Weighting Types

[mbeckr18]

Hello,

I am currently trying to run a WLS regression. For the weight series I like to use the according t-statistics. However, I am unsure which weighting type I should use in this case: the "inverse variance" or the "inverse std. dev.". Could you explain what the different weighting types actually do? (Regarding the scaling options: I used "average" and "E-Views default" for "inverse variance" and "inverse std. dev." respectively) Thank you very much for your help in advance.

[backhaum]

Dear Gareth / Glenn,

I actually have a similar problem.
I am trying to estimate a 2nd stage regression (meaning that my dependent variable is an estimated coefficient from single regressions).
Hence, I would like to run a WLS regression with weights according to the estimation error (either t-values/std error, etc). But the handbooks do not explain really what the differences are regarding the weighting types in the EViews menu. The significance levels change very heavily depending on wether I use weighting type proportional to "inverse variance" or "inverse std. deviation". If I want to weight with t-values, which one would be the correct weighting type?

Additionally, what does EViews default scaling actually mean?

Thank you for your help and best regards

Max

[backhaum]

I am using EViews 7 by the way

[EViews Glenn]

The different settings correspond to how the weights are applied. The easiest way to think about is that the weighting rescales the variances of the error in a different way depending on how you define the weights. If the weights are standard deviation weights, then we effectively divide the data by the weights so that the residual variance is scaled by the inverse of the square of the weights. Alternately, if the weights are inverse standard deviation weights, then we multiply the data by the weights so that the residual variances is scaled by the square of the weights.

EViews weights is a special form of the inverse standard deviation weights in which we scale the entire weight series by a constant to improve numeric stability. It only has an effect in the computation of the weighted summary descriptive statistics reported on the bottom of the output. We added the regular inverse standard deviation weights (even though it yields basically the same results) for people who wanted to match other computations.

As to your particular application, I'm not certain what you should choose as it's not clear what your model of the residual variance is. It sounds like you want to do some sort of importance weighting, but I'm not entirely clear how that would work or what you expect the effect of the weighting to be.