In Confidence regions when the Fisher information is zero - Matteo Bottai (2003) the effect of a singular information matrix on the confidence regions from statistical tests is getting looked at.

In the paper Bottai describes some models he uses to perform simulations. I would like to duplicate these simulations but I get stuck on how to generate the data. He says:

Model 1: Gaussian nonlinear regression. The data consist of $n$ independent observations $Y_i=(X_i,Z_i)$ of a bivariate vector $Y = (X, Z)$, with density: $$f(y) = f(x)f(z | x)$$ where $f(x)$ is fixed and known and $$f(z|x) = \phi(z-e^{\theta x}+\theta x)$$ where $\theta$ is unknown and lies in some open interval and $\phi(\cdot)$ denotes the density of a standard normal variable.

A bit further he denotes:

Figure 2 shows typical loglikelihoods obtained by pseudorandomly generating four samples from Model 1, where $n = 10000$, $\theta =0$ and $X$ is generated from a $Un(- 1, 2)$ distribution. Note that, despite the large sample size, two of the loglikelihood functions are bimodal.

Generating uniformly distributed data is no problem in R. But how can I generate this $Y$ (or the $Z_i$).

Also, I guess $f(x)$ can be chosen at random? Would something like $f(x) = \phi(x)$ be valid?

  • $\begingroup$ You have not correctly transcribed the model. Please fix the question $\endgroup$ – Glen_b -Reinstate Monica Apr 14 '17 at 1:50
  1. Note that you said $Y=(X,Z)$, so as soon as you have pairs of $X$ and $Z$ values, you have your $Y$ simply by putting them together.

  2. You already have $X$ values, since they were drawn from a uniform.

  3. note that $f(z|x) = \phi(z-\mu(x))$ specifies a normal density with mean $\mu(x)$ and standard deviation $1$.

  4. As a result you proceed as follows (after you specify your $\theta$), repeating many times:

    a. generate $X_i$ as you already have it (i.e. from a uniform); denote the observed generated value as $x_i$.

    b. generate $Z_i\sim N(\mu(x_i),1)$, call the observed generated value $z_i$.

    c. call the pair $(x_i,z_i)$ "$y_i$".

Also, I guess $f(x)$ can be chosen at random? Would something like $f(x) = \phi(x)$ be valid?

No, for the example $f(x)$ was specified to be uniform just above.

plots of z vs x for theta = 0, 0.8, 1.5 and -1; the first shows no relationship (random scatter about z=1), the other three show varying amounts of curved relationship according to the nonlinear function with random scatter about it


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.