Performance and detection bias – hiding who got what
Bias can occur if the treatment arm to which a given participant is randomized is known. When reading an RCT report, the term “double-blind” is often not sufficient to help appraise this. We need to know from whom treatment identity was masked, and how.
If clinicians and healthcare caregivers know which intervention a child is receiving, this may influence the overall care that they give. For example, in an RCT comparing two doses of dexamethasone for croup, a child who has received the lower dose may be more likely to be admitted to hospital. In other trials, participants may switch between treatments. Prior assumptions can lead to systematic discrepancies between the groups of participants. If inadequate blinding might influence a patient’s overall care, this is said to cause performance bias. The RCT report should tell us who was blinded, how this was achieved, and whether knowledge of the arm to which participants were randomized may have affected the results.
If patients and families know which intervention they received, this may affect the results for certain trial outcomes, particularly subjective ones such as symptom scores or quality of life. Patients receiving an experimental treatment, for example, may be more likely to give positive responses than those receiving placebo. If inadequate blinding might influence outcome measurement, this is said to cause detection bias.
In some trials, for example studies evaluating surgical interventions, blinding may not be possible. Trialists can attempt to minimize the effects of performance bias by using treatment protocols to minimize difference in care given to the groups of participants.
Incomplete outcome data
Patients may drop out of trials before the primary outcome is measured, for several reasons. Some reasons may be linked to patient characteristics, for example they miss appointments. If a trial is adequately randomized, such patients should be spread across both treatment arms. Some reasons for dropout may be, however, linked to one of the study interventions, either because of lack of benefit, or development of adverse effects. Even if patients do not drop out of the study, they may not follow the path expected of them at the outset (eg they may be given interventions to which they were not randomized), and such situations are known as protocol deviations.
If patients drop out, or deviate from the study protocol, they should nevertheless be included in the final analysis. This is to account for those patients whose drop-out or protocol deviation was linked to an intervention, because excluding them would mean that one group might gain an advantage. If lots of patients receiving one intervention left the trial because of lack of benefit, it is clear to see that excluding them would give a falsely positive view of that intervention.
Exclusion of participants who drop put, or deviate from the protocol, puts the study at high risk of attrition bias. To protect from this, trialists should describe that they used an intention-to-treat (ITT) analysis. This means that in the final analysis they (1) keep participants in the intervention groups to which they were randomized, regardless of the intervention they actually received, (2) measure outcome data on all participants, and (3) include all randomized participants in the analysis.
In per-protocol approach, the analysis is based on the treatment that participants received, rather than that to which they were randomized. This can lead to systematic differences between groups, and renders the study at high risk of attrition bias.
‘Modified ITT’ can mean a multitude of approaches, many of which should make the reader wary of attrition bias.