Network meta-analysis combines studies comparing different treatments to estimate effects across entire treatment networks. Study A compares drugs X and Y. Study B compares Y and Z. The network infers the X-vs-Z comparison indirectly. The method is powerful and increasingly standard. The assumption: if the effect modifier distributions are the same across studies, the treatment effects are combinable.
Note (2026-11-13; read against the paper's full text): the paper is more cautious than the essay about the remedy and its reach. It demonstrates the bias with an illustrative example (in its worked example, a mixed population with an equal split gives a bias factor of 0.850, about 15 percent attenuation of the log odds ratio) and with the Neuhaus-Jewell approximation, and says the standard model's estimates are biased toward the null when studies enroll heterogeneous mixtures of risk groups that vary across studies, even with no difference in effect modifiers. The bookend approach, which treats the studies with the smallest and largest baseline risk as homogeneous and models the others as weighted combinations of them, is proposed as a potential solution that gives an unbiased or perhaps less biased estimate, with guidance to practitioners on whether the bias matters; the essay says it recovers the subgroup-level effects and that the bias is present in every network meta-analysis that used odds ratios on populations with varying baseline risk, which is stronger than the paper's wording.
Campbell and Jansen (arXiv:2603.00749) show this assumption is wrong when the effect measure is an odds ratio. Even with identical effect modifiers across all studies, differences in baseline risk — the probability of the outcome in the control group — create systematic bias toward the null. The treatment looks less effective than it is, and the bias is structural, not random.
The mechanism is non-collapsibility. A collapsible effect measure gives the same value whether you compute it for the whole population or average it across subgroups. Risk ratios and risk differences are collapsible. Odds ratios are not. When a study enrolls a mixed population — some high-risk patients, some low-risk — the study-level odds ratio is not the weighted average of the subgroup odds ratios. It's systematically smaller. Mixing population strata shrinks the apparent effect.
In a meta-analysis, different studies enroll different population mixes. Study A might have 60% high-risk patients. Study B might have 30%. Both studies measure the same treatment in the same way, and the treatment works identically in both. But their odds ratios differ — not because the treatment effect varies, but because the population composition varies. The meta-analysis interprets this variation as heterogeneity and averages toward a smaller effect. The bias compounds across the network.
The fix Campbell and Jansen propose — the “bookend” approach — models each mixed study as a weighted combination of the extreme-baseline-risk subpopulations observed elsewhere in the network. It recovers the subgroup-level effects before combining them, bypassing the collapsibility trap.
The structural lesson: the choice of effect measure is not just a convention. It's a lens that can introduce distortion into what it claims to be measuring. Odds ratios are mathematically convenient (symmetric, unbounded, natural for logistic models) and epidemiologically standard. They are also silently biased by population composition — a property of the measure, not of the biology. The bias has been hiding in plain sight, in every network meta-analysis that used odds ratios on populations with varying baseline risk.
Campbell & Jansen, "Hidden in Plain Sight: How Non-Collapsibility Biases Treatment Effects in (Network) Meta-Analysis," arXiv:2603.00749 (2026).