W05_lecture04_模型推广北大光华计量.pptxVIP

第四章．古典线性回归模型的推广 Outline 数据问题 异方差问题 时间系列数据问题 第一节、数据问题 多重共线性问题 奇异观测值问题 1.多重共线性 Perfect multicollinearity is where there is exact linear relationship between explanatory variables included in the regression Perfect multicollinearity never happens in practice, unless you fall in a dummy variable trap But you may encounter some degree of multicollinearity Feature of sample, not populations 最可能发生的情形：多重交叉变量 比如，不同企业在金融危机前后的saving propensity的不同 一般尽量避免三重交叉 后果：精度、稳定性 因为参数估计方差为 Informal assessment of Multicollinearity (Signs of collinearity): rules of thumb Correlation matrix of regressors, 0.75 (say) Very large SEs Contradiction between t-and F-statistics t-statistics individual coefficient(s) not significant F-statistic coefficient(s) jointly significant Take out one variable and re-estimatebing changes in the value of coefficient(s) Formal assesment: VIF=1/(1-R2) ,一般认为10没问题 proc reg data=one; model eret=beta beta2/vif; run; What to do about multicollinearity Using prior information to explicitly restrict the parameters Sample problem: get more data! multicollinearity = problem of lack of data Time series: 1st differencing the data 1st differences usually less correlated Drop some control variables But do not drop key variables! –Omitted variables problem Ridge regression Principal components Form linear combination of correlated variables such that it maximizes its variability Drawback: may lack economic reasoning Do nothing if it is significant and stable (Woodridge) 做一些估计对数据稳健性的检验，比如分成几个小样本、去掉几个观察值等。 2. 奇异值问题 观察到的数据中可能有些奇异值，即与大多数观测值太不一致 可能是由于数据采集错误 可能是上帝疯狂了几次 比如，作业1中的资产负债率数据，最大值与99%分位数差很远，即有1%的观察值与大多数观测值太不一致。 Exer3.sas PROC IMPORT OUT=two DATAFILE= F:\current_teaching\econometrics_II_2\data\FS_Combas1 DBMS=EXCEL REPLACE; SHEET=FS_Combas1; GETNAMES=YES; MIXED=YES; USEDATE=YES; RUN; PROC IMPORT OUT=one DATAFILE= F:\current_teaching\econometrics_II_2\data\FS_Combas DBMS=EXCEL REPLACE; SHEET=FS_Combas; GETNAMES=YES; MIXED=YES; USEDATE=YES;