Peter Hall
LightNews LightNews
banner
peterhall001.bsky.social
Peter Hall
@peterhall001.bsky.social
220 followers 160 following 17 posts
Breast Oncologist and Professor of Cancer Informatics, Edinburgh | Health Economics | Real World Evidence | https://cancer-data.ecrc.ed.ac.uk/
Posts Replies Media Videos
Reposted by Peter Hall
statsepi.bsky.social
Target trial emulation is an important, useful idea, but we must remain careful to not allow novices to think it's a magical shortcut to causal inference. Nor can we lose sight of the *main* thing that separates randomized from non-randomized studies.
POTENTIAL RISKS OF TARGET TRIAL TERMINOLOGY There are risks that the term target trial emulation (TTE) might mislead readers without methodological training. We believe some readers may assume TTE represents a distinct form of observational study design, rather than a framework to help researchers identify and avoid self-inflicted errors in observational studies. We are particularly concerned that claims to have emulated an RCT will be mistaken for a claim to have successfully emulated an RCT. The latter statement implies that the inference is of a comparable standard to a well-conducted RCT. To be certain, the TTE framework (2), used correctly, will help investigators specify estimands, and sidestep avoidable selection bias, but we are probably more pessimistic than Schwarze et al. (1) about the ability of ‘‘advanced methods’’ to address the threat of other sources of bias. However, the difference in how randomization accounts for confounding vs. ‘‘advanced methods’’ like covariate adjustment, weighting, etc. is a difference in kind, not in degree. This is because randomization is guaranteed to eliminate confounding for some key estimands, and to do so in a manner that predominantly rests on our trust in a potentially verifiable randomization process. In contrast, acting on the effect estimate returned from an NRSI requires us to trust in the decisions and competence of the investigator, and even then, no guarantee is possible, even in relatively simple sets of relationships among a small set of potential covariates. This distinction will be obvious to methodological experts, but they only constitute a small proportion of researchers and research-consumers
August 12, 2025 at 10:15 AM