Introduction

In the three previous chapters in this book, we introduced the process of evidence-based dentistry (EBD)¹ and explained how to search for evidence to inform clinical practice² and how to use an article about therapy.³ In this chapter, we explain how to use an article to inform clinical decisions regarding questions of harm. We introduce and describe the basic concepts needed to understand observational studies, and we explain how to use these concepts to critically appraise such studies. In subsequent chapters in this book, we describe how to use other types of study designs.

Clinical Questions of Harm

Questions regarding potentially harmful exposures, either to dental treatments or external agents, are common in dental practice. Some examples of these questions are the following: Do people who live in areas where the water is fluoridated have a higher risk of having enamel defects? Does smoking increase the risk of having oral cancer? Does the dentist’s use of rubber dams when placing a dental restoration increase the patient’s risk of allergic reactions if the patient has a latex allergy?

The classic Population-Intervention-Comparison-Outcome (PICO) framework requires only minor modifications to address questions related to harm. The population is the patients of interest. In cases that address questions related to harm, the population is those patients who may face the potentially harmful agent. The intervention becomes the exposure, which corresponds to the harmful agent. The comparison is the reference, which is the absence of the exposure to the harmful agent. The outcome is the potential negative consequence of the exposure. Table 4.1 shows examples of questions related to harm and the corresponding PICO components.

What Study Design Best Addresses Questions of Harm?

Owing to the hierarchy of evidence used to answer questions about harm, even though investigators might identify randomized controlled trials as being the best type of study design to answer these types of questions, they generally cannot use this type of study design because of ethical reasons. Therefore, at the level of a primary study, an observational study is usually the most appropriate study design to answer questions regarding harm. This is not always true, however. Note, for example, that investigators could address the question listed in Table 4.1 about rubber dams by using a randomized controlled trial design.

Table 4.1. Examples of Questions Related to Harm and the Corresponding PICO* Framework

* PICO = population, intervention, comparison, outcomes.

Figure 4.1. The Designs of Analytical Observational Studies

A. Cohort study: investigators recruit participants on the basis of the patients’ exposure, and they follow up with patients over time to determine the presence of outcomes in the future. B. Case-control studies: investigators recruit participants on the basis of the presence or absence of an outcome, and investigators assess participants’ histories to determine the presence of exposures in the past. C. Cross-sectional studies: investigators measure the presence of the exposure and the outcome at the same time point.

An observational study is one in which the investigator does not assign an exposure or intervention; rather, these exposures or interventions occur naturally in the study setting. Although investigators have conducted descriptive observational studies in which they recruit only one group of patients and do not compare them with any other group of patients, in this chapter we describe the type of observational studies in which investigators use a comparison group (which can happen either because two groups of patients are recruited and followed, or one large group of patients is divided into two or more, on the basis of the presence of an exposure).

Observational studies can be classified according to the direction in which the exposure or outcomes are measured.⁴ The intuitive design is one in which investigators enroll participants who either are exposed or are not exposed (for example, patients living in a community that has fluoride in the water or patients living in a community that does not have fluoridated water) and follow them over a period, recording whether the outcome of interest (that is, fluorosis) does or does not occur. We call these cohort studies (Figure 4.1, Table 4.2^5–7).⁸

Table 4.2. Examples of Analytical Observational Studies

A less intuitive design is one that involves investigators recruiting samples of study participants in whom the outcome has occurred (for example, they have had fluorosis [we call these participants “cases”]) and comparing them with similar study participants who have not had the outcome of interest (that is, no fluorosis [we call these participants “controls”]). Investigators then determine—by asking questions to participants or by looking at medical records or other information sources—whether participants in either group experienced the exposure of interest (that is, water fluoridation). We call these case-control studies (Figure 4.1, Table 4.2^5–7).⁹

Investigators can use another type of design only when they can assess the exposure and the outcome at the same time. Here, the investigator looks simultaneously at the exposure (for example, the current exposure to fluoridated water) and the outcome (for example, fluorosis). We call such designs cross-sectional studies.4

In general, cohort studies are less susceptible to bias than are case-control studies, and case-control studies are less susceptible to bias than cross-sectional studies. Thus, if available, we would choose cohort studies as our source of evidence.

