把第二个“真模型”写为
y=β1x+β2z+e,
若在估计时遗漏了 z,回归 y 对 x 得到的系数为
β=β1+β2Var(x)Cov(x,z)
(假设 Var(x)>0)。
因此二者差异是遗漏变量偏差:
β−β1=β2Var(x)Cov(x,z).
仅当 β2=0 或 Cov(x,z)=0 时,才有 β=β1。
英文解析
Write the second “true model” as
y=β1x+β2z+e,
If z is omitted from the estimate, the coefficient obtained for returning y to x is
β=β1+β2Var(x)Cov(x,z)
(Let's say Var(x)>0).
So the difference between the two is the missing variable bias:
β−β1=β2Var(x)Cov(x,z).
β=β1 is only available whenβ2=0 orCov(x,z)=0.