Selection Bias and Baseline Imbalances in Randomized Trials
- PDF / 1,839,462 Bytes
- 2 Pages / 573.839 x 774.72 pts Page_size
- 82 Downloads / 171 Views
I
LETTER TO T H E EDITOR
Selection Bias and Baseline Imbalances in Randomized Trials (DrugInformation Journal, 2003;3?293 -308) Voice W. Brrgrr National Cancer Institute, Bethesda, Maryland, and Department of Mathematics and Statistics, University of Maryland Baltimore County, Baltimore, Maryland
Corrrspoidricr Address Vance W Berger, National Cancer Institute, EPN, Suite 3131p6130Executive Boulevard, MSC-7354, Beth-da, MD 20892-7354 [email protected]).
McEntegart (1)did a good job of clarifying what exactly it is that one might hope for in a randomization procedure, and it is not my intention to criticize the article as an entire entity. However, there is a statement in the article which is not only incorrect, but also at odds with other statements in the article, thereby rendering the article internally inconsistent. In fact, this incorrect statement is widely believed in the pharmaceutical industry, with unfortunate results. Specifically,McEntegart (1) echoes the common belief that in randomized trials, "There will be no systematic reason for treatment imbalances on prognostic factors but chance imbalances are possible" and "hypothesis testing of baseline imbalances is inappropriate as observed differences are by definition due to chance." If this were true, then why would one be concerned with the fact that if "the investigator can predict the treatment allocation for the next subject.. .then the trial is subject to selection bias?" The mechanism for selection bias mentioned involves the dependence of the entry decision (ie, enroll the patient or not) on the concordance of the patient characteristics and the upcoming treatment, that is, one can enroll healthier patients when one treatment is next, and sicker patients when the other treatment is next. The result of this form of selection bias is a
potentially serious threat to validity in the form of forced confounding, or systematic (most certainly NOT random) baseline imbalances in prognostic baseline covariates. Note that randomization may actually accomplish what it purports to, specifically the elimination of systematic imbalance across treatment groups in the accession numbers. Yet, even without preferential selection of any specific accession number to any particular treatment group, it is still possible to match patient characteristics to accession numbers in such a way that patient characteristics are actually matched to treatment groups. The key is prediction, which serves as a matching of accession numbers to treatment groups while there is still an opportunity to exploit it (before the patient is selected to fill the accession number in question), that is, once it becomes clear which accession numbers will receive which treatments, it is possible to place patients (and their underlying characteristics) accordingly. For this reason, ensuring baseline balance in patient characteristics (as opposed to accession numbers), even in probability, is not in randomization's job description. While there are a few designs that can eliminate selection bias, thes
Data Loading...
