另一种方法是使用平均误差作为因变量来估算第二阶段方程式中的时不变系数。
β and γt. For simplicity, let's forget about the year-effects. Define the estimation error u^it as before:
u^it≡yit−Xitβ^
The linear predictor u¯i is:
u¯i≡∑Tt=1u^iT=yit¯−x¯iβ^
Now, consider the following second stage equation:
u¯i=δmalei+ci
Assuming that gender is uncorrelated with unobserved factors ci. Then, the OLS estimator of δ is unbiased and time-consistent (this is, it is consistent when T→∞).
To prove the above, replace the original model into the estimator u¯i:
u¯i=x¯iβ−x¯iβ^+δmalei+ci+∑Tt=1ϵitT
The expectation of this estimator is:
E(u¯i)=x¯iβ−x¯iE(β^)+δmalei+E(ci)+∑Tt=1E(ϵit)T
If assumptions for FE consistency hold, β^ is an unbiased estimator of β, and E(ϵit)=0. Thus:
E(u¯i)=δmalei+E(ci)
This is, our predictor is an unbiased estimator of the time-invariant components of the model.
Regarding consistency, the probability limit of this predictor is:
plimT→∞u¯i=plimT→∞(x¯iβ)−plimT→∞(x¯iβ^)+plimT→∞δmalei+plimT→∞ci+plimT→∞(∑Tt=1ϵitT)
Again, given FE assumptions, β^ is a consistent estimator of β, and the error term converges to its mean, which is zero. Therefore:
plimT→∞u¯i=δmalei+ci
Again, our predictor is a consistent estimator of the time-invariant components of the model.