We observe an i.i.d. sample $(X_1, Y_1), \ldots (X_n, Y_n).$ Let $m(x) = E(Y|X=x),$ $\sigma^2(x) = \operatorname{Var}(Y|X=x)$ and let $f(\cdot)$ be the density of $X.$

Under some regularity conditions, the conditional variance of the local linear estimator $\hat m(\cdot)$ of $m(\cdot)$ at $x_0$ is $$ \operatorname{Var}\{\hat m(x_0)|X_1,\ldots,X_n\} = \int\!\! K^2(u) du \, \frac{\sigma^2(x_0)}{f(x_0)} \, \frac{1}{nh} + o_P\left(\frac{1}{nh}\right),\tag{1} $$ as $h\to0$ and $nh \to \infty,$ where $K$ is the employed kernel.

Note that this expression involves the density $f$ of $X.$ Yet the expression is supposed to remain valid under fixed design.*

My understanding is that under fixed design, $X$ is not random but deterministic. But if $X$ is not random, what is its density? How can $(1)$ remain true under fixed design?

*My source for this is page 68 of Fan, J. and Gijbels, I. (1996). Local Polynomial Modelling and its Applications


If $x_0\in [0,1]$ and $X_i=i/n$, it is equivalent to $X_i\sim i.i.d. Uniform[0,1]$. So I believe that, in this case, $f(x_0)=1$.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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