Article type
Year
Abstract
Background: Differences in baseline risk (BR) across studies may be a source of heterogeneity in meta-analysis (1; see attachment for a full list of references); thus, generalizability of findings across subgroups of patients requires constancy of treatment effects across BRs (2).
In the case of risk ratio (RR), its independence from BR cannot hold due to its range limitations (2). Odds ratio (OR), instead, is theoretically independent from BR (3), but this property may not hold in practice (4). Moreover, in the specific case of using OR in network meta-analysis (NMA), even factors affecting baseline risk but not conditional effect may lead to inconsistency (5). Thus, regardless of the outcome measure adopted, in the case of binary outcomes, treatment effects should be considered as likely dependent from BR.
Objectives: To suggest a method making results from NMA for binary outcomes more reliable and generalizable, by expressing their dependence on BR, to equip clinicians and patients with better predictions of potential outcomes for each treatment option.
Methods: The arm-based (AB) model from (6) estimates log-odds of risks for each treatment arm by allowing both for dependence between baseline risk and treatment effect and for random effects (guaranteeing evaluation of heterogeneity and inconsistency). After estimating the logit of the probability of baseline risk and treatment effect, marginal treatment and BRs can be derived and marginal estimates of the desired effect size calculated (4).
Results: We will present an example of use of the AB model from (6) with a dichotomous outcome, showing pitfalls of the comparison-based approach using both OR and RR as outcome measures.
Conclusions: The never-ending debate about whether OR or RR should be used as an effect measure seems to be focused lately on ‘portability’, i.e., on their possible independence from BR (7). Owing to the general implausibility of such assumption in practice for both outcomes, we argue that, in NMA for binary events, BR should not be considered as a possible reason for inconsistency among others, but as a crucial factor to determine the type of analysis: the pitfalls of the AB model may be counterbalanced by a lower risk of inconsistency.
No patient, public or healthcare consumer has been involved.
In the case of risk ratio (RR), its independence from BR cannot hold due to its range limitations (2). Odds ratio (OR), instead, is theoretically independent from BR (3), but this property may not hold in practice (4). Moreover, in the specific case of using OR in network meta-analysis (NMA), even factors affecting baseline risk but not conditional effect may lead to inconsistency (5). Thus, regardless of the outcome measure adopted, in the case of binary outcomes, treatment effects should be considered as likely dependent from BR.
Objectives: To suggest a method making results from NMA for binary outcomes more reliable and generalizable, by expressing their dependence on BR, to equip clinicians and patients with better predictions of potential outcomes for each treatment option.
Methods: The arm-based (AB) model from (6) estimates log-odds of risks for each treatment arm by allowing both for dependence between baseline risk and treatment effect and for random effects (guaranteeing evaluation of heterogeneity and inconsistency). After estimating the logit of the probability of baseline risk and treatment effect, marginal treatment and BRs can be derived and marginal estimates of the desired effect size calculated (4).
Results: We will present an example of use of the AB model from (6) with a dichotomous outcome, showing pitfalls of the comparison-based approach using both OR and RR as outcome measures.
Conclusions: The never-ending debate about whether OR or RR should be used as an effect measure seems to be focused lately on ‘portability’, i.e., on their possible independence from BR (7). Owing to the general implausibility of such assumption in practice for both outcomes, we argue that, in NMA for binary events, BR should not be considered as a possible reason for inconsistency among others, but as a crucial factor to determine the type of analysis: the pitfalls of the AB model may be counterbalanced by a lower risk of inconsistency.
No patient, public or healthcare consumer has been involved.