singular matrix是奇异矩阵的意思。

设A为n阶方阵，若存在另一n阶方阵B，使得AB=BA=I，则称A为非奇异矩阵，若不存在，则为奇异矩阵。

当exogenous variable 中虚拟变量过多，可能产生singular matrix或near singular matrix，表示这些变量间存在较大相关性。

在数据处理时，控制变量个数太多而样本量太小（损耗过多自由度，尤其是在时间序列中，若时间窗口较窄），都有可能出现奇异矩阵的问题。

要改变的话只有增加样本量或减少解释变量的个数。

[此贴子已经被作者于2008-4-23 18:02:51编辑过]

OLS estimator is a weighted average approach. The estimates depends on the independent variables, Xs (i.e., regressors; the ones on the right of the estimated equation). When some of these Xs are linearly dependent (or highly linearly dependent), the inverse of matrix (X'X) used in the OLS is not obtainable. The reason is its determinant is very close to zero (the cause of near singular matrix). The best way to solve the problem is to find these highly dependent variables and drop (exclude them from Xs) some of them. For example, "working experience" and "age" may be equally important to determine the wage people make. However, they basically reprenst similar things and "near singular matrix" phenonemon may happen by including both of them as the regressors. Drop either one would solve this problem. 

受益，我也遇到了这种情况，的确是我的样本量太小，不过，由于数据的可获取性，想增加样本数是不可行的，可能只有减少变量了

受益了，呵呵

谢谢，受益