# Feature selection on a Bayesian hierarchical generalized linear model

I am looking to estimate a hierarchical GLM but with feature selection to determine which covariates are relevant at the population level to include.

Suppose I have $G$ groups with $N$ observations and $K$ possible covariates That is, I have design matrix of covariates $\boldsymbol{x}_{(N\cdot G) \times K}$, outcomes $\boldsymbol{y}_{(N\cdot G) \times 1}$. Coefficients on these covariates are $\beta_{K \times 1}$.

Suppose $Y$~$Bernoulli(p(x,\beta))$

The below is a standard hierarchical bayesian GLM with logit sampling model and normally distributed group coefficients.

$${\cal L}\left(\boldsymbol{y}|\boldsymbol{x},\beta_{1},...\beta_{G}\right)\propto\prod_{g=1}^{G}\prod_{t=1}^{N}\left(\Pr\{j=1|p_{t},\beta^{g}\}\right)^{y_{g,t}}\left(1-\Pr\{j=1|p_{t},\beta^{g}\}\right)^{1-y_{g,t}}$$

$$\beta_{1},...\beta_{G}|\mu,\Sigma\sim^{iid}{\cal N}_{d}\left(\mu,\Sigma\right)$$

$$\mu|\Sigma\sim{\cal N}\left(\mu_{0},a^{-1}\Sigma\right)$$ $$\Sigma\sim{\cal IW}\left(v_{0},V_{0}^{-1}\right)$$

I want to modify this model (or find a paper that does, or work that discusses it) in such a way that there is some sharp feature selection (as in LASSO) on the dimensionality of $\beta$.

(1) The simplest most direct way would be to regularize this at the population level so that we essentially restrict the dimensionality of $\mu$ and all $\beta$ have the same dimension.

(2) The more nuanced model would have shrinkage at the group level, where dimension of $\beta$ depends on the hierarhical unit.

I am interested in solving 1 and 2, but much more important is 1.

The way I'd tackle (1) would be involve a spike and slab model something like:

$\beta_{g,k} = z_{k}m_{g,k}$

$z_k \sim Bern(p)$

$m_{g,k} \sim N(\mu, \Sigma)$

$\mu, \Sigma \sim NIW_{v_0}(\mu_0, V_0^{-1})$

This:

• Retains the flexibility on the $\beta$'s from the NIW prior on $\mu, \Sigma$.
• Models selection of variables for all groups at once.
• Easily extensible by adding a sub-index for group to $z_{g,k}$ and having a common beta prior for each location $k$.

Of course, I think this is the kind of problem where there are a number of valid approaches.

Feature selection is not a great goal to have in an analysis. Unless all the predictors are uncorrelated with each other and your sample size is immense, the data will be unable to reliably tell you the answer. Model specification is more important than model selection. Details are in my RMS Course Notes. But shrinkage, without feature selection (e.g., ridge or $$L_{2}$$ penalized maximum likelihood estimation) can be a good idea. Hierarchical Bayesian models are even better because they allow for statistical inference in the shrunken model whereas we lose most of the inferential tools in the frequentist world after shrinking.