化。一个非常流行的过程是伪z评分，传统上应用于时间频谱功率估计，最近显示它会产生正偏差。偏倚被认为是由于光谱功率值的偏斜分布引起的离群值而产生的。在这里，我们挑战这种观点，因果关系表明，偏见源于一个更普遍的问题，影响了广泛的规范化技术，包括一些常规使用。我们表明，偏见是由相关条款的划分，它直接取决于分子和分母之间的相关性的符号和幅度。相关性要么来自被归一化的数据的属性，要么来自归一化方法的属性。当源数据具有偏斜分布时，z评分会产生偏倚，但当分布对称时，z评分无偏倚，而用于荧光数据的dF/F等方法会导致偏倚，因为分子和分母固有相关。我们提供了一个简单、快速和通用的解决方案，通过将多个试验的基线期焊接（融合）到一个单一的大基线中来减少甚至消除偏差。这种方法是通用的，可用于归一化单个试验，并在足够长的扩展基线下提供无偏倚估计。我们表明，基线融合是上级更复杂的技术，已经提出了。
Baseline normalization procedures are essential for the analysis of brain activity. These use statistics of a reference (baseline) period to normalize data along the entire trial (baseline and stimulus periods). A very popular procedure is pseudo z-scoring, traditionally applied to time-frequency spectral power estimates, where it was recently shown to generate positive bias. Bias was thought to arise because of outliers stemming from the skewed distribution of spectral power values. Here we challenge this view and causally show that bias originates from a more general problem that affects a wide array of normalization techniques, including some that are routinely used. We show that bias is caused by the division of correlated terms and that it depends directly on the sign and magnitude of correlation between the numerator and denominator. Correlation emerges either from the properties of the data being normalized or from the properties of the normalization method. z-scoring produces bias when source data have a skewed distribution but it is bias-free when the distribution is symmetric, while methods such as dF/F for fluorescence data lead to bias because the numerator and denominator are inherently correlated. We provide a simple, fast and general solution to reduce and even eliminate bias by welding (fusing) baseline periods of multiple trials into a single, large baseline. This method is generic, can be used to normalize individual trials and provides bias-free estimates given a long enough extended baseline. We show that baseline fusing is superior to more complex techniques that have been proposed before.