※ 引述《wwwwwww (哪个王八蛋一天上十九次됩》之铭言： : ※ 引述《liton (欧吉桑留学生)》之铭言： : : 这些该念的我都念过了 : : 我是对Time Series 和Cross Section的不同处理方式有疑问 : : 在CrossSection中X=alpha+a*Y+b*Z : : Y和Z的相关性很高的话 : : 我们会用instrument variables等方法来处理 : : 但在AR中X=alpha+a*X(-1)+b*X(-2) 如果ACF和PACF很高的话 : : 我们反倒觉得变数自己的递回性很高 : : 用该变数自己的历史资料便可预测下一期的X : : 那这样不就代表Corr[X,X(-1)]或Corr[X,X(-2)]会很高 : : 在Cross Section中 这是个很严重的问题 : : 但在Time Series中 这怎反倒变成是一个很好的性质? : Instrument variables is mainly used to deal with the difficulty : that the explanatory variables and error terms are correlated. : AR models have no such difficulty. : But ARMA models do have and can be treated by instrument variables. : For example, in the ARMA(1,1) case, you cannot get a consistent estimator of : AR coeff. by regressing x_{t} on x_{t-1}. : But you can get a consistent estimator of the AR coff. by regressing : x_{t} on x_{t-2}. Now x_{t-2} is the instrument variable. Well, I just take one example to overcome the correlation problem in cross section. In practice, there are many methods to handle with the problem. For example, I can drop the independent variables in the regression. A best practice for cross section model always includes testing the correlation between independend variables. The key is that correlation in cross section is a serios problem, no matter in theories or practice. Correlation will result in at least three kinds of trouble: 1.measurement error or errors in variables 2.endogeneity 3.omitted variables However, it seems that time series care more about unit roots. I think that's another problem in time series, but the unit roots theory has not resolve the correlation problem. --

※ 发信站: 批踢踢实业坊(ptt.cc) ◆ From: 59.115.53.251