Background
Channelling bias is a form of Selection Bias (link to that page) that occurs primarily in observational research, particularly studies using electronic health record data. It is most commonly associated with preferential drug prescribing, but can occur in any situation where non-random exposure assignment occurs. In channelling bias, different patients are “channelled” towards one treatment over another based on perceived risk, prognosis, prescriber preference, safety concerns, marketing influences, or post-marketing caution.
Even when treatments have similar indications, different patient groups may be “channelled toward specific treatments. This can work in many different ways, for example:
(1) Younger or apparently healthier patients may be prescribed newer, less established treatments.
(2) Older or more vulnerable patients may be given established treatments perceived as more efficacious, cheaper, or having fewer side effects.
(3) Higher-risk patients may be given alternative or experimental therapies after intolerance or failure of standard treatments.
If channelling effects are not considered in the design or analysis of studies evaluating a treatment, estimates of its safety and effectiveness may be distorted.
Example
Selective cyclooxygenase-2 (COX-2) inhibitors were introduced in the 1990s for the management of osteoarthritis and rheumatoid arthritis, and were perceived as safer than traditional non-steroidal anti-inflammatory (NSIADs) drugs due to reduced gastrointestinal (GI) bleeding. Meloxicam, a semi-selective COX-2 inhibitor, was therefore often prescribed to patients considered at higher risk of GI events [see here].
Martin et al. examined adverse events among patients prescribed meloxicam and found that prescribers may have been preferentially prescribing the medication to patients already at higher risk of GI events. MacDonald et al. [2] subsequently demonstrated that apparent increased GI event rates among users of meloxicam and other selective COX-2 inhibitors could be explained by this preferential prescribing. After adjustment for baseline GI risk factors consistent with preferential prescribing, the apparent excess risk associated with meloxicam was attenuated, and COX-2 inhibitors were associated with a lower risk of GI events. This study highlighted the importance of recognising and adjusting for channelling bias when analysing research data.
Impact
Channelling bias can also occur in surgical research. For example, younger patients requiring hand surgery may be offered more complex reconstructive procedures because of a perceived lower perioperative risk, while older patients are not. In retrospective analyses, such channelling bias may create the appearance of superior outcomes among the surgically treated groups when differences are partly attributable to patient selection rather than treatment effect.
Li et al. showed that accounting for differences in patient characteristics and comparator selection substantially altered estimates of cardiovascular risk associated with testosterone replacement therapy. The attenuation of effect estimates after adjustment demonstrated how channelling bias can exaggerate observed treatment effects in observational research.
Preventive steps
For observational research, minimising channelling bias requires careful study design and appropriate analytical methods. The selection of comparison groups in particular needs careful consideration. Approaches such as active-comparator new-user designs, matching, restriction, standardisation, stratification, propensity score methods, multivariate regression and instrumental variable analysis may reduce bias arising from preferential prescribing. Careful design and analysis may reduce channelling bias, but it is unlikely that the effects of channelling bias can be fully removed, so results should be interpreted with caution, with limitations clearly acknowledged.
Randomised controlled trials largely prevent channelling bias through random allocation but do require careful study design to avoid channelling pre-randomisation. For example, bias may still arise before randomisation if certain patients are selectively enrolled. Trial protocols should minimise opportunities for clinicians to influence treatment allocation or post-randomisation treatment changes from that which a patient is randomised to receive, particularly in unblinded studies.
In cluster randomised trials, where a clinic or hospital may be randomised rather than the patient, differences in patient populations across clusters may also introduce distortions. For example, higher-risk patients may attend a specific hospital, potentially introducing channelling bias. Careful consideration at all stages of trial design is therefore essential.
