关于fmri里的建模，GLM的first level analysis 和second level analysis

Introduce the first level analysis and second level analysis in GLM (genereal linear model )
GLM is used in time series mostly.
ANNOTION: GLM IS USED FOR TASK DESIGN ONLY.
description from wiki:
Following pre-processing, data analysis is often carried in two steps: a separate first-level analysis of data from each individual subject, followed by a second-level analysis in which results from multiple subjects are combined.

Simply speaking, the first-level analysis could be known as the analysis in single-subject, most of this type trials are task fmri design. So that the second level analysis is the analysis in multiple subjects, or can be told the group analysis of single subjects.

For GLM
GLM, general nilearn model.
For single subjec, to investigate the association between volumes and predictor variables, we set up a GLM model to fit that. The format are as follow:

1. y1 - represents y label you named in first time point;
2. X1.1 - the former 1 represents time point, and the latter 1 represents volume 1;
3. β1 - the weight contribution of X1.1 volume for y label prediction; ( beta )
4. ε1 - the constant item, which represents distance between predict y and truly y, it also be known as error. ( epsilon )

The aim of the first-level statistical analysis is to determine how large the contribution of each predictor variable xi is to the observed values of y.

GLM model could be simplified as follow:
Y = Xβ + ε
annotation:

1. Y is an nx1 colmn vevtor and the length equal to time series.
2. X is an nxp matrix, and it always include confunding factors, such as nuisance, head motion.
3. β is an px1 vector of unknown weights setting.
4. ε is an nx1 vector containning the error values in each observation.
   visualize:
   
   the goal in fittting X~Y is to minimize the sum of squared residuals:
   
   second level analysis
   Once single-subject data has been analyzed for a set of participants, individual results are aggregated to assess commonality and stability of effects within or across groups of interest.
   second level analysis can be understood easily as the multiples subjects analysis.
   REMENBER: Prior to group analysis (second analysis), individual results should be warpped into a common reference space ( Talairach or MNI152 ).

FIXED EFFECTS 固定效应
classical statistical hypothesis testing proceeds by comparing the difference between the observed and hypothesized effect against the “yardstick” of variance (Holmes and Friston, 1998).
In a fixed effects (FFX) analysis, the variance considered is that derived from scan-to-scan measurement error, and represents the within-subject variability (sw 2 ). 在固定效应中，变量考虑来源于扫描方式的误差，即被试内差异。
这个固定效应模型感觉一般不常用。其原因是仅在所分析的样本水平上作推论，但是没有考虑到抽样误差，即被试间差异，固定效应假设每个受试者在同一状态下激活模式是相同的。这对于推广健康受试者的核磁标准是难以实行的，所以不常用。
RANDOM EFFECTS, MIXED EFFECTS, AND SUMMARY STATISTICS
随机效应，混合效应，和汇总统计

random and mixed effects 随机和混合效应
要在总体水平上应用推论，就要考虑到被试间有差异这个观点。在随机效应模型中（random effects），统计测试评估这种效应的大小相对于受试者之间的变异性是否显著。这种误差项包括被试的基本差异，不同被试对刺激产生的不同血流动力学及神经反应，头颅解剖上的差异。当然，除非单个被试的β已经知道，否则不能得出精确的随机效应推断，只能是评估。
混合效应模型就是包括session内的固定效应和session间的随机效应，综合起来就是混合效应，mixed effects.
Summary statistics 汇总统计：
A hierarchical two-level linear model of (fMRI) data analysis can be written as follows 数据分析的分层两级线性模型


the error term h is not equal to the random effect component eG, but contains a mixture of both the within and between variability (hence the characterization as a MFX).
Finally, a different approach is to derive b-estimates from a GLM but then assess significance of spatial distribution of activations, rather than individual voxels . This strategy, by switching from a massive-univariate to a (local) multivariate approach has the promising advantage of assessing patterns of information representation, rather than localization of information, something that may be of great interest from a cognitive neuroscience point of view.
最后，另一种方法是从GLM中得出b估计值，然后评估激活的空间分布的重要性，而不是单个体素。这种策略，通过从大规模单变量方法切换到（局部）多变量方法，具有评估信息表征模式的潜在优势，而不是信息的本地化。