Chapter 6. Infectious Diseases Epidemiology

Epidemiology is the study of health and disease in a population. Studying the causes, distribution, risk factors, and treatment of diseases, all of which fall under its scope, provides the cornerstone for research into explaining how and why infectious diseases occur. This chapter is designed to provide a brief introduction to the core principles of clinical research and epidemiology. Our goal is to present clinicians with an overview of the key terms and concepts found throughout published research, and we hope that this chapter will serve as a useful reference for pediatricians seeking to better understand the fundamentals of infectious diseases epidemiology.

In this chapter, we begin with a discussion of study designs, including the strengths and weaknesses of different types of studies. We then explore important concepts used in studies of diagnostic tests: sensitivity, specificity, positive and negative predictive values, and likelihood ratios. Next, we discuss statistical analysis by comparing univariate and multivariate analyses, while explaining terms used to define the magnitude of an effect such as odds ratios (ORs), relative risk (RR), and confidence intervals. Finally, we end the chapter with an examination of relative risk reduction (RRR), absolute risk reduction (ARR), and number needed to treat—measures used to enumerate the benefits of an intervention. We hope that the information presented in this chapter provides a concise overview of the key concepts of epidemiologic and clinical research in infectious diseases.

A variety of study designs are utilized in clinical research and each allows authors to address different questions. Depending on whether an author would like to determine the prevalence of a disease, the natural course of an infection, or the effectiveness of a treatment, for example, will determine which study design the investigator will use. In general, studies can be broken down into two principle types: descriptive and analytic. Descriptive studies include case reports, case series, and survey studies, all of which are used primarily to generate hypotheses.1–4 As the name implies, these types of study designs allow authors to describe the characteristics of a single patient or group of patients with a common disease. However, since these studies do not use a control or comparison group, they are poorly suited to make causal inferences. Instead, they are more useful for characterizing emerging or rare diseases, such as avian influenza A, where little is known about the natural course of disease.

Analytic studies are hypothesis-testing in nature and are better equipped to explore the relationship between cause and effect. Analytic studies can be further subdivided into observational and experimental studies. Observational studies include cross-sectional, case-control, and cohort studies, and provide information about prevalence, incidence, causes, risk factors, and outcomes of disease. Experimental studies test the effects of an intervention and include clinical trials, both randomized and nonrandomized. Randomized clinical trials (RCTs) are often considered the most powerful and scientific study designs, and are the gold standard for evaluating the efficacy of therapies and other interventions. While all study designs are subject to varying amounts of external validity, or generalizability to other populations or to the population at large, RCTs have the highest internal validity, meaning that they are the most capable to conclude that a cause has a specific and real effect. Table 6–1 provides a list of the most common types of study designs and includes information about each design's strengths and limitations.1–7 In this table, study designs are listed in order of increasing strength of evidence.

One of the primary issues that an investigator must face when designing a study is, “How well can this study design answer the question at hand?” Regardless of design, the main obstacles to the internal validity of any study are bias and confounding. Bias is any source of variation that distorts the study findings and creates systematic error, and can only be avoided through careful study design that addresses these issues upfront. Meanwhile, confounding is any extrinsic factor that is both associated with the exposure and a cause of the outcome.5 Unless addressed during the design or statistical analysis phases, confounding variables will alter the association between an exposure and an outcome such that the result is actually an over- or underestimation of the true effect of the exposure on the outcome. Additionally, in any study there is the potential that an investigator's results are false merely owing to random error. This error is unrelated to how the study was designed or conducted, but occurs simply by chance. However, random error will decrease as the sample size increases.

Many epidemiologic studies examine the utility of diagnostic tests. In infectious diseases, these studies are of particular importance. Clinical decision making relies on an understanding of the reliability of test results and knowledge of how these tests should be used in a clinical context. For example, a child may be admitted to the hospital because of what appears to be a viral syndrome. And, the child's physician suspects that influenza is responsible and would like to send a rapid test for the virus. Before sending the test, there are numerous questions that the physician may ask. What is the probability of influenza in this child (pretest probability)? How often will patients with the virus have a positive test result (sensitivity)? How often will patients that do not have the virus have a negative test result (specificity)? Or, how often will a positive/negative test result mean that the patient actually has/does not have an active viral infection (positive/negative predictive value)? Knowledge of these important concepts will affect the ordering of the test, the interpretation of the test results, and the likelihood that the patient does or does not have the disease once the test results are known (posttest probability).8,9

