The architecture of clinical research

Author and Disclosure Information



While randomization is an excellent strategy to ensure baseline similarity between compared groups, randomization can fail, and its effectiveness must be checked. Specifically, in a randomized trial, it is customary to examine the compared groups at baseline on all features that can affect the likelihood of developing the outcome measure. If the groups turn out to be dissimilar at baseline in an important way, then the study is at risk for bias, which is specifically called “susceptibility bias” in Feinstein’s construct. Obviously, the larger number of baseline clinical and demographic features that can condition the likelihood of developing the outcome measure, the more difficult it is to achieve baseline similarity between compared groups and the more important it becomes to ensure that randomization has been effective. In this circumstance, larger numbers of participants in both compared groups are generally needed. More about susceptibility bias later.

There are generally 2 types of randomized trials: the so-called “parallel controlled trials” in which each group receives either the principal or the comparative maneuver and is followed and “crossover trials” in which each compared group receives both the principal maneuver and the co-maneuver at different times after an effective wash-out period. Wash-out was discussed above. Figure 2 shows an example of a crossover trial examining the effects of terbutaline on diaphragmatic function.6 The investigators administered terbutaline for a week, measured transdiaphragmatic pressures, gave the patient a terbutaline vacation (the “wash-out period”), and then crossed over those patients who were initially receiving terbutaline to placebo and initial placebo recipients to terbutaline, having remeasured diaphragmatic function after the wash-out period to assure that the patient’s diaphragmatic function prior to the second crossover was identical to his/her baseline state. If this return to baseline is accomplished, then the criteria from effective wash-out are satisfied.

Types of bias in a clinical trial according to Feinstein
Now, with these basic structural terms of clinical research defined, bias will occupy the remainder of the discussion. By definition, bias in a clinical trial is any factor in the design or conduct of the trial, either external to the trial or internal to the trial, that can alter the results in a way that either threatens the reliability of attributing the differences in outcomes between the compared groups with the principal maneuver (“internal validity”) or limits the ability of the results, however internally valid, to be applied to a specific population beyond the study group (“external validity”) (Table 1).1 This again is because the main goal of cause-and-effect research is to make sure that you can attribute differences between the 2 compared groups at the end of the trial to the intervention under study and nothing else.
A comparison of surgery vs nonsurgical therapy for advanced lung cancer.
Figure 3. A comparison of surgery vs nonsurgical therapy for advanced lung cancer. An example of possible susceptibility bias.1

As we begin to talk about sources of bias, consider a study in which we compare survival of patients allocated to surgery vs nonsurgical therapy for lung cancer (Figure 3).1 This study is subject to the first type of so-called “internal bias” in the Feinsteinian construct—so-called “selection bias.” For example, all patients treated surgically were considered healthy enough by their doctors to undergo surgery, whereas patients treated without surgery may have been deemed inoperable because of comorbidities, lung dysfunction, cardiac dysfunction, and so on. If the results of such a comparison show that the mortality rate among surgical patients in this study was lower, the question then becomes: is the improved survival in surgical candidates due to the superior efficacy of surgery vs other therapy or was the enhanced survival due to the surgical patients being healthier to begin with? You can intuitively sense that the answer to this question is that the enhanced survival may be due to the better health of patients treated surgically rather than to the surgery itself because of how the patients were selected to receive it. So, this is a simple example of what Dr. Feinstein would call “susceptibility bias.” Susceptibility bias occurs when the 2 baseline groups are not comparably at risk or susceptible to developing the outcome measure, leading the naïve investigator in this specific example to attribute the difference in outcomes to the superiority of surgery when in fact it may have nothing to do with the surgery vs. the other maneuver. When susceptibility bias is in play, the difference between the outcomes in the compared groups could be attributed to the baseline imbalance of the groups rather than to the principal maneuver itself.

Turning back to the taxonomy of bias, there are four types that can threaten internal validity—“susceptibility,” “performance,” “detection,” and “transfer” bias—and 1 type of bias (called “external bias”) that can affect the generalizability of the study called “assembly bias” (Table 1).

Potential sources of bias in a randomized, controlled trial according to Feinstein.
Figure 4. Potential sources of bias in a randomized, controlled trial according to Feinstein.1

Figure 4 shows where these various sources of bias appear in the architecture of a clinical trial. As just discussed, susceptibility bias affects the baseline state and the comparability of the groups. Performance bias relates to how effective and how comparably the co-maneuvers are given and whether the primary intervention is potent enough to affect an outcome. Both transfer and detection bias operate in detecting the outcome, especially regarding the rigor and frequency with which they are investigated. Transfer bias has to do with selective loss to follow-up of those included in the trial. If there is a systematic reason for loss to follow-up that is related to the impact of the intervention, then the study is at risk for transfer bias. For example, in a randomized trial of drug A vs placebo for pneumonia, if drug A is effective but all the drug A recipients fail to follow-up because they feel too good to return for follow-up, then transfer bias could be causing the study to show nonefficacy even though the drug works. So, if those who respond favorably are systematically lost to follow-up, and if all the patients who felt lousy wanted to see the doctor and came back for follow-up, such transfer bias would bias towards nonefficacy. Specifically, only patients remaining in the trial would be those who failed to respond and that would dilute any difference between the 2 groups despite the active efficacy of drug A.

Hopefully, you are already beginning to get a sense that one has to be extremely disciplined in thinking about each of these sources of bias because they can have some very subtle nuances in randomized trials that can easily escape attention.

Returning to sources of bias, let’s consider the second type of bias, “performance bias.” Performance bias relates to the administration of the compared maneuvers—the primary or principal maneuver, compared with the comparative maneuver. Performance bias can occur when the main maneuver is not administered adequately or when the co-maneuvers are administered in an imbalanced way between the compared groups. Consider the example of the Long-Term Oxygen Treatment Trial (LOTT) trial, which compared use of supplemental oxygen with no supplemental oxygen in patients with stable COPD and resting or exercise-induced moderate desaturation.9 The principal outcome measure of LOTT was all-cause hospitalization or death. In such a study, many potential sources of performance bias exist. For example, performance bias might exist if none of the patients allocated to oxygen actually used supplemental oxygen. Alternately, to the extent that use of inhaled corticosteroids or antimuscarinic agents lessens the risk of COPD exacerbation, performance bias could occur if use of these co-maneuvers was imbalanced between the compared groups. As a specific extreme circumstance, if all patients in the nonoxygen group used these inhalers but none of the patients in the oxygen group did, then a lack of difference between exacerbation frequency could be related to this imbalance in co-maneuvers (a form of performance bias) rather than to the lack of efficacy of supplemental oxygen.

Next Article:

Basics of study design: Practical considerations

Related Articles