# Critique of paper on generalizability of oncology trials

**URL:** <https://discourse.datamethods.org/t/critique-of-paper-on-generalizability-of-oncology-trials/28523>\
**Category:** journal club\
**Tags:** oncology, simulation, ml, generalizability, rct\
**Created:** [November 9, 2025, 2:32pm UTC](https://discourse.datamethods.org/t/critique-of-paper-on-generalizability-of-oncology-trials/28523 "2025-11-09T14:32:10Z")\
**Posts on this page:** 1\
**Showing post:** 5

<div class="post-metadata">

**Author:** ![Pavlos\_Msaouel](https://discourse.datamethods.org/user_avatar/discourse.datamethods.org/pavlos_msaouel/32/739_2.png) [@Pavlos\_Msaouel](https://discourse.datamethods.org/u/Pavlos_Msaouel)\
**Post date:** [November 10, 2025, 3:55pm UTC](https://discourse.datamethods.org/t/critique-of-paper-on-generalizability-of-oncology-trials/28523/5 "2025-11-10T15:55:48Z")

</div>

> [@PhilGiguere](#):
>
> Question for @Pavlos_Msaouel: would you not agree that biological interactions as you describe them would manifest in an RCT as a statistical interaction? Do you think this is true of PD-1/PDL-1 expression, for example?

Nope, not necessarily. The two concepts of biological and statistical interactions can overlap but are not the same. We elaborate on this [here](https://www.mdpi.com/2072-6694/14/16/3923), focusing on HER2 as a prototypical example. See in particular Section 3.4 and Figure 7. More tailored to a general oncology audience would also be Section 17. “Prognostic and Predictive Effects” [here](https://www.mdpi.com/2072-6694/15/19/4674). For data science nerds, [here is a cool related paper](https://arxiv.org/abs/2507.03884) and [datamethods thread](https://discourse.datamethods.org/t/leo-breiman-the-rashomon-effect-and-the-occam-dilemma/28301) on broader concepts this is all embedded to.

Note also that many of the statistically “predictive” biomarkers oncologists are seeing in clinical datasets are actually ultra-prognostic biomarkers. A great example of this is the beautiful story of ctDNA in muscle-invasive bladder cancer culminating [with IMvigor011](https://www.nejm.org/doi/full/10.1056/NEJMoa2511885). Luckily, the _ **prognostic** _ effects of ctDNA monitoring were so powerful that just about everyone with ctDNA+ would recur and practically all with ctDNA- would remain disease-free post-surgery. Thus, the [prognostic signal persisted](https://pubmed.ncbi.nlm.nih.gov/34135506/) even after using [information-destroying methods](https://pubmed.ncbi.nlm.nih.gov/38546648/) such as looking for “predictive” 2-way interactions at the HR scale in forest plots of subgroup analyses. That may indeed be the indirect value of looking for these interactions: it can be a quick practical filter for “ultra-prognostic” biomarkers such as adjuvant ctDNA in muscle-invasive bladder cancer.

---

_[View the full topic](https://discourse.datamethods.org/t/critique-of-paper-on-generalizability-of-oncology-trials/28523)._