The utility of diagnostic tests can often be thought of in regard to a 2 × 2 table (Table 6–2) and explained using the following terms:

• 1. Sensitivity—the ability of a test to accurately identify those with disease.
  • a. Calculated as the proportion of individuals with the disease in whom the test is positive.
• 2. Specificity—the ability of a test to accurately identify those without disease.
  • a. Calculated as the proportion of individuals without the disease in whom the test is negative.
• 3. Positive-predictive value (PPV)—the likelihood that a positive result represents the identification of an individual that actually has the disease.
  • a. Calculated as the proportion of individuals that test positive and have disease among all individuals that test positive.
• 4. Negative-predictive value (NPV)—the likelihood that a negative result represents the identification of an individual without disease.
  • a. Calculated as the proportion of individuals that test negative and do not have disease among all individuals that test negative.

There are drawbacks to the use of these test parameters in a clinical setting, however.8 In general, highly sensitive tests are good screening tests because they will pick up almost all cases of disease. But the increased sensitivity usually comes at the expense of specificity, so highly sensitive tests sometimes have a high false-positive rate. In that setting (high sensitivity but only a modest specificity), a negative test result provides more certainty to the clinician regarding the disease status of the patient. A test that is 100% sensitive, for example, will detect all cases of disease, but if the specificity is only modest (i.e., 85% or less), may also produce numerous false-positives. So, while a negative test result will effectively rule out disease and allow the clinician to reassure the patient, a positive test result will be less helpful, especially if the disease is uncommon (so most positive test results are false-positives), and may warrant further testing to determine true disease status. A good example of such tiered testing is syphilis where a positive nontreponemal antibody test (e.g., rapid plasma regain [RPR]) is typically confirmed with treponemal antibody testing (e.g., FTA-ABS, fluorescent treponemal antibody absorption). On the other hand, a diagnostic test that is highly specific is much more helpful if the test result is positive since high specificity often comes at the expense of decreasing sensitivity. Since tests with high specificity are used to accurately identify those with disease, they are generally regarded as good confirmatory tests because false-positive results are uncommon. Unfortunately, they will likely have numerous false-negatives. Therefore, a positive result can help direct clinical management, while a negative test doesn't necessarily exclude disease in a patient unless the sensitivity is also quite high.

As the above paragraph alludes to the main issue with both specificity and sensitivity is that they do not help clinicians estimate probability of disease in individual patients.8 They are useful at ruling in or out disease, but cannot be used to determine how likely a patient is to have or not have the disease. Instead, a clinician can use positive and negative predictive values to help predict the likelihood of disease once a test result is known. Both the PPV, which tells the probability of having disease given a positive result, and the NPV, which tells the probability of being disease-free given a negative result, can increase or decrease a clinician's posttest probability. In this manner, test results can be applied to individual patients. However, there is a significant limitation to the use of PPV and NPV: Both are affected by the prevalence of disease. PPV will increase as the prevalence of disease increases, while NPV will decrease, and vice versa when disease prevalence decreases. So, clinicians need to be careful when applying PPV and NPV to test results in their patients, since the prevalence may differ from that of the population used to define the test parameters.

While sensitivity, specificity, PPV, and NPV are useful when interpreting diagnostic test results, they have limitations when assessing the probability of disease in individual patients. Instead, by combining the sensitivity and specificity, likelihood ratios(LRs) provide a more clinically useful measure. LRs describe how many times more (or less) likely patients with disease are to have a positive (or negative) test result than those without disease.

• 1. Likelihood ratio positive (LR+) explains the probability of a patient with disease having a positive test result compared to the probability of a patient without disease having a positive test result.
  • a. Calculated as the true-positive rate (sen-sitivity) divided by the false-positive rate (1.0—specificity).
  • b. LR+ greater than 1.0 indicates an increased probability of disease.
• 2. Likelihood ratio negative (LR–) explains the probability of a negative result in a patient with disease compared with a negative result in a patient without disease.
  • a. Calculated as 1.0—sensitivity of a test divided by the specificity.
  • b. LR– less than 1.0 indicates a decreased probability of disease.

