Presentation (CTI)

OBJECTIVES: Population-adjusted indirect comparisons (PAICs) of binary outcomes are typically quantified in terms of odds ratios (ORs). While conditional log ORs (LORs) are independent of baseline risk and population (hence, transportable) under certain conditions in PAICs, marginal LORs vary by population. In some cases, relative risks (RRs) or risk differences (RDs) may be needed to address the research question. We examine the transportability of population-average conditional and marginal log RRs (LRRs) and RDs in PAICs using a simulated example.
METHODS: We specified a logistic model characterizing the relative effects of treatments from an index study comparing treatment B vs. A and a comparator study comparing C vs. A, corresponding to the type of model commonly fitted in ML-NMR. Each study population was defined by X, a uniformly distributed covariate that is both prognostic and effect modifying (shared effect for B and C). The population-average conditional and marginal effects were estimated for B vs. C in the index population where X has a Uniform(-1,1) distribution, and the comparator’s where X is Uniform(-1.5,0.5).
RESULTS: The conditional LOR for B vs. C was 1 in both populations, but the marginal LOR decreased from 0.725 to 0.526 in the comparator population. Both the conditional (0.876 vs. 0.677) and marginal LRRs (0.636 vs. 0.357) differed in the index and comparator populations. Similarly, the conditional and marginal RDs differed across populations and increased from 0.078 to 0.114 in the comparator population.
CONCLUSIONS: In PAICs of binary outcomes based on logistic models, the conditional LOR is the only effect measure independent of covariate distributions, making it transportable. Marginal LORs and LRRs and RDs vary by population and, therefore, are not transportable. Thus, when these effects are derived in MAICs or STCs, they are only applicable in the comparator population. ML-NMR is preferable as it can compute relative effects in any population.