{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreib7d7gvegu2khfaru343xjxwbtyhwlvy2svpcqu2o6xs6racmntvi",
    "uri": "at://did:plc:wwyqal4cnqhuwyacdj7rqq3n/app.bsky.feed.post/3mj7lan4hv6r2"
  },
  "path": "/t/bayesian-predictive-projection-for-variable-selection/28620#post_14",
  "publishedAt": "2026-04-10T10:39:05.000Z",
  "site": "https://discourse.datamethods.org",
  "tags": [
    "rob-mcculloch.org",
    "nonlinvarsel.pdf"
  ],
  "textContent": "This might also be of interest:\n\nrob-mcculloch.org\n\n### nonlinvarsel.pdf\n\n552.07 KB\n\n> This paper describes a practical procedure for Bayesian variable selection in non- linear regression and classification models. A first stage model is fit in which all variables are included. Typically, this first stage will not include the prior belief that only a small subset of variables is needed, but it may. Given this first stage fit, we look for functions of variable subsets which approximate the predictions from the first stage fit well. A computationally efficient surrogate model is used to search for approximating functions which depend on low numbers of predictors. Rather than assuming there is some true sparcity, we seek sparse approximations to the non-sparse truth. In the case that our first stage fit involves a Bayesian assessement of the uncertainty, we use this to gauge the uncertainty of our approximation error. If we learn that, with high probability, we can obtain a good approximation to the non-sparse truth using a subset of the variables, we deem that subset to be of inter- est. We demonstrate the procedure in empirical examples involving prediction and classification and simulated examples.",
  "title": "Bayesian predictive projection for variable selection"
}