Clinical research involves investigation of the cause of a disease, evaluation of associations between cause and effect, and assessment of the preventive or therapeutic value by isolating and recording potential associations. It is imperative that research findings are interpreted and used optimally. Best conduct, use, and interpretation of research require an understanding of research methodology and study design complemented by adequate and transparent reporting.

Clinical research can be broadly classified into nonrandomized and randomized studies. The nonrandomized category, also known as observational studies, includes mainly nonexperimental studies: nonrandomized clinical trials, cohort studies, case-control studies, cross-sectional studies, case series, and ecological studies ( Fig 1 ). Randomized studies are usually known in biomedical research as randomized controlled trials (RCTs). The purpose of randomization is the creation of groups that differ only randomly at the time of allocation of the intervention. Thus, the key difference between randomized and nonrandomized studies is that in the former, the investigator allocates the interventions to participants randomly: eg, by throwing dice or coins, or by using computer software to generate an unpredictable sequence.

Types of clinical studies based on Cochrane collaboration classification.

There has been considerable debate as to which design is more appropriate to answer clinically important questions. RCTs are considered the gold standard for assessing the efficacy and safety of the intervention of interest. However, it is not always possible or ethical to conduct an RCT. When RCTs are not feasible or unethical, nonrandomized studies, such as cohort or case-control studies, may be undertaken. In general, nonrandomized studies are more prone to systematic and confounding biases than are RCTs; consequently, it is also more difficult to make causal inferences concerning the effect of an intervention.

Observational studies are used extensively to describe the distribution of disease and exposure in populations. They are also useful for hypothesis generation and testing; however, hypotheses may be assessed more adequately, when feasible, with RCTs. On the other hand, RCTs, although highly controlled and often less biased because they are conducted in highly selected settings, may yield less generalizable results that lack relevance to other populations and settings. RCTs, therefore, tend to have high internal validity but low external validity. For example, if we consider the assessment of Class II correction, Class II malocclusion is a multifactorial problem expressed with many variants including maxillary protrusion, mandibular retrognathia, or a combination. A successful intervention for Class II correction demonstrated in an RCT with a highly selected group of participants may not apply to the wider Class II population. Consequently, we can say that the findings of RCTs are less generalizable. Examples of published randomized and nonrandomized (observational) studies are shown in Table I .

Bias and confounding in nonrandomized studies

It is often difficult to establish a causal relationship between the exposure and the outcome based on observational designs because of potential biases and confounding. Bias is the systematic error in the selection of participants, the collection of data, and the ascertainment of exposure and outcome variables, leading to incorrect associations. Methods of reducing potential bias should be carefully considered and planned at the study design stage, and the risk of bias should be accounted for when interpreting the results of such studies. Bias or systematic error differs from random error, which is related to the variability in the sampled population and is expressed by uncertainty (confidence intervals) around the estimate. Increasing the sample size increases the precision of the estimate by reducing the random error but not the systematic error. Selection and information bias are the 2 most prevalent types of systematic bias encountered in observational studies ( Table III ).

Table III

Systematic bias encountered in nonrandomized studies

Systematic bias	Description
Selection bias	Comparison groups differ due to selection error in terms of important baseline characteristics that are important outcome predictors.
Information bias
Recall bias	Recollection of information on exposure is not accurate and can be influenced by experiencing or not experiencing the outcome.
Detection bias	Recording the exposure or the outcome is influenced by knowledge of the outcome or the exposure, respectively.

 

Confounding bias caused by factors that causally affect the exposure and outcome (known as confounders) can blur the associations between exposures and outcomes ( Fig 2 ). Let us assume that we are interested in evaluating the association between oral hygiene status and dental caries in 2 groups of young children with good and less-good hygiene. If all other parameters were equal, as expected in a randomized design, we could be confident that any differences in dental caries between treatment groups would be the result of differences in the levels of hygiene. Now, imagine that another risk factor, dietary sugar, is associated with caries (the outcome). Those subjects with substandard hygiene consumed more sugar than those with good hygiene because the former like to eat sweets. If we accept that sugar consumption promotes caries and that the consumption of sugar is higher among those with substandard hygiene, it is likely that the effect of hygiene on caries may be exaggerated because of another factor that would be counteracting caries prevention in those with substandard hygiene. This factor is called a confounder because it complicates the association between exposure (brushing) and outcome (caries). Sugar consumption is considered a potential confounder, since it is associated with the exposure—hygiene level—and is an independent predictor of the outcome, and it is not an effect of the exposure (the third condition for a parameter to be considered a potential confounder).

Diagrammatic presentation of confounding.