{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreidsgvquyltnel4ko7bmy6ryekd7k5wqdfx7awc5hsc7lafun56hdi",
    "uri": "at://did:plc:wwyqal4cnqhuwyacdj7rqq3n/app.bsky.feed.post/3mfutyeg25e32"
  },
  "path": "/t/never-seen-this-sort-or-primary-endpoint-and-analysis-in-an-rct/28564#post_17",
  "publishedAt": "2026-02-27T10:24:01.000Z",
  "site": "https://discourse.datamethods.org",
  "tags": [
    "@f2harrell",
    "@kertviele",
    "borrow",
    "static1.squarespace.com",
    "REMAP-CAP+Statistical+Analysis+Appendix+-+V4+-+05+NOV+2024_CLEAN.pdf"
  ],
  "textContent": "The Berries also use a slightly modified version of the Markov State Transition Model @f2harrell has proposed for REMAP-CAP (active since November 2024; see page 10 of the PDF below). So I doubt they think ordinal models are bad.\n\nI’ve also seen @kertviele argue against people dichotomizing ordinal scales such as the modified ranking scale, because that way you automatically declare states below, respectively above the cut-off to be equally bad / equally good. Is this worse than the common odds ratio being driven by the frequency of each category? I would tend to think so.\n\nLastly, what appealed to me as a clinician about the longitudinal ordinal model was how easily you can derive clinically relevant estimates. Such as differences in the time alive and out of hospital, or any contrast you really wish for. The issue is of course that under PO, the treatment effect, if present, will be biased because the most prevalent states drive the common odds ratio. On the other hand, as a clinician I’m also often happy to borrow some information for rare outcomes using the PO assumption, but that depends on the research area.\n\nstatic1.squarespace.com \n\n### REMAP-CAP+Statistical+Analysis+Appendix+-+V4+-+05+NOV+2024_CLEAN.pdf\n\n601.84 KB",
  "title": "Never seen this sort or primary endpoint and analysis in an RCT"
}