Why, then, would investigators bother conducting case-control studies? The reason is that if an outcome is rare or if the outcome occurs over a long period, conducting a cohort study may be challenging or not feasible at all and choosing the case-control design might be a better option.

Consider the question of whether smoking increases the risk of oral cancer. Because oral cancer is (fortunately) rare and because it develops over a long period, addressing the smoking issue would involve enrolling thousands of patients and following them for many years. Indeed, the initial studies demonstrating the association between smoking and cancer used a case-control design. Only later did investigators undertake the large cohort studies that definitively reported the association.

Sometimes, it might be completely infeasible to conduct cohort studies. Consider the question of whether pacifier use as an infant is associated with having temporomandibular disorders in adulthood. Following people from infancy to adulthood is likely to be impossible, and thus the only way to address the issue is by using a case-control design. Thus, these study designs may provide the best available evidence.

Critically Appraising Observational Studies to Inform Clinical Decisions

The process of using an article from the dental literature consists of three steps: assessing the risk of bias (that is, determining whether the results are systematically different from the truth), assessing the results (that is, determining the magnitude and precision of the estimates of the association between exposure and outcome), and assessing the applicability of the results (that is, determining the degree to which the results of the study can be applied to the patients who generated the clinical question).¹¹ We describe each of these steps in the sections that follow.

How Serious Is the Risk of Bias?

The extent to which a study’s results are likely to be correct for the sample of patients enrolled depends on how well the study was designed and conducted.^12,13

Are exposed and unexposed study participants sufficiently similar?

Bias—systematic difference from the truth—will occur if exposed and unexposed study participants differ with respect to an important determinant of outcome (which we call a prognostic factor).¹⁴ For example, if we ask whether patients with dental crowding (exposed group) are more likely to have caries than patients without dental crowding (unexposed group), misleading results caused by bias could occur if patients with dental crowding brush their teeth less frequently (the extraneous prognostic factor).

Or consider the question of whether drinking milk at night (the exposure) causes dental caries (the outcome) in children. Parents who are less aware of appropriate oral health care practices for babies may be more likely to give their babies milk at night, and they might be less likely to brush their children’s teeth (the extraneous prognostic factor). As a result, the imbalance in the extraneous prognostic factor (toothbrushing) may create a spurious association between milk at night and caries. We sometimes refer to prognostic imbalance (that is, the extraneous prognostic factor being distributed differently in exposed and unexposed) as selection bias or a confounding factor.

In both cohort and case-control studies, prognostic imbalance is likely to occur. What can investigators do when faced with these situations? Fortunately, there are statistical strategies to deal with the problem that involve comparing like with like. For instance, in the example previously described, investigators could focus first on children whose parents brushed their teeth and, among these children, compare those who did and who did not receive milk at night. Then, the investigators could focus on the children whose parents did not brush their teeth and, among those children, compare those who did and who did not receive milk at night. Finally, the investigators could combine the results across these two comparisons. In this way, they could avoid inadvertently causing the bias that could otherwise result from prognostic imbalance. We call this analytical strategy an adjusted or stratified analysis.¹⁵

Is information collected in the same way in exposed and unexposed study participants?

Biased results will arise when investigators gather data in different ways regarding the presence of an exposure, a prognostic factor, or an outcome in the exposed and unexposed study participants, or the cases and the controls.¹⁶ For example, in a case-control study, researchers may search more thoroughly for the presence of a past exposure if they know that the patient belongs to the group of cases rather than the group of controls. Even if the method of information collection is similar, bias will intrude if patients who are cases are more likely to remember a past exposure than patients who are controls.

Asking the following three questions can help clinicians assess the extent of the risk of bias in an article about harm: Were patients similar for prognostic factors known to be associated with the outcome, or did investigators conduct an adjusted analysis that considered all such factors? Were the circumstances and methods for detecting the exposure, prognostic factors, or outcome similar in both groups? In a cohort study, was the follow-up sufficiently complete? Table 4.3^6,17–20 lists these questions, provides some strategies to reduce bias to consider when answering them, and offers examples from the dental literature.

Stay updated, free dental videos. Join our Telegram channel

Join