I'm currently looking at a paper of Dirichlet process random effects model and the model specification is as follows: $$ \begin{align*}y_{i} &= X_{i}\beta + \psi_{i} + \epsilon_{i}\\ \psi_{i} &\sim G \\ G &\sim \mathcal{DP}\left(\alpha, G_{0}\right) \end{align*}$$ where $\alpha$ is the scale parameter and $G_{0}$ is the base measure. Later on in the paper, it suggests that we integrate a function over the base measure $G_{0}$ such as $$ \int f\left(y_{j}|\theta, \psi_{j}\right)\, dG_{0}\left(\psi_{j}\right).$$ Is the base measure in Dirichlet process a c.d.f. or is it a p.d.f.? What happens if the base measure is a Gaussian?


Denote by $\mathcal{M}$ a measurable space of probability measures, containing the realisations of the Dirichlet process. The random probability measure $G$ is a measurable function $$ G : \omega \mapsto G_\omega \in \mathcal{M} $$ and the integral with respect to $G$ is the random variable $$ \int f(\,\cdot\,|\, \psi) dG(\psi) : \omega \mapsto \int f(\,\cdot\,|\,\psi) dG_\omega(\psi). $$ Thus $\int f(\,\cdot\,|\, \psi) dG(\psi)$ is itself a random p.d.f. (if $f(\cdot| \psi)$ is a p.d.f.).

The idea is that $\psi_i$ follows some unknown distribution $G$. In some cases, you may have reasons to believe that $\psi_i$ is normally distributed and then put a prior on the mean and variance. In other cases, you don't want to make such parametric assumptions. In your model, for instance, the prior on $G$ is a Dirichlet process.

Is the base measure in Dirichlet process a c.d.f. or is it a p.d.f.?

The base measure is any probability measure, usually taken to have full support. In some cases, it can be represented by a probability density function. This is not very important.


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.