Clinical Example

Respiratory syncytial virus (RSV) is a prevalent virus, which is the cause of high levels of morbidity and mortality in infants and young children. Although treatment is largely supportive, rapid detection of RSV may expedite management and prevent the use of additional costly and potentially harmful diagnostic tests. In 2002, Dayan et al. sought to determine the test characteristics of RSV EIA in febrile infants.10 Using tissue and shell viral cultures as the reference standard, the nasal washings of 174 infants were evaluated using RSV EIA. Table 6–3 demonstrates the comparison of RSV EIA to culture.

Based on these findings, the RSV EIA had a sensitivity of 75%, a specificity of 98%, a PPV of 89%, and a NPV of 95%. The LR+ for this test was found to be 35.5, while the LR– was 0.26. Based on these findings, the RSV EIA is a modest screening test and good confirmatory test for the detection of RSV. It will correctly identify three quarters of febrile infants with RSV (sensitivity), and infants without RSV will be accurately identified as they do not have disease 98% of the time (specificity). Additionally, this study found that a febrile infant with a positive RSV EIA was more than 35 times more likely to have RSV than a febrile infant without RSV (LR+). Based on these results, the RSV EIA has significant clinical utility and the results of testing in a febrile infant can assist clinicians in their management.

Once data collection has been completed, the next step is statistical analysis. The first step in this process is to conduct univariate analysis, which explores the relationship between the dependent variable and each independent variable individually. Independent variables are those factors, which may impact the outcome (dependent variable) and can take a variety of forms: demographic and historical data, exposures, treatments, interventions, etc. Univariate analysis provides descriptive information about the frequencies and distributions of the independent variables across subjects in a study. Additionally, univariate analysis sets the stage for further analyses. Variables can be summarized across groups of subjects in the study allowing for comparison across exposed and unexposed groups, or across the intervention and nonintervention groups. Thus, through univariate analysis, investigators can determine whether the exposure or intervention group was more or less likely to experience an outcome or event than the unexposed or nonintervention group.

While it is useful to demonstrate that one group was more or less likely to have experienced an event or outcome, it is often necessary to determine how much the groups differed. Two concepts used to quantify the magnitude of an effect are RR or risk ratio, and the OR. Both the OR and RR describe the likelihood of an event between two groups. And, interpretation of OR and RR is somewhat similar: If the RR or OR equals 1.0, the likelihood of an event is comparable between the two groups. If the RR or OR is greater than 1.0, an event is more likely to occur in the exposed or intervention group than in the unexposed or nonintervention group. The opposite is true if the OR or RR is less than 1.0.11 However, there are important differences between the RR and OR that must be recognized.

The OR describes the relative odds of an outcome or event in the exposed group compared to odds of the outcome or event in the unexposed group. On the other hand, the RR is the ratio of the probability of an outcome or event in the exposed group to the probability of the outcome or event in the unexposed group. Since the difference between these two concepts is not always clear, it is useful to think in terms of a 2 × 2 table (Table 6–4). The primary difference between OR and RR is the discrepancy between odds and probability. Although both OR and RR are ratios, the concept of RR is more straightforward. RR can be simply understood as the increased or decreased likelihood of an outcome or event in the exposed group versus the unexposed group. And, for most individuals, this is the intuitive way to compare the likelihood of an event between two groups. An RR can be stated as “group A is X times more or less likely to experience Y than group B.” However, an OR cannot be understood in this manner. Instead, the OR needs to be interpreted in terms of the prevalence of the event or outcome. As the prevalence of an event or outcome increases, the OR poorly approximates relative risk. In fact, an OR will always overestimate the RR if greater than 1.0 and underestimate a relative risk if less than 1.0. The amount of over- and underestimation will be magnified if the prevalence of the event or outcome is high.12 Therefore, it is only when the prevalence of a disease or outcome is rare that the OR truly approximates the RR and can be interpreted as such.

