A large number of study designs and methods are used to generate data on the uses and risks of new and older drugs. The types of study designs used in pharmacoepidemiology can be classified as experimental and observational. Experimental studies employ control in the assignment of individuals to exposure groups, usually through random assignment of individuals to the exposure under investigation, and then follow-up of individuals to detect the effects of exposure. For example, a recent clinical trial demonstrated that hormone-replacement therapy does not prevent coronary heart disease in women. This randomized, controlled, primary prevention trial, the Women's Health Initiative (WHI), studied 16,608 postmenopausal women aged 50 to 79 years with an intact uterus at baseline recruited by 40 U.S. clinical centers in 1993–1998. Overall health risks exceeded benefits from use of combined estrogen plus progestin for an average 5.2-year follow-up among healthy postmenopausal U.S. women.56
Observational epidemiologic study designs, such as case-control, cohort, and cross-sectional studies, are used extensively. Large automated databases, meta-analyses, and hybrid designs, such as nested case-control studies, also play an important role in pharmacoepidemiology. Epidemiologic studies typically do not use randomization to determine who will receive a particular drug exposure. Rather, associations between exposure(s) and disease(s) under study are determined through the use of observational study designs and statistical analyses. Observational methods are used in most situations because ethics and cost limit the use of experimentation. For example, one would not experimentally subject individuals to certain drugs to determine if they develop cancer. While observational studies are generally quicker and less costly than experimental studies, they have important disadvantages. One such limitation of observational study designs in pharmacoepidemiology is confounding by indication. Confounding by indication occurs when subjects treated with the medication of interest differ from the non-treated group on a characteristic(s) also associated with the outcome. For example, a health user effect is one explanation for why observational studies have reported a decreased risk of cardiovascular disease events among women using hormone therapy compared to non-users.57 A hypothesis that was refuted in the WHI randomized controlled trial of hormone therapy use. While there are design and analytic techniques to cope with confounding, the possibility of distorted effects by confounding should be carefully considered in any observational study.
A number of methods are used to study health events associated with drug exposures. The usual approach to studying ADEs begins with the collection of spontaneous reports of drug-related morbidity or mortality. There has been a growing interest in using computerized databases containing medical care information for pharmacoepidemiologic studies.58 These databases usually consist of patient-level data from two or more separate files (e.g., billing files for pharmacy and medical services reimbursement) that were developed originally for clinical or administrative applications.59 Through record linkage, person-based longitudinal files can be created on an ad hoc basis. Multipurpose databases used for pharmacoepidemiologic studies include data from managed-care organizations, the Medicaid program, the Medicare program, and geographically defined populations. In general, these databases include information on patient demographics, outpatient drugs, hospital discharge diagnoses, and ambulatory care encounters. The advantages and disadvantages of linked databases for pharmacoepidemiologic studies have been the subject of numerous publications.60,61
Case Reports and Case Series
Case reports also referred to as spontaneous case reports or passive surveillance describe a single patient who was exposed to a drug and experienced a particular, usually adverse event. Such reports might be communicated by healthcare professionals or consumers to companies, regulatory agencies, the World Health Organization, or reported in the medical literature. The FDA receives approximately 400,000 reports of suspected adverse events annually. Well-documented case reports can be viewed as a safety signal, alerting to the possibility of a rare adverse event not previously detected in premarketing studies. Spontaneous reports can also provide information on at-risk groups, risk factors, and clinical characteristics of known serious adverse drug reactions. The reporting of AEs is influenced by several factors, including the elapsed time since its introduction into the marketplace, regulatory activity, and media attention.
Case series are collections of patients, all of whom have a single exposure, whose clinical outcomes are then evaluated and described. They are useful for quantifying the incidence of an adverse reaction, particularly for a newly approved drug. Further, case series can be useful for being certain that the incidence rate of any particular adverse effects of concern does not occur in a population that is larger than that studied prior to drug's marketing. It is uncommon for a case report or a series of case reports to be used to make a statement about causation. If the event is rare and the exposure combination is very specific, the cause of the adverse health event may be inferred from a case-series study. In most situations, however, it is necessary to compare cases with a group of controls to identify risk factors. Thus the major disadvantage of a case-series study is the lack of a comparison group.
Active surveillance is the regular, periodic collection of case reports from healthcare providers or sentinel site facilities. Active surveillance, in contrast to passive surveillance, seeks to ascertain completely the number of adverse events via a continuous preorganized process. An example of active surveillance is the follow-up of patients treated with a particular drug through a risk management program. Patients who fill a prescription for this drug may be asked to complete a brief survey form and give permission for later contact. In general, it is more feasible to get comprehensive data on individual adverse event reports through an active surveillance system than through a passive reporting system.
