Other Forms of Randomization
Despite the popularity of random allocation, there are a number of problems with it. First, if the new treatment is less effective than the traditional one, half of the subjects would be receiving less than optimum therapy. Conversely, if a new intervention is more effective than placebo, then the latter group has missed out on a chance to benefit. Second, because not all subjects who are approached agree to participate in a study and we know that volunteers are different from refusers (you’ll know that, too, in another few pages), a high refusal rate can jeopardize the external validity of a trial (i.e., the extent to which the results can be generalized to the world at large). Through the years, a number of variations of randomization have been developed to try to overcome these shortcomings.
One group of procedures is referred to as adaptive sampling. This means simply that the allocation is adapted to match the results from people previously enrolled in the study. The best known procedure is play the winner, which is useful if (1) the outcome is a binary one, and (2) the results are known fairly quickly. We start off by assigning the first patient to either Treatment A or Treatment B at random; let’s assume it was to Treatment B. If the outcome was positive, then the next patient would also be assigned to B, whereas if it were negative, the next patient would be enrolled in A. In this way the more successful treatment will end up with the larger number of patients. In fact the difference in the proportion of patients allocated to each group is directly related to the magnitude of the treatment difference. The effect of this procedure then is to minimize the number of subjects receiving the less effective or more harmful treatment.
Another variant to random assignment is called Zelen randomization, which can be used when a new therapy is being tested against usual practice. Figure 3-10, A Fig.1232, shows what happens with the usual procedure: prospective patients are approached for their consent to be in a trial, and only those who agree are randomized to the treatment conditions. But why should people who end up receiving the usual form of care be approached for consent? If they refuse to participate in the study, they will be getting this anyway; all consent does is perhaps make them concerned about being in a study and makes a larger group of people eligible to refuse. Instead, Zelen randomization follows the procedure shown in Figure 3-10, B Fig.1232. Patients are first randomized, and only those allocated to the new treatment are approached for consent. This cuts down the number of people who can say “No.”
This technique can work only if the control condition is usual practice and not placebo. Even so, there are some ethical problems in that the patients in the control group are in a study but haven’t been told this. For this reason, there have been many articles written about Zelen randomization but few studies that have actually used it.
Figure 1232 – (Figure 3-10) A, Regular and B, Zelen randomization
Streiner DL, Norman GR. PDQ Epidemiology-Second Edition, 1996, BC Decker Inc., Hamilton, Ontario.
Some figures may not display clearly when rendered as a PDF or printed.
MatchingThe term matching can have two meanings: one applies at the level of the individual subject and the other describes the general strategy for selecting a control group.
Matching at the individual level means that a pair of experimental and control subjects are chosen to be as similar as possible in terms of certain key variables, such as age, sex, race, socioeconomic status, number of hospital admissions, or diagnosis. A person from the smaller subject pool is often chosen first (e.g., if there are fewer “exposed” than “nonexposed” people in a case-control design, the pool of potential experimental subjects is smaller than that of the controls). Then a subject from the other pool is selected and matched as closely as possible on the key characteristics. The larger the ratio of potential subjects to the desired number to be chosen, the more matching variables can be used. If there are not too many people to choose from, the number of matching variables must be reduced or the criteria for similarity are relaxed (e.g., matching for age within plus or minus 10 years rather than within 5). The result of matching is two groups that are as similar as possible on these key variables.
At the level of the group, matching refers to selecting a control group that has certain characteristics as an aggregate. For example, subjects in this control group can be (1) patients at the same hospital but with a different diagnosis, (2) drawn from the same community, or (3) working at similar jobs. Control subjects, however, are not matched to experimental subjects on a one-to-one basis.
The purpose of matching on certain variables is to eliminate the effect of those variables on group differences. If the two groups are matched on age, for example, any difference in outcome between the groups cannot result from this factor. The downside is that matching prevents us from examining at some later point the effect of age on the outcome. The moral is to match only when you’re certain that you aren’t ruling out examination of an association in which you may later be interested.
Groups are undermatched if they differ on some variable that is related to the outcome. The effect of undermatching is that group differences at the end may be caused by the variables that aren’t matched. So there is a fine line between overmatching and thus being unable to explore potentially interesting relationships, and undermatching, which may cause your results to be explained by some extraneous variable.
Content on this page was last changed on March 19, 2009.
© 2002 BC Decker Inc. Show Disclaimer
| 5476. | Streiner DL, Norman GR. PDQ Epidemiology. 2nd ed. Hamilton, Ontario: BC Decker Inc.; 1996. |
Next Page: Threats to Validity »