急求！！怎样消除模型中的异方差？

如题。我在毕业论文里研究的是汇率问题，建立了多元线性回归模型。但是由于汇率的时间序列数据之中有0的存在，所以无法用ln模型来消除异方差。我非常希望能采用加权最小二乘的方法来消除异方差，但是权数的确定却毫无头绪。哪位大侠能否指点一二？该怎么确定权数？？

此问题非常紧急，希望好心的各位能尽快帮上忙！！

感激不尽！！！！！！

Do not panic about heteroskedasticity. Although it violates one the key assumptions of OLS (ordianry least square) model, but the estimates from the model is still unbiased and consistent, although the standard errors of the estiamtes are biased.

First, you need to make sure the existence of heteroskedasticity, by using White test, for example.
Second, you can just simply use robust standard error to correct heteroskedasticity. (I am not sure what statistical package you are using for your research, Stata can easily handle it.)
Last, try to use Generalized Least Square (GLS) to correct it.

Hope this helps.

2# dw99

我用EViews检验出来确实有异方差，并且需要修正。
就我现在所掌握的知识里就只有模型变换和加权最小二乘这两种方法来消除异方差，但是因为我的数据（是时间序列数据）里含有0，所以无法按一般的权数（如1/x）来加权。所以希望能告诉我这种情况下该怎么做。

Thank you all the same for your reply.

推荐用 robust-hetero~估计，可以得到方差的一致估计，实在想用GLS，就FGLS，ln(残差^2)=f(x),估计残差^2=exp(f(x)),然后wls

Agree with JNGOD.
Using robust standard error might be the easiest (laziest) way to adjust heteroskedasticity. (sometime, it is called White's approach.) In stata, you just add an option ',robust' at the end of regression command. I know nothing about Eviews, so no comments.

Another technique you can consider is to add a small positive number (0.00001) to your response variable to avoid the issue with log transformation. I use it when I deal with cost data.

Lastly, you need to double check on your model specification. As you know, heteroskedasticity is often caused by misspecification, omitted terms, and outliers. If you are using the model from someone else, try to be dubious. If it is totally new, then be very cautious.

Best luck with your research

学习了。。。