A registry is a type of active surveillance whereby a list of patients presenting with the same characteristic(s) is followed. This characteristic can be a disease (disease registry) or a specific exposure (drug registry) or a type of exposure occurring during a specific life event (pregnancy exposure registry). Registries can collect information using standardized questionnaires in a prospective fashion. Disease registries, such as registries for blood dyscrasias, severe cutaneous reactions, or congenital malformations can help collect data on drug exposure and other factors associated with a clinical condition. A disease registry, such as a cancer registry, might also be used as a base for a case-control study comparing the drug exposure of cases identified from the registry and controls selected from either patients with another condition within the registry, or patients outside the registry. The most common approach to systematically evaluate the postmarketing safety of medicines used during pregnancy is the use of pregnancy exposure registries. These are prospective approaches used to identify and follow pregnant women who are exposed to medications of interest until the end of pregnancy (i.e., before the outcome is known).
Through public–private partnerships, the Sentinel Initiative will facilitate the development of active surveillance methodologies related to signal detection, strengthening, and validation. While signal detection can include observations by clinicians and patients, case reports in the literature, and clusters of spontaneous reporting systems, automated signal detection (or signal generation) involves active monitoring and statistical testing of potential associations between a medical product and an outcome or set of outcomes within automated databases for rapid identification of unwanted side effects. Signal strengthening (or signal evaluation) investigates signals identified through automated signal detection or another source such as MedWatch. It can include seeking the association in a new data environment, checking the nature and quality of the data, individual case report assessment through medical chart review, scanning for temporal clustering, targeted comparisons, biologic plausibility, and strength of the association. Signal validation (confirmation or refutation) is the process of determining if a signal represents a true causal relationship. It involves developing and conducting a full epidemiologic study often using advanced epidemiologic methods such as cohort and case-control design with appropriate control of confounders to further investigate associations between a medical product and outcome(s) of interest.
A case-control study assembles a group of cases (people who have the disease of interest) and controls (people who do not). The exposure histories of the cases and the controls are determined to establish the extent of association between exposure(s) of interest and disease. Case-control studies compare patients with a specific disease with a control group composed of similar people but without the disease. Case-control studies attempt to identify risk factors for a disease by examining differences in antecedent exposure variables between cases and controls. For example, one can select cases of women of childbearing age with ovarian cysts and compare them with controls, looking for differences in prior use of oral contraceptives. Such a study was performed to determine if the then newly introduced triphasic oral contraceptives were associated with functional ovarian cysts.62
Case-control studies have been used extensively to assess the safety of pharmaceuticals. There are many examples of case-control studies that have identified important associations between drugs and adverse health events: vaginal cancer and diethylstilbestrol, Reye syndrome and aspirin, peptic ulcer disease and nonsteroidal anti-inflammatory drugs, and venous thromboembolism and oral contraceptives. Data from case-control studies are used to calculate an odds ratio, which is the ratio of the odds of developing the disease for exposed patients to the odds of developing the disease for unexposed patients.
A classic example is a study of diethylstilbestrol given during pregnancy and the risk of vaginal adenocarcinoma among female offspring nearly a generation later.63 The association between use of antibiotics and the risk of breast cancer was studied in a case-control study among women enrolled in a large, nonprofit health plan. Controls were selected from health plan records and frequency matched to cases on age and length of enrollment. Cases were identified from the Surveillance, Epidemiology, and End Results Cancer Registry, whereas antibiotic use was ascertained from computerized pharmacy records.64 An increased risk of breast cancer was reported for all antibiotic studies and a clear dose-response was observed. A study of serious coronary heart disease risk in relation to the use of COX-2 selective inhibitors exemplifies a nested case-control design.65 Serious coronary heart disease (myocardial infarction and sudden cardiac death) cases and controls, who were similar to cases in age, sex, and health plan region, were chosen from a large health plan. Information on medication use and relevant diagnoses were obtained from health plan data. Mortality status was determined from state death records. The investigators found an increased risk of serious coronary heart disease with rofecoxib use compared to celecoxib use. A nested case-control study is an efficient variation of a case-control and a cohort study, and commonly used when predictor variables are expensive to measure and can be assessed at the end of the study. In a nested case-control study, all cases (or a sample of all cases) and only a random sample of all controls are chosen for study from the same defined population.
