{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreibpvdzktmc373lffzmrptykad2e5kicb4krqy2rrsaw5ixebyly5u",
    "uri": "at://did:plc:wwyqal4cnqhuwyacdj7rqq3n/app.bsky.feed.post/3mkibdwpze6k2"
  },
  "path": "/t/experimenting-with-a-bayesian-partial-proportional-odds-model/28677#post_5",
  "publishedAt": "2026-04-27T12:16:44.000Z",
  "site": "https://discourse.datamethods.org",
  "textContent": "Phil I’m so glad you joined the discussion! I wish we could give an award for helpfulness. I think you’ve won it for this year.\n\nWe often think of X affecting an ordinal scaled response Y as shifting patients along the Y spectrum. When adding non-PO to the fullest, e.g., adding k-2 degrees of freedom for a binary X where Y has k levels, the X effect can be in any direction for any level of Y. When the non-PO is constrained, you’re showing that there are constraints on what X is allowed to do for a specific Y even if the constraint that is lifted is targeted exactly at that level of Y. That was the wakeup call for me, and I still cling to hoping that the constraint is not too severe. The example being discussed here raises concern for me.\n\nIn going to the “no constraints for death” dual models you propose it would be nice if we could constrain all the covariate effects to be the same for both models, other than the one target covariate (such as treatment). That would bring stability as well as ease of getting a predicted unconditional quantity. On the other hand, risk factors for nonfatal outcomes can be much different from those for death in many situations.\n\nWhen using a conditional independence argument as you are using (which is what Kaplan-Meier and Cox PH use) I assume that this is very easy to justify. We do need to combine the models to get causal estimates, i.e., intent to treat estimands such as P(good function and alive | X). I assume that is not a problem, right? Except for confidence intervals being complex, perhaps (but Bayesian uncertainty intervals probably come straight from posterior sampling from a joint stream from the two models).",
  "title": "Experimenting with a Bayesian Partial Proportional Odds Model"
}