Processing math: 100%

1.2. Mean Estimation in the Binary Choice Problem

Another motivating example can be found in the estimation of mean of the binary choice model.


Mean estimation in the binary classification

Our model is from the (case 2) of the previous article.

P(Y=1|Z=z)=F0(z), ZQ, F0Λ,Λ:={F:R[0,1], F is increasing}.

This time the estimand is the population mean

θ0=θF0:=F0(z)dQ(z)=EF0(Z)=P(Y=1)=EY.

By definition, it is clear that E[YiF0(Zi)]=0.

We will only consider the case where Q is known. In this case, the MLE of F0 becomes

ˆθn=θˆFn=ˆFn(z)dQ(z),

where ˆFn is the MLE of F0.

For FΛ, define

θF=F(z)dQ(z),

then by the classical central limit theorem, we get

1nni=1(F(Zi)θF)dN(0,Var(F(Z))).

Again, if we prove some form of functional CLT for all FΛ, then together with the result form the classical CLT, we get the asymptotic distribution of the MLE.

1nni=1(ˆFn(Zi)ˆθn)=1nni=1(F0(Zi)θ0)+oP(1)n(ˆθnθ0)dN(0,F0(z)(1F0(z))dQ(z)).

We will take the fact 1nni=1ˆFn(Zi)=ˉY for granted. By this fact, n(ˆθnθ0)=n(ˉYθ0)n(ˉYˆθn)=n(ˉYθ0)1nni=1(ˆFn(Zi)ˆθn)=n(ˉYθ0)1nni=1(F0(Zi)θ0)+oP(1)=1nni=1(Yiθ0F0(Zi)+θ0)+oP(1). By the classical CLT, 1nni=1(YiF0(Zi))dN(0,Var(YF0(Z))). The variance can be rewritten as Var(YF0(Z))=E[YF0(Z)]2=E[Y22YF0(Z)+F20(Z)]=E[Y2F20(Z)]=EYEF20(Z)=EF0(Z)EF20(Z). Hence the result follows from the Slutsky's theorem.


References

  • van de Geer. 2000. Empirical Processes in M-estimation. Cambridge University Press.
  • Theory of Statistics II (Fall, 2020) @ Seoul National University, Republic of Korea (instructor: Prof. Jaeyong Lee).