The importance of randomization

The goal of good trial design is to enable a clinical question to be answered reliably, in particular to estimate the effect of treatment without bias, and with adequate precision. If a trial is designed appropriately it should always be possible to extract useful information. If the design fails to ensure this, even the most sophisticated analysis cannot guarantee to extract an unbiased estimate.

To identify the ways in which bias can enter a trial design, it is useful to consider why the treatment effect observed within a trial may differ from the true, underlying effect which it aims to estimate.

In fact, observed effect = true effect + 'error', where the 'error' term can be broken down into two components:

♦ Systematic error or bias - this can be controlled if not eliminated by good trial design, the most fundamental aspect of which is the allocation of patients to treatment groups by a means which does not systematically select patients with different characteristics to go in different groups and which is unpredictable to the person deciding whether to invite a patient to enter a trial - randomization.

♦ Random error - this is simply the play of chance which is always present to some degree when the number of patients being studied is finite. In a randomized trial it can be minimized by randomizing large numbers of patients, and to some degree controlled by the method of randomization chosen.

Good design will address systematic error by ensuring that the patient groups being compared differ overall only in the treatment they received - this generally means ensuring that they are randomized appropriately- and that the trial outcome measures are assessed without bias. It will address random error by ensuring, through appropriate sample size calculation, that the impact of random error will be small in relation to the size of effect the trial aims to detect. Sample size issues are addressed in Chapter 5; here though we discuss the potential for systematic bias to be introduced through non-randomized comparisons.

0 0

Post a comment