{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreic36ctdpnuwclcb64anegmjlnfizk2fs2yqe2jucvqmmsiiupnfii",
    "uri": "at://did:plc:wwyqal4cnqhuwyacdj7rqq3n/app.bsky.feed.post/3milb3khaevd2"
  },
  "path": "/t/collider-in-rct-subgroup-analysis/28689#post_7",
  "publishedAt": "2026-04-03T06:52:29.000Z",
  "site": "https://discourse.datamethods.org",
  "textContent": "The DAG encodes **our own** (not other peoples) assumptions and I do not really agree with these authors. Assuming SG is a categorical variable with two levels (present, absent), the DAG on the left below would imply effect modification when presence of SG mediates the effect in which case the direct effect could be zero or partial effect when SG is absent and U then is important. The DAG on the right is the far more common scenario (99%) where SG is a collider (artefact of the sample) in which case U is just prognostic for the outcome.\n\nAddendum: What the conventional framework calls mediation is really induced mediation where T both activates the mechanism and the mechanism exists because of T. What I am describing is better called _conditional mediation_ or a _facilitated mechanism_ , the mechanism exists independently, but T’s effect is entirely channeled through it. Both are legitimately mediation conceptually. The difference is entirely about whether the decomposition arithmetic works, not about whether SG deserves to be called a mechanism. The field has perhaps been too quick to let the limitations of its estimation tools define its concepts.",
  "title": "Collider in RCT Subgroup Analysis"
}