This is an archived version of the Handbook. For the current version, please go to or search for this chapter here.

8.9.1  Rationale for concern about bias

Under the domain of random sequence generation in the Collaboration’s tool for assessing risk of bias, we address whether or not the study used a randomized sequence of assignments. This is the first of two domains in the Collaboration’s tool that address the allocation process, the second being concealment of the allocation sequence (allocation concealment). We start by explaining the distinction between these domains.


The starting point for an unbiased intervention study is the use of a mechanism that ensures that the same sorts of participants receive each intervention. Several interrelated processes need to be considered.  First, an allocation sequence must be used that, if perfectly implemented, would balance prognostic factors, on average, evenly across intervention groups. Randomization plays a fundamental role here. It can be argued that other assignment rules, such as alternation (alternating between two interventions) or rotation (cycling through more than two interventions), can achieve the same thing (Hill 1990). However, a theoretically unbiased rule is insufficient to prevent bias in practice. If future assignments can be anticipated, either by predicting them or by knowing them, then selection bias can arise due to the selective enrolment and non-enrolment of participants into a study in the light of the upcoming intervention assignment.


Future assignments may be anticipated for several reasons. These include (i) knowledge of a deterministic assignment rule, such as by alternation, date of birth or day of admission; (ii) knowledge of the sequence of assignments, whether randomized or not (e.g. if a sequence of random assignments is posted on the wall); (iii) ability to predict assignments successfully, based on previous assignments (which may sometimes be possible when randomization methods are used that attempt to ensure an exact ratio of allocations to different interventions). Complex interrelationships between theoretical and practical aspects of allocation in intervention studies make the assessment of selection bias challenging. Perhaps the most important among the practical aspects is concealment of the allocation sequence, that is the use of mechanisms to prevent foreknowledge of the next assignment. This has historically been assessed in Cochrane reviews, with empirical justification. We address allocation sequence concealment as a separate domain in the tool (see Section 8.10).


Randomization allows for the sequence to be unpredictable. An unpredictable sequence, combined with allocation sequence concealment, should be sufficient to prevent selection bias. However, selection bias may arise despite randomization if the random allocations are not concealed, and selection bias may (in theory at least) arise despite allocation sequence concealment if the underlying sequence is not random. We acknowledge that a randomized sequence is not always completely unpredictable, even if mechanisms for allocation concealment are in place. This may sometimes be the case, for example, if blocked randomization is used, and all allocations are known after enrolment. Nevertheless, we do not consider this special situation under either sequence generation or allocation concealment, and address it as a separate consideration in Section


Methodological studies have assessed the importance of sequence generation. At least four of those studies have avoided confounding by disease or intervention, which is critical to the assessment (Schulz 1995b, Moher 1998, Kjaergard 2001, Siersma 2007). The inadequate generation of allocation sequences was observed to be associated with biased intervention effects across the studies (Als-Nielsen 2004). In one study that restricted the analysis to 79 trials that had reported an adequately concealed allocation sequence, trials with inadequate sequence generation yielded exaggerated estimates of intervention effects, on average, than trials with adequate sequence generation (relative odds ratio of 0.75; 95% CI of 0.55 to 1.02; P=0.07). These results suggest that if assignments are non-random, some deciphering of the sequence can occur, even with apparently adequate concealment of the allocation sequence (Schulz 1995b).