So, why would one ever use ORs if their interpretation is difficult? Firstly, the direction of the effect is useful. An OR greater than or less than 1.0 will demonstrate which group is more or less likely to experience an outcome or event. Often, that has as much utility as the magnitude of the effect itself. Secondly, ORs are necessary in case-control studies, since relative risks cannot be calculated in these studies. Because case-control studies are designed to identify risk factors for development of disease, cases and controls are selected based on the presence or absence of disease, respectively. An RR cannot be determined because cases overestimate the likelihood of disease in the population. Therefore, a case-control study cannot estimate the risk of disease and, simply because of the design, cannot calculate the RR of developing disease based on the presence or absence of a risk factor. Instead, an investigator can use ORs to express the relative likelihood of a risk factor in cases and controls. And, since case-control studies study rare diseases, the OR will actually approximate the RR mathematically (in Table 6–4, cells a and b will be significantly smaller than c and d, and the calculated OR will equate the RR). Finally, ORs are practical when performing multivariate analyses. It is far easier to adjust the ORs for confounding variables than it is to adjust RRs. Thus, determining ORs will allow for the measurement of the independent effects of variables on the outcome of interest.

Multivariate analysis allows investigators to determine the independent effects of multiple factors on a single event or outcome.13 When numerous independent variables, or contributing factors, have an effect on a dependent variable, or outcome, multivariate analysis allows investigators to determine the individual contribution of each. As mentioned above, confounding variables are factors associated to the exposure and causally related to the outcome. Multivariate analysis enables researchers to determine the true effect of an exposure on an outcome by adjusting for known or suspected confounders.

There are numerous types of multivariate analysis that are available for researchers: multiple linear regression, multiple logistic regression, and Cox proportional-hazards regression, for example. Linear regression is used when the outcome is a continuous variable (i.e., HIV-viral load), and logistic regression is used when the outcome is a dichotomous, often yes or no, variable (i.e., HIV seroconversion or death). Cox regression is useful when the outcome of interest is a period of time, such as time to positivity for blood cultures.

Once the results of a multivariate model are determined, the adjusted ORs relate to the independent effects of each variable on the outcome of interest. And, while the investigator may have adjusted for all known confounders, there is still the possibility that either random error or unmeasured confounders played a role in the point estimate of the effect between the variable and the outcome. Confidence intervals allow the investigator to present a range of values within which the true value of the effect likely lies. Ninety-five percent confidence intervals, for example, which are most commonly used in clinical research, provide a range of values that contains the true effect with 95% certainty. Wider confidence intervals represent more uncertainty, while more narrow confidence intervals signify greater precision. When the confidence intervals surrounding an OR or an RR include 1.0, the author cannot be certain whether there is actually a difference between the two groups.11 Therefore, when the confidence intervals include 1.0, a causal relationship between the independent and dependent variables cannot be established.

ORs and RRs are informative when summarizing case-control and cohort studies. However, when an investigator conducts a clinical trial, it is important to enumerate the effects of the intervention being studied. Patients and clinicians alike want to know when the benefits of an intervention outweigh the harms. One way to express the effects of the intervention is in terms of the RRR: the relative difference in the rate of an event (i.e., death, hospitalization, etc.) between the intervention and the nonintervention group. RRR is calculated by dividing the difference between rate of an event in the intervention group and nonintervention group by the rate of the event in the nonintervention group. For example, in a hypothetical trial, if an event occurs in 40% of the nonintervention group and in 20% of the intervention group, the RRR from the intervention is 0.50, or 50% (calculated as 40% minus 20%, divided by 40%). Another measure of the benefit or harm of an intervention is the ARR. This is the absolute difference in the rate of an event between the intervention and nonintervention group. To calculate ARR, the event rate in the intervention group is subtracted from the event rate in the nonintervention group. In the hypothetical trial above, the ARR would be 20%.11,14,15

Because the rate of an event in the nonintervention group serves as the denominator for calculating RRR, the RRR will always be greater than the ARR. And, as the event rate becomes smaller, the difference between the relative and ARR will widen, making the RRR appear more dramatic.14 However, RRR should be interpreted with caution. Because it is a relative measure, it does not take into account the base-line event rate in the control group. Therefore, when the event is very rare or very common, RRR will dramatically overestimate or underestimate, respectively, the true effects of the intervention. Contrarily, ARR does account for the event rate in the nonintervention group at baseline and, thus, provides a more clinically practical measure of the effect of the intervention.15 Hence, the ARR is often referred to as the attributable risk of an intervention.

