3
$\begingroup$

Suppose I have an i.i.d. sample $\{(Y_i, X_i)\}_{i=1}^n$ on which I am trying to estimate a conditional expectation model:

$$Y = g(X) + \varepsilon,\quad \mathbb E[\varepsilon | X] = 0$$

There is a lot of literature on how different estimators of $g$ will imply different rates of convergence to the true conditional expectation function for different regularity assumptions. However, for my application, I am not interested in $g$ directly but instead care about a single summary statistic about $g$, namely

$$\theta = \mathbb E[(g(X) - \mathbb E[Y])^2] \quad (\text{alternatively}, \eta = \mathbb E[\varepsilon ^2])$$

My question is the following: are there settings in which $\theta$ or $\eta$ can be estimated at a faster rate than $g$ itself?

$\endgroup$

1 Answer 1

1
$\begingroup$

We get some interesting results when we ask this question within a semi-parametric statistics framework. In particular, I will focus on estimating $\eta$, since by the law of total variance, once we can estimate $\eta$, we can estimate $\theta = \mathbb V\mathrm{ar}[Y] - \eta$. Define the influence function for $\eta$ as

$$\psi_i = (Y_i - g(X_i))^2$$

so that if $g$ was known, then we could easily estimate $\eta$ via

$$\tilde\eta = \frac1n\sum_{i=1}^n \psi_i$$

However, we have to estimate $g(X_i)$. Therefore, a reasonable "plug-in" estimate of $\psi_i$ would be $\hat\psi_i = (Y_i - \hat g(X_i))^2$ and our updated estimate of $\eta$ would be

$$\hat\eta = \frac1n\sum_{i=1}^n \hat \psi_i$$

It is a simple exercise to check that $\psi_i$ is first order insensitive to $g$ around the true $g$ in the sense of the Gateaux derivative:

$$\partial_r \mathbb E[\psi(g + r(h - g))] = - \mathbb E[(h(X)-g(X))\cdot (Y - g(X))] = 0$$

This is the "Neyman Orthogonality" condition as described, for example, in Chernozhukov et al (2017). It implies that even if $g$ is not estimated at the parametric $O\left(n^{-1/2}\right)$ rate, it is possible that $\hat\eta$ will not only converge to $\eta_0$ at the parametric rate, but that it will be asymptotically normal with feasibly estimated standard errors. In particular, assuming $g$ is smooth enough that it is estimated at a rate $o\left(n^{-1/4}\right)$, then we should expect that (with some sample-splitting):

$$\frac1{\hat\sigma \sqrt n}\sum_{i=1}^n \hat \psi_i - \eta_0 \Rightarrow \mathcal N(0,1),\quad \hat\sigma = \frac1n\sum_{i=1}^n \left(\hat \psi_i - \frac1n\sum_{i=1}^n \psi_i\right)^2$$

The $O\left(n^{-1/2}\right)$ convergence is not too surprising since we mechanically have that if $\hat g \to g$ in $L^2$ norm, we would have that the norm squared (which is essentially what $\hat\eta$ is estimating) must be converging at twice the rate, although the asymptotic normality is a potentially useful addition.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.