The aim of a randomized clinical trial is to provide an unbiased assessment of the effects and safety of a medical or dental intervention. Unfortunately, it is not easy to have completely unbiased trial results because errors can ensue at all levels of the study. Those errors can be classified in the 3 broad categories of random error, bias, and confounding.
Random error can arise when the results of the clinical trial, however well conducted, could be different from the truth because of the uncertainties related to obtaining a random sample. In small trials, the probability of reporting a clinically important treatment effect due to chance is relatively high, whereas, in large trials, it is low. Imprecision, another term for random error, means that multiple replications of the same study will produce different effect estimates because of sampling variations, even if they would give the correct answer on average. This imprecision is reflected in the confidence interval around the effect of the trial, and the effect of random error on the estimates decreases as the trial becomes larger. Precision and confidence intervals will be explained in detail in a later article.
Bias is systematic error that leads to distortion of the true treatment effects and can arise at various stages of the trial: during design, conduct, analysis, and reporting. A key focus of every trial is to decrease bias because it might question the results. The most common types of bias are shown in the Table .
|Type of bias||Example||Remedies|
|Selection bias||Assigning patients with better oral hygiene to the treatment group favored by the investigator||Appropriate randomization|
|Study management or performance bias||Following more closely the patients in the treatment group favored by the investigator||Standardization of procedures
Blinding, when feasible
|Detection bias||Recording outcomes in a way that proves the investigator’s or the participant’s beliefs||Blinding, when feasible|
|Attrition or postrandomization bias||Participant loss related to the outcome (eg, severe side effects)||Intention to treat analysis|
|Publication or reporting bias||Selective reporting of only statistically significant results||Trial registration, prepublication of trial protocol, reporting not only interesting or positive results, but also negative results|
Selection bias is likely to occur in the early stages of the trial during participant recruitment. Suppose a trial is assessing the effect of deciduous canine extraction vs nonextraction for alleviating permanent canine impaction. If an investigator favors deciduous canine extraction, he or she might consciously or subconsciously select, for the extraction group, participants with canines in a more favorable position for eruption and vice versa. Selecting patients to include in the extraction group who are more likely to show canine eruption introduces bias by overestimating the effect of deciduous canine extraction (the intervention) on the resolution of canine impaction (the outcome).
Bias in study management (o performance bias) is related to differential treatment, care, and follow-up during the trial. For example, let’s say that we wish to evaluate plaque and periodontal indexes in patients wearing self-ligating brackets compared with patients wearing conventional brackets. An investigator who believes that self-ligating brackets collect less plaque and lead to lower plaque and periodontal index scores might follow patients with self-ligating brackets closer and offer more exhaustive oral-hygiene instructions. This behavior of the investigator could result in distortion of the true effect of the appliance on plaque accumulation and consequently on the recorded index scores. In this example, selection bias might have also occurred if the investigator included the better brushers in the self-ligating group.
Detection bias can occur during outcome recording and might be related to both investigators and participants. To continue with the previous example, investigators favoring 1 type of bracket compared with the other might consciously or subconsciously round up or down the plaque and periodontal index scores, depending on their preconceptions. At the participant level, observer bias is of particular importance when the outcome to be recorded is subjective and involves a response from each participant. Outcomes such as reporting of pain levels on a visual analog scale after placement of different wire types or after taking pain medication could be modified if the patient knows the treatment group that he or she belongs to and if for some reason believes that 1 therapy is superior to the other.
Attrition or postrandomization bias is encountered after patients are assigned to treatment groups and during follow-up. It is often related to unequal numbers of patients lost to follow-up because of biased patient exclusions, noncompliance, or losses related to the intervention. For example, suppose the participants are allocated to receive either a standard drug or a new drug with the objective to assess side effects. If the patients with the worst side effects from the new drug did not return for follow-up and outcome recording, the results would most likely be biased. Excluding the patients lost from the data analysis creates treatment groups that contain participants with less severe side effects, thus underestimating the side effects of the new drug. To remedy this problem, an intention to treat (ITT) analysis is recommended compared with a per-protocol analysis. The ITT analysis includes all patients in the final analysis, even the lost ones, according to which group they were randomized. The per-protocol analysis includes only patients who were followed according to the study protocol—ie, have no missing data. The ITT scenario is more conservative, closer to real-life treatment effects, and tends not to exaggerate the effects of treatment. The updated CONSORT statement has modified the requirements for reporting ITT analysis; however, the key point is that exclusion of participants lost for reasons related to the outcome (eg, lost because of severe side effects) might bias the trial results and should not be ignored.
Finally, during the publication stage of the trial, bias can ensue if investigators or editors preferentially report or publish only positive, statistically significant, or interesting trials. To better understand publication or reporting bias, let us assume that a treatment in reality is ineffective, but a few small studies are published that show incorrectly that the treatment of interest is effective. However, suppose that larger studies, showing that the treatment is ineffective, are not published. A systematic review, which would combine the available evidence, might come to misleading conclusions if the authors find only the small published trials which show that the treatment is effective. The implications of publication bias are important for health care, since treatment recommendations are derived from published evidence. If only part of the evidence is accessible, then health care recommendations can be compromised. Publication bias has been reported in medicine and also recently in orthodontics. Publication bias could be limited by trial registration and protocol publication, and by encouraging authors, reviewers, and editors not to publish studies with only significant findings. Publication bias can be reduced with mandatory trial registration and protocol publication that would make it difficult for authors to report only “interesting” findings, since this will create a noticeable discrepancy between what was initially intended and what was actually published.
Confounding will be discussed in a future article. The next 3 articles will discuss randomization.
In clinical trials, we must always guard against selection, performance, detection, attrition, and reporting biases.
Direction and size of bias are hard to calculate at the end of the study; therefore, the only reliable strategy to prevent bias is by using a good study design.