An advantage of the case-control design for the study of drug-outcome relationships is its efficiency for the study of rare or delayed outcomes. Compared with other strategies, the case-control study is relatively inexpensive. One potential problem with case-control studies is their susceptibility to certain types of bias, including selection bias and information bias. Selection bias refers to systematic differences between those selected for study and those who are not, whereas information bias is systematic differences in the quality of information gathered for study and comparison groups.
A cohort study assembles a group of persons without the disease(s) of interest at the onset of the study, ascertains the exposure status of each person, and then follows the cohort over time to determine the development of disease in exposed and nonexposed persons. Cohort studies involve a comparison of the incidence of one or more outcome events among those who received a drug or some other exposure of interest compared with the incidence of the event(s) for a comparison group. For example, much information about the risk of fatal cardiovascular diseases among oral contraceptive users has come from the Royal College of General Practitioners Oral Contraception Study, in which 23,000 oral contraceptive users were compared with 23,000 nonusers chosen from the same British general practices.66 Death certificate records were used to ascertain instances of fatal events during the follow-up period.
Cohort studies can be prospective, as the Royal College of General Practitioners study illustrates, or retrospective. Some prospective cohort studies follow a large population over decades. For example, the Nurses' Health Study was begun in 1976 to investigate the potential long-term consequences of the use of oral contraceptives and was later expanded to include diet and nutrition and their relationship with the development of chronic diseases.67 Prospective cohort studies are one of the most valid types of observational study designs because exposure is measured and recorded prior to the development of the health outcome(s) of interest. Using prospectively collected data from the Nurses' Health Study, Chan and colleagues evaluated the association between long term use of aspirin and NSAIDs and risk of colorectal cancer.68 The investigators confirmed previous results that long term aspirin therapy (and non-aspirin NSAIDs) are associated with a reduced risk of colorectal cancer compared to non-users.
An alternative to the prospective cohort design is the retrospective cohort study. Retrospective cohort studies are useful when comparison cohorts of persons exposed and not exposed to drugs of interest can be identified at some time in the past from large preexisting databases and followed from that time to the present with regard to the incidence of a given outcome. Raebel and associates used a retrospective cohort design to describe the proportion of patients with poor serum drug concentration monitoring of drugs with narrow therapeutic ranges and factors associated with poor monitoring at 10 HMO Research Network sites.69 Retrospective cohort studies are commonly used to evaluate the risks and benefits of marketed medications in large populations, especially with the availability of longitudinal electronic databases.
Prospective cohort studies can provide strong evidence of associations between drugs and diseases because the exposure is assessed before the outcome occurs. However, because many cohort studies require large numbers of people followed for long periods of time, they can be expensive and, in some instances, infeasible. Retrospective or historical cohort studies can overcome these limitations if high-quality data have been collected already.
Experimental and Quasi-Experimental Study Designs
Phase IV clinical trials might be used to assess the risk or benefit in subpopulations that are inadequately studied in premarketing clinical trials (e.g., the elderly, children), to better determine the benefit-risk profile of a drug. Another rationale for phase IV trials is to evaluate the health risks and benefits of chronically used medications that were approved on the basis of short-term trials of surrogate endpoints (e.g., blood pressure and lipid and hemoglobin A1c levels), and for comparisons against other medications. One approach to the conduct of Phase IV trials is the use of large simple trials.
One of the opportunities that has emerged with increased computerization in healthcare is the use of large, linked databases for exploring pharmaceutical outcomes. The ability to use transaction or claims data from an insurance company or state Medicaid agency and link these data to files containing diagnostic and other patient-specific information has allowed researchers to explore outcomes questions at relatively low expense. Because these studies do not rely on random assignment of subjects, they have been described as quasi-experimental.70 The typical design includes a treatment (exposed) group, a control (unexposed) group, and some type of posttest assessment for both. Although efforts may be made to match treatment and control groups for important patient characteristics, the groups are not equivalent in the sense of an RCT. A refinement to this design is one where an analysis of underlying trends—factors that could influence study outcomes and progress independent of the study—is made using time-series methods. These studies often are used to evaluate the consequences of a change of policy, such as a prescription limit, or addition or removal of a drug from the marketplace. For instance, Soumerai and associates71 studied the effect of a prescription cap on the use of psychotropic drugs and emergency mental health services using claims data. They used pharmacy claims data collected over a 42-month period, including the 11 months that the prescription cap was in effect, and found that drug use decreased while costs to the state Medicaid program increased during the period of the cap.
A quasi-experimental design was used to study British Columbia's reference pricing policy for five therapeutic classes of drugs to determine if a worsening of health outcomes could be detected after implementation of the reference pricing policy. The authors reported that there has been no worsening of health outcomes associated with implementing the reference pricing policy.72