Although both RRR and ARR are ways to understand the benefits of an intervention, they are not easily applied to individual patients. Instead, clinicians can employ a measure known as the number needed to treat (NNT) to better enumerate these effects in a clinically useful way. The NNT is defined as the number of patients that need to receive a treatment in order to prevent one patient from experiencing an adverse outcome over a specified period of time (the study period).16 It is calculated as the inverse of the ARR (1/ARR) and the smaller the NNT, the more effective the treatment. In the hypothetical trial above, the ARR was 20%. Therefore the NNT would be five, which is to say that five patients would need to receive the treatment to avoid one adverse outcome. However, the NNT can also be applied to individual patients: In the example, an individual patient has a one in five chance of benefitting from the treatment.14

Clinical Example

In January, 2006, Vesikari, et al. published a study examining the safety and efficacy of a live pentavalent human-bovine reassortant vaccine against rotavirus gastroenteritis.17 This study was a double-blind, placebo-controlled, randomized trial involving more than 69,000 children in 11 countries. Healthy infants between 6 and 12 weeks of age were randomized to receive three doses of oral vaccine or placebo. All subjects were followed for the development of acute gastroenteritis requiring hospitalization or emergency department care, as well as the development of complications such as intussusception.

In this study, 57,134 children were involved in an analysis of health care service use attributed to rotavirus gastroenteritis. For our discussion, we will look simply at the efficacy of the vaccine in reducing hospitalizations resulting from acute rotavirus gastroenteritis. Among the infants who received the vaccine, 6 of 28,646 developed acute rotavirus gastroenteritis requiring hospitalization for 14 days or more following receipt of the last dose of the vaccine. In the placebo group, 138 of 28,488 infants were hospitalized. Children in the vaccine group had an RR of 0.043 for developing rotavirus gastroenteritis requiring hospitalization (calculated as 6/28,646 divided by 138/28,488), meaning that vaccinated children had 4.3% of the risk of being hospitalized as a result of rotavirus gastroenteritis compared to unvaccinated children. The RRR from vaccination was 95.7% (calculated as 138/28,488 minus 6/28,646 divided by 138/28,488), while because of the low incidence of hospitalization overall, the ARR from vaccination was just 0.46% (calculated as 138/28,488 minus 6/28,646 times 100). However, the NNT of 216 (calculated as 1 divided by the difference of 138/28,488 and 6/28,646, or 0.0046) means that vaccination of 216 children prevented 1 case of acute rotavirus gastroenteritis requiring hospitalization over the first year of life.

An NNT of 216 may seem high to some readers. However, one needs to take into account the prevalence of disease, as well as the risk-benefit ratio of the intervention. Rotavirus infection affects more than 100 million children annually worldwide and is the leading cause of deaths resulting from acute gastroenteritis.18 Even in the United States, rotavirus gastroenteritis is responsible for more than 60,000 hospitalizations per year.19 Therefore, given the cost associated with hospitalization (parental missed days of work, health care costs, etc.), the significant morbidity and mortality associated with severe rotavirus gastroenteritis, the incidence of this disease worldwide, and the relative safety and efficacy of this vaccine, widespread vaccination has a considerable potential benefit.

1. Study designs vary significantly in regard to the strength of evidence. While descriptive studies provide useful anecdotes and often serve as the basis for further investigation, analytic studies have the capacity to have tremendous internal validity and demonstrate true cause and effect.

2. Sensitivity and specificity relate to the ability of a test to identify an individual with and without disease, respectively.

3. The positive and negative predictive values relate to the likelihood that a positive/negative result has accurately identified an individual with or without disease, respectively.

4. LRs describe how many times more (or less) likely patients with disease are to have a positive (or negative) test result than those without disease.

5. ORs and RR are two ways to quantify differences between study groups. An OR or RR greater than 1.0 signifies an increased likelihood of the event in the comparison group, while OR or RR less than 1.0 signifies an increased likelihood of the event in the control group.

6. The NNT, calculated as 1 divided by the ARR, represents the number of individuals that would need to receive the treatment/intervention in order to avoid one adverse outcome.