Withdrawal bias

Patients who are withdrawn from an experiment may differ systematically from those who remain Index of biases | Editing advice |

Background
An important part of any trial of a treatment is to see what harms result from the intervention, as well as what benefits. This information allows clinicians and patients to weight up whether that treatment might be useful or not in a given clinical situation. It is therefore crucial that if patients receiving an active treatment experience a side effect or die from the treatment, this needs to be included in the analysis and trial report. If harms are not included, the trial report will give an overly-optimistic view of the treatment involved. This will make patients much more likely to accept the treatment in the belief that it is safer than it really is.

Withdrawal of patients' data from the analysis might be done to enable an analysis of the data as an explanatory trial (see below). This needs to be explicitly stated in the trial method. It might also be done out of ignorance of how to handle outlying, unusual or unexpected data. More worryingly, it might be done in order to make the treatment under investigation look better than in reality in order to increase the chances of a positive outcome, and thus increase the chances of publication. The first group is acceptable, if so stated in the method & report; the second is the result of inadequate training in research methods; the latter is a form of research fraud.

'Explanatory' and 'management' trials
Explanatory trials are intended to elucidate mechanisms of action of the intervention, and focus on those patients who comply with the protocol as planned. A management trial focuses on the observed effects of intervention when it is prescribed as in the normal bounds of clinical practice - i.e. with an 'intention to treat'.

Withdrawal bias is a particular for of attrition bias, a general term meaning that unequal numbers of patients in each group entered into a trial end up being included in the analysis and final report.

Example
In a neurosurgical trial of surgical versus medical therapy of cerebrovascular disease, patients who died or stroked-out during surgery were withdrawn as ‘unavailable for follow-up’ and excluded from early analyses

Impact
At an individual level, this results in misleading individual patients' choices. On a wider level, once the treatment is more widely used the real incidence of harmful effects will become apparent. The mismatch of real life experience with the trial results would then justifiably undermine people's trust in the research process, and researchers' probity overall. This would then

Preventive steps
Good trials will have a trial profile. This is a written account or flow diagram showing an audit trail of what happened to all patients, from screening for entry into the trial to completion of trial. Here is an example from the PROSPER trial of pravastatin in older patients:

This diagram makes clear exactly what happens to all patients at each stage in the trial process (click on it for a larger, more easily visible image).

In addition, using the 'intention to treat' method of analysis, all trial subjects are followed up, irrespective of what subsequently did happen to them. This reduces the chances of the people doing the trial conveniently omitting a subgroup of patients from the analysis because of a change in circumstances after the start of the trial.

Cite as
Catalogue of Bias Collaboration. Byatt K. Withdrawal bias. In: Catalogue Of Bias 2018: www.catalogofbiases.org/biases/withdrawal-bias

Related biases
attrition bias